diff --git a/CITATION.cff b/CITATION.cff index 8fbb437..12b30cc 100644 --- a/CITATION.cff +++ b/CITATION.cff @@ -16,13 +16,14 @@ abstract: >- say which of its inputs its predictions depend on. Its four instruments -- typed evidence propagation under a checked soundness invariant, input-dependence testing against an ordered ladder of degraded inputs with a - prespecified equivalence margin and multiplicity control, an ablation ladder + descriptive median-gap tolerance and multiplicity adjustment, an ablation ladder scored against a matched reference distribution, and specification-family robustness with variance attribution -- are not specific to any domain. They are demonstrated on cuts of the public MaleCNS v1.0 connectome, where the headline result is methodological: a dependence verdict obtained on one extract of a network can invert on another extract of the same network, and - the extract's recurrence rather than its size predicts which way. + node count alone does not order the observed classifications. Recurrence is + a candidate associated feature, not an isolated cause. Concretely, FlyLab maps a compound and a free concentration onto a typed engagement panel for insect receptors and their vertebrate counterparts at the same dose, converts the insect side through a fixed mechanism table into gain @@ -36,9 +37,9 @@ abstract: >- N/A rather than a small number. Around that it ships the analyses that test its own predictions: connectome-dependence analysis (empirical permutation probabilities against degraded graph models, with a three-way verdict - against a prespecified equivalence margin and the weakest graph model the + against a descriptive median-gap tolerance and the weakest graph model the test could not distinguish from the real cut), an ablation ladder over the - pipeline's layers, a conclusion-stability matrix over a prespecified family + pipeline's layers, a conclusion-stability matrix over a fixed family of gain specifications, a variance-based uncertainty budget with a value-of-information ranking, a receptor-versus-circuit selectivity landscape, literature concordance against published orderings, published @@ -60,7 +61,6 @@ keywords: - sensitivity analysis - specification robustness - reproducible research - - 3Rs identifiers: - type: other value: "Zenodo DOI to be minted on release" diff --git a/HANDOFF.md b/HANDOFF.md index 2d5556f..738778e 100644 --- a/HANDOFF.md +++ b/HANDOFF.md @@ -48,15 +48,15 @@ flylab serve # http://127.0.0.1:8765 - Nicotinic key split by subunit: spinosad → α6, imidacloprid → β1; the rest stay on a documented aggregate because their sources used hybrid Dα/chicken-β2 constructs. Source correction: **Dederer 2011 is a cat-flea α1/α2 + chicken β2 hybrid, not a *Drosophila* receptor**, and the row now says so. **Analysis — five new modules** -- `analysis/dependence.py` — connectome-dependence analysis: permutation *p* first, the profile, the dependence class, and the **necessary information level** (the weakest degraded graph the test could **not** distinguish from the real cut — never "reproduces"). v0.6.1 adds a three-way verdict against a prespecified equivalence margin, a weight-matched transmitter null on the ladder's rank-3 rung, a relative effect floor, Benjamini-Hochberg across a landscape's structural tests with confirmatory/exploratory labelling, `cut_census` for the substrate's own statistics, and `synthetic_cut` / `ladder_recovery` / `ladder_power` for ground-truth validation of the instrument. +- `analysis/dependence.py` — connectome-dependence analysis: permutation *p* first, the profile, the dependence class, and the **necessary information level** (the weakest degraded graph the test could **not** distinguish from the real cut — never "reproduces"). v0.6.1 adds a three-way verdict against a configured median-gap tolerance, a weight-matched transmitter null on the ladder's rank-3 rung, a relative effect floor, Benjamini-Hochberg across a landscape's structural tests as an exploratory sensitivity analysis, `cut_census` for the substrate's own statistics, and `synthetic_cut` / `ladder_recovery` / `ladder_power` for limited synthetic recovery checks. - `analysis/baselines.py` — the ablation ladder: receptor-only / composition-only / topology-only / full, compared by ordering. -- `analysis/robustness.py` — 25 prespecified gain specifications × 7 conclusions, and the 3×3 amplify/buffer threshold grid. +- `analysis/robustness.py` — 25 fixed gain specifications × 7 conclusions, and the 3×3 amplify/buffer threshold grid. - `analysis/uncertainty_global.py` — Sobol' budget with Jansen estimators, a null factor that measures the noise floor, and Ishigami validation. - `analysis/voi.py` — `VOI_j = S_j · Var(Y)` mapped onto concrete experiments. - `analysis/claims.py` (added by the dashboard agent, used by the pipeline) — `claim_audit()` walks the nine-link dependency chain behind a readout and labels each link OBSERVED / LITERATURE-DERIVED / MODEL-ASSUMPTION / COMPUTED; exactly one link (the MaleCNS edges) is a measurement of the system being simulated. `fact_inference_unknown()` gives the facts / model-inference / unknown split. Table T18. **Paper and pipeline** -- The manuscript is now a **research article** (Introduction / Methods / Results / **Discussion** / Conclusion) organised around four research questions, with `papers/SUPPLEMENT.md` holding the mechanism rationale, the null-model definitions, the LIF calibration, the supporting capabilities and the prospective predictions. Body is inside the 5000–7000 target and a test enforces it. +- The manuscript is a **research article** (Introduction / Methods / Results / **Discussion** / Conclusion) organised around three research questions. Reproduction and backend parity are supporting engineering evidence, not a fourth research question. `papers/SUPPLEMENT.md` holds the mechanism rationale, null-model definitions, LIF calibration, supporting capabilities and prospective predictions. The concise revision is below IJRC's typical article-specific 6000–8000 words but within its general 5000–10000 guidance; confirm with the editor rather than pad it. The test's 5000–8000 editorial guard does not certify venue compliance. The Word preparation copy and remaining submission checks are in `papers/submission/`. - `scripts/reproduce_paper.py` now produces 16 figures and T0–T25, including T12b (the landscape on `taste_motor`), T19 (both cuts' structure), T20 (the transformation rule), T21 (the transmitter signs), T22 (instrument validation), T23 (transmitter-null balance), T24 (the composition reference distribution) and T25 (the normalisation sweep). Default ~35–45 min, `--fast` ~4 min. A test asserts the committed record was produced at the shipped statistical effort, not merely that it is self-consistent. - "Pre-registered" is gone everywhere: the predictions are **prospective** until a tagged release is externally archived. @@ -64,16 +64,16 @@ flylab serve # http://127.0.0.1:8765 Read these before writing any new claim into the paper, the README or a PR description. -1. **The dependence verdict is a joint property of the prediction and the extract, and recurrence predicts it. That is the headline.** The same landscape, same instrument, no parameter changed: on the 1-hop `named` cut (mean degree 1.208, 85% of edges onto four seed cells, 184 of 1126 nodes with any input) 11 of 84 cells are topology-dependent, 42 composition-dominated and 31 no-effect at n = 1000 with Benjamini-Hochberg and a 1 % relative effect floor. On `taste_motor` (mean degree 10.4) the same cells give 54 / 0 / 30. The **scaling study** (`papers/scale_study.json`, T27, `flylab.analysis.scale`) then settles what the inversion tracks: over seven extracts the composition-dominated verdict survives on **exactly one**, the in-star, and `scale_1k` — *fewer* nodes than `named` (1000 vs 1126) but mean degree 22.9 — is already topology-dependent. **Recurrence, not size.** At 5k and 10k imidacloprid is distinguishable from *every* null; the top rungs run at n = 20 (resolution 0.048) and so confirm rather than establish. "Most predictions do not need the connectome" and "the topology-dependent set is exactly the chloride-channel blockers" are **withdrawn**. Never state a dependence result without the extract, its structure, the compound, the concentration, the readout, the *p* and the correction. - 1a. **Non-rejection is not equivalence.** Each mode carries a three-way verdict against a prespecified margin (0.332 Hz, 5% of the vehicle readout on `named`). For imidacloprid at n = 1000: equivalent within tolerance for rewire_degree_preserving, sign_permute_weight_matched, weight_permute, indeterminate for sign_permute. Never write "reproduces". +1. **The dependence verdict is a joint property of the prediction and the extract.** Classes differ between the sparse named and denser taste-motor cuts. Node count alone does not order the scale-ladder classes; recurrence is not isolated from density, composition or extraction recipe. The largest rungs use 20 permutations, so matching labels do not establish convergence. Report the extract, readout, concentration, permutation effort and adjustment with every dependence claim. The scale table is replayed from a saved study, not recomputed by the standard paper pipeline. + 1a. **Non-rejection is not equivalence.** The legacy within-tolerance label means a small point gap from the null median after non-rejection, not confidence-bounded statistical equivalence. Never write "reproduces" on that basis. 1b. **The transmitter null was biased.** Plain label permutation moves the cholinergic share of out-weight from 0.619 to 0.544 ± 0.027, putting the real graph at the 100th percentile of its own null. It is now reported as a joint target-set-and-sign null; the ladder's rank-3 rung is the weight-matched shuffle. - 1c. **The instrument is validated.** Planted ground truth on a synthetic cut: the ladder recovered 2 of 3 planted topology-dependent effects at n = 200 and returned no false positive on the unplanted control. False-positive rate 0 of 18; detection by permutation count n=50: 0.61, n=200: 0.61, n=1000: 0.67, so effect size dominates. -2. **The bitter-veto arm is a properly powered negative** (n = 300: p = 0.42 / 0.88 / 0.20 / 0.24, all |z| < 0.25). H6 is a pharmacology prediction, not a wiring prediction. Imidacloprid's veto ratio there is **undefined** (MN9 silenced), not zero. -3. **The suppression result is a property of the gain rule — and the old stability matrix was vacuous.** "A nicotinic agonist suppresses circuit activity" survives 15/25 specifications and is reversed by all 10 monotone ones, where the same drug at the same engagement *excites* the network. The v0.6 topology rows were computed at 6 shuffles (p ≤ 0.05 unattainable) **and** under a context that never rebound the gain function on the engine the permutation path resolves, so every specification fed default gains to its nulls. Recomputed at 100 shuffles with the specification pushed into the null engine and FDR across 100 tests, C3 holds 25/25 and C4 25/25 — but only 17 of C3's retentions reach equivalence and 8 are indeterminate. Retained by every specification: C2_rdl_disinhibition, C3_imidacloprid_topology_not_distinguishable, C4_fipronil_topology_exceeds, C7_map_bitter_veto. -4. **The ablation conclusion is withdrawn, and the composition correlation needs its reference.** Level C's old generic multiplier could only depress, so it could not express disinhibition; with a direction-aware rule its rank correlation at 1 µM rises from 0.242 to 0.961, and "the connectome without the pharmacology is not a cheap substitute" is retracted. The composition-versus-full correlation is 0.988, but pharmacology-free pseudo-compounds already reach a median of 0.847 and shuffling compound labels leaves it identically 0.988 — it carries no compound-level information. It is also not robust to the engine: 0.666 at 1 µM and -0.372 at 10 nM under a degree-corrected denominator. Never quote it without its reference distribution and the normalisation it was computed under. -5. **The uncertainty lives in two assumptions**: gain_transform, weight_threshold are the only factors whose bootstrap first-order interval excludes zero (S1 0.410 and 0.294). The estimator's noise floor is 0.046, set by gain_coef and **not** by the declared null factor `lif_seed` (-0.007); `drive` sits inside it and is no longer a ranked finding. Potency and Hill *n* are invisible at a saturating dose. Interactions carry 0.344 raw and 0.272 with negative estimates clipped. + 1c. **The instrument has limited synthetic positive-control evidence.** Six networks per setting give imprecise detection rates, and the zero-strength control differs in cycle edges and labels. This does not establish endpoint-specific power or general false-positive calibration. +2. **The bitter-veto arm is a non-rejection, with no endpoint-specific power analysis.** Do not call it a properly powered negative or proof that wiring is unnecessary. Imidacloprid's veto ratio is undefined when MN9 is silenced. +3. **The suppression result is a property of the gain rule — and the old stability matrix was vacuous.** "A nicotinic agonist suppresses circuit activity" survives 15/25 specifications and is reversed by all 10 monotone ones, where the same drug at the same engagement *excites* the network. The v0.6 topology rows were computed at 6 shuffles (p ≤ 0.05 unattainable) **and** under a context that never rebound the gain function on the engine the permutation path resolves, so every specification fed default gains to its nulls. Recomputed at 100 shuffles with the specification pushed into the null engine and BH adjustment across 100 component tests, C3 holds 25/25 and C4 25/25 — but only 17 of C3's retentions pass the descriptive point-gap check and 8 are indeterminate. Retained by every specification: C2_rdl_disinhibition, C3_imidacloprid_topology_not_distinguishable, C4_fipronil_topology_exceeds, C7_map_bitter_veto. +4. **The ablation comparison needs its reference.** Composition-versus-full correlation is high under row normalisation and depends on shared assumptions, but exceeds the matched pseudo-compound reference 95th percentile. Jointly permuting both score vectors preserves correlation by construction and does not prove absence of compound information. Report the reference and normalisation. +5. **The uncertainty lives in two assumptions**: gain_transform, weight_threshold are the only factors whose bootstrap first-order interval excludes zero (S1 0.410 and 0.294). The largest observed negative-estimate magnitude is 0.046, set by gain_coef and **not** by the declared null factor `lif_seed` (-0.007); `drive` sits inside it and is no longer a ranked finding. Potency and Hill *n* are invisible at a saturating dose. Interactions carry 0.344 raw and 0.272 with negative estimates clipped. 6. **The amplify/buffer split is threshold-sensitive** (`stable = False`): quote the range over the 3×3 grid, never the point estimate. -7. **The rate and LIF engines agree on direction only** (sugar-driven MN9 ≈ 0.29 Hz vs ≈ 105 Hz). That is not cross-model validation and must not be called one. +7. **The rate and LIF engines agree on direction only** (sugar-driven MN9 ≈ 0.29 rate-model units vs ≈ 105 LIF spikes/s). That is not cross-model validation and must not be called one. 8. **There is no independent out-of-sample validation of the circuit model.** Rank comparisons are literature *concordance*, and the pipeline flags which of them share a source with the library. 9. **`data/literature/README.md` lists ten literature-vs-library contradictions.** Report them; do not silently edit the library to match. 10. **Five transmitter sign magnitudes are asserted** (glutamate -0.4, histamine -0.5, dopamine +0.2, serotonin +0.2, octopamine +0.2), vary in no specification family, and are not bounded by the label-permuting null. They are published as T21; quote them as assertions. diff --git a/README.md b/README.md index fd6c286..1837b74 100644 --- a/README.md +++ b/README.md @@ -68,11 +68,11 @@ The analysis asks whether the real-graph effect is distinguishable from null ens A plain label permutation is also run, but it moves the weighted excitation/inhibition balance as well as transmitter identity, so it is reported as a joint null rather than as a rung of the ladder. -This makes it possible to separate effects that are sensitive to detailed wiring from effects that are already explained by much coarser network information. Failing to distinguish an effect from a null is reported as exactly that: each mode carries a three-way verdict against a prespecified equivalence margin, and "not distinguishable" is never rendered as "the degraded graph reproduces the effect". +This makes it possible to separate effects that are sensitive to detailed wiring from effects that are already explained by much coarser network information. Failing to distinguish an effect from a null is reported as exactly that: each mode carries a three-way verdict against a descriptive median-gap tolerance (not formal statistical equivalence), and "not distinguishable" is never rendered as "the degraded graph reproduces the effect". The ladder itself is validated rather than assumed to work. A recurrent loop of known strength is planted in a synthetic graph of the same size, density and transmitter composition as a real cut, and the analysis is asked to find it; the unplanted control gives an empirical false-positive rate, and a power surface maps detection against effect size and permutation count. -**The verdict this analysis returns depends on the extract it is measured on, and recurrence — not size — predicts which way.** The repository's two committed cuts disagree: the same landscape classifies most cells composition-dominated on the sparse 1-hop `named` cut and none of them on the denser `taste_motor` cut. Run across a ladder of nested cuts of 1k to 50k cells (`flylab.analysis.scale`), the composition-dominated verdict survives on exactly one extract — the in-star — while a cut with *fewer* nodes but twenty times the mean degree is already topology-dependent. A dependence result is therefore only quotable together with the structure of the graph it was measured on. +**The verdict this analysis returns depends on the extract it is measured on.** The repository's two committed cuts disagree: the same landscape classifies most cells composition-dominated on the sparse 1-hop `named` cut and none of them on the denser `taste_motor` cut. Run across a ladder of nested cuts of 1k to 50k cells (`flylab.analysis.scale`), the composition-dominated verdict survives on exactly one extract — the in-star — while a cut with *fewer* nodes but twenty times the mean degree is already topology-dependent. This does not isolate recurrence from density, composition or extraction recipe, and the largest rungs use only 20 permutations. A dependence result is therefore only quotable together with the structure of the graph it was measured on. ### 3. Model conclusions are stress-tested diff --git a/artifact-manifest.json b/artifact-manifest.json index f58d6f4..030e20c 100644 --- a/artifact-manifest.json +++ b/artifact-manifest.json @@ -8,8 +8,8 @@ ], "flylab_manifest_version": "1.0", "flylab_version": "0.6.0", - "generated_utc": "2026-09-20T16:42:40+00:00", - "git_sha": "b82b8edaaabbab7be1f2b6aec639dd8a076feb27", + "generated_utc": "2026-09-23T04:43:37+00:00", + "git_sha": "24431be9597646988c2862a18ea55f3d87d0db05", "git_tracked": true, "kind": "flylab-artifact-manifest", "note": "Hashes of the inputs the committed results depend on. Regenerate with `flylab manifest --write` whenever code or data changes, and before tagging a release: `flylab manifest --check` is what the release workflow gates on.", @@ -21,13 +21,13 @@ "flylab/analysis/__init__.py": "a6029eeec0258ce9799bde5f62b00fa03452c886abdb9881d38505f3f576ab1b", "flylab/analysis/baselines.py": "600093fb62a79d0dfe21077649cbda8016e4e9d30139638582936cf72c127939", "flylab/analysis/claims.py": "d657444619f3c4218b8224ad09df140494a504b50b25fabc8aa118d3a3e6e391", - "flylab/analysis/dependence.py": "6b6a98ea8e008318509e68e6a513d3bd162d7e7f173d13bcd41b8968043ce9d7", + "flylab/analysis/dependence.py": "3b5ac341ab3e21c84988d04249cc5b54bbbb247b0a86f53156ace1ff02de3769", "flylab/analysis/fit.py": "b6946fffa209c7444a054198e462a1f43cdf0dd0d12c084803579e2ef48f0ddf", "flylab/analysis/impact.py": "72f68e33369c92f3e47e957e7d50294941ba931441e56d1486f59a80f167dcd6", "flylab/analysis/layout.py": "771f168fac412e6747e98ae0f8bf0a39221a07694a41fa641e4b84a226f9a773", "flylab/analysis/nullmodels.py": "53f81c146483d26d5ec8649d43f3e483560d8f856d1bb30a6ec13281da9ddb9e", "flylab/analysis/predictions.py": "3382c24628274db35d60199a0be428b89a890efce250e944e2651245c1105e83", - "flylab/analysis/robustness.py": "31978a1371fb0863793bd0c4e6191ff201b1582b901de6c334f838b30857ec8e", + "flylab/analysis/robustness.py": "be69befe932c24ab9b9c920a49bdfada61027c66885c1bcc1e833f07096a46f6", "flylab/analysis/scale.py": "6264d2ce3497b1cbc075617b3d2ec4f532e092e5b8646aa665cdd76d6615f581", "flylab/analysis/selectivity.py": "622c881f58d50806a2f344a1c721a2a07dacc896e6718210fedbfc01250f2ef6", "flylab/analysis/uncertainty_global.py": "6adab31073fa0b7f933c2054af9e70f070abc0a69e641626f48421c6e0ae2d02", @@ -75,74 +75,78 @@ "pyproject.toml": "eab14bbee64da107f20259927bd9325c29e376286fef569423c50cc02b127140", "scripts/__init__.py": "8e742cba903af8699473573f6c8253b581f30fe9d14e6162e7f74c5ac1b6c5fa", "scripts/build_pages.py": "56185a50fd8d69736d6433f0eb4b0d26b032e0da0b88b3ceee3a5240c7311e34", - "scripts/manifest.py": "b73a4954d86d3b49c4cc2033bf2c9d0dcba31420a2a36fe77b8743a7f6927fc6", - "scripts/reproduce_paper.py": "45d0cd592a3e24971593c2587601aa4afd6e5e48c0ce8f639ba5ed38f5007f42" + "scripts/check_commit_attribution.py": "63f90c7d3bd4ef5ee9dbd5bfc4197fc51500fd92d1e5e34f0498e7f4a66316bc", + "scripts/export_paper_word.py": "06ff06bafd0bc81d6e11adfff71bd95de309bea426a635a143d3046e5733e448", + "scripts/manifest.py": "da4f0438a0c144e510dbcd829c70b3bd9d396481024610e9a1ad68ba172a1aab", + "scripts/paper_review_metadata.py": "1a7fae6a3fae6080bca184839c937c5102cf88050f57d0bd6e14da4b89a25906", + "scripts/reproduce_paper.py": "6502cd7ce4098fd5d1b829b9b36f2e37293423a5c0c542418cc581e4c40528de" }, "fatal": true, "kind": "files", - "n_entries": 60, + "n_entries": 63, "patterns": [ "flylab/**/*.py", "scripts/*.py", "pyproject.toml" ], - "sha256": "4a40094ca024eede4b8d5cdf540459c7ec1f9381218411ec9f82d44488c9535f" + "sha256": "8001fc13c701d33870b3b4231c7aeb1f34d8027b6e4b68b388f512819cb6438d" }, "dependencies": { "description": "resolved versions of the project's dependency closure", "entries": { - "PyYAML": "6.0.1", + "PyYAML": "6.0.3", "Pygments": "2.21.0", "annotated-doc": "0.0.5", "annotated-types": "0.8.0", "anyio": "4.15.1", - "certifi": "2026.2.25", + "certifi": "2026.7.22", "click": "8.5.0", - "colorama": "0.4.6", "contourpy": "1.3.3", "cycler": "0.12.1", "fastapi": "0.141.1", - "fonttools": "4.65.0", + "fonttools": "4.61.1", "h11": "0.16.0", "httpcore": "1.0.9", "httpx": "0.28.1", - "idna": "3.11", + "idna": "3.19", "iniconfig": "2.3.0", - "kiwisolver": "1.5.1", + "kiwisolver": "1.4.9", "markdown-it-py": "4.2.0", - "matplotlib": "3.11.2", + "matplotlib": "3.10.8", "mdurl": "0.1.2", - "numpy": "2.4.6", - "packaging": "24.0", - "pandas": "3.0.6", + "numpy": "2.3.5", + "packaging": "26.3", + "pandas": "2.2.3", "pillow": "12.3.0", "pluggy": "1.6.0", "pyarrow": "25.0.1", "pydantic": "2.13.5", "pydantic_core": "2.46.5", - "pyparsing": "3.1.1", + "pyparsing": "3.3.2", "pytest": "9.1.1", "python-dateutil": "2.9.0.post0", + "pytz": "2026.3.post1", "rich": "15.0.0", "shellingham": "1.5.4", - "six": "1.16.0", + "six": "1.17.0", "starlette": "1.6.0", "typer": "0.27.2", "typing-inspection": "0.4.4", "typing_extensions": "4.16.0", + "tzdata": "2026.3", "uvicorn": "0.53.0" }, "fatal": false, "kind": "packages", "missing": [ + "colorama", "exceptiongroup", - "tomli", - "tzdata" + "tomli" ], - "n_entries": 40, + "n_entries": 41, "note": "Informational: a manifest is verified on machines other than the one that wrote it, so a version difference is reported and never fails a check.", "patterns": [], - "python": "3.11.15", + "python": "3.12.14", "roots": [ "fastapi", "httpx", @@ -156,14 +160,14 @@ "typer", "uvicorn" ], - "sha256": "60a916447ed6449fa0c371058ca74bd68ae97df26bdf3a1517eaf8ded595ddf5" + "sha256": "56de6647d273599d4efabd84241a89b35de2dc26a999eca166424290824d5989" }, "graphs": { "description": "the derived MaleCNS cuts; CI products, never hand-edited", "entries": { "data/derived/malecns_census_v1.json": "7aebb432cea83eebce8f4f6a2d0967a79f782bad6bee20c05898cfc1526479f9", "data/derived/malecns_gustatory_seeds.json": "616adbe67df4034b4f1e0b1a3d363fd46c5f5a1b9e402bb44699cec6ae783b03", - "data/derived/malecns_named_neighborhood.json": "7f296ec5ec19490c46560732be83119a872668a87b70932e07aecf64443f49ae", + "data/derived/malecns_named_neighborhood.json": "8ae12b354372d1a4fcc428cca3eab88ea72b7027e8108fcf7abbaa966c0c63d1", "data/derived/malecns_scale_1k.json": "3ff5224ec586a9da2a9ea9ab007736c357ca76324f52d8f953e4085cb21437a9", "data/derived/malecns_taste_motor_neighborhood.json": "e4a28229b2c469a39a962aac15a01693effd9dd18d65694968be39c216cc0473" }, @@ -173,7 +177,7 @@ "patterns": [ "data/derived/*.json" ], - "sha256": "f7bc155e879a17c6a15349c746f5f4439ee357d9d35b0dfb5bacb2deeffb9f2d" + "sha256": "8357617f74bec25ec3fb2c02349ad8989c4801dcdc11261ada591f962f52c1ca" }, "library": { "description": "the typed compound library (schema v3)", @@ -209,10 +213,12 @@ "paper": { "description": "the committed paper artifacts regenerated by scripts/reproduce_paper.py", "entries": { - "papers/IJRC_FlyLab_draft.md": "640671d5c83bdb58fbd4782a81fcf230bc6b96a42f62716fdd9bf6feba26c606", - "papers/STATUS.md": "b9024f67ddc100a5e6683510b4caa9443b17f2666f780e1975ebd79760999c67", - "papers/SUBMISSION_CHECKLIST.md": "b54b5a6e38ceb3e7a4b810544d8606fe6e52fce7a358fabf08a1e18f767008f6", - "papers/SUPPLEMENT.md": "ff391afbe3a3651cb510ed82060f63264611c426cb5c8f3a045b0eda9f92ce4d", + "papers/IJRC_FlyLab_draft.md": "daf31354c7b8bff83c7319aab05d7ec915731aad03e105fdc2d1a44338bea917", + "papers/IJRC_FlyLab_draft.md.in": "4ab9479b5d04f2c437fcc7fe1289e881c905cb1bab32567e095fdee36adfca71", + "papers/STATUS.md": "637815a124ae9fd92440df9a33c6c9f6f0dd85c4053dc9881c09cf001110be79", + "papers/SUBMISSION_CHECKLIST.md": "226f27b52928120fd64139171f3d022995d5dba04c802daf3b1d9630178dabb2", + "papers/SUPPLEMENT.md": "66da51477d9cf42e3deec1e9e39484a6f646cbf6759f9d4719e76cd4720424fa", + "papers/SUPPLEMENT.md.in": "0bd2c1569616b6ad9c3db9f089fc2ca13654318b92f05ad99d9e6c426d55314f", "papers/figures/F10_rank_validation.png": "4943edf7eedeccebd5d8c642a3629add43997ece4ad83b19a634f11c7abb1222", "papers/figures/F11_evidence_types.png": "71c27c1084780bd0d502d9584386ba006aaab28a74bfdbfc690d4dbb8b2b7287", "papers/figures/F12_dependence_landscape.png": "2dc05bbb7a2fc49b61a4481ed5c9984927583d747caaa43c3403791977438c89", @@ -220,19 +226,28 @@ "papers/figures/F14_conclusion_stability.png": "f4a8474416dc694c703696343717436ae3b2ad39abcdb23142fbb39b0510c2f2", "papers/figures/F15_uncertainty_voi.png": "17855019835772b11757de863b8bae71d200524d4ec39f3508278f1029f2bc4b", "papers/figures/F16_instrument_validation.png": "7cc64976c52d975dfbd3c25d0f9279cdd6b3ce744709cbd8cc819253bf3657fc", - "papers/figures/F1_architecture.png": "649251040d5e94e209901c6da13eb96715496b458a72d2bcf7858f30006b324b", - "papers/figures/F1_architecture.svg": "f499b2a41f2016a74b4424b37e75e587fdf80bdfa283ee9a158ff0580850da87", - "papers/figures/F2_dual_scorecard.png": "22e67e15642afc6ef30a31d07c28eb663d0b9b2b013567cf741f888940e1b3df", - "papers/figures/F3_occupancy_curves.png": "45ed0c6f4f3107ff350ddeb4808947c145556a6ba74bf8f1f970094d7b5e5b0b", + "papers/figures/F1_architecture.png": "75b8ec5279eea12d2adf73f8a551c996e1d2a3946e727b746be13c5b84f51227", + "papers/figures/F1_architecture.svg": "82ebf528e9525193e6936353e0df3c285472667e73848e1312474ff94bee8f08", + "papers/figures/F2_dual_scorecard.png": "1bb991f277b60bebf806cabc7950cdc20c983eff43015fcb7976c07cdb25dd3c", + "papers/figures/F3_occupancy_curves.png": "59348fbbb51b068a525f9949c37afbc44ea77d100b6c2b33838cf74c727b3bd8", "papers/figures/F4_taste_motor_graph.png": "7474e1a85dcadcc4296b77912f517b5635fc034ea58d63511a97c8c072d1c319", - "papers/figures/F5_bitter_veto.png": "879920de976d9ac6d0ebe402c20cf0fedc2dea9a69b1fe7d339c751f665f2dc2", + "papers/figures/F5_bitter_veto.png": "86983cb01b646f651cfd53722a320d60af10c066d62938b816a6774fbb973885", "papers/figures/F6_dependence_profile.png": "838219d044a94200adbf6b04270c510a4ec3b3e625bb23b262c2fd4e8b942918", "papers/figures/F7_selectivity_landscape.png": "57d091db0d150e8e11d472b8ac5f5d283d0efe6706f742e330c3c1ea2436424d", "papers/figures/F8_dose_response_sensitivity.png": "f014b43b16ab8d1324604d0a1e8db1581f39c8dd9b102e7428be237e1ba190e4", - "papers/figures/F9_genotype_panel.png": "f51b3f7a33c7b949b98c35cd2d40d4dc4606aca5a4dd7262b59da0d27f56f3fd", - "papers/figures/captions.md": "697969d5df58bf2bce1eed8f33a137b9ca5d90013d51aab0858da53d7dede7bb", - "papers/results.json": "fab4a68fce4a1e3fd2aa472485191c492511e4a8c149660425efdd80ddb86513", + "papers/figures/F9_genotype_panel.png": "266a48da15eb96e1f00008a88c5f240bc783100975108b8bca9c8d2e874a4338", + "papers/figures/captions.md": "0e1ffc3302939aee0efd000f040c301464980c63f55ebc0497762659abd4d94b", + "papers/results.json": "e44976f3b180cce388183a61a8ef4705c168a5f62d5b7fdddd681879d89503c6", + "papers/reviews/current_reference_audit.md": "8e2fc0b66605ede04e49c47745d14e41ec70db162a1e989a76b4f1acc5499afe", + "papers/reviews/current_revision_decisions.md": "4ad5557d8a86767ecb2c3a2e0c184a14fa813d834283562fe39c83ed12ee10c9", + "papers/reviews/current_round1_statistical.md": "5244b774d71d55a7237d4678ab7242df6e69a4dd72658ebb74bc7a93dcc7f5cc", + "papers/reviews/current_round2_statistical.md": "5cdfcb66a68ae1931e9731104a2259b4508fb7ee8656c95a8938538e4ace54d6", + "papers/reviews/current_verification.md": "f9061dd62713efbd8fb8f3432b37139959790049f28b1bfa0488a9085a8a84c7", "papers/reviews/round2_referee_report.md": "b58e5c64701ad591959f5b3096a97ab3b1c6f767c3d626c4e0e01b2f3eff44cd", + "papers/reviews/round4_editorial_decisions.md": "61114cafeb8b7206c83c843acb40bc96fda2dc34baa025234b04269719fc39ea", + "papers/submission/FlyLab_manuscript.docx": "4c6d5da13b9acad3254b0af2d1ddf162d7f83094441c416b57a6322dd217e8ed", + "papers/submission/FlyLab_manuscript.json": "d7279749cedc85935430e5e9aab3ae1c1dfc5dcd46c0635e795c8fac3ddec11e", + "papers/submission/README.md": "20f2fa9c2f39901d1e9cd37f74ebc68a2984e8a64a79e64a335884ed9504c0b9", "papers/tables/T0_labellar_types.csv": "d49e71e9dfa9dcc10a38a2d062619d8f834dbec631a2e59d86c73b1f97259298", "papers/tables/T0_labellar_types.md": "3c09c0c2e46e0f54c9542a043604295a208a5705f03b8ee4872b7178aa5ca6e2", "papers/tables/T10_evidence_types.csv": "d772d1ee5abc3ad54d693ece39b2f1d0df8c1ac9661c6ac350f7d0309c2dc2ad", @@ -240,23 +255,23 @@ "papers/tables/T11_permutation_sweep.csv": "8ff2159495b6e153d502da3af1fe0b4b60a895e6454ed8cf10664d2f0ccf1cd1", "papers/tables/T11_permutation_sweep.md": "d853419ef147ce80ef1081584ce8bb2afcea8af0f0a5ab568b8a326cf6088e51", "papers/tables/T12_dependence_landscape.csv": "913c269f1011a0c161aa2375b04efa6b7e9f0c0264e9b6c1d33c6f50cc3a60e2", - "papers/tables/T12_dependence_landscape.md": "1c00f1d0d6053d06345f99a5d303620ca8d6a98a32cf898978deb37c0e69f8db", + "papers/tables/T12_dependence_landscape.md": "cfc014521ba420a8fc7f87593e504c6d0379012fc44bbc726310018bb456b0d2", "papers/tables/T12b_dependence_landscape_taste_motor.csv": "87af2cc0d5da3e390358a8156ce134b3f8f45c1e2e70b24063f662f12a08158b", - "papers/tables/T12b_dependence_landscape_taste_motor.md": "f558db4a5fca7d410fb97637d2e9fc1c9e0ab7445abedf57ad7db27c38f5df17", + "papers/tables/T12b_dependence_landscape_taste_motor.md": "1cd4df9a7fb078d9c0720ef7962b0e1bcb76c3a6a42a50e2c85eb4976f4e3f1a", "papers/tables/T13_ablation.csv": "68e3fbe9192b84d3398e103f8f9ab37aae56e5538db488dc81e53139df8950c1", - "papers/tables/T13_ablation.md": "826c1aeeca9fde8c3c53884e883fe7f8099a37ff0762e312abbb8bc2dcb1eedb", + "papers/tables/T13_ablation.md": "2ba1950c2f26861893434e590ba7ac4551b5f095a9c4cff640e910d11188fcdd", "papers/tables/T13b_ablation_information.csv": "5b54afed1d64d30a820d29030864d9bfeb344e3e5f575ee4a243f813642192da", - "papers/tables/T13b_ablation_information.md": "7de9b02ad7f919d8e5c3a1ef3037418cd0138040e3a0402614df1fda8cbaefbd", + "papers/tables/T13b_ablation_information.md": "c9160e14e76c359b23e291fc7505c5a49786d721cc73ddae5da76151d189fd5f", "papers/tables/T14_conclusion_stability.csv": "514f1245a9de829445fec6edbf0341fd62e559454541b7bdd35cb7a426bceb7e", - "papers/tables/T14_conclusion_stability.md": "a0bdb6653b01456f079416beda7497e05589091a201c88b354ac52b88d7ad8e4", + "papers/tables/T14_conclusion_stability.md": "7116a3ee78d96d5b71e65bbf4007582aaf325d411a5174c8356bc7fee2d4585c", "papers/tables/T14b_specification_diagnostics.csv": "7669deb2887fb69f6a6bbfa0f3affb80055ad72ca55b9da0b2a56c724885e586", "papers/tables/T14b_specification_diagnostics.md": "8afdc57d37a4e25b9c96881091482464ad93d094c25243505419a0be06d2a3e2", "papers/tables/T15_threshold_grid.csv": "84601c9a63851da30268ead0789a03d88c930dacd282d5b42c6b1d1ee59b56ca", "papers/tables/T15_threshold_grid.md": "da2a3e642512b52156b7e3fceb145e3c568b67b1acd9c9702d1686ecc5460ddc", "papers/tables/T16_uncertainty_budget.csv": "e7b14de72c72916c027b039e5e07394be473f7699445b56866c10931bec7d73a", - "papers/tables/T16_uncertainty_budget.md": "cdff58f5e9127ca4bcaf8cb267980aa7f13643254411ae4cea134fffb80ce2e8", + "papers/tables/T16_uncertainty_budget.md": "e1293704e54aecea97ebd472c35cfb5c4436f2104e9e61cdc8f0f76927911e65", "papers/tables/T17_value_of_information.csv": "0383106c6b196411536ed5f4dc4844f0052e23d299fd40db5e59b5c3d1052e63", - "papers/tables/T17_value_of_information.md": "0e4aa934ddf448ede83535f3b8944a0f2bd0e002efd73368fb5337ce59ea4daf", + "papers/tables/T17_value_of_information.md": "3d6cd0ece3f32e97f49bf0b363832a1551be9f9428ed2eb41991f8bd2957a875", "papers/tables/T18_claim_provenance.csv": "a9ce64585a892ae6917bd92a4863f2c8d13d170d8e350cabaead19089c144587", "papers/tables/T18_claim_provenance.md": "c6ffe932b0fea59efea6dc5668f692b73221a19d4f1e385a4baffbeaf167b043", "papers/tables/T19_cut_census.csv": "baa387d8621ed7cc0978e707d03ec304a2b92bb78fdbc950dcc9638b819e438b", @@ -264,11 +279,11 @@ "papers/tables/T1_library.csv": "e910be5808eabb7d73ed06583722c16619ad15abed4aee351a7909e22b94f8ea", "papers/tables/T1_library.md": "a2f5ca4b388ebf6d07f2845fa43e5ed1a1f3f49c3ac0056d1bf0421fefa7436c", "papers/tables/T20_transformation_rule.csv": "204aa33dd9d26703768025fb07bb98afa8ded052683110e03fd30fdcb4bcaf53", - "papers/tables/T20_transformation_rule.md": "5b47011470411ee6647ae181f2c6ffe44d4c9967f117a7bf5e38d6180c75cadb", + "papers/tables/T20_transformation_rule.md": "c98948df8d0fa85f70e924b742debf85d8a2f8db1c6c37f80f3f2377f5473709", "papers/tables/T21_transmitter_signs.csv": "a116af54def25de5e2bb86a0b133460aee2b2e15db7d424115cd3d8d7b9c133b", "papers/tables/T21_transmitter_signs.md": "92f018dd28843692f00d08037192930b53ee2cc524597241120be914ff5cedb8", "papers/tables/T22_instrument_validation.csv": "f8c3949515d6c250e6edc3f0a6f46ffcdff8a77b9d378d87a02d7948ef4a302c", - "papers/tables/T22_instrument_validation.md": "3c6ea0d076897d92cbbf97878234fa92f2f6f92ba48bcec756fe15144f0c8f3f", + "papers/tables/T22_instrument_validation.md": "556ba00deb758329c8c7b25ed882b427330143fecf723a67b17ae1577d0ac635", "papers/tables/T23_transmitter_null_balance.csv": "e4637660da0612c15fe5ab4138d757c1382a6b4bc697a530e584711b343ba8e6", "papers/tables/T23_transmitter_null_balance.md": "190f7af5782f8169875f3fef5f151db15352a3a0f518d1ab690eb83b0a4042e9", "papers/tables/T24_composition_reference.csv": "cc66a68aa48ef82c1597fed6b257dd4ebf123f85fd14a785448d07bb066bdd9d", @@ -278,7 +293,7 @@ "papers/tables/T26_scale_frontier.csv": "f2a497c778e69ff6906f263bf664c18be1bf9657e8d339b94844e4d6f0bee52a", "papers/tables/T26_scale_frontier.md": "c942dc98d36c845b22f9da433e1f6d3dfa1d1925c7576118e3473ae15cf3a594", "papers/tables/T27_scale_study.csv": "71493be2c495cbebe866a187f0a380dbc1e9c9b2e4f00de3983bd48290b78ad5", - "papers/tables/T27_scale_study.md": "9a01845e385d98af33de3073123dc5422cdd97779bf773ec67cf526ef921ab0a", + "papers/tables/T27_scale_study.md": "1579c53bc03df6edf4aa0bd9c2bca8d1430fa3d84fc2ea485100ce52d954cc68", "papers/tables/T2_mechanisms.csv": "0a7bb3009746100d51946e590d5a7eb23a628e67ea4d99967c3504a1f3deed59", "papers/tables/T2_mechanisms.md": "01adf76048f99904933dae79fb8a5e4e41625c477cf344eff889db73982ba102", "papers/tables/T3_predictions.csv": "07c2b18ae2d0dfc01fbc74046ce80db268a4fa607eb86660e5de180402a63dcb", @@ -288,7 +303,7 @@ "papers/tables/T5_mixtures.csv": "89faa9e76bc1313d27aa1369973ef404362c0f52f34830521bd71b3dd9ca91a6", "papers/tables/T5_mixtures.md": "3e6f65096e6f9b07b686089f4f76c4b332191deeef846197cdd0a707775bc3d8", "papers/tables/T6_dependence_profile.csv": "6fefaee56cfb66c84e9128b3a198da4f2236dcd9ad01495f7518c605f31c93be", - "papers/tables/T6_dependence_profile.md": "fd383f7e44430e3abde00e7601df427c90a1cf81df6e42c7aeae152b1b4febe0", + "papers/tables/T6_dependence_profile.md": "e98c1c2fe2504b3a4fae44b76267cda532a417da45ec97279436857afd4e3d9f", "papers/tables/T7_top_paths.csv": "42c9802e0901a9ec55b5c86eb0c5aa81d548dc2807c7ad67c4d73bb50e1dafd9", "papers/tables/T7_top_paths.md": "898e2c0fc95b89c9c57b75fe8338c9a80248a0915febdc650c3058a86596aa52", "papers/tables/T8_landscape.csv": "415ec69baf7cef49aa8bedfb87aa46c503e3e6fb5f16464bf82d46756ada5c1b", @@ -298,15 +313,17 @@ }, "fatal": true, "kind": "files", - "n_entries": 86, + "n_entries": 97, "patterns": [ "papers/results.json", "papers/*.md", + "papers/*.md.in", "papers/figures/*", - "papers/tables/*" + "papers/tables/*", + "papers/submission/*" ], - "sha256": "90db07af3e8418052c470b06c8cdd57bd19fb9081b571c7a39090aeed15b0b74" + "sha256": "622162378b0524433bda6154765233011256533f897f7f209fc8980529483975" } }, - "sha256": "011a7800b53b5fde49eae1ae4111245b9d2597dce6a00c97a9da0c4443e180fb" + "sha256": "68d5285559c527bb4b88d95751ad20047b03152d471a3e8d2e35522d8864553b" } diff --git a/docs/ARCHITECTURE.md b/docs/ARCHITECTURE.md index 2a6aa4d..47865c1 100644 --- a/docs/ARCHITECTURE.md +++ b/docs/ARCHITECTURE.md @@ -178,10 +178,10 @@ flylab/analysis/dependence.py compares a real-graph drug contrast with degraded- It reports one of three verdicts per null mode, never two: - **distinguishable**: the empirical permutation probability is at or below alpha, so the real graph and that graph model give different drug effects, -- **equivalent within tolerance**: the test did not reject and the gap between the real effect and the null ensemble's median is below a prespecified margin, +- **equivalent within tolerance**: the test did not reject and the gap between the real effect and the null ensemble's median is below a configured margin, - **indeterminate**: the test did not reject and the gap is not below that margin, so the result is consistent with a difference this study cannot resolve. -The margin is prespecified rather than fitted, and is by default a small fraction of the vehicle readout of the real graph. Only the equivalence verdict licenses saying a graph model *reproduces* an effect; a bare non-rejection is indeterminate and must be written as such. +The margin defaults to a small fraction of the vehicle readout. The legacy equivalence label is a descriptive point-gap diagnostic: it supplies neither a confidence interval nor a formal equivalence test. It does not license saying that a degraded graph reproduces the effect. The instrument itself is tested against planted ground truth rather than assumed to work: dependence.py provides synthetic_cut, ladder_recovery and ladder_power, which plant a recurrent loop of known strength, check that the ladder calls it topology-dependent and the unplanted control not, and map detection rate against effect size and permutation count. @@ -223,7 +223,7 @@ This is value with respect to model uncertainty, not biological or clinical valu flylab/analysis/scale.py runs the *same* dependence profile across a ladder of nested cuts of increasing size, so a verdict can be compared across scale without the construction changing underneath it. It reimplements no statistic: it calls dependence_profile unchanged and adds the bookkeeping that makes the comparison honest — a permutation budget that falls with the edge count and refuses to drop below the resolution floor, a per-rung record of the n actually used, a recipe filter so that hops-limited neighbourhoods are not compared against nested ranked-BFS cuts as if the difference were scale, and a stability criterion under which one rung of agreement does not count as settled. -Its result is the paper's main finding: the dependence verdict is a joint property of the prediction and the extract, and the extract's recurrence rather than its node count predicts which way it goes. +The observed dependence verdict changes across extracts and is not ordered by node count alone. Recurrence covaries with density, composition and extraction recipe; the study does not isolate it as a cause or validated predictor. Matching labels at the two largest rungs (20 permutations each) do not establish convergence. The paper pipeline imports the saved scale study; it does not recompute its uncommitted large cuts. ### Claim provenance diff --git a/docs/NOVELTY.md b/docs/NOVELTY.md index 17c8ea5..c340929 100644 --- a/docs/NOVELTY.md +++ b/docs/NOVELTY.md @@ -62,7 +62,7 @@ The dependence framework uses degraded graph ensembles and empirical permutation flylab/analysis/dependence.py reports one of three verdicts per null mode, and the safe interpretation follows them: - a small empirical p means the real-graph effect is **distinguishable** from that null ensemble at the chosen permutation effort, -- a non-rejection whose gap from the null ensemble's median is below the prespecified margin is reported as **equivalent within tolerance**; the margin is declared in advance, not fitted, and defaults to a small fraction of the vehicle readout, +- a non-rejection whose gap from the null ensemble's median is below the configured margin is reported as **equivalent within tolerance**; the margin is a reporting convention, not an externally registered equivalence bound, and defaults to a small fraction of the vehicle readout, - any other non-rejection is **indeterminate**: consistent with a difference this study cannot resolve, and not evidence of equality or biological equivalence. Accordingly, prefer: @@ -73,19 +73,19 @@ over: > "The degree-preserving graph reproduces the effect." -Only the equivalent-within-tolerance verdict licenses the word "reproduces", and even then it means "within the prespecified margin", never "the shuffle gave the same effect". Check the verdict against a run rather than assuming: for imidacloprid at the headline dose and n = 1000, the weight, degree-preserving and weight-matched transmitter modes fall inside the margin and report **equivalent_within_tolerance**, while the plain transmitter permutation reports **indeterminate** (gap 2.65 Hz against a margin of 0.33 Hz) and the Erdos-Renyi control is **distinguishable**. +The legacy equivalent-within-tolerance label is only a non-rejection plus a small point gap. It is not a confidence-bounded equivalence test and does not license the word "reproduces". Check the verdict against a run rather than assuming: for imidacloprid at the headline dose and n = 1000, the weight, degree-preserving and weight-matched transmitter modes fall inside the margin and report **equivalent_within_tolerance**, while the plain transmitter permutation reports **indeterminate** (gap 2.65 Hz against a margin of 0.33 Hz) and the Erdos-Renyi control is **distinguishable**. -### The verdict depends on the extract, recurrence predicts it, and that is now the result +### The verdict depends on the extract The single most important thing to know before writing any dependence sentence: **the same analysis, with the same instrument and no parameter changed, gives opposite answers on different extracts of one connectome.** On the sparse 1-hop `named` cut — an in-star, most of whose edges terminate on four seed cells and most of whose nodes receive no input — most cells come back composition-dominated. On the denser `taste_motor` cut, none does. -The scaling study (`flylab.analysis.scale.dependence_vs_scale`, T27) then says what the inversion tracks, and the answer is **not** size. Across seven extracts the composition-dominated verdict survives on exactly one — the in-star — while `scale_1k` has *fewer* nodes than `named` (1000 against 1126) and a mean degree of 22.9 against 1.21, and is already topology-dependent. At the 5k and 10k rungs the prediction is distinguishable from every null on the ladder. A degree-preserving rewire of an in-star is close to the identity, so a negative topology verdict there is weak evidence rather than a finding. Three consequences: +The scaling study (`flylab.analysis.scale.dependence_vs_scale`, T27) shows that node count alone does not order the observed classes. It does not separate recurrence from density, composition or extraction recipe. Across seven extracts the composition-dominated verdict survives on exactly one — the in-star — while `scale_1k` has *fewer* nodes than `named` (1000 against 1126) and a mean degree of 22.9 against 1.21, and is already topology-dependent. At the 5k and 10k rungs the prediction is distinguishable from every null on the ladder. A degree-preserving rewire of an in-star is close to the identity, so a negative topology verdict there is weak evidence rather than a finding. Three consequences: 1. **Never state a dependence class without the extract it was measured on**, and quote `flylab.analysis.dependence.cut_census` (mean degree, share of edges onto seeds, share of nodes with any input, recurrence budget) beside it. -2. **"Most predictions do not need the connectome" is withdrawn**, along with "the topology-dependent set is exactly the chloride-channel blockers". What replaces them is a methodological claim: the dependence verdict is a joint property of the prediction and the extract, **recurrence rather than size predicts which way it goes**, and that is the transferable result. -3. **Do not quote the top of the ladder as strong evidence.** The 25k and 50k rungs run at n = 20, where the smallest attainable probability is 0.048 against α = 0.05; they confirm a verdict settled at better-powered rungs rather than establishing it. +2. **"Most predictions do not need the connectome" is withdrawn**, along with "the topology-dependent set is exactly the chloride-channel blockers". What replaces them is a methodological claim: the dependence verdict is a joint property of the prediction and the extract, **node count alone does not order the observed classifications**, and that is the transferable result. +3. **Do not quote the top of the ladder as strong evidence.** The 25k and 50k rungs run at n = 20, where the smallest attainable probability is 0.048 against α = 0.05; their matching labels do not establish convergence or independently confirm the lower-rung result. -The instrument itself is not in doubt here: `ladder_recovery` and `ladder_power` plant a recurrent loop of known strength in a synthetic graph of matched size, density and composition, recover it, and give an empirical false-positive rate on the unplanted control. The reversal is therefore about the graphs, not about power. +The synthetic recovery grid is a limited positive-control check, with six networks per setting and broad uncertainty. The unplanted control also differs in cycle edges and transmitter labels. It neither validates all null models nor establishes power for the taste-motor ratio endpoint. ## Current headline comparisons @@ -111,7 +111,7 @@ Do **not** summarize it as: That statement is biologically inaccurate and too categorical for the current exploratory classification. The topology-sensitive set is chlorpyrifos_oxon, dieldrin, fipronil, gaba, ivermectin and picrotoxin, and three of those are not chloride-channel blockers: gaba is an RDL agonist, ivermectin a GluCl opener, and chlorpyrifos_oxon an AChE inhibitor that appears at one concentration only. -Multiplicity is handled rather than pending. The landscape applies Benjamini-Hochberg across its structural tests, records p_adjusted, q_value and fdr_alpha per cell, classifies on the adjusted values while keeping the raw ones beside them, and labels the run **confirmatory** or **exploratory** according to whether it was prespecified and whether its permutation resolution can support a rejection at the adjusted threshold at all. Any count taken from the landscape must be quoted with that label. +The exploratory landscape applies Benjamini-Hochberg across two structural mode tests per cell and retains both raw and adjusted probabilities. Classes based on adjusted component tests are not themselves guaranteed cell-level FDR control. The dependence conditions for BH have not been established. Permutation effort does not turn a retrospective analysis into a confirmatory study. Describe the landscape as: diff --git a/docs/RESEARCH_MAP.md b/docs/RESEARCH_MAP.md index f58e92c..51d4bef 100644 --- a/docs/RESEARCH_MAP.md +++ b/docs/RESEARCH_MAP.md @@ -44,12 +44,12 @@ Public feathers (`body-annotations…`, `body-neurotransmitters…`, `connectome | 5 Live fly | PER / climbing *n* | **blocked — protocol only** | | 6 IJRC | research article with a Discussion, or a live table | **v0.6.1 draft complete after a second referee round, not submitted** | | 6a Typed evidence | parameter type + ordered evidence distance decide the transformation, under a checked soundness invariant | **done in v0.6, graded and proved in v0.6.1** | -| 6b Connectome dependence | permutation *p*, dependence class, necessary information level | **done in v0.6**; v0.6.1 adds the three-way equivalence verdict, a weight-matched transmitter null, a relative effect floor, FDR with confirmatory/exploratory labelling and ground-truth validation of the ladder | +| 6b Connectome dependence | permutation *p*, dependence class, necessary information level | **done in v0.6**; v0.6.1 adds a three-way descriptive median-gap diagnostic, a weight-matched transmitter null, a relative effect floor, BH-adjusted component tests in an exploratory landscape and ground-truth validation of the ladder | | 6c Ablation ladder | which layer carries the ordering information | **done in v0.6**; v0.6.1 gives level C a direction-aware rule, scores on signed ρ and adds a matched reference distribution | | 6d Specification robustness | 25 gain specifications x 7 conclusions + threshold grid | **done in v0.6**, but its topology rows were vacuous until **v0.6.1** | | 6e Global uncertainty + VOI | Sobol' budget, validated estimator, ranked experiments | **done in v0.6** | | **6f Substrate check** | repeat every dependence claim on a cut that is not an in-star | **done in v0.6.1 — and the verdict reversed** | -| **6i Scale ladder (new)** | the same profile across nested extracts of 1k-50k cells | **done in v0.6.1** — the verdict settles as topology-dependent, and recurrence rather than size predicts the inversion (T27) | +| **6i Scale ladder (new)** | the same profile across nested extracts of 1k-50k cells | **done in v0.6.1** — the observed class remains topology-dependent across the tested clean rungs; recurrence is not isolated causally (T27) | | 7 Expression coverage | adult motor-neuron receptor expression, or a declared bound | **open — 0.205, MN9 unresolved** | | 7b Better power at the top of the ladder | the 25k/50k rungs run at n = 20 (resolution 0.048) | **open** — they confirm the settled verdict rather than establishing it | | 8 Archival | Zenodo deposit; only then may "prospective" become "pre-registered" | **open** | @@ -86,10 +86,10 @@ H1 dual panel gap at 1 µM · H2 imidacloprid lowers the neighbourhood mean rate Every number here comes from `papers/results.json`. -1. **The dependence verdict is a joint property of the prediction and the extract, and recurrence predicts which way it goes.** Over 21 compounds × 4 concentrations at n = 1000 with Benjamini-Hochberg and a 1 % relative effect floor, the 1-hop `named` cut gives 11 topology-dependent, 42 composition-dominated and 31 no-effect; repeated unchanged on the denser `taste_motor` cut it gives 54 / 0 / 30. The scaling study over seven extracts (T27) shows the composition-dominated verdict surviving on exactly one of them — the in-star — while `scale_1k`, with *fewer* nodes than `named` and twenty times the mean degree, is already topology-dependent. **"Most predictions do not need the connectome" is withdrawn**, and so is "exactly the chloride-channel blockers". Report the extract's structure (T19, T27) with any dependence claim. -2. **The bitter-veto arm is a properly powered negative.** At n = 300 on the taste-motor cut, fipronil's veto ratio gives p = 0.42 / 0.88 / 0.20 / 0.24 with all |z| < 0.25, replacing the ten-shuffle hand-wave of v0.5. H6 is a pharmacology prediction, not a wiring prediction, and that is now measured rather than assumed. Imidacloprid's veto ratio there is **undefined** (MN9 silenced), not zero. +1. **The dependence verdict is a joint property of the prediction and the extract, with recurrence a candidate associated feature.** Over 21 compounds × 4 concentrations at n = 1000 with Benjamini-Hochberg and a 1 % relative effect floor, the 1-hop `named` cut gives 11 topology-dependent, 42 composition-dominated and 31 no-effect; repeated unchanged on the denser `taste_motor` cut it gives 54 / 0 / 30. The scaling study over seven extracts (T27) shows the composition-dominated verdict surviving on exactly one of them — the in-star — while `scale_1k`, with *fewer* nodes than `named` and twenty times the mean degree, is already topology-dependent. **"Most predictions do not need the connectome" is withdrawn**, and so is "exactly the chloride-channel blockers". Report the extract's structure (T19, T27) with any dependence claim. +2. **The bitter-veto arm is a non-rejection without endpoint-specific power calibration.** At n = 300 on the taste-motor cut, fipronil's veto ratio gives p = 0.42 / 0.88 / 0.20 / 0.24 with all |z| < 0.25, replacing the ten-shuffle hand-wave of v0.5. H6 has no demonstrated wiring dependence on this endpoint; wiring independence is not established. Imidacloprid's veto ratio there is **undefined** (MN9 silenced), not zero. 3. **The ordering information is in the mechanism rules — but the number that says so is nearly an identity.** Composition-only tracks the full model's ordering (ρ 0.988 at 1 µM), yet pharmacology-free pseudo-compounds reach a median of 0.847 and permuting compound labels leaves it identically 0.988; under a degree-corrected engine it falls to 0.666 (and -0.372 at 10 nM). Receptor-only orders the library backwards (ρ -0.506 to -0.046). Topology-only with a **direction-aware** generic rule reaches ρ 0.961, against 0.242 for the old depression-only rule — so "the connectome without the pharmacology is not a cheap substitute" is **withdrawn**. -4. **The suppression result is specification-dependent, and the old 25/25 was vacuous.** C1 is retained by 15 of 25 and reversed by all 10 monotone rules. The v0.6 topology rows were computed at 6 shuffles and under a specification context that never reached the null engine; recomputed at 100 shuffles with FDR across 100 structural tests, C3 holds 25/25 of which only 17 reach equivalence and 8 are indeterminate. Nicotinic buffering 18/25; Nav/AChE amplification 19/25. Retained by every specification: C2_rdl_disinhibition, C3_imidacloprid_topology_not_distinguishable, C4_fipronil_topology_exceeds, C7_map_bitter_veto. +4. **The suppression result is specification-dependent, and the old 25/25 was vacuous.** C1 is retained by 15 of 25 and reversed by all 10 monotone rules. The v0.6 topology rows were computed at 6 shuffles and under a specification context that never reached the null engine; recomputed at 100 shuffles with FDR across 100 structural tests, C3 holds 25/25 of which only 17 pass the descriptive median-gap check and 8 are indeterminate. Nicotinic buffering 18/25; Nav/AChE amplification 19/25. Retained by every specification: C2_rdl_disinhibition, C3_imidacloprid_topology_not_distinguishable, C4_fipronil_topology_exceeds, C7_map_bitter_veto. 5. **The uncertainty is in two places only.** Sobol' at n_base 1024 (11 264 evaluations, Var(Y) = 3.149 Hz²): the only factors whose bootstrap first-order interval excludes zero are gain_transform, weight_threshold (S1 0.410 and 0.294). The estimator's noise floor is 0.046, set by gain_coef rather than by the null factor `lif_seed` (-0.007); `drive` sits inside it. Interactions carry 0.344 raw, 0.272 clipped. 6. **The next experiment is named by the model, not by taste.** VOI: `gain_transform` 1.29 Hz² (a synaptic-gain calibration at a known synapse) and `weight_threshold` 0.925 Hz² — which needs **no experiment at all**, only a re-run against synapse-confidence strata. 7. **The amplify/buffer split is threshold-sensitive.** Nicotinic buffering holds 9/9 on the 3×3 grid (gap −0.362 to −1.224); Nav/AChE amplification holds 8/9 and flips at the strictest corner (gap −0.075). Quote ranges, never the point estimate. diff --git a/flylab/analysis/dependence.py b/flylab/analysis/dependence.py index a722bfd..4e172a1 100644 --- a/flylab/analysis/dependence.py +++ b/flylab/analysis/dependence.py @@ -16,9 +16,9 @@ The information ladder ---------------------- -Each null keeps a different amount of the real connectome, so a drug effect -that is distinguishable from a given null tells you that *that* level of -information is necessary: +Each null alters a different part of the real connectome. A drug effect that +is distinguishable from one null establishes a contrast with that degradation; +it does not by itself prove that a general level of information is necessary: ====== ============================ ================================================ rank graph model (null mode) what the model still knows @@ -49,11 +49,12 @@ ``distinguishable`` the empirical two-sided permutation probability is at or below ``alpha``. The real graph and this graph model give different drug effects. -``equivalent_within_tolerance`` +``equivalent_within_tolerance`` (legacy identifier) the test did not reject **and** the absolute gap between the real effect and the null ensemble's median is smaller than a prespecified margin - ``delta``. Only this verdict licenses the phrase "this graph model - reproduces the effect". + ``delta``. This is a descriptive point-gap check, not a formal + equivalence test and not evidence that the graph model reproduces the + effect. The identifier is retained for output compatibility. ``indeterminate`` the test did not reject and the gap is not smaller than ``delta``: the data are consistent with a difference this study cannot resolve. This is @@ -74,21 +75,23 @@ readout is undefined or zero no margin exists, equivalence cannot be claimed, and every non-rejected mode is ``indeterminate``. -Multiplicity ------------- -A single prespecified profile (:func:`dependence_profile` for imidacloprid or -fipronil at ``PAPER_N``) is a **confirmatory** test and needs no correction. -The 84-cell landscape is not: it runs 21 compounds x 4 concentrations x 4 -modes, and turning the resulting count into a claim without correction inflates -it. :func:`dependence_landscape` therefore applies Benjamini-Hochberg FDR -across the structural tests of the run (:data:`STRUCTURAL_MODES` x cells), -reports ``p_adjusted`` / ``q_value`` / ``fdr_alpha`` per cell, classifies on -the adjusted values and keeps the raw ones beside them, and reports **both** -counts so that the change in the headline number is visible. A landscape is -labelled ``exploratory`` unless its permutation resolution ``1/(n+1)`` is fine -enough for the adjusted threshold ``fdr_alpha / m`` to be reachable by a single -test; when the resolution cannot support any rejection at all, the result says -so instead of silently reporting zero topology-dependent cells. +Multiplicity and design status +------------------------------ +Profiles are exploratory by default, including the high-effort imidacloprid +and fipronil reference profiles at ``PAPER_N``. Computational effort and +compound identity do not make a retrospective analysis confirmatory. A caller +may retain the legacy ``confirmatory=True`` flag, but the result records that +this is an unverified caller assertion; this module does not verify a dated +protocol or preregistration. + +The landscape is always exploratory. It applies Benjamini-Hochberg across +the structural **component tests** of the run (:data:`STRUCTURAL_MODES` x +cells), reports adjusted component probabilities and reclassifies cells from +those component decisions. That sensitivity analysis does not establish +cell-level FDR control for the compound-level OR decision, and reuse of shuffle +draws does not verify the dependence conditions required by ordinary BH. +Permutation resolution describes what the component tests can resolve; it +does not change the design into a confirmatory one. Classification rule (a label, not a hypothesis test) ---------------------------------------------------- @@ -101,11 +104,10 @@ * ``no-effect`` - ``|e| <= effect_floor``. * ``topology-dependent`` - at least one *structural* null (``weight_permute`` or ``rewire_degree_preserving``) is distinguishable from - the real graph: the prediction needs the wiring pattern, not just the - graph's size or census. + the real graph: the effect differs from at least one selected structural + degradation under this cut, engine and readout. * ``mixed`` - no structural null is distinguishable but ``sign_permute`` is: - the prediction needs to know *which* cells carry which transmitter, but not - the wiring pattern itself. + the joint sign/target-assignment null differs on this readout. * ``composition-dominated`` - neither, whether or not ``erdos_renyi`` is distinguishable: no structure-preserving degradation could be told apart from the real graph on this readout. Whether that is *equivalence* or @@ -204,6 +206,7 @@ "DEFAULT_EFFECT_FLOOR_FRAC", "effective_effect_floor", "DEFAULT_DELTA_FRAC", + "REFERENCE_COMPOUNDS", "CONFIRMATORY_COMPOUNDS", "FAST_N", "DEFAULT_N", @@ -362,11 +365,11 @@ #: change :mod:`flylab.analysis.selectivity` treats as a circuit response. DEFAULT_DELTA_FRAC = 0.05 -#: The compounds whose profiles the paper prespecifies as confirmatory tests. -#: A profile is ``confirmatory`` only for one of these, at ``n >= PAPER_N``; -#: everything else (in particular the landscape) is exploratory and is -#: multiplicity-corrected instead. -CONFIRMATORY_COMPOUNDS: tuple[str, ...] = ("imidacloprid", "fipronil") +#: Retrospectively selected high-effort reference profiles. The historical +#: ``CONFIRMATORY_COMPOUNDS`` name is retained as an API alias only; membership +#: never changes design status automatically. +REFERENCE_COMPOUNDS: tuple[str, ...] = ("imidacloprid", "fipronil") +CONFIRMATORY_COMPOUNDS: tuple[str, ...] = REFERENCE_COMPOUNDS #: ``--fast``-style n ladder, so one landscape function serves three callers. #: ``FAST_N`` sits just above the smallest n at which anything can be @@ -608,8 +611,8 @@ def benjamini_hochberg( permutation budget rather than a result. ``resolution_supports_single_rejection`` True when ``resolution <= alpha / m``, i.e. one isolated test could be - rejected on its own. This is the condition a *confirmatory* landscape - has to meet. + rejected on its own. This is a resolution diagnostic and does not + confer confirmatory status on a landscape. """ import math @@ -701,12 +704,17 @@ def mode_verdict( alpha: float = DEFAULT_ALPHA, delta: float | None = None, ) -> dict[str, Any]: - """One mode's three-way verdict - the only claim this design supports. + """One mode's three-way descriptive diagnostic. ``distinguishable`` when the permutation test rejects; otherwise ``equivalent_within_tolerance`` when ``|real - median(null)| < delta``; - otherwise ``indeterminate``. A rejected test that *also* sits inside the - margin is still reported as ``distinguishable`` (the test is the decision + otherwise ``indeterminate``. + + This is not an equivalence test: no uncertainty interval for the gap is + tested against the margin. The legacy label is retained for API compatibility. + A missing p-value is unassessed, even if the point gap is small. + + A rejected result inside the margin is still ``distinguishable`` (the decision rule), but ``within_tolerance`` stays True so that "statistically distinguishable, practically small" is visible rather than hidden. """ @@ -724,17 +732,20 @@ def mode_verdict( f"alpha = {float(alpha):.3g}: the real graph and this graph model give " "different drug effects." ) - elif within: + elif within and p is not None: verdict = "equivalent_within_tolerance" why = ( f"the test did not reject (p = {'n/a' if p is None else format(float(p), '.4g')}) " f"and the gap between the real effect and the null median, {gap:.4g}, is " - f"below the prespecified margin delta = {float(delta):.4g}: this graph " - "model gives the same effect to within the margin." + f"below the configured tolerance delta = {float(delta):.4g}. This is a " + "descriptive point-gap check, not statistical equivalence or evidence " + "that individual null graphs give sufficiently close effects." ) else: verdict = "indeterminate" - if gap is None or delta is None: + if p is None: + why = "no permutation probability is available; this comparison is unassessed." + elif gap is None or delta is None: why = ( f"the test did not reject (p = {'n/a' if p is None else format(float(p), '.4g')}) " "and no equivalence margin could be formed, so nothing stronger than " @@ -879,15 +890,15 @@ def classify( why = ( "the effect is distinguishable from " + " and ".join(structural) - + ", so it needs the wiring pattern and not only the graph's size, " - "degree sequence and transmitter census." + + " under the selected engine, cut and threshold. This establishes " + "a contrast with those degradations, not a general necessity claim." ) elif sign_distinguishable: label = "mixed" why = ( "no structure-preserving null is distinguishable from the real " - "graph but sign_permute is: the prediction needs to know which " - "cells carry which transmitter, not the wiring pattern itself." + "graph but sign_permute is: the joint sign/target-assignment null " + "differs on this readout." ) else: label = "composition-dominated" @@ -904,15 +915,14 @@ def classify( ) why = ( "no structure-preserving degradation could be distinguished from " - "the real graph on this readout, so nothing here requires its " - "topology" + "the real graph on this readout; this is a set of non-rejections, " + "not evidence that topology is unnecessary" + ("; " + "; ".join(detail) if detail else "") + ( - ". A same-size random graph IS distinguishable (erdos_renyi), " - "so the prediction is not wholly network-insensitive." + ". A same-size random graph IS distinguishable (erdos_renyi)." if er_distinguishable else ". Even a same-size random graph is not distinguishable, " - "so the prediction is network-insensitive on this readout." + "which remains a non-rejection rather than equivalence." ) ) return { @@ -1076,11 +1086,12 @@ def verdict_of(row: dict[str, Any]) -> dict[str, Any]: equivalent = v["verdict"] == "equivalent_within_tolerance" if equivalent: description = ( - f"a graph model that keeps only {info['keeps']} is equivalent to " - "the real cut on this readout within the prespecified margin " + f"a graph model that keeps only {info['keeps']} has a median " + "point gap below the prespecified descriptive margin " f"delta = {float(v['delta']):.4g} (gap " - f"{float(v['abs_gap_from_null_median']):.4g}), so nothing more " - "detailed is necessary for this prediction." + f"{float(v['abs_gap_from_null_median']):.4g}). The legacy " + "equivalent_within_tolerance identifier does not denote a " + "formal equivalence test or prove that this level suffices." ) else: description = ( @@ -1191,12 +1202,12 @@ def dependence_profile( :func:`equivalence_margin` and the module docstring. The absolute value used is always recorded in ``delta``. - ``confirmatory`` marks a prespecified test. Left at ``None`` it is - inferred: True for one of :data:`CONFIRMATORY_COMPOUNDS` at - ``n >= PAPER_N``, False otherwise. A confirmatory profile is a single - planned comparison and carries no multiplicity correction; an exploratory - one says so in its warnings, and the corrected form of the exploratory - question is :func:`dependence_landscape`. + ``confirmatory`` is a legacy caller assertion. ``None`` always means + exploratory; neither compound identity nor ``n`` can confer prospective + design status. Passing ``True`` preserves API compatibility but the + result explicitly records that no dated protocol was verified. The + retrospective high-effort reference profiles are listed in + :data:`REFERENCE_COMPOUNDS`. Runtime is ``len(modes) * n`` shuffles; on the 1126-node ``named`` cut that is about 4.6 s per 100 shuffles across all four modes (the degree- @@ -1269,10 +1280,11 @@ def dependence_profile( warnings.append(w) rows.sort(key=lambda r: [m for m in modes].index(r["mode"])) - if confirmatory is None: - confirmatory = bool( - str(compound) in CONFIRMATORY_COMPOUNDS and int(n) >= PAPER_N - ) + caller_asserted_confirmatory = bool(confirmatory is True) + confirmatory = caller_asserted_confirmatory + high_effort_reference = bool( + str(compound) in REFERENCE_COMPOUNDS and int(n) >= PAPER_N + ) return _assemble_profile( compound=compound, conc_M=conc_M, @@ -1292,6 +1304,10 @@ def dependence_profile( warnings=warnings, runtime_s=time.perf_counter() - t0, confirmatory=bool(confirmatory), + confirmatory_basis=( + "caller_asserted_unverified" if caller_asserted_confirmatory else None + ), + high_effort_reference=high_effort_reference, ) @@ -1378,6 +1394,8 @@ def _assemble_profile( delta: float | None = None, real_vehicle: float | None = None, confirmatory: bool = False, + confirmatory_basis: str | None = None, + high_effort_reference: bool = False, ) -> dict[str, Any]: margin = equivalence_margin(real_vehicle, delta_frac=delta_frac, delta=delta) floors = effective_effect_floor(real_vehicle, effect_floor, effect_floor_frac) @@ -1426,13 +1444,19 @@ def _assemble_profile( + ". These modes support 'not distinguishable from this null " "ensemble' and nothing stronger." ) - if not confirmatory: + if confirmatory: + warnings.append( + "CONFIRMATORY STATUS IS CALLER-ASSERTED AND UNVERIFIED: the API " + "does not verify a dated protocol or preregistration. Treat this " + "profile as confirmatory only when external design documentation " + "supports the assertion." + ) + else: warnings.append( - "exploratory: this profile is not one of the prespecified " - "confirmatory tests (" + ", ".join(CONFIRMATORY_COMPOUNDS) - + f" at n >= {PAPER_N}) and carries no multiplicity correction of " - "its own. Counting verdicts over many such profiles needs the " - "FDR-controlled dependence_landscape instead." + "exploratory: this profile carries no independently verified " + "prospective design status and no multiplicity correction of its " + "own. Counting verdicts over many profiles requires an explicitly " + "defined multiplicity analysis." ) profile = { "compound": compound, @@ -1455,8 +1479,18 @@ def _assemble_profile( "delta_frac": margin["delta_frac"], "delta_scale": margin["delta_scale"], "confirmatory": bool(confirmatory), + "confirmatory_basis": confirmatory_basis, + "high_effort_reference": bool(high_effort_reference), "multiplicity_correction": None, - "design": "confirmatory (prespecified)" if confirmatory else "exploratory", + "design": ( + "confirmatory (caller asserted; protocol unverified)" + if confirmatory + else ( + "exploratory high-effort reference profile" + if high_effort_reference + else "exploratory" + ) + ), # permutation probabilities first, then the z profile D "p": P, "q": Q, @@ -1670,7 +1704,7 @@ def _ladder_on_state( "and exists to test the instrument rather than to describe a fly.", ], runtime_s=time.perf_counter() - t0, - confirmatory=True, + confirmatory=False, ) @@ -1977,7 +2011,7 @@ def dependence_landscape( n_jobs: int = 1, **kw: Any, ) -> dict[str, Any]: - """The compound x concentration dependence landscape, FDR-controlled. + """The exploratory compound x concentration dependence landscape. One cell per ``(compound, conc_M)``: its real effect, the per-mode permutation p and z, the three-way verdict, the dependence class and the @@ -1986,10 +2020,15 @@ def dependence_landscape( **Multiplicity.** A landscape is a screen, not a planned comparison: the default grid is 21 compounds x 4 concentrations x 4 modes, and the count of topology-dependent cells is the number the paper quotes. Benjamini- - Hochberg FDR is therefore applied across the structural tests of the run + Hochberg adjustment is therefore applied across the structural component + tests of the run (``fdr_modes`` x cells, ``fdr_alpha`` defaulting to ``alpha``), every cell carries ``p_adjusted`` / ``q_value`` / ``fdr_alpha``, and the class is - decided on the adjusted probabilities while the raw ones stay beside them. + decided on the adjusted component probabilities while the raw ones stay + beside them. BH control for those component hypotheses is not a guarantee + of cell-level FDR control for the compound-level OR decisions, and the + dependence assumptions of ordinary BH are not verified here. Treat this + as a multiplicity sensitivity analysis. The summary reports **both** counts (``n_topology_dependent`` and ``n_topology_dependent_raw``) so the effect of the correction on the headline number is visible rather than silent. @@ -2000,8 +2039,8 @@ def dependence_landscape( before anything can be rejected) and ``fdr.can_reject``; when nothing can be rejected the summary says so instead of reporting zero topology-dependent cells as if that were a finding. The landscape is - ``confirmatory`` only when its resolution is fine enough for one isolated - test to clear ``fdr_alpha / m``; otherwise it is labelled ``exploratory``. + always labelled ``exploratory``. Fine permutation resolution improves the + numerical component tests; it does not change retrospective grid design. ``gains_by_compound`` replaces the gain patch this landscape would otherwise compute from the library for the named compounds. It exists so @@ -2297,12 +2336,14 @@ def _apply_fdr( fdr_modes: Sequence[str], resolution: float, ) -> dict[str, Any]: - """Benjamini-Hochberg across the structural tests of a whole landscape. + """Benjamini-Hochberg across structural component tests of a landscape. Writes ``p_adjusted`` / ``q_value`` / ``fdr_alpha`` into every row of the family, records each cell's own ``q_value`` (its smallest adjusted structural probability), keeps the uncorrected class as ``class_raw`` and - re-decides ``class`` on the adjusted values. + re-decides ``class`` on the adjusted values. The resulting cell classes + are based on BH-adjusted component tests; this does not establish FDR + control for cell-level OR decisions. """ family = [m for m in fdr_modes if m in MODES] index: list[tuple[int, str]] = [] @@ -2346,6 +2387,9 @@ def _apply_fdr( cell.get("delta"), ) bh["modes"] = list(family) + bh["scope"] = "structural_component_tests" + bh["cell_level_fdr_controlled"] = False + bh["dependence_assumptions_verified"] = False bh.pop("adjusted", None) bh.pop("rejected", None) return bh @@ -2486,17 +2530,17 @@ def _landscape_result( "any is rejected, so an isolated strong cell cannot survive the " "correction at this n." ) - confirmatory = bool(bh.get("resolution_supports_single_rejection")) - if not confirmatory: - warnings.append( - "EXPLORATORY: this landscape's permutation resolution " - f"({res:.5f}) does not reach the adjusted threshold fdr_alpha/m = " - f"{q_alpha / max(bh['m'], 1):.6f}, so its counts are a screen to be " - "confirmed, not a confirmatory result. The prespecified " - "confirmatory tests are the " - + "/".join(CONFIRMATORY_COMPOUNDS) - + f" profiles at n >= {PAPER_N}." - ) + # Resolution is a property of the component tests, not of study design. + # A retrospectively assembled compound x concentration grid remains + # exploratory at every n. + confirmatory = False + warnings.append( + "EXPLORATORY LANDSCAPE: permutation resolution does not confer " + "confirmatory status. Benjamini-Hochberg is applied to structural " + "component tests as a sensitivity analysis; its dependence assumptions " + "are unverified, and the resulting compound-level OR classes are not " + "guaranteed cell-level FDR control." + ) unstable = [r for r in table if not r["stabilised"]] if unstable: warnings.append( @@ -2534,24 +2578,19 @@ def _landscape_result( f"multiplicity changes the headline count: {n_topo_raw} cells are " f"topology-dependent on the raw permutation p and {n_topo} survive " f"Benjamini-Hochberg at fdr_alpha = {q_alpha} over {bh['m']} " - "structural tests. The corrected count is the one to quote." + "structural component tests. Report both as an exploratory " + "sensitivity analysis; neither is a cell-level FDR guarantee." ) topo = sorted({r["compound"] for r in table if r["class"] == "topology-dependent"}) topo_raw = sorted({r["compound"] for r in table if r["class_raw"] == "topology-dependent"}) comp = sorted({r["compound"] for r in table if r["class"] == "composition-dominated"}) statement = ( - "exploratory screen, FDR-controlled: " + "exploratory classes based on BH-adjusted component tests: " f"{n_topo} of {len(cells)} cells are topology-dependent after " - f"Benjamini-Hochberg at fdr_alpha = {q_alpha} across {bh['m']} structural " - f"tests ({n_topo_raw} before correction)." - if not confirmatory - else ( - f"confirmatory-resolution landscape: {n_topo} of {len(cells)} cells " - f"are topology-dependent after Benjamini-Hochberg at fdr_alpha = " - f"{q_alpha} across {bh['m']} structural tests ({n_topo_raw} before " - "correction); the permutation resolution supports the adjusted " - "threshold for a single isolated test." - ) + f"Benjamini-Hochberg at alpha = {q_alpha} across {bh['m']} structural " + f"component tests ({n_topo_raw} before adjustment). This is a " + "multiplicity sensitivity analysis under unverified dependence, not a " + "guarantee of cell-level FDR control." ) if not bh["can_reject"]: statement += ( @@ -2579,17 +2618,24 @@ def _landscape_result( "p_resolution": res, "resolution_coarser_than_alpha": bool(res > float(alpha)), "confirmatory": confirmatory, - "design": "confirmatory" if confirmatory else "exploratory", + "design": "exploratory", + "design_basis": "retrospective compound-by-concentration landscape", "design_statement": statement, "multiplicity_correction": bh["method"], + "multiplicity_scope": "structural_component_tests", + "cell_level_fdr_controlled": False, + "bh_dependence_assumptions_verified": False, "fdr": bh, "shape": [len(compounds), len(concs)], "n_cells": len(cells), "cells": cells, "table": table, "summary": { - "design": "confirmatory" if confirmatory else "exploratory", + "design": "exploratory", "statement": statement, + "multiplicity_scope": "structural_component_tests", + "cell_level_fdr_controlled": False, + "bh_dependence_assumptions_verified": False, "class_counts": counts, "class_counts_raw": counts_raw, "class_counts_absolute_floor": counts_absolute_floor, diff --git a/flylab/analysis/robustness.py b/flylab/analysis/robustness.py index 1bc8cbc..2cd922b 100644 --- a/flylab/analysis/robustness.py +++ b/flylab/analysis/robustness.py @@ -68,9 +68,12 @@ Because 25 specifications x 2 compounds x 2 structural modes is a family of tests and not a planned comparison, Benjamini-Hochberg is applied across the -whole run and the verdicts are re-decided on the adjusted probabilities; +component tests as a multiplicity sensitivity analysis and the verdicts are +re-decided on the adjusted probabilities; ``matrix_uncorrected`` keeps the uncorrected view and every topology row carries ``fraction_retained_uncorrected`` beside its headline fraction. +This does not establish FDR control for the compound-level OR decisions, and +reuse of a shuffle stream does not verify ordinary BH's dependence conditions. One consequence is worth stating plainly, because it invalidates the v0.6 result rather than refining it: :func:`mechanism_spec` rebinds @@ -977,8 +980,9 @@ def _c_fipronil_topology_exceeds(run: SpecRun) -> bool | None: on the same engine, the same null draws and the same criterion that imidacloprid is measured against in C3. - The pair (C3 False-positive-free, C4 positive) is the contrast the paper - rests on: same cut, same readout, same shuffles, opposite verdict. + C4 is a thresholded rejection predicate: at least one of the two + structural component tests must reject. It is not a threshold-free + effect-size contrast. """ ev = run.topology_evidence("fipronil") run.evidence["C4_fipronil_topology_exceeds"] = ev @@ -1038,7 +1042,7 @@ def _c_map_bitter_veto(run: SpecRun) -> bool | None: "structural tests of this run; a failure to reject, not a claim of " "equivalence" ), - "kind": "permutation classification (failure to reject), FDR-corrected", + "kind": "permutation classification (failure to reject), based on BH-adjusted component tests", "topology": True, "expect_topology_dependent": False, "compound": "imidacloprid", @@ -1057,7 +1061,7 @@ def _c_map_bitter_veto(run: SpecRun) -> bool | None: "{alpha:g}, Benjamini-Hochberg across the {n_tests} structural " "tests of this run" ), - "kind": "permutation classification (rejection), FDR-corrected", + "kind": "permutation classification (thresholded OR rejection), based on BH-adjusted component tests", "topology": True, "expect_topology_dependent": True, "compound": "fipronil", @@ -1134,7 +1138,7 @@ def _topology_fdr( alpha: float, resolution: float, ) -> dict[str, Any]: - """Benjamini-Hochberg across every structural test of a stability run. + """Benjamini-Hochberg across structural component tests of a stability run. A single specification's topology cell is a planned comparison; 25 of them are not. The family is therefore all @@ -1142,7 +1146,8 @@ def _topology_fdr( correction, at the same alpha, that :func:`flylab.analysis.dependence.dependence_landscape` applies to its own grid -- and each cell's topology verdict is re-decided on the adjusted - probabilities. + probabilities. This is a sensitivity analysis; it does not guarantee FDR + control for the cell-level OR decisions. Returns the adjusted view keyed ``(spec, compound) -> bool`` plus the family's own diagnostics, including whether its resolution can support a @@ -1179,6 +1184,9 @@ def _topology_fdr( "q_value": q_by_cell, "family": [list(k) for k in index], "fdr": bh, + "scope": "structural_component_tests", + "cell_level_fdr_controlled": False, + "dependence_assumptions_verified": False, } @@ -1257,9 +1265,10 @@ def conclusion_stability( warnings.append( f"the topology conclusions were evaluated at n_shuffles = {shuffles} " f"per specification (permutation resolution 1/(n+1) = {resolution:.4f}). " - "C3 is an empirical two-sided permutation test at alpha = " - f"{TOPOLOGY_ALPHA}; C4 is an effect-size contrast with no threshold. " - "Neither uses a Gaussian z." + "C3 is retained when neither structural component test rejects at " + f"alpha = {TOPOLOGY_ALPHA}; C4 is a thresholded OR predicate retained " + "when at least one structural component test rejects after the stated " + "Benjamini-Hochberg adjustment. Neither uses a Gaussian z." ) if resolution > TOPOLOGY_ALPHA: warnings.append( @@ -1333,18 +1342,28 @@ def conclusion_stability( warnings.append( "every specification's topology cell is run on the SAME shuffle " f"stream (seed {int(seed)}), so the {n_structural_tests} structural " - "tests of this run are positively dependent rather than " - "independent. Benjamini-Hochberg controls the FDR under positive " - "dependence, but the corrected fractions should be read as a " - "property of this shuffle stream, not as 25 independent replications." + "tests of this run are dependent rather than independent. Reusing " + "draws does not verify the positive-dependence conditions required " + "by ordinary Benjamini-Hochberg. Treat the adjusted component-test " + "fractions as a multiplicity sensitivity analysis, not cell-level " + "FDR control or 25 independent replications." + ) + warnings.append( + "the adjustment is asymmetric across the topology conclusions: " + "raising component probabilities can support C4 only when a " + "rejection remains, but it can only make C3's failure-to-reject " + "predicate easier to retain. Do not describe C3 as strengthened " + "by or surviving correction; report its raw non-rejections and " + "descriptive margin breakdown separately." ) if topo_names and not topo_fdr["fdr"]["can_reject"]: warnings.append( - "FDR cannot reject anything across this stability run: with " + "the BH-adjusted component tests cannot reject anything across " + "this stability run: with " f"{n_structural_tests} structural tests at permutation resolution " f"{resolution:.4f}, at least {topo_fdr['fdr']['min_rejections']} of " "them would have to sit at the resolution floor together. The " - "corrected topology fractions below are therefore an artefact of " + "adjusted topology fractions below are therefore an artefact of " "the shuffle budget, not a result; raise n_shuffles." ) @@ -1385,7 +1404,7 @@ def conclusion_stability( rows[-1].update( { "multiplicity": ( - "benjamini-hochberg across the " + "benjamini-hochberg sensitivity adjustment across the " f"{n_structural_tests} structural tests of this run " f"({len(spec_names)} specifications x " f"{len(TOPOLOGY_COMPOUNDS)} compounds x " @@ -1412,8 +1431,9 @@ def conclusion_stability( f"{r['conclusion']}: multiplicity changes the fraction - " f"{r['n_retained_uncorrected']}/{r['n_specs']} retained on the raw " f"permutation p, {r['n_retained']}/{r['n_specs']} after " - "Benjamini-Hochberg across the run. The corrected fraction is " - "the one to quote." + "Benjamini-Hochberg across the component tests. Report both " + "as a sensitivity analysis; adjustment does not strengthen a " + "failure-to-reject conclusion." ) if any(r["fragile"] for r in rows): warnings.append( @@ -1432,6 +1452,11 @@ def conclusion_stability( "topology_compounds": list(TOPOLOGY_COMPOUNDS), "topology_engine": "flylab.analysis.dependence.dependence_landscape", "topology_fdr": topo_fdr["fdr"], + "topology_multiplicity_scope": topo_fdr["scope"], + "cell_level_fdr_controlled": topo_fdr["cell_level_fdr_controlled"], + "bh_dependence_assumptions_verified": topo_fdr[ + "dependence_assumptions_verified" + ], "n_structural_tests": n_structural_tests, "matrix_uncorrected": matrix_uncorrected, "topology_criteria": { diff --git a/papers/IJRC_FlyLab_draft.md b/papers/IJRC_FlyLab_draft.md index d474b51..0389887 100644 --- a/papers/IJRC_FlyLab_draft.md +++ b/papers/IJRC_FlyLab_draft.md @@ -1,55 +1,62 @@ -# Which Inputs Does a Network Simulation Depend On? Typed Evidence Propagation and Input-Dependence Testing, Evaluated on Cuts of a Synapse-Resolution *Drosophila* Connectome +# Auditing Input Dependence in Literature Parameterised Network Simulations -**Research article, submitted draft for the International Journal of Research in Computing (IJRC)** +**Computational-methods Research Article. Working draft for the International Journal of Research in Computing (IJRC); not submitted.** + +**Michél Nguyen** + +University of the People + +ORCID: 0000-0001-6834-4422 + +Corresponding author: michel_ng@icloud.com Software: https://github.com/pinkysworld/FlyLab -Version 0.6.0, commit `f95fd88a2165`, compound library SHA-256 `594b8d32c5e6`, map `male-cns:v1.0`. +Software version 0.6.0, compound library SHA-256 `594b8d32c5e6`, map `male-cns:v1.0`. Numerical source provenance is recorded in `papers/results.json`; a tagged archival release remains pending. -> Every numeric span in this document is generated by `scripts/reproduce_paper.py` from `papers/results.json`; the prose is written around the numbers, not the other way round. No number in this paper was measured in a living animal. +> Simulation-result numeric spans in this document are generated by `scripts/reproduce_paper.py` from `papers/results.json`; the prose is written around the numbers, not the other way round. No number in this paper was measured in a living animal. ## Abstract -A simulation built on a measured network and parameterised from published literature cannot, on its own, say which of its inputs its predictions depend on. We present a method that makes that question routine and testable, and an open implementation, FlyLab, that answers it per prediction. The method has four parts: **typed evidence propagation**, in which the parameter a source measured and its graded distance from the modelled target jointly decide which transformation is admissible, enforced by a soundness invariant checked exhaustively over the parameter library (56 of 102 rows refused rather than approximated); **input-dependence testing** against an ordered ladder of degraded copies of the network, reporting — with a prespecified equivalence margin and multiplicity control — the weakest input the test cannot distinguish from the real one; an **ablation ladder** over pipeline layers scored against a matched reference distribution; and **specification-family robustness** with variance attribution that ranks what to measure next. We evaluate the method rather than the domain: the dependence instrument is validated against planted ground truth with a power surface, its permutation budget is characterised by a convergence sweep, and the analysis is deterministic across two backends. **The principal result is methodological: a dependence verdict obtained on one extract of a network can invert on another extract of the same network, and the extract's recurrence, not its size, predicts which way.** Demonstrated on cuts of the public MaleCNS v1.0 *Drosophila* connectome, a prediction is composition-dominated on a sparse 1-hop extract (mean degree 1.208, 85% of edges onto four seed cells) and topology-dependent on every extract with real recurrence, including one with *fewer* nodes; across nested extracts of 1k, 5k, 10k, 25k, 50k cells the verdict settles as topology-dependent. Such a verdict is therefore a joint property of the prediction and the extract, and must be reported with the extract's structure. The central limitation is unchanged: the pipeline evaluates a model's behaviour, and its parameter-to-gain rules are asserted rather than measured. +Applied to cuts of the MaleCNS v1.0 *Drosophila* connectome, the same dependence analysis gives opposite answers: an imidacloprid prediction is composition-dominated on the sparse `named` cut but topology-dependent on denser extracts, including a smaller one. Node count therefore does not order the observed labels, and the study does not isolate recurrence as their cause. The result is also runtime-conditional: composition-only and full-model rankings agree under row normalisation but not under a degree-corrected denominator. Two instruments had previously flattered the conclusions. A transmitter null changed the weighted excitation/inhibition balance it was meant to probe, while alternative gain rules did not reach the engine used by their topology tests. FlyLab addresses these failures with provenance-gated evidence transformations, graph degradations, matched ablation references, specification-family tests and variance attribution. A finite-library audit refuses inadmissible evidence (56 of 102 rows), and synthetic cases provide limited positive-control evidence for the dependence ladder. The corrected audit finds that nicotinic suppression reverses under monotone gain rules. These are results about model behaviour, not animal responses. Gain rules remain asserted, the large-scale table is imported from a saved study, and no independent circuit-level validation exists. -**Index terms:** computational method, scientific software, provenance, evidence typing, network simulation, null models, equivalence testing, multiple comparisons, specification robustness, sensitivity analysis, reproducible research, connectomics. +**Index terms:** computational method, scientific software, provenance, evidence typing, network simulation, null models, tolerance diagnostics, multiple comparisons, specification robustness, sensitivity analysis, reproducible research, connectomics. ## 1. Introduction -A large class of scientific software has the same shape: a measured network, a table of parameters compiled from published literature, a rule that turns those parameters into perturbations of the network, and a numerical runtime that produces a prediction. Power grids, metabolic models, epidemic contact networks and connectome simulations are all built this way. Such a pipeline readily produces a number; what it does not produce is an answer to the question a reviewer asks first — *which of its inputs does this number actually depend on?* Without that answer the prediction cannot be audited, because a result carried entirely by an asserted coefficient is indistinguishable, at the output, from one carried by the measured structure. +A large class of scientific software has the same shape: a measured network, a table of parameters compiled from published literature, a rule that turns those parameters into perturbations of the network, and a numerical runtime that produces a prediction. Power grids, metabolic models, epidemic contact networks and connectome simulations are all built this way. Such a pipeline readily produces a number; what it does not produce is an answer to the question a reviewer asks first: *which of its inputs does this number actually depend on?* Without that answer the prediction cannot be audited, because a result carried entirely by an asserted coefficient is indistinguishable, at the output, from one carried by the measured structure. -Three properties make the question hard to answer in practice. The parameters are heterogeneous: sources measure different quantities, in different preparations, and are often simply absent, so a pipeline that coerces them into one numeric field silently manufactures evidence. The network enters through a runtime whose own conventions — normalisation, sign, iteration — are modelling choices that can carry part of the answer. And the usual test of whether structure matters is a permutation against a shuffled network, whose *non-rejection* is routinely written up as though the shuffle had reproduced the effect. +Three properties make the question hard to answer in practice. The parameters are heterogeneous: sources measure different quantities, in different preparations, and are often simply absent, so a pipeline that coerces them into one numeric field silently manufactures evidence. The network enters through a runtime whose own conventions (normalisation, sign and iteration) are modelling choices that can carry part of the answer. And the usual test of whether structure matters is a permutation against a shuffled network, whose *non-rejection* is routinely written up as though the shuffle had reproduced the effect. -This paper contributes a method addressing all three (§2), an implementation that runs it per prediction rather than once per paper, and — the part we think is missing from this literature — an evaluation of the instruments themselves (§3). We demonstrate the method on cuts of a public synapse-resolution *Drosophila* connectome, a domain supplying exactly the hard case: a measured network, a literature parameter table with large gaps, and asserted transformation rules. The headline finding is about the method rather than the domain: **a dependence verdict obtained on one extract of a network can invert on another extract of the same network.** +This paper contributes a method addressing all three (§2), an implementation that runs it per prediction, and an evaluation of the instruments themselves (§3). Cuts of a public synapse-resolution *Drosophila* connectome supply the hard case: a measured network, a literature parameter table with large gaps, and asserted transformation rules. The central result is methodological: **a dependence verdict obtained on one extract of a network can invert on another extract of the same network.** -We organise the work around four research questions. +We organise the work around three research questions. * **RQ1.** Can heterogeneous published parameters be turned into computational perturbations *without* erasing their provenance or their parameter semantics, and can that discipline be stated as a checkable invariant rather than a convention? -* **RQ2.** For a given prediction, which inputs does it depend on — and is that answer a property of the prediction, or of the extract of the network it was measured on? +* **RQ2.** For a given prediction, which inputs does it depend on, and is that answer a property of the prediction or of the extract of the network on which it was measured? * **RQ3.** Which conclusions survive alternative admissible specifications, and which assumptions dominate the output's variance? -* **RQ4.** Can the entire analysis be reproduced identically, across backends, from auditable artifacts? -The answers are: yes, by grading the evidence, refusing invalid transformations and checking the refusal exhaustively; **it is a property of the extract**, which is the paper's main result; not all of them, and an undocumented normalisation inside the runtime carries part of RQ2's answer; and yes. An earlier version of this work answered RQ2 with "most predictions do not need the network", measured on a single extract — a conclusion this revision withdraws and explains. +The results support evidence gating and finite regression checks for RQ1. For RQ2, labels vary with the extract, prediction and runtime normalisation. For RQ3, some conclusions reverse under admissible specifications. An earlier version answered RQ2 with "most predictions do not need the network", measured on a single extract; this revision withdraws that conclusion. Recomputability and backend parity are reported separately as engineering support, not as a fourth research question. ### 1.1 Related work and what is new here -**Network runtimes without parameterised perturbation.** Teaching simulators [1]–[4] model dose–response with no network at all. At the other extreme, FlyWire [5] and MaleCNS v1.0 [6] publish netlists with per-node annotations [10], [12]; Shiu *et al.* [7] show a literature-parameterised model over such a netlist reproduces a sensorimotor control, and downstream ports repeat the exercise with varying rigour [11]. FlyBrainLab [68] builds executable circuits directly from connectome data, but a perturbation, where present, enters as a hand-set multiplier with nothing behind it — exactly the case typed propagation is meant to replace. +**Network runtimes without parameterised perturbation.** The implemented teaching simulators RatCVS and Virtual Rat Web [1], [2] model dose–response with no network; PharmVR [3] is a non-peer-reviewed concept paper and is not evidence of an implemented capability. At the other extreme, FlyWire [4] and MaleCNS v1.0 [5] publish netlists with per-node annotations [9], [11]; Shiu *et al.* [6] show a literature-parameterised model over such a netlist reproduces a sensorimotor control, while an informal newsletter essay catalogs downstream ports with varying rigour [10]. FlyBrainLab [67] builds executable circuits directly from connectome data. The cited implementation does not describe the per-row pharmacological evidence gating evaluated here. -**Parameter-map-informed network models.** On human structural connectomes, intervention is simulated by scaling regional responses with PET density maps [69], [70] — the closest published relative, whose most relevant result is also a negative one: in Mindlin *et al.* [69] the simulated improvement tracked the *mean* density of the activated receptors rather than where they sit. Methodologically, §2.4 is a multiverse analysis [74], [75] applied to a simulation, and §2.6 uses Saltelli cross-sampling [71] with Jansen estimators [72] validated on Ishigami [73] and ranked by value of information [76]. +**Parameter-map-informed network models.** The closest published relatives simulate intervention on human structural connectomes by scaling regional responses with PET density maps [68], [69]. In Mindlin *et al.* [68], the simulated improvement tracked the *mean* density of the activated receptors rather than where they sit. Methodologically, §2.4 is a multiverse analysis [73], [74] applied to a simulation, and §2.6 uses Saltelli cross-sampling [70] with Jansen estimators [71] validated on Ishigami [72] and ranked by value of information [75]. -**The domain evidence base.** The parameter library, and the supporting datasets for resistance alleles, expression, mixtures, exposure and behavioural endpoints, are compiled from the insect-versus-vertebrate pharmacology literature [8], [9], [13]–[67] and documented in §S4 and §S6; what matters methodologically is that it is typical of such tables — incomplete, heterogeneous in what each source measured, and internally inconsistent with its own citations in ten places that `data/literature/README.md` records without editing the library to match. Those gaps are the reason the type system changes results rather than documentation. +**The domain evidence base.** The parameter library, and the supporting datasets for resistance alleles, expression, mixtures, exposure and behavioural endpoints, are compiled from the insect-versus-vertebrate pharmacology literature [7], [8], [12]–[66] and documented in §S4 and §S6. What matters methodologically is that the library is typical of such tables: incomplete, heterogeneous in what each source measured, and internally inconsistent with its own citations in ten places that `data/literature/README.md` records without editing the library to match. Those gaps make evidence gating affect numerical outputs. -**Position.** Table I states what is new narrowly. The transferable part is not the domain coupling but the instrumentation around it: routine, per-prediction, *validated* testing of which inputs a prediction depends on, with non-rejection separated from equivalence, multiplicity controlled, and the test's own power characterised against planted ground truth. +**Position.** Table I states what is new narrowly. The transferable part is not the domain coupling but the instrumentation around it: routine per-prediction diagnostics combining provenance, graph degradations and specification checks. Synthetic recovery tests are an initial instrument evaluation. A median-distance tolerance is distinct from statistical equivalence, and multiplicity guarantees depend on the assumptions of the adjustment. -**Table I. What each system provides.** +**Table I. Capabilities described in the cited implementations.** A negative entry means not described in that cited application; it is not a claim about every version or extension. -| Capability | Teaching sims [1]–[4] | Connectome LIF runtimes [7], [11] | FlyBrainLab [68] | Receptor-map whole-brain models [69], [70] | FlyLab | +| Capability | Implemented teaching sims [1], [2] | Connectome LIF runtimes [6], [10] | FlyBrainLab [67] | Receptor-map whole-brain models [68], [69] | FlyLab | |---|---|---|---|---|---| | Named cells at synapse resolution | no | yes | yes | no (regions) | yes | | Compound **and free concentration** as the input | yes (organ level) | no | no | partly (receptor-map scaling) | yes | | Typed evidence with per-row provenance and distance | no | no | no | no | yes | -| Two species scored at the same dose | no | no | no | no | yes | -| Dependence of each prediction on its inputs, tested | no | rarely | no | partly [69] | yes, routinely | -| The dependence test itself validated against ground truth | n/a | no | no | no | yes (§3.1, F16) | +| Insect engagement compared with a vertebrate receptor panel | no | no | no | no | yes; no vertebrate circuit | +| Dependence of each prediction on its inputs, tested | no | rarely | no | partly [68] | yes, routinely | +| Synthetic recovery cases evaluated | n/a | not assessed | not assessed | not assessed | yes (§3.1, F16) | | Specification robustness of the stated conclusions | no | no | no | no | yes | | Whole analysis re-runs in a browser | some | no | no | no | endpoint parity verified | @@ -61,19 +68,19 @@ We claim no new reconstruction, no fitted *in vivo* potency, no replacement for The common failure this replaces is a pipeline that applies one transform to whatever number a source reported and gives the result one name. That erases two distinctions at once: a normalised functional response is not a fraction of occupied binding sites, and a row with no sourced number must not yield a small response at all. -Each row instead records the **parameter type** its source measured and an ordered **evidence distance** between that source and the modelled target: E0 (on-target); E1 (cross-species); E2 (related receptor); E3 (class extrapolation); E4 (unsupported). The admissible **transformation** is a total function of that pair (T20): a binding constant yields a physical occupancy only at E0 and a labelled proxy beyond it; a functional potency yields a direct engagement at E0 and a proxy beyond; anything at E4 is `not_modelled` and returns N/A. The result carries the name it earned, so a transferred number cannot be read as a measured one downstream. +Each row records the **parameter type** its source measured and an ordered **evidence distance** from the modelled target: E0 (on-target); E1 (cross-species); E2 (related receptor); E3 (class extrapolation); E4 (unsupported). The admissible transformation is a total function of that pair (T20). Example legal triples are `(Kd, E0, binding_occupancy)`, `(Kd, E1, binding_engagement_proxy)` and `(EC50, E0, functional_engagement)`; transferred functional potencies use `functional_engagement_proxy`, while any E4 row is `not_modelled`. An attempted `(EC50, E0, binding_occupancy)` transformation raises `EvidenceTypeError`. Every numeric branch retains the same Hill algebra, `C^n/(x^n + C^n)`: the exact *Drosophila* Kd row licenses a binding-occupancy label but introduces no new kinetic binding model. -Two properties make this a type discipline rather than a lookup table, and both are what we claim as the computing contribution. The refusal is **enforced**: requesting a value from an inadmissible row raises rather than returning one. And we state and check a **soundness invariant** over the whole pipeline — *no numeric engagement reaching a readout may originate from a row whose (parameter type, distance) pair does not admit one* — exhaustively, over every row and every entry point, with a companion check that dropping the inadmissible rows changes no downstream value, and a mutation check that makes the test fail with the right message when the rule table is altered. +The enforceable invariant is narrow: *no numeric engagement reaching a readout may originate from a row whose (parameter type, distance) pair does not admit one*. A finite regression audit exercises every shipped row through the enumerated comparison, curve, uncertainty, assay and dashboard paths, checks that dropping inadmissible rows changes no gain vector, and verifies deliberate invalid requests. This is coverage of the shipped library and enumerated paths, not proof about an unenumerated future entry point. ### 2.2 The extracts, and what the runtime does to them -Two committed extracts of MaleCNS v1.0 are used, streamed out of the ~1.1 GB public weight matrix on a CI runner and each under about 1 MB of JSON: **`named`**, a 1-hop neighbourhood of two seed cells (1126 nodes, 1360 edges), and **`taste_motor`**, the same seeds plus sensory types and an induced closure (1841 nodes, 19 066 edges). §3.2 adds a ladder of nested extracts of 1k, 5k, 10k, 25k, 50k cells grown by one fixed rule. +Two committed extracts of MaleCNS v1.0 are used, streamed out of the ~1.1 GB public weight matrix on a CI runner and each under about 1 MB of JSON: **`named`**, a 1-hop neighbourhood of two named cell types (four seed cells) (1126 nodes, 1360 edges), and **`taste_motor`**, the same seeds plus sensory types and an induced closure (1841 nodes, 19 066 edges). §3.2 adds a ladder of nested extracts of 1k, 5k, 10k, 25k, 50k cells grown by one fixed rule. -**The extracts are structurally very different, and the difference turns out to be load-bearing (T19).** `named` has mean degree 1.208: 85% of its edges terminate on the 4 seed cells, only 184 of 1126 nodes receive any input, and only 25% of edges leave a node that itself receives input. It is an in-star, not a circuit. `taste_motor` has mean degree 10.4, 10% of edges onto seeds and 97% of edges leaving a node with input. Composition differs too, and one class is absent from both (0 cells), so a perturbation acting only through it cannot move either readout for substrate rather than domain reasons; the full census is in T19 and §S9(v). +**The extracts are structurally very different, and the difference turns out to be load-bearing (T19).** `named` has mean degree 1.208: 85% of its edges terminate on the 4 seed cells, only 184 of 1126 nodes receive any input, and only 25% of edges leave a node that itself receives input. It is dominated by incoming edges to seeds, with limited recurrent connectivity. `taste_motor` has mean degree 10.4, 10% of edges onto seeds and 97% of edges leaving a node with input. Composition differs too, and octopamine is absent from both (0 cells), so a perturbation acting only through it cannot move either readout for substrate rather than domain reasons; the full census is in T19 and §S9(v). -Two runtimes consume a gain patch: a deterministic clipped leaky rate iteration (~50 ms per evaluation of `named`) and a current-based LIF network [7] whose synaptic scale and background drive are tuned constants, not fits (§S3). +Two runtimes consume a gain patch: a deterministic clipped leaky rate iteration (~50 ms per evaluation of `named`) and a current-based LIF network [6] whose synaptic scale and background drive are tuned constants, not fits (§S3). -**The rate runtime row-normalises its weight matrix, and previous versions of this paper did not say so.** Each row of the signed matrix is divided by `max(Σ|W|, 1)` before any gain is applied, and gains then act as a presynaptic column scale, so every node's recurrent input is a *composition-weighted average of its presynaptic gains*. That builds part of the answer to RQ2 into the readout itself — an example of the general hazard that a runtime convention can carry a result — so it is documented here, in T2 and §S3, and swept in §3.3 against `none` and a degree-corrected denominator. The sign table the runtime iterates is published as T21: 5 of seven magnitudes (glutamate -0.4, histamine -0.5, dopamine +0.2, serotonin +0.2, octopamine +0.2) are asserted, varied by no member of the specification family and not bounded by the label-permuting null. +**The rate runtime row-normalises its weight matrix, and previous versions did not report this.** Each row of the signed matrix is divided by `max(Σ|W|, 1)` before gains act as presynaptic column scales. Recurrent input is therefore a composition-weighted average of presynaptic gains, making the convention part of the answer to RQ2. Section 3.3 sweeps it against `none` and a degree-corrected denominator. T21 publishes the sign table: 5 of seven magnitudes (glutamate -0.4, histamine -0.5, dopamine +0.2, serotonin +0.2, octopamine +0.2) are asserted, absent from the specification family and not bounded by the label-permuting null. ### 2.3 Input-dependence testing (RQ2) @@ -87,93 +94,95 @@ The instrument we consider most transferable asks, for one `compound × concentr | `sign_permute_weight_matched` | + the real weights on the real edges, and each transmitter's share of total outgoing weight | | the MaleCNS cut | everything | -The rank-3 rung changed in this revision. Because a gain patch acts *through* transmitter identity, a null that moves the weighted balance moves the drug's target set and the sign matrix at once, so the plain label permutation is relabelled a **joint target-set-and-sign null** and replaced by a constrained shuffle holding every tracked transmitter's share of outgoing weight within 0.002 of the real graph's. Permuting the node transmitter-label array preserves the transmitter histogram *by count* but not *by weight*: out-strength is heterogeneous, so on the `named` cut the cholinergic share of total outgoing synaptic weight is 0.619 on the real graph against 0.544 ± 0.027 across 1000 permutations, putting the real graph at the 100th percentile of its own null (T23). +The rank-3 rung uses a constrained shuffle holding every tracked transmitter's outgoing-weight share within 0.002 of the real graph. A plain label permutation is instead a **joint target-set-and-sign null**: it preserves label counts but moves weighted balance because out-strength is heterogeneous. On `named`, the cholinergic outgoing-weight share is 0.619 on the real graph versus 0.544 ± 0.027 across 1000 permutations, placing the real graph at the 100th percentile of that null (T23). -**Non-rejection is not equivalence, and the software no longer lets the prose say it is.** Each mode receives one of three verdicts against a margin prespecified as 5% of the vehicle readout (0.332 Hz on `named`): `distinguishable` when the empirical two-sided permutation probability `(k+1)/(n+1)` is at or below α; `equivalent_within_tolerance` when the test does not reject *and* the gap from the null median is below the margin; `indeterminate` otherwise. From the verdicts follow a **class** and a **necessary information level** — the weakest graph model the test could not distinguish from the real cut, never a claim that it reproduces the effect. Where a weaker model survives while a richer one is rejected the ladder is non-monotone, which the software reports and §3.2 counts. +Each mode is first assessed with the empirical two-sided permutation probability `(k+1)/(n+1)`. After non-rejection, the implementation compares the real effect with the null median using a tolerance of 5% of the vehicle readout (0.332 rate-model units on `named`). The legacy label `equivalent_within_tolerance` records only that point-gap check, not statistical equivalence; other non-rejections are `indeterminate`. Classes and the legacy **necessary information level** are descriptive summaries, not proofs of necessary or sufficient information. The nulls are not nested. -Three further conventions are prespecified and §S2 states them in full: a *relative* effect floor, a separation of **confirmatory** from **exploratory** tests with Benjamini-Hochberg applied across the latter, and a check of how many tests sit at the resolution floor, so a corrected count of zero is never mistaken for a finding. +The implementation also fixes a relative effect floor, separates detailed profiles from library-wide screens, and applies Benjamini-Hochberg across the screens' component structural tests (§S2). A cell is topology-dependent if either adjusted structural test rejects. This is not a demonstrated cell-level FDR guarantee, and the conventions were refined during analysis rather than externally registered. ### 2.4 The ablation ladder and the specification family (RQ3) -*Which layer carries the information?* Four models predict the same cell with one layer denied: **A** receptor engagement alone; **B** mechanism gains on the graph's transmitter composition only, no edges; **C** the real cut and runtime with one generic multiplier; **D** the full model. They carry different units, so they are compared by how well each reproduces the full model's *ordering* of the library, scored on the **signed** Spearman ρ because a perfect inversion reproduces nothing. +Four models predict the same cell with one layer denied: **A** receptor engagement alone; **B** mechanism gains on transmitter composition only; **C** the real cut and runtime with one generic multiplier; **D** the full model. Because their units differ, signed Spearman ρ compares each library ordering with D. -Level C's multiplier changed in this revision: it could previously only *depress*, so no entry could come out positive and the level could not order a library half of whose entries raise the readout. Its magnitude is unchanged and its **sign** now comes from the mechanism table — one bit per compound — with the old rule retained as a floor (§S1). Levels B and D are not independent models, both being functions of the same gain vector, so §3.3 reports a **matched reference distribution** saying what agreement would have been for structural reasons alone. +Level C previously could only depress activity and could not order compounds that raise the readout. Its magnitude is unchanged, but its sign now comes from one mechanism-table bit per compound; the old rule remains as a floor (§S1). Because B and D share the same gain vector, §3.3 reports a matched reference distribution. -*Which conclusions are properties of the data rather than of a modelling choice?* The parameter-to-gain mapping is the least defensible part of the pipeline — a textbook shape, but an asserted functional form. We prespecify 25 admissible specifications (five shapes at five coefficient scales, §S1.1) and re-derive every principal conclusion under every member; undecidable cells count against the conclusion and are reported separately, and the fraction is a fraction of *that family*, not a posterior probability. The topology conclusions are decided at 100 permutations per specification (resolution 0.010), corrected across the run's 100 structural tests. The two thresholds are likewise conventions, so the verdict is recomputed over the full 25%, 40%, 50% × 10%, 20%, 30% grid on 10 compounds. +The parameter-to-gain mapping is an asserted functional form. We therefore re-derive each principal conclusion under 25 specifications (five shapes at five coefficient scales, §S1.1); undecidable cells count against a conclusion, and fractions describe this family rather than posterior probabilities. Topology predicates use 100 permutations per specification and adjusted component probabilities across 100 tests, without a conclusion-level FDR guarantee. Threshold verdicts are recomputed over the 25%, 40%, 50% × 10%, 20%, 30% grid on 10 compounds. ### 2.5 Evaluation taxonomy -Agreement with published work is worth little unless its kind is named, so we separate three. **Implementation consistency** — the same inputs give the same outputs across transports, runtimes and runs (§3.7). **Literature concordance** — the model's *ordering* agrees with a published ordering, scored per entry and never as a bare mean, since an entry ordering two items can only return ±1; where the published ordering shares a source with the library, agreement is internal consistency, and the pipeline computes that overlap from DOIs and PMIDs rather than asserting it. **Independent out-of-sample validation** — against data entering neither the library nor the model, of which there is **none** at the circuit level. The ground-truth experiment of §3.1 validates the *instrument*, not the domain model, and we are careful not to let the first stand in for the second. +We distinguish **implementation consistency**, **literature concordance** and **independent out-of-sample validation**. Concordance is scored per published ordering, with shared library sources detected from DOIs and PMIDs; a two-item ordering is never allowed to dominate a bare mean. There is no independent circuit-level validation. The synthetic experiment assesses selected instrument behaviours, not the domain model or general calibration. ### 2.6 Global uncertainty and the value of information -A one-at-a-time tornado cannot separate a factor's own effect from its interactions, so uncertainty is attributed with a variance-based (Sobol') analysis over every assumption the implementation exposes — potency, Hill coefficient, the engagement-to-gain transformation and its coefficient, the external drive, the cut's synapse-weight threshold, the transmitter-sign convention, the expression-weighting layer — plus a deliberately null factor, the spiking RNG seed, which the deterministic rate engine ignores. The design is Saltelli cross-sampling with Jansen estimators [71], [72], validated against the analytic Ishigami benchmark [73] in the same run. +A one-at-a-time tornado cannot separate a factor's own effect from its interactions, so uncertainty is attributed with a variance-based (Sobol') analysis over selected assumptions the implementation exposes; potency, Hill coefficient, the engagement-to-gain transformation and its coefficient, the external drive, the cut's synapse-weight threshold, the transmitter-sign convention, the expression-weighting layer; plus a deliberately null factor, the spiking RNG seed, which the deterministic rate engine ignores. The design is Saltelli cross-sampling with Jansen estimators [70], [71], validated against the analytic Ishigami benchmark [72] in the same run. -Estimates are left unclipped, so a factor whose true index is zero can be estimated below zero. A first-order index cannot be negative, so the **largest magnitude among the negative estimates is a measured lower bound on the estimator's error**, and that — not the null factor's own draw — is the noise floor we report. The budget becomes a ranking by `VOI_j = max(0, S_j) · Var(Y)` [76] — variance of a model output under assumed input ranges, not an expected gain in accuracy about a living fly. +Estimates are left unclipped, so a factor whose true index is zero can be estimated below zero. A first-order index cannot be negative, so the **largest magnitude among the negative estimates is a measured lower bound on the estimator's error**, and we report it as an error diagnostic, not a calibrated noise floor or universal bound. The budget becomes a ranking by `VOI_j = max(0, S_j) · Var(Y)` [75]; variance of a model output under assumed input ranges, not an expected gain in accuracy about a living fly. ## 3. Results -This section evaluates the instruments, not the domain. All values are regenerated by `scripts/reproduce_paper.py` and stored keyed in `papers/results.json`; the parameter library and the domain readouts they are demonstrated on are described in §3.6 and §S6. +This section evaluates the instruments, not the domain. Core committed-cut values are generated by `scripts/reproduce_paper.py` and stored keyed in `papers/results.json`; T27 is imported from the cached scale study described in §3.7. The parameter library and demonstration readouts are described in §3.6 and §S6. -### 3.1 Is the dependence instrument sound? Ground truth, power and convergence (Fig. F16, Tables T11, T22) +### 3.1 Synthetic recovery, detection frequency and convergence (Fig. F16, Tables T11, T22) -A negative from a dependence test is uninterpretable unless the test can be shown to return a positive when one exists, so the instrument is characterised against **planted ground truth** before it is used on anything. A recurrent cycle of known strength is planted in a synthetic network of the same size, density and composition as the real extract, so that the perturbation removes an amplification existing only while the cycle is intact: the *contrast*, not merely the level, depends on the wiring by construction. Strength 0 plants nothing and is the negative control. +The ladder recovered 2 of 3 planted effects at *n* = 200, with no detection in its unplanted control. Detection rates by planted strength are strength 0.25: 0.00-0.17; strength 0.5: 0.67-0.83; strength 1: 1.00-1.00, with all sampled runs detected at 1 and low detection at 0.25; by permutation count they are n=50: 0.61, n=200: 0.61, n=1000: 0.67. The test can therefore return a positive when wiring is constructed to matter, but weak effects are often missed. -The ladder recovered 2 of 3 planted effects at *n* = 200 with no false positive on the control. Over a power grid (6 independently generated networks per cell, permutation counts 50, 200, 1000) the detection rate by planted strength is strength 0.25: 0.00-0.17; strength 0.5: 0.67-0.83; strength 1: 1.00-1.00 — full power at strength 1, and at or below one in four at 0.25, so the instrument is **blind to weak planted structure and must not be read as ruling it out**. The empirical false-positive rate on the unplanted control is 0 of 18. Averaged over strengths, detection by permutation count is n=50: 0.61, n=200: 0.61, n=1000: 0.67: **effect size dominates, and permutation count buys little beyond a few hundred draws.** That is a design rule for this class of test — budget spent on permutations past the point of convergence is wasted, and it is the reason the scaling study of §3.2 can trade permutations for network size at all. +This is limited positive-control evidence. The network has 250 nodes and 1500 background edges and is not matched to either real extract. Strength 0 omits both cycle edges and associated transmitter relabelling. The grid uses only 6 networks per cell at counts 50, 200, 1000, with 0 detections in 18 control runs; networks and shuffle prefixes are reused. These frequencies neither estimate general power or type-I error precisely nor establish power for the taste-motor veto-ratio endpoint. -The convergence sweep (T11) separates the two quantities that permutation budget buys: a *verdict* at α = 0.05 settles from *n* = 50–400, while a quotable mid-range probability needs 25–400 draws to reach ±0.02 of its final value. The sweep re-reads prefixes of the same draws, so every row is a valid smaller sample and no extra runtime was spent producing it. +The prefix sweep (T11) distinguishes label agreement from probability precision. Observed labels remain unchanged from *n* = 50–400, while probabilities first remain within ±0.02 of their final estimates after 25–400 draws. These are descriptive comparisons with the final estimate, not confidence bounds on the true permutation probability. Prefixes reuse the same draws and are not independent replications. -### 3.2 The verdict inverts with the extract, and recurrence predicts it (Figs. F6, F12; Tables T6, T12, T12b, T23, T27) +### 3.2 Dependence labels change with the extract (Figs. F6, F12; Tables T6, T12, T12b, T23, T27) -**The confirmatory profiles** (full per-mode table: T6, §S2.1). At 1000 permutations per mode (resolution 0.0010), imidacloprid's effect on the sparse `named` extract is -6.173 Hz and the cell is **composition-dominated**: against the prespecified margin of 0.332 Hz it is equivalent within tolerance for rewire_degree_preserving, sign_permute_weight_matched, weight_permute, merely indeterminate for sign_permute, and distinguishable only from the Erdős–Rényi control (*p* = 9.99e-04 against 0.472 for degree-preserving rewiring). Fipronil on the same extract is the contrast — 0.900 Hz, **topology-dependent**, *p* = 0.006 and 0.007 for weight permutation and rewiring, necessary level `wiring_with_weighted_transmitter_balance` — so the instrument separates two predictions on one extract. Correcting the transmitter null moves no probability but changes what may be claimed: imidacloprid's gap from the null median falls from 2.652 Hz to 0.236 Hz, turning an indeterminate result into an equivalence one. +**The landscape inverts between the two extracts.** Over 21 compounds × 4 concentrations = 84 cells at 1000 permutations, with Benjamini-Hochberg across the 168 component structural tests and the relative effect floor applied, `named` gives **11 topology-dependent, 42 composition-dominated, 31 no-effect, 0 mixed** (uncorrected 14 / 39 / 31; on the absolute floor alone 16 / 55 / 13, the 18 cells the relative floor removes named in T12). Repeated unchanged on the denser `taste_motor` extract at 300 permutations, the same 84 cells give **54 / 0 / 30**: every cell above the floor is topology-dependent, imidacloprid included (*p* = 0.003, adjusted 0.006, class topology-dependent). These are cell classes formed from adjusted component tests, not a cell-level FDR-controlled count; shared shuffle draws also leave the usual Benjamini-Hochberg dependence condition unverified. -**The landscape inverts between the two extracts.** Over 21 compounds × 4 concentrations = 84 cells at 1000 permutations, with Benjamini-Hochberg across the 168 structural tests and the relative effect floor applied, `named` gives **11 topology-dependent, 42 composition-dominated, 31 no-effect, 0 mixed** (uncorrected 14 / 39 / 31; on the absolute floor alone 16 / 55 / 13, the 18 cells the relative floor removes named in T12). Repeated unchanged on the denser `taste_motor` extract at 300 permutations, the same 84 cells give **54 / 0 / 30**: every cell above the floor is topology-dependent, imidacloprid included (*p* = 0.003, adjusted 0.006, class topology-dependent). +**The detailed reference profiles** show that the instrument also separates predictions within one extract (T6, §S2.1). At 1000 permutations per mode (resolution 0.0010), imidacloprid on `named` is **composition-dominated** with an effect of -6.173 rate-model units. It passes the descriptive gap criterion for rewire_degree_preserving, sign_permute_weight_matched, weight_permute, is indeterminate for sign_permute, and is distinguishable only from Erdős–Rényi (*p* = 9.99e-04 versus 0.472 for degree-preserving rewiring). Fipronil is **topology-dependent** at 0.900 rate-model units, with *p* = 0.006 and 0.007 for weight permutation and rewiring, necessary level `wiring_with_weighted_transmitter_balance`. Weight matching reduces imidacloprid's transmitter-null median gap from 2.652 to 0.236 rate-model units; passing the gap criterion is not statistical equivalence. -**Recurrence, not size, predicts the inversion.** Repeating the same profile across a ladder of nested extracts grown from one seed set by a single fixed rule (7 extracts in all, scale_1k, named, taste_motor, scale_5k, scale_10k, scale_25k, scale_50k; T27) settles the question of what the reversal tracks. The composition-dominated verdict appears on exactly 1 of them, `named`, and the inversion does not track the number of cells: `scale_1k` has 1000 nodes against `named`'s 1126 and a mean degree of 22.9 against 1.21, and it is already topology-dependent. Imidacloprid is topology-dependent on every rung but the in-star, and fipronil on all seven (per-rung sequences in §S2.4, full table T27). Both compounds' verdicts are settled on the clean ladder by the two-largest-rungs criterion of §S2.4, and at the denser middle rungs (scale_5k, scale_10k) the prediction is distinguishable from **every** null, its necessary level being the real network itself. The relative effect size is stable over this range (0.52-0.93 of the vehicle readout), so the change of verdict is not a change of what is being measured. +Across 7 extracts (scale_1k, named, taste_motor, scale_5k, scale_10k, scale_25k, scale_50k; T27), composition-dominated labels occur on 1 extract, `named`, and the inversion does not track the number of cells: `scale_1k` has 1000 nodes against `named`'s 1126 and a mean degree of 22.9 against 1.21, and it is already topology-dependent. This excludes a simple monotone ordering by node count. At scale_5k, scale_10k every tested null is rejected, and the imidacloprid relative effect spans 0.52-0.93 of vehicle. Construction, density, composition, paths, readout support and effort also vary, however, so recurrence is associated with but not isolated as the cause. Agreement of the final two nested-rung labels is not convergence to a whole-network verdict. -Two honesty constraints travel with that result. The permutation budget must fall as the network grows (20-1000 draws, resolution 0.0010-0.0476), and at scale_10k, scale_25k, scale_50k the permutation budget falls to n = 20, where the smallest attainable probability is 0.048 against alpha = 0.05: a rejection there is the smallest the test can express — so the top rungs confirm the settled verdict rather than establishing it. And the two hops-limited cuts are built by a different rule from the nested rungs, so only comparisons *within* the ladder separate scale from construction. +The permutation budget falls as networks grow (20-1000 draws, resolution 0.0010-0.0476), and at scale_10k, scale_25k, scale_50k the permutation budget falls to n = 20, where the smallest attainable probability is 0.048 against alpha = 0.05: a rejection there is the smallest the test can express. Top-rung labels are therefore resolution-limited, nominal observations. The hops-limited cuts also use a different construction from the nested rungs. -**What is withdrawn, and what replaces it.** A degree-preserving rewire of an in-star is close to the identity, so the null had nothing to destroy; an extract with genuine recurrence does, and destroying it changes the answer. We withdraw the generalisation that most of what such a model predicts is carried by composition rather than topology, and the domain claim that went with it. What replaces it is stronger and methodological: **a dependence verdict is a joint property of the prediction and the extract, the extract's recurrence predicts which way it goes, and a negative measured on a sparse extract is not evidence about the network it was cut from.** +The earlier claim that most predictions are carried by composition does not survive a change of extract. A label is conditional on the model, extract, readout, null family and statistical effort. Testing recurrence causally would require changing it while holding the correlated properties fixed. -**Multiplicity, working.** As a prespecified confirmatory profile, fipronil at 1 µM stands comfortably (*p* = 0.006 at *n* = 1000). Inside the exploratory landscape the same cell's adjusted probability is 0.050 and it does **not** survive correction; its corrected class there is composition-dominated against an uncorrected topology-dependent. +**Multiplicity sensitivity.** In the detailed reference profile, fipronil at 1 µM has a nominal result (*p* = 0.006 at *n* = 1000). Inside the exploratory landscape the same cell's adjusted component probability is 0.050 and no structural mode remains rejected; its adjusted-test class there is composition-dominated against an unadjusted topology-dependent. -**Two things the class label does not say.** The ladder is **non-monotone** in 11 of 84 cells on `named` and 0 of 84 on `taste_motor`, because adjacent rungs destroy different things; the level therefore means "the cheapest degraded input this test cannot tell apart from the real one", never "everything above it is indistinguishable too". And `composition-dominated` reports a failure to distinguish, never a reproduction: across the `named` landscape's mode verdicts, 281 are equivalent within tolerance, 50 indeterminate and 89 distinguishable. +**Two things the class label does not say.** The ladder is **non-monotone** in 11 of 84 cells on `named` and 0 of 84 on `taste_motor`, because adjacent rungs destroy different things; the level therefore means "the cheapest degraded input this test cannot tell apart from the real one", never "everything above it is indistinguishable too". And `composition-dominated` reports a failure to distinguish, never a reproduction: across the `named` landscape's mode verdicts, 281 pass the descriptive median-gap tolerance, 50 indeterminate and 89 distinguishable. ### 3.3 Which pipeline layer carries the information, and what agreement would have been anyway (Fig. F13, Tables T13, T13b, T24, T25) -Receptor engagement alone orders the library **backwards** (signed ρ -0.506 to -0.046) and is now the worst level at every concentration — constant across the table: yes (10 nM: A_receptor_only (-0.046); 100 nM: A_receptor_only (-0.348); 1 µM: A_receptor_only (-0.431); 10 µM: A_receptor_only (-0.506)). Under the historical level-C rule the worst level was *not* constant (10 nM: C_topology_only; 100 nM: A_receptor_only; 1 µM: A_receptor_only; 10 µM: A_receptor_only), so the previous claim that the topology-only level is the worst of the four was true of one column and false of the others. +**The matched reference and runtime sweep change the reading of the raw correlation.** Pharmacology-free pseudo-compounds with the gain shape produced by the shipped rules give a median composition-versus-full correlation of 0.847 over 50 draws (5th–95th percentile 0.676 to 0.953). The observed value exceeds that 95th percentile, so shared algebra does not explain its entire magnitude. Yet the correlation is 0.988 under the shipped row normalisation and 0.906 with none, but 0.666 under a degree-corrected denominator at 1 µM and -0.372 at 10 nM. It survives row_abs, none and not a degree correction, making it conditional on the runtime's readout. -**The ablation conclusion is withdrawn.** Level C's multiplier applied a depression-only factor to every transmitter, so every entry was at most zero and the level could not represent disinhibition. With a direction-aware rule — same generic magnitude, one bit of sign — its rank correlation with the full model at 1 µM rises from 0.242 to 0.961. The sentence "the connectome without the pharmacology is not a cheap substitute for the connectome with it" was a statement about a sign-broken baseline and is retracted. +Against that reference, the raw composition-versus-full correlation is 0.988 at 1 µM and 0.989 after dropping compounds on the 0.05 gain floor, so it is not a tie artefact. Levels B and D nevertheless share the same gain vector. Jointly permuting their paired labels leaves ρ = 0.988 unchanged by construction and is not an independence test. A sign disagreement between the levels moves ρ by at most 0.057 and does not produce the agreement (§S9(i)). -**The composition-versus-full correlation carries no compound-level information.** It is 0.988 at 1 µM and 0.989 after dropping the compounds on the 0.05 gain floor, so it is not a tie artefact. But levels B and D are both functions of the same gain vector, and the matched reference says what would have been unsurprising: pharmacology-free pseudo-compounds with the gain shape the shipped rules produce give a median of 0.847 over 50 draws (5th–95th percentile 0.676 to 0.953), and permuting the compound labels leaves the observed value identically 0.988, as it must, because the permutation moves the (B, D) pairs as a block. +Receptor engagement alone orders the library **backwards** (signed ρ -0.506 to -0.046) and is the worst level at every concentration: yes (10 nM: A_receptor_only (-0.046); 100 nM: A_receptor_only (-0.348); 1 µM: A_receptor_only (-0.431); 10 µM: A_receptor_only (-0.506)). Under the historical level-C rule the worst level was not constant (10 nM: C_topology_only; 100 nM: A_receptor_only; 1 µM: A_receptor_only; 10 µM: A_receptor_only). -**And it is not robust to the runtime.** The correlation is 0.988 under the shipped normalisation and 0.906 with none, but 0.666 under a degree-corrected denominator at 1 µM and -0.372 at 10 nM: it survives row_abs, none and not a degree correction, so it is a property of *this runtime's readout* as much as of the perturbation. A sign disagreement between the two levels moves it by at most 0.057 and so is not what produces the agreement (§S9(i)). +**The earlier ablation conclusion is withdrawn.** Level C applied a depression-only multiplier to every transmitter and could not represent disinhibition. With the same generic magnitude but one mechanism-table sign bit, its correlation with the full model at 1 µM rises from 0.242 to 0.961. The claim that the connectome without pharmacology was not a cheap substitute was therefore a statement about a sign-broken baseline. ### 3.4 Specification robustness: which conclusions survive the modelling choices (Fig. F14, Tables T14, T14b, T15) -**The previous 25/25 topology result was vacuous, and the failure mode is general.** Two faults compounded. The matrix was computed at 6 shuffles per specification, where the smallest attainable permutation probability is 1/7 = 0.143, so `p ≤ 0.05` was arithmetically unattainable and the predicate silently fell back on a Gaussian criterion. And the specification context rebound the transformation on the assay modules but not on the engine the permutation path resolves, so **all 25 specifications fed default values to their nulls and produced identical effects**. A robustness matrix can thus report perfect stability because its alternatives never reached the code under test — a bug no determinism check can catch, since the run is perfectly reproducible either way. +Recomputed at 100 shuffles per specification (resolution 0.010), with the specification reaching the null engine and Benjamini-Hochberg applied across 100 component tests, C4 remains topology-dependent in 25/25 specifications. C3 has no structural rejection in 25/25, unchanged from 25 unadjusted retentions; 17 pass the descriptive gap criterion and 8 remain indeterminate. Under the 25 specifications: -Recomputed at 100 shuffles per specification (resolution 0.010), with the specification pushed into the null engine and Benjamini-Hochberg applied across the 100 structural tests, both topology conclusions still hold 25/25 and 25/25. The nuance the old matrix could not show is that a retention rate for a conclusion shaped as a *failure to reject* is only as strong as the equivalence behind it: of C3's 25, 17 reach equivalence within the prespecified margin and 8 are merely indeterminate. Under the 25 specifications: +The adjustment acts asymmetrically on these predicates: it can remove the rejection that retains C4, but it cannot strengthen C3's failure to reject into evidence of equivalence. -* The domain's most quotable claim — that the agonist suppresses activity — is retained by only 15 and **reversed** by all 10 monotone specifications: the same compound at the same engagement takes the network from 6.631 to 10.1 Hz (+52%) instead of from 6.631 to 0.458 Hz, with `g_ach` at 1.396 against 0.050. The direction is a property of one asserted term, not of the integration. -* A buffering claim is retained 18 of 25 and reversed by none, its 7 failures being undecidable. The previous explanation — that the monotone rules buffer "at least as much, never less" — is **false**: over the 6 reversing specifications with a defined index the mean gap runs -0.199 to -1.734 log₁₀ units, 3 buffer *less* (linear@1.50x, saturating@1.25x, saturating@1.50x), and the aggregate is carried by the extreme — an aggregate standing in for a range it does not represent. +* The claim that the agonist suppresses activity is retained by only 15 and **reversed** by all 10 monotone specifications: the same compound at the same engagement takes the network from 6.631 to 10.1 rate-model units (+52%) instead of from 6.631 to 0.458 rate-model units, with `g_ach` at 1.396 against 0.050. The direction is a property of one asserted term, not of the integration. +* A buffering claim is retained 18 of 25 and reversed by none, its 7 failures being undecidable. The previous claim that monotone rules buffer "at least as much, never less" is **false**: over the 6 reversing specifications with a defined index the mean gap runs -0.199 to -1.734 log₁₀ units, 3 buffer *less* (linear@1.50x, saturating@1.25x, saturating@1.50x), and the aggregate is carried by the extreme; an aggregate standing in for a range it does not represent. * The matching amplification claim is weakest at 19 of 25. Retained by every specification: C2_rdl_disinhibition, C3_imidacloprid_topology_not_distinguishable, C4_fipronil_topology_exceeds, C7_map_bitter_veto. -The threshold grid makes the same point about conventions rather than transformations: the split is **not** stable (`stable = no`), buffering holding in 9 of 9 cells (-0.362 to -1.224 log₁₀ units) and amplification in 8 of 9, flipping at the strictest corner (circuit_frac=0.50,vert_limit=0.10, gap -0.075), so ranges are quoted and never the point estimate. Two tables that appear to disagree in fact score different sets — T15 the 10 compounds of two mechanism classes, T8 the whole library — so acetylcholine is scored in one and outside the other's scope. +The earlier 25/25 topology result was vacuous for two reasons. At 6 shuffles, the smallest probability was 1/7 = 0.143, so `p ≤ 0.05` was unattainable and the predicate fell back on a Gaussian criterion. More seriously, alternative transformations were rebound on assay modules but not on the engine resolved by the permutation path, so all 25 specifications fed default values to their nulls. The run was deterministic but unable to make the intended comparison. + +The threshold split is also unstable (`stable = no`): buffering holds in 9 of 9 cells (-0.362 to -1.224 log₁₀ units), while amplification holds in 8 of 9 and flips at circuit_frac=0.50,vert_limit=0.10 (gap -0.075). T15 covers 10 compounds in two mechanism classes; acetylcholine is scored in the whole-library T8 but lies outside T15's scope. ### 3.5 Variance attribution, and what to measure next (Fig. F15, Tables T16, T17) -The Sobol' analysis (11 264 model evaluations at n_base = 1024, Var(Y) = 3.149 Hz²) resolves 2 of 9 factors in the strict sense of a bootstrap first-order interval excluding zero: gain_transform, weight_threshold — the **engagement-to-gain transformation** (S1 0.410, ST 0.626) and the cut's **synapse-weight threshold** (S1 0.294, ST 0.542). Everything else is unresolved at this sample size (drive, expression, gain_coef, hill_n, lif_seed, potency, transmitter), including the external drive, whose S1 of 0.019 sits below the estimator's noise floor of 0.046 on an interval of (-0.072, +0.112). The floor is set by gain_coef, not by the declared null factor `lif_seed` at -0.007; the previous version of this section quoted the null factor and then ranked a factor inside the resulting noise. Interactions carry 0.344 of the variance raw and 0.272 with negatives clipped. On the analytic benchmark the estimator returns 0.3163, 0.4495, 0.0072 against the exact 0.3139, 0.4424, 0.0000 at n_base = 16 384 (max error 0.007). +The Sobol' analysis (11 264 model evaluations at n_base = 1024, Var(Y) = 3.149 squared rate-model units) resolves 2 of 9 factors in the strict sense of a bootstrap first-order interval excluding zero: gain_transform, weight_threshold; the **engagement-to-gain transformation** (S1 0.410, ST 0.626) and the cut's **synapse-weight threshold** (S1 0.294, ST 0.542). Everything else is unresolved at this sample size (drive, expression, gain_coef, hill_n, lif_seed, potency, transmitter), including the external drive, whose S1 of 0.019 sits below the largest observed negative-index magnitude of 0.046 on an interval of (-0.072, +0.112). This finite-sample diagnostic is set by gain_coef, not by the declared null factor `lif_seed` at -0.007; the previous version of this section quoted the null factor and then ranked a factor inside the resulting noise. The residual `1 − sum(S1)` is 0.344 raw and 0.272 with negatives clipped; estimator error contributes to this difference, so neither is a precise interaction fraction. On the analytic benchmark the estimator returns 0.3163, 0.4495, 0.0072 against the exact 0.3139, 0.4424, 0.0000 at n_base = 16 384 (max error 0.007). -The top factor is **gain_transform** (1.290 of 3.149 Hz²), resolved by *calibration of receptor engagement against synaptic gain* — reading the *shape* of the response-versus-engagement curve at a known synapse, which is precisely the assumption §3.4 shows the suppression conclusion turns on. The second, **weight_threshold** (0.925 Hz²), needs **no experiment at all**: it is resolved by re-analysing synapse-confidence strata of data already held. +The top factor is **gain_transform** (1.290 of 3.149 squared rate-model units), resolved by *calibration of receptor engagement against synaptic gain*; reading the *shape* of the response-versus-engagement curve at a known synapse, which is precisely the assumption §3.4 shows the suppression conclusion turns on. The second, **weight_threshold** (0.925 squared rate-model units), motivates re-analysis of synapse-confidence strata. Such analysis measures sensitivity to extraction choices; it does not resolve missing connectivity or guarantee the stated variance reduction in practice. ### 3.6 What the refusal rule actually does (Fig. F11, Tables T10, T20) -The type system changes results rather than documentation. Of 102 library rows, 56 carry no admissible evidence and return N/A instead of a small plausible number, and the admissible transformation is decided by the (parameter type, distance) pair rather than by the author: 1 row earns a physical binding model, 1 a labelled binding proxy, 9 a direct functional engagement and 35 a transferred one. The single physical row is imidacloprid at insect_nAChR_native_dmel (Kd 2e-09 M) [77]; the library's previous binding constant, imidacloprid at insect_nAChR_beta1 (Kd 8.3e-11 M), is demoted to a cross-species proxy — the same number under a name that says it was transferred. Removing that one row would leave a branch of the type system with no instances, which is the kind of fact a lookup table hides and an exhaustively checked invariant surfaces. The full census, the domain readouts the engagements feed, and the literature-concordance analysis are in §S6 and §S10. +Evidence gating changes numerical outputs and their interpretation. Of 102 library rows, 56 carry no admissible evidence and return N/A instead of a small plausible number. Dispatch on the (parameter type, distance) pair assigns 1 row a physical binding-occupancy model, 1 a binding proxy, 9 a direct functional engagement and 35 a transferred one. The physical row is imidacloprid at insect_nAChR_native_dmel (Kd 2e-09 M) [76]; the previous binding constant, imidacloprid at insect_nAChR_beta1 (Kd 8.3e-11 M), is retained numerically but labelled as a cross-species proxy. Without the new exact-target row, the shipped library would have no example exercising the physical-occupancy branch. The finite audit makes that coverage visible; it does not introduce new Hill algebra. The full census, downstream domain readouts and literature-concordance analysis are in §S6 and §S10. -### 3.7 Reproduction, determinism and cross-backend parity (Fig. F1) +### 3.7 Availability and reproducibility (engineering support; Fig. F1) -One command regenerates every figure, table and number in this paper from committed artifacts, with no download and no access to the 1.1 GB weight matrix: +The standard command recomputes the core analyses on committed cuts and renders the paper. The scale table is an exception: it imports `papers/scale_study.json`, a committed earlier result, rather than recomputing the large-rung profiles. Only the smallest scale cut is committed; a fresh scale study requires the missing cuts and their upstream data. Re-rendering T27 is therefore not an independent reproduction of that experiment: ```bash python -m pip install -e ".[dev,viz]" @@ -182,37 +191,45 @@ python scripts/reproduce_paper.py --fast # reduced permutation and replicate c python scripts/reproduce_paper.py --only dependence --outdir /tmp/x ``` -Beyond the paper, `flylab run` turns a declarative spec into a self-describing run directory whose manifest records the spec, code and library hashes, the seeds and two hashes per artifact; the same spec and seed reproduce it byte for byte (`docs/WORKFLOW.md`). `papers/results.json` carries every scalar quoted above with its value, rendered text, unit and producing step; two runs at one commit and seed produce identical values, runtimes excepted, and a test asserts it. Every numeric span in the manuscript is a substitution from that record, so a number cannot enter the prose unless the pipeline produced it, and the test suite asserts that every key either document quotes exists in the record, that the record did not come from a reduced-effort run, and — new here, because the previous record's shuffle count did not match the shipped default and nothing caught it — that **every statistical parameter in the record equals the shipped default**. +This section records engineering support rather than a fourth research question. `flylab run` produces a manifest containing spec, code and library hashes, seeds and two hashes per artifact; a test reproduces that artifact byte for byte (`docs/WORKFLOW.md`). The paper determinism test covers only the reduced provenance-and-census path, not every full analysis. `papers/results.json` records each quoted value, unit and producing step, and guards reject reduced-effort records or mismatched declared defaults. + +The same wheel runs on Pyodide. `tests/test_browser_bridge.py` asserts **endpoint parity** for a representative call to each route, including JSON readouts, warnings and provenance; it does not establish whole-analysis equality. An archival DOI and release tag remain outstanding. -The same wheel runs in a browser tab on Pyodide and is published as a static site, so these assays can be re-run by opening one link with nothing installed; `tests/test_browser_bridge.py` asserts **endpoint parity** — equal JSON, every readout, warning and provenance field, for a representative call to each route, which is route-level equality rather than a whole-analysis equality claim. MaleCNS v1.0 and FlyWire are CC-BY 4.0 from their consortia [5], [6] and must be cited before this software; licences are listed under Statements, and an archival Zenodo deposit will carry the DOI **[to be minted on acceptance]**. +The latest attempted full numerical verification was interrupted during dependence analysis. This revision preserves the recorded numerical source and edits its prose; it is not presented as a completed independent rerun. ## 4. Discussion -**What transfers.** *First*, the dependence verdict is a joint property of a prediction and the extract it is measured on, and the extract's recurrence predicts which way it goes. The same validated instrument, with no parameter changed, classifies 42 of 84 cells composition-dominated on a sparse extract (mean degree 1.208) and 0 of 84 on a denser one (mean degree 10.4), and across 7 extracts the composition-dominated verdict survives on 1 — the in-star — while an extract with *fewer* nodes and twenty times the mean degree is already topology-dependent. The practical rule generalises to any pipeline of this shape: **a dependence result must be reported with the structural statistics of the extract it was measured on, and a negative measured on a sparse extract is not evidence about the network it was cut from.** Because the structural statistic that predicts the flip is computable before any permutation is run (T19, T27), this is a cheap check, not a research programme. +**What transfers.** Dependence labels change across extracts: 42 of 84 cells are composition-dominated on `named`, versus 0 of 84 on `taste_motor`. Only one extract has that label, and several structural properties differ at once, so recurrence is not isolated. Evidence typing also changes outputs by refusing unsupported numbers and labelling transferred constants. Finally, the biased null and disconnected specification context show why reproducibility alone is insufficient: both broken comparisons were deterministic and both flattered the conclusions. The gain transformation is consequently both the source of the direction reversal and the largest resolved variance factor. -*Second*, typing the evidence changes results rather than documentation: refusing inadmissible rows removed a class of small, plausible, invented numbers no reader could have told apart from measurements, and grading a source's distance demoted a cross-species constant to a labelled proxy without changing its value. Stated as an invariant and checked exhaustively, this is a property of the pipeline rather than a convention its authors intend to follow. +Other literature-parameterised network domains can apply the same audit without adopting this biological model: -*Third*, two of the instruments were themselves broken in ways that flattered the conclusions, and both were found by pointing the instruments inward: a null that moved the quantity it claimed to preserve exactly, and a specification family whose alternatives never reached the engine its predicates ran on, so a uniform robustness rate was guaranteed by construction. Both failures share a signature worth naming — **a test whose result was determined by its own implementation rather than by the data** — and both were invisible to determinism checks, because a record can be perfectly reproducible and still be produced by a test that could not have returned another answer. The specification family and the variance decomposition then converge on one weak point: the direction of the headline effect is set by an asserted functional form, which is also the largest resolved factor in the variance budget and the top-ranked measurement. +1. type each external parameter and reject transformations its provenance does not license; +2. state the extract, readout, runtime normalisation and statistical effort with every dependence result; +3. design each null around the quantity it must preserve, then measure whether it actually preserves it; +4. trace every alternative specification into the engine used by the test, including null-generation paths; +5. compare extracts and runtime conventions before generalising, and use variance attribution to select the next measurement. -**Encoded assumptions versus generated observations.** The separation is itself generated rather than asserted: an audit of the dependency chain behind one readout returns 9 links (T18), of which exactly 1 link is a measurement of the system being simulated, 1 reuses a measurement of something else, 4 are asserted modelling choices and 3 are arithmetic on the links above. A companion decomposition — a different decomposition, not a repartition of those links — reports 5 facts, 1 model inference and 5 unknowns. That one link in nine is a measurement of the simulated system is worth stating plainly for any pipeline of this shape. +These are proposed audit steps, not evidence that FlyLab has been implemented in power, metabolic or epidemic networks. -**Relation to adjacent work.** Our version of the [69], [70] result survives neither a change of extract (§3.2) nor a change of the runtime's normalisation (§3.3), and both models normalise or regionally average, so the parallel is at least as consistent with a shared modelling artefact as with a property of receptor-weighted network perturbation. That is itself a methodological observation: a result reproduced across two pipelines that share a normalisation convention has not been independently confirmed. +**Encoded assumptions versus generated observations.** The audit returns 9 links (T18): 1 link measures the simulated system, 1 measures something else, 4 are asserted and 3 are computed. A separate decomposition, not a repartition of those links, reports 5 facts, 1 model inference and 5 unknowns. -**What the pipeline cannot support**, and these limits are themselves computing results about what such a pipeline can establish. It cannot predict a dose–response in an animal: with no pharmacokinetics available, the free concentration is an *input*, not a derived quantity. It cannot support a cross-species safety claim, nor a subunit-level claim where the sources used hybrid constructs. It cannot support a general statement about whether perturbation on networks of this kind needs topology — only a statement per extract with that extract's structure reported. And it cannot produce a live-animal result, which no code path may write. +**Relation to adjacent work.** Our version of the [68], [69] result survives neither a change of extract (§3.2) nor runtime normalisation (§3.3). Because both models normalise or regionally average, the parallel may reflect a shared convention rather than independent confirmation. -**Five model-specific caveats** are given in full in §S9: an asserted sign table exempt from the specification family and unbounded by the label-permuting null; the load-bearing row normalisation; perturbations applied across transmitter-defined rather than receptor-defined synapses, which the expression layer bounds rather than corrects; hops-limited extracts, whose missing inhibition can flip the sign of any readout and whose cost §3.2 now quantifies; and agreement of direction only between the two runtimes. +**What the pipeline cannot support.** Without pharmacokinetics, free concentration is an input rather than an animal dose. The model supports neither cross-species safety nor subunit claims from hybrid constructs, and it cannot generalise topology dependence beyond the reported extract. No code path may write a live-animal result. -**Prospective, not pre-registered.** The hypotheses in §S5 are written before any experiment and labelled as software predictions with no live result; a mutable repository is not a registry, so they become pre-registered only once a tagged release is externally archived. Two of them predict a readout whose *sign* is a modelling choice over this paper's own specification family, and the supplement now says so beside them. +**Five model-specific caveats** are given in full in §S9: an asserted sign table exempt from the specification family and unbounded by the label-permuting null; the load-bearing row normalisation; perturbations applied across transmitter-defined rather than receptor-defined synapses, which the expression layer bounds rather than corrects; hops-limited extracts with unobserved boundary inputs and extract-sensitive results; and agreement of direction only between the two runtimes. Expression mapping coverage is 0.205, and the principal MN9 readout remains unresolved. + +**Prospective, not pre-registered.** The hypotheses in §S5 are written before any experiment and labelled as software predictions with no live result; a mutable repository is not a registry, and an archive DOI alone does not establish preregistration. A dated, frozen experimental protocol specifying hypotheses, outcomes and analyses before collection is also needed. Two of them predict a readout whose *sign* is a modelling choice over this paper's own specification family, and the supplement now says so beside them. ## 5. Conclusion -We set out to make a network simulation say which of its inputs its predictions depend on, and to evaluate that machinery rather than assert it. The method is four instruments — typed evidence propagation under a checked soundness invariant, input-dependence testing against an ordered ladder with an explicit equivalence margin and multiplicity control, an ablation ladder scored against a matched reference, and specification-family robustness with variance attribution — none specific to the domain we demonstrate them on. Pointed at its own results, the method establishes that the headline effect's direction is a property of an asserted functional form; that a layer-agreement score near 1 can be an algebraic consequence of the pipeline's own normalisation; and — the result we would most want another group to break — that **a dependence verdict obtained on one extract of a network can invert on another extract of the same network, with recurrence rather than size predicting which way**. A "this prediction does not need the structure" result is only as good as the extract it was measured on, and the statistic that says so is cheap to compute in advance. +We set out to make a network simulation say which of its inputs its predictions depend on, and to evaluate that machinery rather than assert it. The method combines four components: typed evidence propagation under an enforced transformation rule and finite-library invariant audit, input-dependence diagnostics with an explicit point-gap tolerance and component-test multiplicity adjustment, an ablation ladder scored against a matched reference, and specification-family robustness with variance attribution; none specific to the domain we demonstrate them on. Pointed at its own results, the method establishes that the headline effect's direction is a property of an asserted functional form; that layer-agreement scores depend on shared gains and runtime normalisation; and that **dependence labels can change across extracts of the same network**. The present study does not isolate recurrence from correlated extraction properties. A "this prediction does not need the structure" result is only as good as the extract it was measured on, and structural summaries can flag the need for additional checks; they are not yet validated substitutes for those checks. What would change the picture is short and specific, and the variance budget rather than taste names it: a calibration of the parameter-to-gain transformation at a known synapse; a reconstruction-confidence re-analysis, which needs no new measurement; a full landscape on an extract that is neither an in-star nor hops-limited; and one live assay. Until then this is a description of a method, and of a tool that tests itself. ## Statements -**AI assistance.** The author defined the research questions, methods and analyses; verified the source literature, checking each cited publication against the value attributed to it; executed and reviewed every analysis; and accepts full responsibility for the scientific content. AI assistance was used for software implementation and for drafting and editing prose. No data were generated, imputed, estimated or extrapolated by a language model; every number in the results, tables and figures was produced by the reproduction pipeline from committed artifacts and is keyed in `papers/results.json`. +**AI assistance.** AI tools assisted implementation, statistical checking, literature lookup, peer review and drafting. The reported simulation outputs were computed by the analysis code. The author is responsible for verifying the source literature, analysis choices and final scientific claims before submission; automated checks do not establish that this author verification has occurred. **Animal research.** No animal experiments were performed and no live-animal data are reported. The `live_lab` field of every notebook produced by this pipeline is `null`. The protocols in `protocols/` are unexecuted drafts. @@ -224,157 +241,155 @@ What would change the picture is short and specific, and the variance budget rat [1] J. Dempster, *RatCVS*, University of Strathclyde, Glasgow, UK. Simulated pithed-rat cardiovascular pharmacology teaching package. -[2] S. Okabe *et al.*, "A web-based simulation of cardiovascular pharmacology for undergraduate teaching," *Eur. J. Pharmacol.*, vol. 997, art. 177618, 2025. - -[3] *PharmVR*: immersive virtual-reality pharmacology training, preprint, 2025. +[2] S. Okabe, T. Goto, D. Hirayama and Y. Nishimura, "Virtual rat web: A versatile simulation tool for pharmacology education in a variety of settings," *Eur. J. Pharmacol.*, vol. 997, art. 177618, 2025. DOI 10.1016/j.ejphar.2025.177618. PMID 40222443. -[4] "Virtual simulation in pharmacology and physiology education: an evaluation," *J. Educ. Eval. Health Prof.*, 2020. +[3] K. Thiyagarajan, "PharmVR, Immersive Virtual Reality based Animal Simulator Software," *Preprints.org*, concept paper, 2025. DOI 10.20944/preprints202507.0959.v1. Not peer reviewed. -[5] S. Dorkenwald *et al.*, "Neuronal wiring diagram of an adult brain," *Nature*, vol. 634, pp. 124–138, 2024. DOI 10.1038/s41586-024-07558-y. +[4] S. Dorkenwald *et al.*, "Neuronal wiring diagram of an adult brain," *Nature*, vol. 634, pp. 124–138, 2024. DOI 10.1038/s41586-024-07558-y. -[6] HHMI Janelia FlyEM, University of Cambridge, MRC LMB and Google Research, *MaleCNS v1.0: a synapse-resolution connectome of the adult male* Drosophila *central nervous system*, 2026. CC-BY 4.0. +[5] HHMI Janelia FlyEM, University of Cambridge, MRC LMB and Google Research, *MaleCNS v1.0: a synapse-resolution connectome of the adult male* Drosophila *central nervous system*, 2026. CC-BY 4.0. -[7] P. K. Shiu, G. R. Sterne, N. Spiller *et al.*, "A leaky integrate-and-fire computational model based on the connectome of the entire adult *Drosophila* brain reveals insights into sensorimotor processing," *Nature*, vol. 634, 2024. +[6] P. K. Shiu, G. R. Sterne, N. Spiller *et al.*, "A leaky integrate-and-fire computational model based on the connectome of the entire adult *Drosophila* brain reveals insights into sensorimotor processing," *Nature*, vol. 634, 2024. -[8] C. Bass *et al.*, "Mutation analysis and functional characterisation of insect nicotinic acetylcholine receptors," *Insect Biochem. Mol. Biol.*, vol. 36, pp. 86–96, 2006. +[7] C. Bass *et al.*, "Mutation analysis and functional characterisation of insect nicotinic acetylcholine receptors," *Insect Biochem. Mol. Biol.*, vol. 36, pp. 86–96, 2006. -[9] H. Dederer *et al.*, "Selectivity of neonicotinoids at hybrid insect nicotinic acetylcholine receptors," *Insect Biochem. Mol. Biol.*, vol. 41, p. 51, 2011. +[8] H. Dederer *et al.*, "Selectivity of neonicotinoids at hybrid insect nicotinic acetylcholine receptors," *Insect Biochem. Mol. Biol.*, vol. 41, p. 51, 2011. -[10] FlyConnectome, *Neurotransmitter ground-truth repository for* Drosophila *connectomes*, 2024. +[9] N. Eckstein *et al.*, "Neurotransmitter classification from electron microscopy images at synaptic sites in *Drosophila melanogaster*," *Cell*, vol. 187, pp. 2574–2594.e23, 2024. DOI 10.1016/j.cell.2024.03.016. PMID 38729112. PMCID PMC11106717. -[11] P. Mineault, "Deconstructing viral fly simulations," technical note, 2026. +[10] P. Mineault, "Deconstructing viral fly sims," *The NeuroAI archive*, 14 September 2026. https://www.neuroai.science/p/are-flies-playing-beat-saber. Newsletter essay, not peer reviewed. -[12] Y. Deng *et al.*, "Chemoconnectomics: mapping chemical transmission in *Drosophila*," *Neuron*, vol. 101, pp. 876–893, 2019. +[11] Y. Deng *et al.*, "Chemoconnectomics: mapping chemical transmission in *Drosophila*," *Neuron*, vol. 101, pp. 876–893, 2019. -[13] C. Bass, A. M. Puinean, M. Andrews *et al.*, "Mutation of a nicotinic acetylcholine receptor β subunit is associated with resistance to neonicotinoid insecticides in the aphid *Myzus persicae*," *BMC Neurosci.*, vol. 12, art. 51, 2011. DOI 10.1186/1471-2202-12-51. PMID 21627790. +[12] C. Bass, A. M. Puinean, M. Andrews *et al.*, "Mutation of a nicotinic acetylcholine receptor β subunit is associated with resistance to neonicotinoid insecticides in the aphid *Myzus persicae*," *BMC Neurosci.*, vol. 12, art. 51, 2011. DOI 10.1186/1471-2202-12-51. PMID 21627790. -[14] R. H. ffrench-Constant, T. A. Rocheleau, J. C. Steichen and A. E. Chalmers, "A point mutation in a *Drosophila* GABA receptor confers insecticide resistance," *Nature*, vol. 363, pp. 449–451, 1993. DOI 10.1038/363449a0. +[13] R. H. ffrench-Constant, T. A. Rocheleau, J. C. Steichen and A. E. Chalmers, "A point mutation in a *Drosophila* GABA receptor confers insecticide resistance," *Nature*, vol. 363, pp. 449–451, 1993. DOI 10.1038/363449a0. -[15] W. T. Garrood, C. T. Zimmer, O. Gutbrod *et al.*, "Influence of the RDL A301S mutation in the brown planthopper *Nilaparvata lugens* on the activity of phenylpyrazole insecticides," *Pestic. Biochem. Physiol.*, vol. 142, pp. 1–8, 2017. DOI 10.1016/j.pestbp.2017.01.007. PMID 29107231. +[14] W. T. Garrood, C. T. Zimmer, O. Gutbrod *et al.*, "Influence of the RDL A301S mutation in the brown planthopper *Nilaparvata lugens* on the activity of phenylpyrazole insecticides," *Pestic. Biochem. Physiol.*, vol. 142, pp. 1–8, 2017. DOI 10.1016/j.pestbp.2017.01.007. PMID 29107231. -[16] A. M. Hosie, H. A. Baylis, S. D. Buckingham and D. B. Sattelle, "Actions of the insecticide fipronil on dieldrin-sensitive and -resistant GABA receptors of *Drosophila melanogaster*," *Br. J. Pharmacol.*, vol. 115, no. 6, pp. 909–912, 1995. DOI 10.1111/j.1476-5381.1995.tb15896.x. PMID 7582519. +[15] A. M. Hosie, H. A. Baylis, S. D. Buckingham and D. B. Sattelle, "Actions of the insecticide fipronil on dieldrin-sensitive and -resistant GABA receptors of *Drosophila melanogaster*," *Br. J. Pharmacol.*, vol. 115, no. 6, pp. 909–912, 1995. DOI 10.1111/j.1476-5381.1995.tb15896.x. PMID 7582519. -[17] A. M. Hosie and D. B. Sattelle, "Allosteric modulation of an expressed homo-oligomeric GABA-gated chloride channel of *Drosophila melanogaster*," *Br. J. Pharmacol.*, vol. 117, no. 6, pp. 1229–1237, 1996. DOI 10.1111/j.1476-5381.1996.tb16720.x. PMID 8882620. +[16] A. M. Hosie and D. B. Sattelle, "Allosteric modulation of an expressed homo-oligomeric GABA-gated chloride channel of *Drosophila melanogaster*," *Br. J. Pharmacol.*, vol. 117, no. 6, pp. 1229–1237, 1996. DOI 10.1111/j.1476-5381.1996.tb16720.x. PMID 8882620. -[18] D. H. Kwon, K. S. Yoon, J. M. Clark and S. H. Lee, "A point mutation in a glutamate-gated chloride channel confers abamectin resistance in the two-spotted spider mite, *Tetranychus urticae* Koch," *Insect Mol. Biol.*, vol. 19, pp. 583–591, 2010. DOI 10.1111/j.1365-2583.2010.01017.x. PMID 20522121. +[17] D. H. Kwon, K. S. Yoon, J. M. Clark and S. H. Lee, "A point mutation in a glutamate-gated chloride channel confers abamectin resistance in the two-spotted spider mite, *Tetranychus urticae* Koch," *Insect Mol. Biol.*, vol. 19, pp. 583–591, 2010. DOI 10.1111/j.1365-2583.2010.01017.x. PMID 20522121. -[19] G. Le Goff, A. Hamon, J.-B. Bergé and M. Amichot, "Resistance to fipronil in *Drosophila simulans*: influence of two point mutations in the RDL GABA receptor subunit," *J. Neurochem.*, vol. 92, pp. 1295–1305, 2005. DOI 10.1111/j.1471-4159.2004.02922.x. PMID 15748149. +[18] G. Le Goff, A. Hamon, J.-B. Bergé and M. Amichot, "Resistance to fipronil in *Drosophila simulans*: influence of two point mutations in the RDL GABA receptor subunit," *J. Neurochem.*, vol. 92, pp. 1295–1305, 2005. DOI 10.1111/j.1471-4159.2004.02922.x. PMID 15748149. -[20] P. Menozzi, M. A. Shi, A. Lougarre, Z. H. Tang and D. Fournier, "Mutations of acetylcholinesterase which confer insecticide resistance in *Drosophila melanogaster* populations," *BMC Evol. Biol.*, vol. 4, art. 4, 2004. DOI 10.1186/1471-2148-4-4. PMID 15018651. +[19] P. Menozzi, M. A. Shi, A. Lougarre, Z. H. Tang and D. Fournier, "Mutations of acetylcholinesterase which confer insecticide resistance in *Drosophila melanogaster* populations," *BMC Evol. Biol.*, vol. 4, art. 4, 2004. DOI 10.1186/1471-2148-4-4. PMID 15018651. -[21] A. Mutero, M. Pralavorio, J. M. Bride and D. Fournier, "Resistance-associated point mutations in insecticide-insensitive acetylcholinesterase," *Proc. Natl. Acad. Sci. USA*, vol. 91, no. 13, pp. 5922–5926, 1994. DOI 10.1073/pnas.91.13.5922. PMID 8016090. +[20] A. Mutero, M. Pralavorio, J. M. Bride and D. Fournier, "Resistance-associated point mutations in insecticide-insensitive acetylcholinesterase," *Proc. Natl. Acad. Sci. USA*, vol. 91, no. 13, pp. 5922–5926, 1994. DOI 10.1073/pnas.91.13.5922. PMID 8016090. -[22] P. N. R. Usherwood, H. Vais, B. P. S. Khambay, T. G. E. Davies and M. S. Williamson, "Sensitivity of the *Drosophila para* sodium channel to DDT is not lowered by the super-kdr mutation M918T on the IIS4–S5 linker that profoundly reduces sensitivity to permethrin and deltamethrin," *FEBS Lett.*, vol. 579, no. 28, pp. 6317–6325, 2005. DOI 10.1016/j.febslet.2005.09.096. PMID 16263118. +[21] P. N. R. Usherwood, H. Vais, B. P. S. Khambay, T. G. E. Davies and M. S. Williamson, "Sensitivity of the *Drosophila para* sodium channel to DDT is not lowered by the super-kdr mutation M918T on the IIS4–S5 linker that profoundly reduces sensitivity to permethrin and deltamethrin," *FEBS Lett.*, vol. 579, no. 28, pp. 6317–6325, 2005. DOI 10.1016/j.febslet.2005.09.096. PMID 16263118. -[23] H. Vais, M. S. Williamson, S. J. Goodson *et al.*, "Activation of *Drosophila* sodium channels promotes modification by deltamethrin: reductions in affinity caused by knock-down resistance mutations," *J. Gen. Physiol.*, vol. 115, no. 3, pp. 305–318, 2000. DOI 10.1085/jgp.115.3.305. PMID 10694259. +[22] H. Vais, M. S. Williamson, S. J. Goodson *et al.*, "Activation of *Drosophila* sodium channels promotes modification by deltamethrin: reductions in affinity caused by knock-down resistance mutations," *J. Gen. Physiol.*, vol. 115, no. 3, pp. 305–318, 2000. DOI 10.1085/jgp.115.3.305. PMID 10694259. -[24] X. Wang, A. M. Puinean, A. O. O'Reilly *et al.*, "Mutations on M3 helix of *Plutella xylostella* glutamate-gated chloride channel confer unequal resistance to abamectin by two different mechanisms," *Insect Biochem. Mol. Biol.*, vol. 86, pp. 50–57, 2017. DOI 10.1016/j.ibmb.2017.05.006. PMID 28576654. +[23] X. Wang, A. M. Puinean, A. O. O'Reilly *et al.*, "Mutations on M3 helix of *Plutella xylostella* glutamate-gated chloride channel confer unequal resistance to abamectin by two different mechanisms," *Insect Biochem. Mol. Biol.*, vol. 86, pp. 50–57, 2017. DOI 10.1016/j.ibmb.2017.05.006. PMID 28576654. -[25] H. G. Zhang, R. H. ffrench-Constant and M. B. Jackson, "A unique amino acid of the *Drosophila* GABA receptor with influence on drug sensitivity by two mechanisms," *J. Physiol.*, vol. 479, no. 1, pp. 65–75, 1994. DOI 10.1113/jphysiol.1994.sp020278. PMID 7527461. +[24] H. G. Zhang, R. H. ffrench-Constant and M. B. Jackson, "A unique amino acid of the *Drosophila* GABA receptor with influence on drug sensitivity by two mechanisms," *J. Physiol.*, vol. 479, no. 1, pp. 65–75, 1994. DOI 10.1113/jphysiol.1994.sp020278. PMID 7527461. -[26] A. Cartereau, C. Martin and S. H. Thany, "Neonicotinoid insecticides differently modulate acetylcholine-induced currents on mammalian α7 nicotinic acetylcholine receptors," *Br. J. Pharmacol.*, vol. 175, no. 11, pp. 1987–1998, 2018. DOI 10.1111/bph.14018. +[25] A. Cartereau, C. Martin and S. H. Thany, "Neonicotinoid insecticides differently modulate acetylcholine-induced currents on mammalian α7 nicotinic acetylcholine receptors," *Br. J. Pharmacol.*, vol. 175, no. 11, pp. 1987–1998, 2018. DOI 10.1111/bph.14018. -[27] L. M. Cole, R. A. Nicholson and J. E. Casida, "Action of phenylpyrazole insecticides at the GABA-gated chloride channel," *Pestic. Biochem. Physiol.*, vol. 46, no. 1, pp. 47–54, 1993. DOI 10.1006/pest.1993.1035. +[26] L. M. Cole, R. A. Nicholson and J. E. Casida, "Action of phenylpyrazole insecticides at the GABA-gated chloride channel," *Pestic. Biochem. Physiol.*, vol. 46, no. 1, pp. 47–54, 1993. DOI 10.1006/pest.1993.1035. -[28] G. R. Dawson, K. A. Wafford, A. Smith *et al.*, "Anticonvulsant and adverse effects of avermectin analogs in mice are mediated through the γ-aminobutyric acid A receptor," *J. Pharmacol. Exp. Ther.*, vol. 295, no. 3, pp. 1051–1060, 2000. PMID 11082440. +[27] G. R. Dawson, K. A. Wafford, A. Smith *et al.*, "Anticonvulsant and adverse effects of avermectin analogs in mice are mediated through the γ-aminobutyric acid A receptor," *J. Pharmacol. Exp. Ther.*, vol. 295, no. 3, pp. 1051–1060, 2000. PMID 11082440. -[29] K. Lees, M. Musgaard, S. Suwanmanee *et al.*, "Actions of agonists, fipronil and ivermectin on the predominant in vivo splice and edit variant (RDLbd, I/V) of the *Drosophila* GABA receptor expressed in *Xenopus laevis* oocytes," *PLoS ONE*, vol. 9, no. 5, art. e97468, 2014. DOI 10.1371/journal.pone.0097468. +[28] K. Lees, M. Musgaard, S. Suwanmanee *et al.*, "Actions of agonists, fipronil and ivermectin on the predominant in vivo splice and edit variant (RDLbd, I/V) of the *Drosophila* GABA receptor expressed in *Xenopus laevis* oocytes," *PLoS ONE*, vol. 9, no. 5, art. e97468, 2014. DOI 10.1371/journal.pone.0097468. -[30] T. Perry, D. G. Heckel, J. A. McKenzie and P. Batterham, "Effects of mutations in *Drosophila* nicotinic acetylcholine receptor subunits on sensitivity to insecticides targeting nicotinic acetylcholine receptors," *Pestic. Biochem. Physiol.*, vol. 102, no. 1, pp. 56–60, 2012. DOI 10.1016/j.pestbp.2011.10.010. +[29] T. Perry, D. G. Heckel, J. A. McKenzie and P. Batterham, "Effects of mutations in *Drosophila* nicotinic acetylcholine receptor subunits on sensitivity to insecticides targeting nicotinic acetylcholine receptors," *Pestic. Biochem. Physiol.*, vol. 102, no. 1, pp. 56–60, 2012. DOI 10.1016/j.pestbp.2011.10.010. -[31] G. S. Ratra and J. E. Casida, "GABA receptor subunit composition relative to insecticide potency and selectivity," *Toxicol. Lett.*, vol. 122, no. 3, pp. 215–222, 2001. DOI 10.1016/s0378-4274(01)00366-6. PMID 11489356. +[30] G. S. Ratra and J. E. Casida, "GABA receptor subunit composition relative to insecticide potency and selectivity," *Toxicol. Lett.*, vol. 122, no. 3, pp. 215–222, 2001. DOI 10.1016/s0378-4274(01)00366-6. PMID 11489356. -[32] G. S. Ratra, S. G. Kamita and J. E. Casida, "Role of human GABA_A receptor β3 subunit in insecticide toxicity," *Toxicol. Appl. Pharmacol.*, vol. 172, no. 3, pp. 233–240, 2001. DOI 10.1006/taap.2001.9154. PMID 11312652. +[31] G. S. Ratra, S. G. Kamita and J. E. Casida, "Role of human GABA_A receptor β3 subunit in insecticide toxicity," *Toxicol. Appl. Pharmacol.*, vol. 172, no. 3, pp. 233–240, 2001. DOI 10.1006/taap.2001.9154. PMID 11312652. -[33] Q. Shan, J. L. Haddrill and J. W. Lynch, "Ivermectin, an unconventional agonist of the glycine receptor chloride channel," *J. Biol. Chem.*, vol. 276, no. 16, pp. 12556–12564, 2001. DOI 10.1074/jbc.M011264200. +[32] Q. Shan, J. L. Haddrill and J. W. Lynch, "Ivermectin, an unconventional agonist of the glycine receptor chloride channel," *J. Biol. Chem.*, vol. 276, no. 16, pp. 12556–12564, 2001. DOI 10.1074/jbc.M011264200. -[34] K. Tasman, S. A. Rands and J. J. L. Hodge, "The power of *Drosophila melanogaster* for modeling neonicotinoid effects on pollinators and identifying novel mechanisms," *Front. Physiol.*, vol. 12, art. 659440, 2021. DOI 10.3389/fphys.2021.659440. +[33] K. Tasman, S. A. Rands and J. J. L. Hodge, "The power of *Drosophila melanogaster* for modeling neonicotinoid effects on pollinators and identifying novel mechanisms," *Front. Physiol.*, vol. 12, art. 659440, 2021. DOI 10.3389/fphys.2021.659440. -[35] M. Tomizawa and J. E. Casida, "Selective toxicity of neonicotinoids attributable to specificity of insect and mammalian nicotinic receptors," *Annu. Rev. Entomol.*, vol. 48, pp. 339–364, 2003. DOI 10.1146/annurev.ento.48.091801.112731. PMID 12208819. +[34] M. Tomizawa and J. E. Casida, "Selective toxicity of neonicotinoids attributable to specificity of insect and mammalian nicotinic receptors," *Annu. Rev. Entomol.*, vol. 48, pp. 339–364, 2003. DOI 10.1146/annurev.ento.48.091801.112731. PMID 12208819. -[36] M. Tomizawa and J. E. Casida, "Neonicotinoid insecticide toxicology: mechanisms of selective action," *Annu. Rev. Pharmacol. Toxicol.*, vol. 45, pp. 247–268, 2005. DOI 10.1146/annurev.pharmtox.45.120403.095930. +[35] M. Tomizawa and J. E. Casida, "Neonicotinoid insecticide toxicology: mechanisms of selective action," *Annu. Rev. Pharmacol. Toxicol.*, vol. 45, pp. 247–268, 2005. DOI 10.1146/annurev.pharmtox.45.120403.095930. -[37] M. Tomizawa, D. L. Lee and J. E. Casida, "Neonicotinoid insecticides: molecular features conferring selectivity for insect versus mammalian nicotinic receptors," *J. Agric. Food Chem.*, vol. 48, no. 12, pp. 6016–6024, 2000. DOI 10.1021/jf000873c. PMID 11312774. +[36] M. Tomizawa, D. L. Lee and J. E. Casida, "Neonicotinoid insecticides: molecular features conferring selectivity for insect versus mammalian nicotinic receptors," *J. Agric. Food Chem.*, vol. 48, no. 12, pp. 6016–6024, 2000. DOI 10.1021/jf000873c. PMID 11312774. -[38] A. J. Wolstenholme, "Glutamate-gated chloride channels," *J. Biol. Chem.*, vol. 287, no. 48, pp. 40232–40238, 2012. DOI 10.1074/jbc.R112.406280. +[37] A. J. Wolstenholme, "Glutamate-gated chloride channels," *J. Biol. Chem.*, vol. 287, no. 48, pp. 40232–40238, 2012. DOI 10.1074/jbc.R112.406280. -[39] A. M. Allen, M. C. Neville, S. Birtles *et al.*, "A single-cell transcriptomic atlas of the adult *Drosophila* ventral nerve cord," *eLife*, vol. 9, art. e54074, 2020. DOI 10.7554/eLife.54074. PMID 32314735. +[38] A. M. Allen, M. C. Neville, S. Birtles *et al.*, "A single-cell transcriptomic atlas of the adult *Drosophila* ventral nerve cord," *eLife*, vol. 9, art. e54074, 2020. DOI 10.7554/eLife.54074. PMID 32314735. -[40] V. Croset, C. D. Treiber and S. Waddell, "Cellular diversity in the *Drosophila* midbrain revealed by single-cell transcriptomics," *eLife*, vol. 7, art. e34550, 2018. DOI 10.7554/eLife.34550. PMID 29671739. +[39] V. Croset, C. D. Treiber and S. Waddell, "Cellular diversity in the *Drosophila* midbrain revealed by single-cell transcriptomics," *eLife*, vol. 7, art. e34550, 2018. DOI 10.7554/eLife.34550. PMID 29671739. -[41] K. Davie, J. Janssens, D. Koldere *et al.*, "A single-cell transcriptome atlas of the aging *Drosophila* brain," *Cell*, vol. 174, no. 4, pp. 982–998.e20, 2018. DOI 10.1016/j.cell.2018.05.057. PMID 29909982. +[40] K. Davie, J. Janssens, D. Koldere *et al.*, "A single-cell transcriptome atlas of the aging *Drosophila* brain," *Cell*, vol. 174, no. 4, pp. 982–998.e20, 2018. DOI 10.1016/j.cell.2018.05.057. PMID 29909982. -[42] L. Enell, Y. Hamasaka, A. Kolodziejczyk and D. R. Nässel, "γ-Aminobutyric acid (GABA) signaling components in *Drosophila*: immunocytochemical localization of GABA_B receptors in relation to the GABA_A receptor subunit RDL and a vesicular GABA transporter," *J. Comp. Neurol.*, vol. 505, no. 1, pp. 18–31, 2007. DOI 10.1002/cne.21472. PMID 17729251. +[41] L. Enell, Y. Hamasaka, A. Kolodziejczyk and D. R. Nässel, "γ-Aminobutyric acid (GABA) signaling components in *Drosophila*: immunocytochemical localization of GABA_B receptors in relation to the GABA_A receptor subunit RDL and a vesicular GABA transporter," *J. Comp. Neurol.*, vol. 505, no. 1, pp. 18–31, 2007. DOI 10.1002/cne.21472. PMID 17729251. -[43] N. Konstantinides, K. Kapuralin, C. Fadil *et al.*, "Phenotypic convergence: distinct transcription factors regulate common terminal features," *Cell*, vol. 174, no. 3, pp. 622–635.e13, 2018. DOI 10.1016/j.cell.2018.05.021. PMID 29909983. +[42] N. Konstantinides, K. Kapuralin, C. Fadil *et al.*, "Phenotypic convergence: distinct transcription factors regulate common terminal features," *Cell*, vol. 174, no. 3, pp. 622–635.e13, 2018. DOI 10.1016/j.cell.2018.05.021. PMID 29909983. -[44] H. Li, J. Janssens, M. De Waegeneer *et al.*, "Fly Cell Atlas: a single-nucleus transcriptomic atlas of the adult fruit fly," *Science*, vol. 375, no. 6584, art. eabk2432, 2022. DOI 10.1126/science.abk2432. PMID 35239393. +[43] H. Li, J. Janssens, M. De Waegeneer *et al.*, "Fly Cell Atlas: a single-nucleus transcriptomic atlas of the adult fruit fly," *Science*, vol. 375, no. 6584, art. eabk2432, 2022. DOI 10.1126/science.abk2432. PMID 35239393. -[45] C. I. Bliss, "The toxicity of poisons applied jointly," *Ann. Appl. Biol.*, vol. 26, no. 3, pp. 585–615, 1939. DOI 10.1111/j.1744-7348.1939.tb06990.x. +[44] C. I. Bliss, "The toxicity of poisons applied jointly," *Ann. Appl. Biol.*, vol. 26, no. 3, pp. 585–615, 1939. DOI 10.1111/j.1744-7348.1939.tb06990.x. -[46] W. R. Greco, G. Bravo and J. C. Parsons, "The search for synergy: a critical review from a response surface perspective," *Pharmacol. Rev.*, vol. 47, no. 2, pp. 331–385, 1995. PMID 7568331. +[45] W. R. Greco, G. Bravo and J. C. Parsons, "The search for synergy: a critical review from a response surface perspective," *Pharmacol. Rev.*, vol. 47, no. 2, pp. 331–385, 1995. PMID 7568331. -[47] Q. Huang, Y. Deng, T. Zhan and Y. He, "Synergistic and antagonistic effects of piperonyl butoxide in fipronil-susceptible and resistant rice stem borers, *Chilo suppressalis*," *J. Insect Sci.*, vol. 10, art. 182, 2010. DOI 10.1673/031.010.14142. PMID 21062143. +[46] Q. Huang, Y. Deng, T. Zhan and Y. He, "Synergistic and antagonistic effects of piperonyl butoxide in fipronil-susceptible and resistant rice stem borers, *Chilo suppressalis*," *J. Insect Sci.*, vol. 10, art. 182, 2010. DOI 10.1673/031.010.14142. PMID 21062143. -[48] S. Loewe and H. Muischnek, "Über Kombinationswirkungen: 1. Mitteilung: Hilfsmittel der Fragestellung," *Naunyn-Schmiedebergs Arch. Exp. Pathol. Pharmakol.*, vol. 114, pp. 313–326, 1926. DOI 10.1007/BF01952257. +[47] S. Loewe and H. Muischnek, "Über Kombinationswirkungen: 1. Mitteilung: Hilfsmittel der Fragestellung," *Naunyn-Schmiedebergs Arch. Exp. Pathol. Pharmakol.*, vol. 114, pp. 313–326, 1926. DOI 10.1007/BF01952257. -[49] S. Loewe, "The problem of synergism and antagonism of combined drugs," *Arzneimittelforschung*, vol. 3, no. 6, pp. 285–290, 1953. PMID 13081480. +[48] S. Loewe, "The problem of synergism and antagonism of combined drugs," *Arzneimittelforschung*, vol. 3, no. 6, pp. 285–290, 1953. PMID 13081480. -[50] D. Philippou, V. Borzatta, E. Capparella *et al.*, "The use of substituted alkynyl phenoxy derivatives of piperonyl butoxide to control insecticide-resistant pests," *Pest Manag. Sci.*, vol. 72, no. 10, pp. 1946–1950, 2016. DOI 10.1002/ps.4234. PMID 26800141. +[49] D. Philippou, V. Borzatta, E. Capparella *et al.*, "The use of substituted alkynyl phenoxy derivatives of piperonyl butoxide to control insecticide-resistant pests," *Pest Manag. Sci.*, vol. 72, no. 10, pp. 1946–1950, 2016. DOI 10.1002/ps.4234. PMID 26800141. -[51] Y. C. Zhu, J. Yao, J. Adamczyk and R. Luttrell, "Synergistic toxicity and physiological impact of imidacloprid alone and binary mixtures with seven representative pesticides on honey bee (*Apis mellifera*)," *PLoS ONE*, vol. 12, no. 5, art. e0176837, 2017. DOI 10.1371/journal.pone.0176837. PMID 28467462. +[50] Y. C. Zhu, J. Yao, J. Adamczyk and R. Luttrell, "Synergistic toxicity and physiological impact of imidacloprid alone and binary mixtures with seven representative pesticides on honey bee (*Apis mellifera*)," *PLoS ONE*, vol. 12, no. 5, art. e0176837, 2017. DOI 10.1371/journal.pone.0176837. PMID 28467462. -[52] P. J. Daborn, J. L. Yen, M. R. Bogwitz *et al.*, "A single P450 allele associated with insecticide resistance in *Drosophila*," *Science*, vol. 297, no. 5590, pp. 2253–2256, 2002. DOI 10.1126/science.1074170. PMID 12351787. +[51] P. J. Daborn, J. L. Yen, M. R. Bogwitz *et al.*, "A single P450 allele associated with insecticide resistance in *Drosophila*," *Science*, vol. 297, no. 5590, pp. 2253–2256, 2002. DOI 10.1126/science.1074170. PMID 12351787. -[53] D. G. Folk, C. Han and T. J. Bradley, "Water acquisition and partitioning in *Drosophila melanogaster*: effects of selection for desiccation-resistance," *J. Exp. Biol.*, vol. 204, no. 19, pp. 3323–3331, 2001. DOI 10.1242/jeb.204.19.3323. PMID 11606606. +[52] D. G. Folk, C. Han and T. J. Bradley, "Water acquisition and partitioning in *Drosophila melanogaster*: effects of selection for desiccation-resistance," *J. Exp. Biol.*, vol. 204, no. 19, pp. 3323–3331, 2001. DOI 10.1242/jeb.204.19.3323. PMID 11606606. -[54] R. Fusetto, S. Denecke, T. Perry, R. A. J. O'Hair and P. Batterham, "Partitioning the roles of CYP6G1 and gut microbes in the metabolism of the insecticide imidacloprid in *Drosophila melanogaster*," *Sci. Rep.*, vol. 7, art. 11339, 2017. DOI 10.1038/s41598-017-09800-2. PMID 28900131. +[53] R. Fusetto, S. Denecke, T. Perry, R. A. J. O'Hair and P. Batterham, "Partitioning the roles of CYP6G1 and gut microbes in the metabolism of the insecticide imidacloprid in *Drosophila melanogaster*," *Sci. Rep.*, vol. 7, art. 11339, 2017. DOI 10.1038/s41598-017-09800-2. PMID 28900131. -[55] N. Joußen, D. G. Heckel, M. Haas, I. Schuphan and B. Schmidt, "Metabolism of imidacloprid and DDT by P450 CYP6G1 expressed in cell cultures of *Nicotiana tabacum*," *Pest Manag. Sci.*, vol. 64, no. 1, pp. 65–73, 2008. DOI 10.1002/ps.1472. PMID 17912692. +[54] N. Joußen, D. G. Heckel, M. Haas, I. Schuphan and B. Schmidt, "Metabolism of imidacloprid and DDT by P450 CYP6G1 expressed in cell cultures of *Nicotiana tabacum*," *Pest Manag. Sci.*, vol. 64, no. 1, pp. 65–73, 2008. DOI 10.1002/ps.1472. PMID 17912692. -[56] S. C. Piyankarage, H. Augustin, Y. Grosjean, D. E. Featherstone and S. A. Shippy, "Hemolymph amino acid analysis of individual *Drosophila* larvae," *Anal. Chem.*, vol. 80, no. 4, pp. 1201–1207, 2008. DOI 10.1021/ac701785z. +[55] S. C. Piyankarage, H. Augustin, Y. Grosjean, D. E. Featherstone and S. A. Shippy, "Hemolymph amino acid analysis of individual *Drosophila* larvae," *Anal. Chem.*, vol. 80, no. 4, pp. 1201–1207, 2008. DOI 10.1021/ac701785z. -[57] S. C. Piyankarage, D. E. Featherstone and S. A. Shippy, "Nanoliter hemolymph sampling and analysis of individual adult *Drosophila melanogaster*," *Anal. Chem.*, vol. 84, no. 10, pp. 4460–4466, 2012. DOI 10.1021/ac3002319. +[56] S. C. Piyankarage, D. E. Featherstone and S. A. Shippy, "Nanoliter hemolymph sampling and analysis of individual adult *Drosophila melanogaster*," *Anal. Chem.*, vol. 84, no. 10, pp. 4460–4466, 2012. DOI 10.1021/ac3002319. -[58] S. Suchail, L. Debrauwer and L. P. Belzunces, "Metabolism of imidacloprid in *Apis mellifera*," *Pest Manag. Sci.*, vol. 60, no. 3, pp. 291–296, 2004. DOI 10.1002/ps.772. PMID 15025241. +[57] S. Suchail, L. Debrauwer and L. P. Belzunces, "Metabolism of imidacloprid in *Apis mellifera*," *Pest Manag. Sci.*, vol. 60, no. 3, pp. 291–296, 2004. DOI 10.1002/ps.772. PMID 15025241. -[59] S. Suchail, G. De Sousa, R. Rahmani and L. P. Belzunces, "In vivo distribution and metabolisation of ¹⁴C-imidacloprid in different compartments of *Apis mellifera* L.," *Pest Manag. Sci.*, vol. 60, no. 11, pp. 1056–1062, 2004. DOI 10.1002/ps.895. +[58] S. Suchail, G. De Sousa, R. Rahmani and L. P. Belzunces, "In vivo distribution and metabolisation of ¹⁴C-imidacloprid in different compartments of *Apis mellifera* L.," *Pest Manag. Sci.*, vol. 60, no. 11, pp. 1056–1062, 2004. DOI 10.1002/ps.895. -[60] W. Cao, L. Song, J. Cheng *et al.*, "An automated rapid iterative negative geotaxis assay for analyzing adult climbing behavior in a *Drosophila* model of neurodegeneration," *J. Vis. Exp.*, no. 127, art. 56507, 2017. DOI 10.3791/56507. PMID 28931001. +[59] W. Cao, L. Song, J. Cheng *et al.*, "An automated rapid iterative negative geotaxis assay for analyzing adult climbing behavior in a *Drosophila* model of neurodegeneration," *J. Vis. Exp.*, no. 127, art. 56507, 2017. DOI 10.3791/56507. PMID 28931001. -[61] M. Ö. Çevik and A. Erden, "The course of habituation of the proboscis extension reflex can be predicted by sucrose responsiveness in *Drosophila*," *PLoS ONE*, vol. 7, no. 6, art. e39863, 2012. DOI 10.1371/journal.pone.0039863. PMID 22761915. +[60] M. Ö. Çevik and A. Erden, "The course of habituation of the proboscis extension reflex can be predicted by sucrose responsiveness in *Drosophila*," *PLoS ONE*, vol. 7, no. 6, art. e39863, 2012. DOI 10.1371/journal.pone.0039863. PMID 22761915. -[62] W. Chen, X. Gu, Y. T. Yang, P. Batterham and T. Perry, "Dual nicotinic acetylcholine receptor subunit gene knockouts reveal limits to functional redundancy," *Pestic. Biochem. Physiol.*, vol. 184, art. 105118, 2022. DOI 10.1016/j.pestbp.2022.105118. PMID 35715057. +[61] W. Chen, X. Gu, Y. T. Yang, P. Batterham and T. Perry, "Dual nicotinic acetylcholine receptor subunit gene knockouts reveal limits to functional redundancy," *Pestic. Biochem. Physiol.*, vol. 184, art. 105118, 2022. DOI 10.1016/j.pestbp.2022.105118. PMID 35715057. -[63] A. S. French, M. J. Sellier, M. Ali Agha *et al.*, "Dual mechanism for bitter avoidance in *Drosophila*," *J. Neurosci.*, vol. 35, no. 9, pp. 3990–4004, 2015. DOI 10.1523/JNEUROSCI.1312-14.2015. PMID 25740527. +[62] A. S. French, M. J. Sellier, M. Ali Agha *et al.*, "Dual mechanism for bitter avoidance in *Drosophila*," *J. Neurosci.*, vol. 35, no. 9, pp. 3990–4004, 2015. DOI 10.1523/JNEUROSCI.1312-14.2015. PMID 25740527. -[64] J. W. Gargano, I. Martin, P. Bhandari and M. S. Grotewiel, "Rapid iterative negative geotaxis (RING): a new method for assessing age-related locomotor decline in *Drosophila*," *Exp. Gerontol.*, vol. 40, no. 5, pp. 386–395, 2005. DOI 10.1016/j.exger.2005.02.005. PMID 15919590. +[63] J. W. Gargano, I. Martin, P. Bhandari and M. S. Grotewiel, "Rapid iterative negative geotaxis (RING): a new method for assessing age-related locomotor decline in *Drosophila*," *Exp. Gerontol.*, vol. 40, no. 5, pp. 386–395, 2005. DOI 10.1016/j.exger.2005.02.005. PMID 15919590. -[65] H. Y. Lee, L. Zhou, S. Ghanta *et al.*, "Mechanisms of naturally evolved ethanol resistance in *Drosophila melanogaster*," *J. Exp. Biol.*, 2014. DOI 10.1242/jeb.110510. +[64] H. Y. Lee, L. Zhou, S. Ghanta *et al.*, "Mechanisms of naturally evolved ethanol resistance in *Drosophila melanogaster*," *J. Exp. Biol.*, 2014. DOI 10.1242/jeb.110510. -[66] F. Martelli, Z. Zhongyuan, J. Wang *et al.*, "Low doses of the neonicotinoid insecticide imidacloprid induce ROS triggering neurological and metabolic impairments in *Drosophila*," *Proc. Natl. Acad. Sci. USA*, vol. 117, no. 41, pp. 25840–25850, 2020. DOI 10.1073/pnas.2011828117. PMID 32989137. +[65] F. Martelli, Z. Zhongyuan, J. Wang *et al.*, "Low doses of the neonicotinoid insecticide imidacloprid induce ROS triggering neurological and metabolic impairments in *Drosophila*," *Proc. Natl. Acad. Sci. USA*, vol. 117, no. 41, pp. 25840–25850, 2020. DOI 10.1073/pnas.2011828117. PMID 32989137. -[67] M. J. Burton, I. R. Mellor, I. R. Duce *et al.*, "Differential resistance of insect sodium channels with kdr mutations to deltamethrin, permethrin and DDT," *Insect Biochem. Mol. Biol.*, vol. 41, no. 9, pp. 723–732, 2011. DOI 10.1016/j.ibmb.2011.05.004. +[66] M. J. Burton, I. R. Mellor, I. R. Duce *et al.*, "Differential resistance of insect sodium channels with kdr mutations to deltamethrin, permethrin and DDT," *Insect Biochem. Mol. Biol.*, vol. 41, no. 9, pp. 723–732, 2011. DOI 10.1016/j.ibmb.2011.05.004. -[68] A. A. Lazar, T. Liu, M. K. Turkcan and Y. Zhou, "Accelerating with FlyBrainLab the discovery of the functional logic of the *Drosophila* brain in the connectomic and synaptomic era," *eLife*, vol. 10, art. e62362, 2021. DOI 10.7554/eLife.62362. PMID 33616035. +[67] A. A. Lazar, T. Liu, M. K. Turkcan and Y. Zhou, "Accelerating with FlyBrainLab the discovery of the functional logic of the *Drosophila* brain in the connectomic and synaptomic era," *eLife*, vol. 10, art. e62362, 2021. DOI 10.7554/eLife.62362. PMID 33616035. -[69] I. Mindlin, R. Herzog, L. Belloli *et al.*, "Whole brain modelling for simulating pharmacological interventions on patients with disorders of consciousness," *Commun. Biol.*, vol. 7, art. 1176, 2024. DOI 10.1038/s42003-024-06852-9. PMID 39300281. +[68] I. Mindlin, R. Herzog, L. Belloli *et al.*, "Whole brain modelling for simulating pharmacological interventions on patients with disorders of consciousness," *Commun. Biol.*, vol. 7, art. 1176, 2024. DOI 10.1038/s42003-024-06852-9. PMID 39300281. -[70] G. Deco, J. Cruzat, J. Cabral *et al.*, "Whole-brain multimodal neuroimaging model using serotonin receptor maps explains non-linear functional effects of LSD," *Curr. Biol.*, vol. 28, no. 19, pp. 3065–3074.e6, 2018. DOI 10.1016/j.cub.2018.07.083. PMID 30270185. +[69] G. Deco, J. Cruzat, J. Cabral *et al.*, "Whole-brain multimodal neuroimaging model using serotonin receptor maps explains non-linear functional effects of LSD," *Curr. Biol.*, vol. 28, no. 19, pp. 3065–3074.e6, 2018. DOI 10.1016/j.cub.2018.07.083. PMID 30270185. -[71] A. Saltelli, P. Annoni, I. Azzini, F. Campolongo, M. Ratto and S. Tarantola, "Variance based sensitivity analysis of model output: design and estimator for the total sensitivity index," *Comput. Phys. Commun.*, vol. 181, no. 2, pp. 259–270, 2010. DOI 10.1016/j.cpc.2009.09.018. +[70] A. Saltelli, P. Annoni, I. Azzini, F. Campolongo, M. Ratto and S. Tarantola, "Variance based sensitivity analysis of model output: design and estimator for the total sensitivity index," *Comput. Phys. Commun.*, vol. 181, no. 2, pp. 259–270, 2010. DOI 10.1016/j.cpc.2009.09.018. -[72] M. J. W. Jansen, "Analysis of variance designs for model output," *Comput. Phys. Commun.*, vol. 117, no. 1–2, pp. 35–43, 1999. DOI 10.1016/S0010-4655(98)00154-4. +[71] M. J. W. Jansen, "Analysis of variance designs for model output," *Comput. Phys. Commun.*, vol. 117, no. 1–2, pp. 35–43, 1999. DOI 10.1016/S0010-4655(98)00154-4. -[73] T. Ishigami and T. Homma, "An importance quantification technique in uncertainty analysis for computer models," in *Proc. 1st Int. Symp. Uncertainty Modeling and Analysis (ISUMA)*, 1990, pp. 398–403. DOI 10.1109/ISUMA.1990.151285. +[72] T. Ishigami and T. Homma, "An importance quantification technique in uncertainty analysis for computer models," in *Proc. 1st Int. Symp. Uncertainty Modeling and Analysis (ISUMA)*, 1990, pp. 398–403. DOI 10.1109/ISUMA.1990.151285. -[74] S. Steegen, F. Tuerlinckx, A. Gelman and W. Vanpaemel, "Increasing transparency through a multiverse analysis," *Perspect. Psychol. Sci.*, vol. 11, no. 5, pp. 702–712, 2016. DOI 10.1177/1745691616658637. PMID 27694465. +[73] S. Steegen, F. Tuerlinckx, A. Gelman and W. Vanpaemel, "Increasing transparency through a multiverse analysis," *Perspect. Psychol. Sci.*, vol. 11, no. 5, pp. 702–712, 2016. DOI 10.1177/1745691616658637. PMID 27694465. -[75] U. Simonsohn, J. P. Simmons and L. D. Nelson, "Specification curve analysis," *Nat. Hum. Behav.*, vol. 4, pp. 1208–1214, 2020. DOI 10.1038/s41562-020-0912-z. PMID 32719546. +[74] U. Simonsohn, J. P. Simmons and L. D. Nelson, "Specification curve analysis," *Nat. Hum. Behav.*, vol. 4, pp. 1208–1214, 2020. DOI 10.1038/s41562-020-0912-z. PMID 32719546. -[76] R. A. Howard, "Information value theory," *IEEE Trans. Syst. Sci. Cybern.*, vol. 2, no. 1, pp. 22–26, 1966. DOI 10.1109/TSSC.1966.300074. +[75] R. A. Howard, "Information value theory," *IEEE Trans. Syst. Sci. Cybern.*, vol. 2, no. 1, pp. 22–26, 1966. DOI 10.1109/TSSC.1966.300074. -[77] M. Tomizawa, B. Latli and J. E. Casida, "Novel neonicotinoid-agarose affinity column for *Drosophila* and *Musca* nicotinic acetylcholine receptors," *J. Neurochem.*, vol. 67, no. 4, pp. 1669–1676, 1996. DOI 10.1046/j.1471-4159.1996.67041669.x. PMID 8858952. +[76] M. Tomizawa, B. Latli and J. E. Casida, "Novel neonicotinoid-agarose affinity column for *Drosophila* and *Musca* nicotinic acetylcholine receptors," *J. Neurochem.*, vol. 67, no. 4, pp. 1669–1676, 1996. DOI 10.1046/j.1471-4159.1996.67041669.x. PMID 8858952. --- diff --git a/papers/IJRC_FlyLab_draft.md.in b/papers/IJRC_FlyLab_draft.md.in index bf8950d..a1bbac5 100644 --- a/papers/IJRC_FlyLab_draft.md.in +++ b/papers/IJRC_FlyLab_draft.md.in @@ -7,58 +7,65 @@ TEMPLATE. Do not quote numbers from this file. never hand-edit a number into the rendered draft. The supplement is generated the same way from `papers/SUPPLEMENT.md.in`. --> -# Which Inputs Does a Network Simulation Depend On? Typed Evidence Propagation and Input-Dependence Testing, Evaluated on Cuts of a Synapse-Resolution *Drosophila* Connectome +# Auditing Input Dependence in Literature Parameterised Network Simulations -**Research article, submitted draft for the International Journal of Research in Computing (IJRC)** +**Computational-methods Research Article. Working draft for the International Journal of Research in Computing (IJRC); not submitted.** + +**Michél Nguyen** + +University of the People + +ORCID: 0000-0001-6834-4422 + +Corresponding author: michel_ng@icloud.com Software: https://github.com/pinkysworld/FlyLab -Version {{flylab_version}}, commit `{{git_sha}}`, compound library SHA-256 `{{library_sha256}}`, map `{{map_id}}`. +Software version {{flylab_version}}, compound library SHA-256 `{{library_sha256}}`, map `{{map_id}}`. Numerical source provenance is recorded in `papers/results.json`; a tagged archival release remains pending. -> Every numeric span in this document is generated by `scripts/reproduce_paper.py` from `papers/results.json`; the prose is written around the numbers, not the other way round. No number in this paper was measured in a living animal. +> Simulation-result numeric spans in this document are generated by `scripts/reproduce_paper.py` from `papers/results.json`; the prose is written around the numbers, not the other way round. No number in this paper was measured in a living animal. ## Abstract -A simulation built on a measured network and parameterised from published literature cannot, on its own, say which of its inputs its predictions depend on. We present a method that makes that question routine and testable, and an open implementation, FlyLab, that answers it per prediction. The method has four parts: **typed evidence propagation**, in which the parameter a source measured and its graded distance from the modelled target jointly decide which transformation is admissible, enforced by a soundness invariant checked exhaustively over the parameter library ({{ev_rows_not_modelled}} of {{ev_rows}} rows refused rather than approximated); **input-dependence testing** against an ordered ladder of degraded copies of the network, reporting — with a prespecified equivalence margin and multiplicity control — the weakest input the test cannot distinguish from the real one; an **ablation ladder** over pipeline layers scored against a matched reference distribution; and **specification-family robustness** with variance attribution that ranks what to measure next. We evaluate the method rather than the domain: the dependence instrument is validated against planted ground truth with a power surface, its permutation budget is characterised by a convergence sweep, and the analysis is deterministic across two backends. **The principal result is methodological: a dependence verdict obtained on one extract of a network can invert on another extract of the same network, and the extract's recurrence, not its size, predicts which way.** Demonstrated on cuts of the public MaleCNS v1.0 *Drosophila* connectome, a prediction is composition-dominated on a sparse 1-hop extract (mean degree {{named_mean_degree}}, {{named_share_onto_seeds}} of edges onto four seed cells) and topology-dependent on every extract with real recurrence, including one with *fewer* nodes; across nested extracts of {{scale_ladder_sizes}} cells the verdict settles as topology-dependent. Such a verdict is therefore a joint property of the prediction and the extract, and must be reported with the extract's structure. The central limitation is unchanged: the pipeline evaluates a model's behaviour, and its parameter-to-gain rules are asserted rather than measured. +Applied to cuts of the MaleCNS v1.0 *Drosophila* connectome, the same dependence analysis gives opposite answers: an imidacloprid prediction is composition-dominated on the sparse `named` cut but topology-dependent on denser extracts, including a smaller one. Node count therefore does not order the observed labels, and the study does not isolate recurrence as their cause. The result is also runtime-conditional: composition-only and full-model rankings agree under row normalisation but not under a degree-corrected denominator. Two instruments had previously flattered the conclusions. A transmitter null changed the weighted excitation/inhibition balance it was meant to probe, while alternative gain rules did not reach the engine used by their topology tests. FlyLab addresses these failures with provenance-gated evidence transformations, graph degradations, matched ablation references, specification-family tests and variance attribution. A finite-library audit refuses inadmissible evidence ({{ev_rows_not_modelled}} of {{ev_rows}} rows), and synthetic cases provide limited positive-control evidence for the dependence ladder. The corrected audit finds that nicotinic suppression reverses under monotone gain rules. These are results about model behaviour, not animal responses. Gain rules remain asserted, the large-scale table is imported from a saved study, and no independent circuit-level validation exists. -**Index terms:** computational method, scientific software, provenance, evidence typing, network simulation, null models, equivalence testing, multiple comparisons, specification robustness, sensitivity analysis, reproducible research, connectomics. +**Index terms:** computational method, scientific software, provenance, evidence typing, network simulation, null models, tolerance diagnostics, multiple comparisons, specification robustness, sensitivity analysis, reproducible research, connectomics. ## 1. Introduction -A large class of scientific software has the same shape: a measured network, a table of parameters compiled from published literature, a rule that turns those parameters into perturbations of the network, and a numerical runtime that produces a prediction. Power grids, metabolic models, epidemic contact networks and connectome simulations are all built this way. Such a pipeline readily produces a number; what it does not produce is an answer to the question a reviewer asks first — *which of its inputs does this number actually depend on?* Without that answer the prediction cannot be audited, because a result carried entirely by an asserted coefficient is indistinguishable, at the output, from one carried by the measured structure. +A large class of scientific software has the same shape: a measured network, a table of parameters compiled from published literature, a rule that turns those parameters into perturbations of the network, and a numerical runtime that produces a prediction. Power grids, metabolic models, epidemic contact networks and connectome simulations are all built this way. Such a pipeline readily produces a number; what it does not produce is an answer to the question a reviewer asks first: *which of its inputs does this number actually depend on?* Without that answer the prediction cannot be audited, because a result carried entirely by an asserted coefficient is indistinguishable, at the output, from one carried by the measured structure. -Three properties make the question hard to answer in practice. The parameters are heterogeneous: sources measure different quantities, in different preparations, and are often simply absent, so a pipeline that coerces them into one numeric field silently manufactures evidence. The network enters through a runtime whose own conventions — normalisation, sign, iteration — are modelling choices that can carry part of the answer. And the usual test of whether structure matters is a permutation against a shuffled network, whose *non-rejection* is routinely written up as though the shuffle had reproduced the effect. +Three properties make the question hard to answer in practice. The parameters are heterogeneous: sources measure different quantities, in different preparations, and are often simply absent, so a pipeline that coerces them into one numeric field silently manufactures evidence. The network enters through a runtime whose own conventions (normalisation, sign and iteration) are modelling choices that can carry part of the answer. And the usual test of whether structure matters is a permutation against a shuffled network, whose *non-rejection* is routinely written up as though the shuffle had reproduced the effect. -This paper contributes a method addressing all three (§2), an implementation that runs it per prediction rather than once per paper, and — the part we think is missing from this literature — an evaluation of the instruments themselves (§3). We demonstrate the method on cuts of a public synapse-resolution *Drosophila* connectome, a domain supplying exactly the hard case: a measured network, a literature parameter table with large gaps, and asserted transformation rules. The headline finding is about the method rather than the domain: **a dependence verdict obtained on one extract of a network can invert on another extract of the same network.** +This paper contributes a method addressing all three (§2), an implementation that runs it per prediction, and an evaluation of the instruments themselves (§3). Cuts of a public synapse-resolution *Drosophila* connectome supply the hard case: a measured network, a literature parameter table with large gaps, and asserted transformation rules. The central result is methodological: **a dependence verdict obtained on one extract of a network can invert on another extract of the same network.** -We organise the work around four research questions. +We organise the work around three research questions. * **RQ1.** Can heterogeneous published parameters be turned into computational perturbations *without* erasing their provenance or their parameter semantics, and can that discipline be stated as a checkable invariant rather than a convention? -* **RQ2.** For a given prediction, which inputs does it depend on — and is that answer a property of the prediction, or of the extract of the network it was measured on? +* **RQ2.** For a given prediction, which inputs does it depend on, and is that answer a property of the prediction or of the extract of the network on which it was measured? * **RQ3.** Which conclusions survive alternative admissible specifications, and which assumptions dominate the output's variance? -* **RQ4.** Can the entire analysis be reproduced identically, across backends, from auditable artifacts? -The answers are: yes, by grading the evidence, refusing invalid transformations and checking the refusal exhaustively; **it is a property of the extract**, which is the paper's main result; not all of them, and an undocumented normalisation inside the runtime carries part of RQ2's answer; and yes. An earlier version of this work answered RQ2 with "most predictions do not need the network", measured on a single extract — a conclusion this revision withdraws and explains. +The results support evidence gating and finite regression checks for RQ1. For RQ2, labels vary with the extract, prediction and runtime normalisation. For RQ3, some conclusions reverse under admissible specifications. An earlier version answered RQ2 with "most predictions do not need the network", measured on a single extract; this revision withdraws that conclusion. Recomputability and backend parity are reported separately as engineering support, not as a fourth research question. ### 1.1 Related work and what is new here -**Network runtimes without parameterised perturbation.** Teaching simulators [1]–[4] model dose–response with no network at all. At the other extreme, FlyWire [5] and MaleCNS v1.0 [6] publish netlists with per-node annotations [10], [12]; Shiu *et al.* [7] show a literature-parameterised model over such a netlist reproduces a sensorimotor control, and downstream ports repeat the exercise with varying rigour [11]. FlyBrainLab [68] builds executable circuits directly from connectome data, but a perturbation, where present, enters as a hand-set multiplier with nothing behind it — exactly the case typed propagation is meant to replace. +**Network runtimes without parameterised perturbation.** The implemented teaching simulators RatCVS and Virtual Rat Web [1], [2] model dose–response with no network; PharmVR [3] is a non-peer-reviewed concept paper and is not evidence of an implemented capability. At the other extreme, FlyWire [4] and MaleCNS v1.0 [5] publish netlists with per-node annotations [9], [11]; Shiu *et al.* [6] show a literature-parameterised model over such a netlist reproduces a sensorimotor control, while an informal newsletter essay catalogs downstream ports with varying rigour [10]. FlyBrainLab [67] builds executable circuits directly from connectome data. The cited implementation does not describe the per-row pharmacological evidence gating evaluated here. -**Parameter-map-informed network models.** On human structural connectomes, intervention is simulated by scaling regional responses with PET density maps [69], [70] — the closest published relative, whose most relevant result is also a negative one: in Mindlin *et al.* [69] the simulated improvement tracked the *mean* density of the activated receptors rather than where they sit. Methodologically, §2.4 is a multiverse analysis [74], [75] applied to a simulation, and §2.6 uses Saltelli cross-sampling [71] with Jansen estimators [72] validated on Ishigami [73] and ranked by value of information [76]. +**Parameter-map-informed network models.** The closest published relatives simulate intervention on human structural connectomes by scaling regional responses with PET density maps [68], [69]. In Mindlin *et al.* [68], the simulated improvement tracked the *mean* density of the activated receptors rather than where they sit. Methodologically, §2.4 is a multiverse analysis [73], [74] applied to a simulation, and §2.6 uses Saltelli cross-sampling [70] with Jansen estimators [71] validated on Ishigami [72] and ranked by value of information [75]. -**The domain evidence base.** The parameter library, and the supporting datasets for resistance alleles, expression, mixtures, exposure and behavioural endpoints, are compiled from the insect-versus-vertebrate pharmacology literature [8], [9], [13]–[67] and documented in §S4 and §S6; what matters methodologically is that it is typical of such tables — incomplete, heterogeneous in what each source measured, and internally inconsistent with its own citations in ten places that `data/literature/README.md` records without editing the library to match. Those gaps are the reason the type system changes results rather than documentation. +**The domain evidence base.** The parameter library, and the supporting datasets for resistance alleles, expression, mixtures, exposure and behavioural endpoints, are compiled from the insect-versus-vertebrate pharmacology literature [7], [8], [12]–[66] and documented in §S4 and §S6. What matters methodologically is that the library is typical of such tables: incomplete, heterogeneous in what each source measured, and internally inconsistent with its own citations in ten places that `data/literature/README.md` records without editing the library to match. Those gaps make evidence gating affect numerical outputs. -**Position.** Table I states what is new narrowly. The transferable part is not the domain coupling but the instrumentation around it: routine, per-prediction, *validated* testing of which inputs a prediction depends on, with non-rejection separated from equivalence, multiplicity controlled, and the test's own power characterised against planted ground truth. +**Position.** Table I states what is new narrowly. The transferable part is not the domain coupling but the instrumentation around it: routine per-prediction diagnostics combining provenance, graph degradations and specification checks. Synthetic recovery tests are an initial instrument evaluation. A median-distance tolerance is distinct from statistical equivalence, and multiplicity guarantees depend on the assumptions of the adjustment. -**Table I. What each system provides.** +**Table I. Capabilities described in the cited implementations.** A negative entry means not described in that cited application; it is not a claim about every version or extension. -| Capability | Teaching sims [1]–[4] | Connectome LIF runtimes [7], [11] | FlyBrainLab [68] | Receptor-map whole-brain models [69], [70] | FlyLab | +| Capability | Implemented teaching sims [1], [2] | Connectome LIF runtimes [6], [10] | FlyBrainLab [67] | Receptor-map whole-brain models [68], [69] | FlyLab | |---|---|---|---|---|---| | Named cells at synapse resolution | no | yes | yes | no (regions) | yes | | Compound **and free concentration** as the input | yes (organ level) | no | no | partly (receptor-map scaling) | yes | | Typed evidence with per-row provenance and distance | no | no | no | no | yes | -| Two species scored at the same dose | no | no | no | no | yes | -| Dependence of each prediction on its inputs, tested | no | rarely | no | partly [69] | yes, routinely | -| The dependence test itself validated against ground truth | n/a | no | no | no | yes (§3.1, F16) | +| Insect engagement compared with a vertebrate receptor panel | no | no | no | no | yes; no vertebrate circuit | +| Dependence of each prediction on its inputs, tested | no | rarely | no | partly [68] | yes, routinely | +| Synthetic recovery cases evaluated | n/a | not assessed | not assessed | not assessed | yes (§3.1, F16) | | Specification robustness of the stated conclusions | no | no | no | no | yes | | Whole analysis re-runs in a browser | some | no | no | no | endpoint parity verified | @@ -70,19 +77,19 @@ We claim no new reconstruction, no fitted *in vivo* potency, no replacement for The common failure this replaces is a pipeline that applies one transform to whatever number a source reported and gives the result one name. That erases two distinctions at once: a normalised functional response is not a fraction of occupied binding sites, and a row with no sourced number must not yield a small response at all. -Each row instead records the **parameter type** its source measured and an ordered **evidence distance** between that source and the modelled target: {{ev_distance_labels}}. The admissible **transformation** is a total function of that pair (T20): a binding constant yields a physical occupancy only at E0 and a labelled proxy beyond it; a functional potency yields a direct engagement at E0 and a proxy beyond; anything at E4 is `not_modelled` and returns N/A. The result carries the name it earned, so a transferred number cannot be read as a measured one downstream. +Each row records the **parameter type** its source measured and an ordered **evidence distance** from the modelled target: {{ev_distance_labels}}. The admissible transformation is a total function of that pair (T20). Example legal triples are `(Kd, E0, binding_occupancy)`, `(Kd, E1, binding_engagement_proxy)` and `(EC50, E0, functional_engagement)`; transferred functional potencies use `functional_engagement_proxy`, while any E4 row is `not_modelled`. An attempted `(EC50, E0, binding_occupancy)` transformation raises `EvidenceTypeError`. Every numeric branch retains the same Hill algebra, `C^n/(x^n + C^n)`: the exact *Drosophila* Kd row licenses a binding-occupancy label but introduces no new kinetic binding model. -Two properties make this a type discipline rather than a lookup table, and both are what we claim as the computing contribution. The refusal is **enforced**: requesting a value from an inadmissible row raises rather than returning one. And we state and check a **soundness invariant** over the whole pipeline — *no numeric engagement reaching a readout may originate from a row whose (parameter type, distance) pair does not admit one* — exhaustively, over every row and every entry point, with a companion check that dropping the inadmissible rows changes no downstream value, and a mutation check that makes the test fail with the right message when the rule table is altered. +The enforceable invariant is narrow: *no numeric engagement reaching a readout may originate from a row whose (parameter type, distance) pair does not admit one*. A finite regression audit exercises every shipped row through the enumerated comparison, curve, uncertainty, assay and dashboard paths, checks that dropping inadmissible rows changes no gain vector, and verifies deliberate invalid requests. This is coverage of the shipped library and enumerated paths, not proof about an unenumerated future entry point. ### 2.2 The extracts, and what the runtime does to them -Two committed extracts of MaleCNS v1.0 are used, streamed out of the ~1.1 GB public weight matrix on a CI runner and each under about 1 MB of JSON: **`named`**, a 1-hop neighbourhood of two seed cells ({{named_nodes}} nodes, {{named_edges}} edges), and **`taste_motor`**, the same seeds plus sensory types and an induced closure ({{taste_nodes}} nodes, {{taste_edges}} edges). §3.2 adds a ladder of nested extracts of {{scale_ladder_sizes}} cells grown by one fixed rule. +Two committed extracts of MaleCNS v1.0 are used, streamed out of the ~1.1 GB public weight matrix on a CI runner and each under about 1 MB of JSON: **`named`**, a 1-hop neighbourhood of two named cell types (four seed cells) ({{named_nodes}} nodes, {{named_edges}} edges), and **`taste_motor`**, the same seeds plus sensory types and an induced closure ({{taste_nodes}} nodes, {{taste_edges}} edges). §3.2 adds a ladder of nested extracts of {{scale_ladder_sizes}} cells grown by one fixed rule. -**The extracts are structurally very different, and the difference turns out to be load-bearing (T19).** `named` has mean degree {{named_mean_degree}}: {{named_share_onto_seeds}} of its edges terminate on the {{named_n_seed_nodes}} seed cells, only {{named_nodes_with_input}} of {{named_nodes}} nodes receive any input, and only {{named_share_recurrent_edges}} of edges leave a node that itself receives input. It is an in-star, not a circuit. `taste_motor` has mean degree {{taste_motor_mean_degree}}, {{taste_motor_share_onto_seeds}} of edges onto seeds and {{taste_motor_share_recurrent_edges}} of edges leaving a node with input. Composition differs too, and one class is absent from both ({{named_cells_octopamine}} cells), so a perturbation acting only through it cannot move either readout for substrate rather than domain reasons; the full census is in T19 and §S9(v). +**The extracts are structurally very different, and the difference turns out to be load-bearing (T19).** `named` has mean degree {{named_mean_degree}}: {{named_share_onto_seeds}} of its edges terminate on the {{named_n_seed_nodes}} seed cells, only {{named_nodes_with_input}} of {{named_nodes}} nodes receive any input, and only {{named_share_recurrent_edges}} of edges leave a node that itself receives input. It is dominated by incoming edges to seeds, with limited recurrent connectivity. `taste_motor` has mean degree {{taste_motor_mean_degree}}, {{taste_motor_share_onto_seeds}} of edges onto seeds and {{taste_motor_share_recurrent_edges}} of edges leaving a node with input. Composition differs too, and octopamine is absent from both ({{named_cells_octopamine}} cells), so a perturbation acting only through it cannot move either readout for substrate rather than domain reasons; the full census is in T19 and §S9(v). -Two runtimes consume a gain patch: a deterministic clipped leaky rate iteration (~50 ms per evaluation of `named`) and a current-based LIF network [7] whose synaptic scale and background drive are tuned constants, not fits (§S3). +Two runtimes consume a gain patch: a deterministic clipped leaky rate iteration (~50 ms per evaluation of `named`) and a current-based LIF network [6] whose synaptic scale and background drive are tuned constants, not fits (§S3). -**The rate runtime row-normalises its weight matrix, and previous versions of this paper did not say so.** Each row of the signed matrix is divided by `max(Σ|W|, 1)` before any gain is applied, and gains then act as a presynaptic column scale, so every node's recurrent input is a *composition-weighted average of its presynaptic gains*. That builds part of the answer to RQ2 into the readout itself — an example of the general hazard that a runtime convention can carry a result — so it is documented here, in T2 and §S3, and swept in §3.3 against `none` and a degree-corrected denominator. The sign table the runtime iterates is published as T21: {{sign_asserted}} of seven magnitudes ({{sign_asserted_transmitters}}) are asserted, varied by no member of the specification family and not bounded by the label-permuting null. +**The rate runtime row-normalises its weight matrix, and previous versions did not report this.** Each row of the signed matrix is divided by `max(Σ|W|, 1)` before gains act as presynaptic column scales. Recurrent input is therefore a composition-weighted average of presynaptic gains, making the convention part of the answer to RQ2. Section 3.3 sweeps it against `none` and a degree-corrected denominator. T21 publishes the sign table: {{sign_asserted}} of seven magnitudes ({{sign_asserted_transmitters}}) are asserted, absent from the specification family and not bounded by the label-permuting null. ### 2.3 Input-dependence testing (RQ2) @@ -96,93 +103,95 @@ The instrument we consider most transferable asks, for one `compound × concentr | `sign_permute_weight_matched` | + the real weights on the real edges, and each transmitter's share of total outgoing weight | | the MaleCNS cut | everything | -The rank-3 rung changed in this revision. Because a gain patch acts *through* transmitter identity, a null that moves the weighted balance moves the drug's target set and the sign matrix at once, so the plain label permutation is relabelled a **joint target-set-and-sign null** and replaced by a constrained shuffle holding every tracked transmitter's share of outgoing weight within {{bal_matched_max_deviation}} of the real graph's. Permuting the node transmitter-label array preserves the transmitter histogram *by count* but not *by weight*: out-strength is heterogeneous, so on the `named` cut the cholinergic share of total outgoing synaptic weight is {{bal_real_ach_share}} on the real graph against {{bal_plain_ach_mean}} ± {{bal_plain_ach_sd}} across {{bal_n}} permutations, putting the real graph at the {{bal_plain_ach_percentile}}th percentile of its own null (T23). +The rank-3 rung uses a constrained shuffle holding every tracked transmitter's outgoing-weight share within {{bal_matched_max_deviation}} of the real graph. A plain label permutation is instead a **joint target-set-and-sign null**: it preserves label counts but moves weighted balance because out-strength is heterogeneous. On `named`, the cholinergic outgoing-weight share is {{bal_real_ach_share}} on the real graph versus {{bal_plain_ach_mean}} ± {{bal_plain_ach_sd}} across {{bal_n}} permutations, placing the real graph at the {{bal_plain_ach_percentile}}th percentile of that null (T23). -**Non-rejection is not equivalence, and the software no longer lets the prose say it is.** Each mode receives one of three verdicts against a margin prespecified as {{dep_named_imidacloprid_delta_frac}} of the vehicle readout ({{dep_named_imidacloprid_delta}} Hz on `named`): `distinguishable` when the empirical two-sided permutation probability `(k+1)/(n+1)` is at or below α; `equivalent_within_tolerance` when the test does not reject *and* the gap from the null median is below the margin; `indeterminate` otherwise. From the verdicts follow a **class** and a **necessary information level** — the weakest graph model the test could not distinguish from the real cut, never a claim that it reproduces the effect. Where a weaker model survives while a richer one is rejected the ladder is non-monotone, which the software reports and §3.2 counts. +Each mode is first assessed with the empirical two-sided permutation probability `(k+1)/(n+1)`. After non-rejection, the implementation compares the real effect with the null median using a tolerance of 5% of the vehicle readout ({{dep_named_imidacloprid_delta}} rate-model units on `named`). The legacy label `equivalent_within_tolerance` records only that point-gap check, not statistical equivalence; other non-rejections are `indeterminate`. Classes and the legacy **necessary information level** are descriptive summaries, not proofs of necessary or sufficient information. The nulls are not nested. -Three further conventions are prespecified and §S2 states them in full: a *relative* effect floor, a separation of **confirmatory** from **exploratory** tests with Benjamini-Hochberg applied across the latter, and a check of how many tests sit at the resolution floor, so a corrected count of zero is never mistaken for a finding. +The implementation also fixes a relative effect floor, separates detailed profiles from library-wide screens, and applies Benjamini-Hochberg across the screens' component structural tests (§S2). A cell is topology-dependent if either adjusted structural test rejects. This is not a demonstrated cell-level FDR guarantee, and the conventions were refined during analysis rather than externally registered. ### 2.4 The ablation ladder and the specification family (RQ3) -*Which layer carries the information?* Four models predict the same cell with one layer denied: **A** receptor engagement alone; **B** mechanism gains on the graph's transmitter composition only, no edges; **C** the real cut and runtime with one generic multiplier; **D** the full model. They carry different units, so they are compared by how well each reproduces the full model's *ordering* of the library, scored on the **signed** Spearman ρ because a perfect inversion reproduces nothing. +Four models predict the same cell with one layer denied: **A** receptor engagement alone; **B** mechanism gains on transmitter composition only; **C** the real cut and runtime with one generic multiplier; **D** the full model. Because their units differ, signed Spearman ρ compares each library ordering with D. -Level C's multiplier changed in this revision: it could previously only *depress*, so no entry could come out positive and the level could not order a library half of whose entries raise the readout. Its magnitude is unchanged and its **sign** now comes from the mechanism table — one bit per compound — with the old rule retained as a floor (§S1). Levels B and D are not independent models, both being functions of the same gain vector, so §3.3 reports a **matched reference distribution** saying what agreement would have been for structural reasons alone. +Level C previously could only depress activity and could not order compounds that raise the readout. Its magnitude is unchanged, but its sign now comes from one mechanism-table bit per compound; the old rule remains as a floor (§S1). Because B and D share the same gain vector, §3.3 reports a matched reference distribution. -*Which conclusions are properties of the data rather than of a modelling choice?* The parameter-to-gain mapping is the least defensible part of the pipeline — a textbook shape, but an asserted functional form. We prespecify {{stab_family_size}} admissible specifications (five shapes at five coefficient scales, §S1.1) and re-derive every principal conclusion under every member; undecidable cells count against the conclusion and are reported separately, and the fraction is a fraction of *that family*, not a posterior probability. The topology conclusions are decided at {{stab_n_shuffles}} permutations per specification (resolution {{stab_shuffle_resolution}}), corrected across the run's {{stab_n_structural_tests}} structural tests. The two thresholds are likewise conventions, so the verdict is recomputed over the full {{thr_circuit_fracs}} × {{thr_vert_limits}} grid on {{thr_n_compounds}} compounds. +The parameter-to-gain mapping is an asserted functional form. We therefore re-derive each principal conclusion under {{stab_family_size}} specifications (five shapes at five coefficient scales, §S1.1); undecidable cells count against a conclusion, and fractions describe this family rather than posterior probabilities. Topology predicates use {{stab_n_shuffles}} permutations per specification and adjusted component probabilities across {{stab_n_structural_tests}} tests, without a conclusion-level FDR guarantee. Threshold verdicts are recomputed over the {{thr_circuit_fracs}} × {{thr_vert_limits}} grid on {{thr_n_compounds}} compounds. ### 2.5 Evaluation taxonomy -Agreement with published work is worth little unless its kind is named, so we separate three. **Implementation consistency** — the same inputs give the same outputs across transports, runtimes and runs (§3.7). **Literature concordance** — the model's *ordering* agrees with a published ordering, scored per entry and never as a bare mean, since an entry ordering two items can only return ±1; where the published ordering shares a source with the library, agreement is internal consistency, and the pipeline computes that overlap from DOIs and PMIDs rather than asserting it. **Independent out-of-sample validation** — against data entering neither the library nor the model, of which there is **none** at the circuit level. The ground-truth experiment of §3.1 validates the *instrument*, not the domain model, and we are careful not to let the first stand in for the second. +We distinguish **implementation consistency**, **literature concordance** and **independent out-of-sample validation**. Concordance is scored per published ordering, with shared library sources detected from DOIs and PMIDs; a two-item ordering is never allowed to dominate a bare mean. There is no independent circuit-level validation. The synthetic experiment assesses selected instrument behaviours, not the domain model or general calibration. ### 2.6 Global uncertainty and the value of information -A one-at-a-time tornado cannot separate a factor's own effect from its interactions, so uncertainty is attributed with a variance-based (Sobol') analysis over every assumption the implementation exposes — potency, Hill coefficient, the engagement-to-gain transformation and its coefficient, the external drive, the cut's synapse-weight threshold, the transmitter-sign convention, the expression-weighting layer — plus a deliberately null factor, the spiking RNG seed, which the deterministic rate engine ignores. The design is Saltelli cross-sampling with Jansen estimators [71], [72], validated against the analytic Ishigami benchmark [73] in the same run. +A one-at-a-time tornado cannot separate a factor's own effect from its interactions, so uncertainty is attributed with a variance-based (Sobol') analysis over selected assumptions the implementation exposes; potency, Hill coefficient, the engagement-to-gain transformation and its coefficient, the external drive, the cut's synapse-weight threshold, the transmitter-sign convention, the expression-weighting layer; plus a deliberately null factor, the spiking RNG seed, which the deterministic rate engine ignores. The design is Saltelli cross-sampling with Jansen estimators [70], [71], validated against the analytic Ishigami benchmark [72] in the same run. -Estimates are left unclipped, so a factor whose true index is zero can be estimated below zero. A first-order index cannot be negative, so the **largest magnitude among the negative estimates is a measured lower bound on the estimator's error**, and that — not the null factor's own draw — is the noise floor we report. The budget becomes a ranking by `VOI_j = max(0, S_j) · Var(Y)` [76] — variance of a model output under assumed input ranges, not an expected gain in accuracy about a living fly. +Estimates are left unclipped, so a factor whose true index is zero can be estimated below zero. A first-order index cannot be negative, so the **largest magnitude among the negative estimates is a measured lower bound on the estimator's error**, and we report it as an error diagnostic, not a calibrated noise floor or universal bound. The budget becomes a ranking by `VOI_j = max(0, S_j) · Var(Y)` [75]; variance of a model output under assumed input ranges, not an expected gain in accuracy about a living fly. ## 3. Results -This section evaluates the instruments, not the domain. All values are regenerated by `scripts/reproduce_paper.py` and stored keyed in `papers/results.json`; the parameter library and the domain readouts they are demonstrated on are described in §3.6 and §S6. +This section evaluates the instruments, not the domain. Core committed-cut values are generated by `scripts/reproduce_paper.py` and stored keyed in `papers/results.json`; T27 is imported from the cached scale study described in §3.7. The parameter library and demonstration readouts are described in §3.6 and §S6. -### 3.1 Is the dependence instrument sound? Ground truth, power and convergence (Fig. F16, Tables T11, T22) +### 3.1 Synthetic recovery, detection frequency and convergence (Fig. F16, Tables T11, T22) -A negative from a dependence test is uninterpretable unless the test can be shown to return a positive when one exists, so the instrument is characterised against **planted ground truth** before it is used on anything. A recurrent cycle of known strength is planted in a synthetic network of the same size, density and composition as the real extract, so that the perturbation removes an amplification existing only while the cycle is intact: the *contrast*, not merely the level, depends on the wiring by construction. Strength 0 plants nothing and is the negative control. +The ladder recovered {{val_recovery_detected}} of {{val_recovery_positive}} planted effects at *n* = {{val_recovery_n}}, with no detection in its unplanted control. Detection rates by planted strength are {{val_power_by_strength}}, with all sampled runs detected at {{val_power_full_strengths}} and low detection at {{val_power_weak_strengths}}; by permutation count they are {{val_power_by_n}}. The test can therefore return a positive when wiring is constructed to matter, but weak effects are often missed. -The ladder recovered {{val_recovery_detected}} of {{val_recovery_positive}} planted effects at *n* = {{val_recovery_n}} with no false positive on the control. Over a power grid ({{val_power_replicates}} independently generated networks per cell, permutation counts {{val_power_ns}}) the detection rate by planted strength is {{val_power_by_strength}} — full power at strength {{val_power_full_strengths}}, and at or below one in four at {{val_power_weak_strengths}}, so the instrument is **blind to weak planted structure and must not be read as ruling it out**. The empirical false-positive rate on the unplanted control is {{val_power_control_hits}} of {{val_power_control_runs}}. Averaged over strengths, detection by permutation count is {{val_power_by_n}}: **effect size dominates, and permutation count buys little beyond a few hundred draws.** That is a design rule for this class of test — budget spent on permutations past the point of convergence is wasted, and it is the reason the scaling study of §3.2 can trade permutations for network size at all. +This is limited positive-control evidence. The network has {{val_recovery_nodes}} nodes and {{val_recovery_edges}} background edges and is not matched to either real extract. Strength 0 omits both cycle edges and associated transmitter relabelling. The grid uses only {{val_power_replicates}} networks per cell at counts {{val_power_ns}}, with {{val_power_control_hits}} detections in {{val_power_control_runs}} control runs; networks and shuffle prefixes are reused. These frequencies neither estimate general power or type-I error precisely nor establish power for the taste-motor veto-ratio endpoint. -The convergence sweep (T11) separates the two quantities that permutation budget buys: a *verdict* at α = 0.05 settles from *n* = {{dep_verdict_stable_from_n}}–{{dep_verdict_stable_to_n}}, while a quotable mid-range probability needs {{dep_p_within_0p02_from_n}}–{{dep_p_within_0p02_worst_n}} draws to reach ±0.02 of its final value. The sweep re-reads prefixes of the same draws, so every row is a valid smaller sample and no extra runtime was spent producing it. +The prefix sweep (T11) distinguishes label agreement from probability precision. Observed labels remain unchanged from *n* = {{dep_verdict_stable_from_n}}–{{dep_verdict_stable_to_n}}, while probabilities first remain within ±0.02 of their final estimates after {{dep_p_within_0p02_from_n}}–{{dep_p_within_0p02_worst_n}} draws. These are descriptive comparisons with the final estimate, not confidence bounds on the true permutation probability. Prefixes reuse the same draws and are not independent replications. -### 3.2 The verdict inverts with the extract, and recurrence predicts it (Figs. F6, F12; Tables T6, T12, T12b, T23, T27) +### 3.2 Dependence labels change with the extract (Figs. F6, F12; Tables T6, T12, T12b, T23, T27) -**The confirmatory profiles** (full per-mode table: T6, §S2.1). At {{dep_n}} permutations per mode (resolution {{dep_p_floor}}), imidacloprid's effect on the sparse `named` extract is {{dep_named_imidacloprid_effect}} Hz and the cell is **{{dep_named_imidacloprid_class}}**: against the prespecified margin of {{dep_named_imidacloprid_delta}} Hz it is equivalent within tolerance for {{dep_named_imidacloprid_equivalent_modes}}, merely indeterminate for {{dep_named_imidacloprid_indeterminate_modes}}, and distinguishable only from the Erdős–Rényi control (*p* = {{dep_named_imidacloprid_p_erdos_renyi}} against {{dep_named_imidacloprid_p_rewire_degree_preserving}} for degree-preserving rewiring). Fipronil on the same extract is the contrast — {{dep_named_fipronil_effect}} Hz, **{{dep_named_fipronil_class}}**, *p* = {{dep_named_fipronil_p_weight_permute}} and {{dep_named_fipronil_p_rewire_degree_preserving}} for weight permutation and rewiring, necessary level `{{dep_named_fipronil_level}}` — so the instrument separates two predictions on one extract. Correcting the transmitter null moves no probability but changes what may be claimed: imidacloprid's gap from the null median falls from {{dep_named_imidacloprid_gap_sign_permute}} Hz to {{dep_named_imidacloprid_gap_sign_permute_weight_matched}} Hz, turning an indeterminate result into an equivalence one. +**The landscape inverts between the two extracts.** Over {{dep_land_compounds}} compounds × {{dep_land_concs}} concentrations = {{dep_land_cells}} cells at {{dep_n_landscape}} permutations, with Benjamini-Hochberg across the {{dep_land_n_structural_tests}} component structural tests and the relative effect floor applied, `named` gives **{{dep_land_topology_dependent}} topology-dependent, {{dep_land_composition_dominated}} composition-dominated, {{dep_land_no_effect}} no-effect, {{dep_land_mixed}} mixed** (uncorrected {{dep_land_topology_dependent_raw}} / {{dep_land_composition_dominated_raw}} / {{dep_land_no_effect_raw}}; on the absolute floor alone {{dep_land_topology_dependent_absfloor}} / {{dep_land_composition_dominated_absfloor}} / {{dep_land_no_effect_absfloor}}, the {{dep_land_below_relative_floor}} cells the relative floor removes named in T12). Repeated unchanged on the denser `taste_motor` extract at {{dep_n_landscape_taste}} permutations, the same {{dep_land_taste_cells}} cells give **{{dep_land_taste_topology_dependent}} / {{dep_land_taste_composition_dominated}} / {{dep_land_taste_no_effect}}**: every cell above the floor is topology-dependent, imidacloprid included (*p* = {{dep_land_taste_imidacloprid_p_rewire_degree_preserving}}, adjusted {{dep_land_taste_imidacloprid_q}}, class {{dep_land_taste_imidacloprid_class}}). These are cell classes formed from adjusted component tests, not a cell-level FDR-controlled count; shared shuffle draws also leave the usual Benjamini-Hochberg dependence condition unverified. -**The landscape inverts between the two extracts.** Over {{dep_land_compounds}} compounds × {{dep_land_concs}} concentrations = {{dep_land_cells}} cells at {{dep_n_landscape}} permutations, with Benjamini-Hochberg across the {{dep_land_n_structural_tests}} structural tests and the relative effect floor applied, `named` gives **{{dep_land_topology_dependent}} topology-dependent, {{dep_land_composition_dominated}} composition-dominated, {{dep_land_no_effect}} no-effect, {{dep_land_mixed}} mixed** (uncorrected {{dep_land_topology_dependent_raw}} / {{dep_land_composition_dominated_raw}} / {{dep_land_no_effect_raw}}; on the absolute floor alone {{dep_land_topology_dependent_absfloor}} / {{dep_land_composition_dominated_absfloor}} / {{dep_land_no_effect_absfloor}}, the {{dep_land_below_relative_floor}} cells the relative floor removes named in T12). Repeated unchanged on the denser `taste_motor` extract at {{dep_n_landscape_taste}} permutations, the same {{dep_land_taste_cells}} cells give **{{dep_land_taste_topology_dependent}} / {{dep_land_taste_composition_dominated}} / {{dep_land_taste_no_effect}}**: every cell above the floor is topology-dependent, imidacloprid included (*p* = {{dep_land_taste_imidacloprid_p_rewire_degree_preserving}}, adjusted {{dep_land_taste_imidacloprid_q}}, class {{dep_land_taste_imidacloprid_class}}). +**The detailed reference profiles** show that the instrument also separates predictions within one extract (T6, §S2.1). At {{dep_n}} permutations per mode (resolution {{dep_p_floor}}), imidacloprid on `named` is **{{dep_named_imidacloprid_class}}** with an effect of {{dep_named_imidacloprid_effect}} rate-model units. It passes the descriptive gap criterion for {{dep_named_imidacloprid_equivalent_modes}}, is indeterminate for {{dep_named_imidacloprid_indeterminate_modes}}, and is distinguishable only from Erdős–Rényi (*p* = {{dep_named_imidacloprid_p_erdos_renyi}} versus {{dep_named_imidacloprid_p_rewire_degree_preserving}} for degree-preserving rewiring). Fipronil is **{{dep_named_fipronil_class}}** at {{dep_named_fipronil_effect}} rate-model units, with *p* = {{dep_named_fipronil_p_weight_permute}} and {{dep_named_fipronil_p_rewire_degree_preserving}} for weight permutation and rewiring, necessary level `{{dep_named_fipronil_level}}`. Weight matching reduces imidacloprid's transmitter-null median gap from {{dep_named_imidacloprid_gap_sign_permute}} to {{dep_named_imidacloprid_gap_sign_permute_weight_matched}} rate-model units; passing the gap criterion is not statistical equivalence. -**Recurrence, not size, predicts the inversion.** Repeating the same profile across a ladder of nested extracts grown from one seed set by a single fixed rule ({{scale_study_n_cuts}} extracts in all, {{scale_study_cuts}}; T27) settles the question of what the reversal tracks. The composition-dominated verdict appears on exactly {{scale_study_n_composition_dominated}} of them, `{{scale_study_composition_dominated_cuts}}`, and {{scale_study_density_not_size}}. Imidacloprid is topology-dependent on every rung but the in-star, and fipronil on all seven (per-rung sequences in §S2.4, full table T27). Both compounds' verdicts are settled on the clean ladder by the two-largest-rungs criterion of §S2.4, and at the denser middle rungs ({{scale_study_every_null_cuts}}) the prediction is distinguishable from **every** null, its necessary level being the real network itself. The relative effect size is stable over this range ({{scale_study_imidacloprid_relative_range}} of the vehicle readout), so the change of verdict is not a change of what is being measured. +Across {{scale_study_n_cuts}} extracts ({{scale_study_cuts}}; T27), composition-dominated labels occur on {{scale_study_n_composition_dominated}} extract, `{{scale_study_composition_dominated_cuts}}`, and {{scale_study_density_not_size}}. This excludes a simple monotone ordering by node count. At {{scale_study_every_null_cuts}} every tested null is rejected, and the imidacloprid relative effect spans {{scale_study_imidacloprid_relative_range}} of vehicle. Construction, density, composition, paths, readout support and effort also vary, however, so recurrence is associated with but not isolated as the cause. Agreement of the final two nested-rung labels is not convergence to a whole-network verdict. -Two honesty constraints travel with that result. The permutation budget must fall as the network grows ({{scale_study_n_range}} draws, resolution {{scale_study_resolution_range}}), and {{scale_study_coarse_note}} — so the top rungs confirm the settled verdict rather than establishing it. And the two hops-limited cuts are built by a different rule from the nested rungs, so only comparisons *within* the ladder separate scale from construction. +The permutation budget falls as networks grow ({{scale_study_n_range}} draws, resolution {{scale_study_resolution_range}}), and {{scale_study_coarse_note}}. Top-rung labels are therefore resolution-limited, nominal observations. The hops-limited cuts also use a different construction from the nested rungs. -**What is withdrawn, and what replaces it.** A degree-preserving rewire of an in-star is close to the identity, so the null had nothing to destroy; an extract with genuine recurrence does, and destroying it changes the answer. We withdraw the generalisation that most of what such a model predicts is carried by composition rather than topology, and the domain claim that went with it. What replaces it is stronger and methodological: **a dependence verdict is a joint property of the prediction and the extract, the extract's recurrence predicts which way it goes, and a negative measured on a sparse extract is not evidence about the network it was cut from.** +The earlier claim that most predictions are carried by composition does not survive a change of extract. A label is conditional on the model, extract, readout, null family and statistical effort. Testing recurrence causally would require changing it while holding the correlated properties fixed. -**Multiplicity, working.** As a prespecified confirmatory profile, fipronil at 1 µM stands comfortably (*p* = {{dep_named_fipronil_min_structural_p}} at *n* = {{dep_n}}). Inside the exploratory landscape the same cell's adjusted probability is {{dep_land_fipronil_q}} and it does **not** survive correction; its corrected class there is {{dep_land_fipronil_class}} against an uncorrected {{dep_land_fipronil_class_raw}}. +**Multiplicity sensitivity.** In the detailed reference profile, fipronil at 1 µM has a nominal result (*p* = {{dep_named_fipronil_min_structural_p}} at *n* = {{dep_n}}). Inside the exploratory landscape the same cell's adjusted component probability is {{dep_land_fipronil_q}} and no structural mode remains rejected; its adjusted-test class there is {{dep_land_fipronil_class}} against an unadjusted {{dep_land_fipronil_class_raw}}. -**Two things the class label does not say.** The ladder is **non-monotone** in {{dep_land_non_monotone}} of {{dep_land_cells}} cells on `named` and {{dep_land_taste_non_monotone}} of {{dep_land_taste_cells}} on `taste_motor`, because adjacent rungs destroy different things; the level therefore means "the cheapest degraded input this test cannot tell apart from the real one", never "everything above it is indistinguishable too". And `composition-dominated` reports a failure to distinguish, never a reproduction: across the `named` landscape's mode verdicts, {{dep_land_verdicts_equivalent_within_tolerance}} are equivalent within tolerance, {{dep_land_verdicts_indeterminate}} indeterminate and {{dep_land_verdicts_distinguishable}} distinguishable. +**Two things the class label does not say.** The ladder is **non-monotone** in {{dep_land_non_monotone}} of {{dep_land_cells}} cells on `named` and {{dep_land_taste_non_monotone}} of {{dep_land_taste_cells}} on `taste_motor`, because adjacent rungs destroy different things; the level therefore means "the cheapest degraded input this test cannot tell apart from the real one", never "everything above it is indistinguishable too". And `composition-dominated` reports a failure to distinguish, never a reproduction: across the `named` landscape's mode verdicts, {{dep_land_verdicts_equivalent_within_tolerance}} pass the descriptive median-gap tolerance, {{dep_land_verdicts_indeterminate}} indeterminate and {{dep_land_verdicts_distinguishable}} distinguishable. ### 3.3 Which pipeline layer carries the information, and what agreement would have been anyway (Fig. F13, Tables T13, T13b, T24, T25) -Receptor engagement alone orders the library **backwards** (signed ρ {{abl_rho_receptor_min}} to {{abl_rho_receptor_max}}) and is now the worst level at every concentration — constant across the table: {{abl_worst_level_constant}} ({{abl_worst_level_by_conc}}). Under the historical level-C rule the worst level was *not* constant ({{abl_worst_level_by_conc_floor_rule}}), so the previous claim that the topology-only level is the worst of the four was true of one column and false of the others. +**The matched reference and runtime sweep change the reading of the raw correlation.** Pharmacology-free pseudo-compounds with the gain shape produced by the shipped rules give a median composition-versus-full correlation of {{abl_ref_matched_median}} over {{abl_ref_n_draws}} draws (5th–95th percentile {{abl_ref_matched_p05}} to {{abl_ref_matched_p95}}). The observed value exceeds that 95th percentile, so shared algebra does not explain its entire magnitude. Yet the correlation is {{abl_rho_composition_row_abs}} under the shipped row normalisation and {{abl_rho_composition_none}} with none, but {{abl_rho_composition_degree}} under a degree-corrected denominator at 1 µM and {{abl_rho_composition_degree_1e8}} at 10 nM. It survives {{abl_composition_survives_normalisations}} and not a degree correction, making it conditional on the runtime's readout. -**The ablation conclusion is withdrawn.** Level C's multiplier applied a depression-only factor to every transmitter, so every entry was at most zero and the level could not represent disinhibition. With a direction-aware rule — same generic magnitude, one bit of sign — its rank correlation with the full model at 1 µM rises from {{abl_rho_topology_floor_paper}} to {{abl_rho_topology_paper}}. The sentence "the connectome without the pharmacology is not a cheap substitute for the connectome with it" was a statement about a sign-broken baseline and is retracted. +Against that reference, the raw composition-versus-full correlation is {{abl_rho_composition_paper}} at 1 µM and {{abl_ref_no_floor}} after dropping compounds on the 0.05 gain floor, so it is not a tie artefact. Levels B and D nevertheless share the same gain vector. Jointly permuting their paired labels leaves ρ = {{abl_ref_shuffled}} unchanged by construction and is not an independence test. A sign disagreement between the levels moves ρ by at most {{abl_glutamate_max_delta_rho}} and does not produce the agreement (§S9(i)). -**The composition-versus-full correlation carries no compound-level information.** It is {{abl_rho_composition_paper}} at 1 µM and {{abl_ref_no_floor}} after dropping the compounds on the 0.05 gain floor, so it is not a tie artefact. But levels B and D are both functions of the same gain vector, and the matched reference says what would have been unsurprising: pharmacology-free pseudo-compounds with the gain shape the shipped rules produce give a median of {{abl_ref_matched_median}} over {{abl_ref_n_draws}} draws (5th–95th percentile {{abl_ref_matched_p05}} to {{abl_ref_matched_p95}}), and permuting the compound labels leaves the observed value identically {{abl_ref_shuffled}}, as it must, because the permutation moves the (B, D) pairs as a block. +Receptor engagement alone orders the library **backwards** (signed ρ {{abl_rho_receptor_min}} to {{abl_rho_receptor_max}}) and is the worst level at every concentration: {{abl_worst_level_constant}} ({{abl_worst_level_by_conc}}). Under the historical level-C rule the worst level was not constant ({{abl_worst_level_by_conc_floor_rule}}). -**And it is not robust to the runtime.** The correlation is {{abl_rho_composition_row_abs}} under the shipped normalisation and {{abl_rho_composition_none}} with none, but {{abl_rho_composition_degree}} under a degree-corrected denominator at 1 µM and {{abl_rho_composition_degree_1e8}} at 10 nM: it survives {{abl_composition_survives_normalisations}} and not a degree correction, so it is a property of *this runtime's readout* as much as of the perturbation. A sign disagreement between the two levels moves it by at most {{abl_glutamate_max_delta_rho}} and so is not what produces the agreement (§S9(i)). +**The earlier ablation conclusion is withdrawn.** Level C applied a depression-only multiplier to every transmitter and could not represent disinhibition. With the same generic magnitude but one mechanism-table sign bit, its correlation with the full model at 1 µM rises from {{abl_rho_topology_floor_paper}} to {{abl_rho_topology_paper}}. The claim that the connectome without pharmacology was not a cheap substitute was therefore a statement about a sign-broken baseline. ### 3.4 Specification robustness: which conclusions survive the modelling choices (Fig. F14, Tables T14, T14b, T15) -**The previous 25/25 topology result was vacuous, and the failure mode is general.** Two faults compounded. The matrix was computed at 6 shuffles per specification, where the smallest attainable permutation probability is 1/7 = 0.143, so `p ≤ 0.05` was arithmetically unattainable and the predicate silently fell back on a Gaussian criterion. And the specification context rebound the transformation on the assay modules but not on the engine the permutation path resolves, so **all {{stab_family_size}} specifications fed default values to their nulls and produced identical effects**. A robustness matrix can thus report perfect stability because its alternatives never reached the code under test — a bug no determinism check can catch, since the run is perfectly reproducible either way. +Recomputed at {{stab_n_shuffles}} shuffles per specification (resolution {{stab_shuffle_resolution}}), with the specification reaching the null engine and Benjamini-Hochberg applied across {{stab_n_structural_tests}} component tests, C4 remains topology-dependent in {{stab_C4_retained}}/{{stab_family_size}} specifications. C3 has no structural rejection in {{stab_C3_retained}}/{{stab_family_size}}, unchanged from {{stab_C3_retained_uncorrected}} unadjusted retentions; {{stab_C3_equivalent}} pass the descriptive gap criterion and {{stab_C3_indeterminate}} remain indeterminate. Under the {{stab_family_size}} specifications: -Recomputed at {{stab_n_shuffles}} shuffles per specification (resolution {{stab_shuffle_resolution}}), with the specification pushed into the null engine and Benjamini-Hochberg applied across the {{stab_n_structural_tests}} structural tests, both topology conclusions still hold {{stab_C3_retained}}/{{stab_family_size}} and {{stab_C4_retained}}/{{stab_family_size}}. The nuance the old matrix could not show is that a retention rate for a conclusion shaped as a *failure to reject* is only as strong as the equivalence behind it: of C3's {{stab_C3_retained}}, {{stab_C3_equivalent}} reach equivalence within the prespecified margin and {{stab_C3_indeterminate}} are merely indeterminate. Under the {{stab_family_size}} specifications: +The adjustment acts asymmetrically on these predicates: it can remove the rejection that retains C4, but it cannot strengthen C3's failure to reject into evidence of equivalence. -* The domain's most quotable claim — that the agonist suppresses activity — is retained by only {{stab_C1_retained}} and **reversed** by all {{stab_C1_lost}} monotone specifications: the same compound at the same engagement takes the network from {{stab_monotone_vehicle_hz}} to {{stab_monotone_treated_hz}} Hz ({{stab_monotone_pct_change}}) instead of from {{stab_default_vehicle_hz}} to {{stab_default_treated_hz}} Hz, with `g_ach` at {{stab_monotone_g_ach}} against {{stab_default_g_ach}}. The direction is a property of one asserted term, not of the integration. -* A buffering claim is retained {{stab_C5_retained}} of {{stab_family_size}} and reversed by none, its {{stab_C5_undecidable}} failures being undecidable. The previous explanation — that the monotone rules buffer "at least as much, never less" — is **false**: over the {{stab_mean_gap_nicotinic_monotone_n}} reversing specifications with a defined index the mean gap runs {{stab_mean_gap_nicotinic_monotone_max}} to {{stab_mean_gap_nicotinic_monotone_min}} log₁₀ units, {{stab_monotone_buffer_less_n}} buffer *less* ({{stab_monotone_buffer_less}}), and the aggregate is carried by the extreme — an aggregate standing in for a range it does not represent. +* The claim that the agonist suppresses activity is retained by only {{stab_C1_retained}} and **reversed** by all {{stab_C1_lost}} monotone specifications: the same compound at the same engagement takes the network from {{stab_monotone_vehicle_hz}} to {{stab_monotone_treated_hz}} rate-model units ({{stab_monotone_pct_change}}) instead of from {{stab_default_vehicle_hz}} to {{stab_default_treated_hz}} rate-model units, with `g_ach` at {{stab_monotone_g_ach}} against {{stab_default_g_ach}}. The direction is a property of one asserted term, not of the integration. +* A buffering claim is retained {{stab_C5_retained}} of {{stab_family_size}} and reversed by none, its {{stab_C5_undecidable}} failures being undecidable. The previous claim that monotone rules buffer "at least as much, never less" is **false**: over the {{stab_mean_gap_nicotinic_monotone_n}} reversing specifications with a defined index the mean gap runs {{stab_mean_gap_nicotinic_monotone_max}} to {{stab_mean_gap_nicotinic_monotone_min}} log₁₀ units, {{stab_monotone_buffer_less_n}} buffer *less* ({{stab_monotone_buffer_less}}), and the aggregate is carried by the extreme; an aggregate standing in for a range it does not represent. * The matching amplification claim is weakest at {{stab_C6_retained}} of {{stab_family_size}}. Retained by every specification: {{stab_specification_independent}}. -The threshold grid makes the same point about conventions rather than transformations: the split is **not** stable (`stable = {{thr_stable}}`), buffering holding in {{thr_nicotinic_matching}} of {{thr_nicotinic_cells}} cells ({{thr_nicotinic_gap_max}} to {{thr_nicotinic_gap_min}} log₁₀ units) and amplification in {{thr_nav_ache_matching}} of {{thr_nav_ache_cells}}, flipping at the strictest corner ({{thr_nav_ache_failing}}, gap {{thr_nav_ache_gap_min}}), so ranges are quoted and never the point estimate. Two tables that appear to disagree in fact score different sets — T15 the {{thr_n_compounds}} compounds of two mechanism classes, T8 the whole library — so {{thr_outside_t15_scope}} is scored in one and outside the other's scope. +The earlier 25/25 topology result was vacuous for two reasons. At 6 shuffles, the smallest probability was 1/7 = 0.143, so `p ≤ 0.05` was unattainable and the predicate fell back on a Gaussian criterion. More seriously, alternative transformations were rebound on assay modules but not on the engine resolved by the permutation path, so all {{stab_family_size}} specifications fed default values to their nulls. The run was deterministic but unable to make the intended comparison. + +The threshold split is also unstable (`stable = {{thr_stable}}`): buffering holds in {{thr_nicotinic_matching}} of {{thr_nicotinic_cells}} cells ({{thr_nicotinic_gap_max}} to {{thr_nicotinic_gap_min}} log₁₀ units), while amplification holds in {{thr_nav_ache_matching}} of {{thr_nav_ache_cells}} and flips at {{thr_nav_ache_failing}} (gap {{thr_nav_ache_gap_min}}). T15 covers {{thr_n_compounds}} compounds in two mechanism classes; {{thr_outside_t15_scope}} is scored in the whole-library T8 but lies outside T15's scope. ### 3.5 Variance attribution, and what to measure next (Fig. F15, Tables T16, T17) -The Sobol' analysis ({{unc_evaluations}} model evaluations at n_base = {{unc_n_base}}, Var(Y) = {{unc_variance}} Hz²) resolves {{unc_n_resolved}} of {{unc_n_factors}} factors in the strict sense of a bootstrap first-order interval excluding zero: {{unc_resolved}} — the **engagement-to-gain transformation** (S1 {{unc_S_gain_transform}}, ST {{unc_ST_gain_transform}}) and the cut's **synapse-weight threshold** (S1 {{unc_S_weight_threshold}}, ST {{unc_ST_weight_threshold}}). Everything else is unresolved at this sample size ({{unc_at_noise_floor}}), including the external drive, whose S1 of {{unc_S_drive}} sits below the estimator's noise floor of {{unc_noise_floor}} on an interval of {{unc_CI_drive}}. The floor is set by {{unc_noise_floor_factor}}, not by the declared null factor `lif_seed` at {{unc_null_factor_S}}; the previous version of this section quoted the null factor and then ranked a factor inside the resulting noise. Interactions carry {{unc_interaction_share}} of the variance raw and {{unc_interaction_share_clipped}} with negatives clipped. On the analytic benchmark the estimator returns {{ish_estimated}} against the exact {{ish_analytic}} at n_base = {{ish_n_base}} (max error {{ish_max_abs_error}}). +The Sobol' analysis ({{unc_evaluations}} model evaluations at n_base = {{unc_n_base}}, Var(Y) = {{unc_variance}} squared rate-model units) resolves {{unc_n_resolved}} of {{unc_n_factors}} factors in the strict sense of a bootstrap first-order interval excluding zero: {{unc_resolved}}; the **engagement-to-gain transformation** (S1 {{unc_S_gain_transform}}, ST {{unc_ST_gain_transform}}) and the cut's **synapse-weight threshold** (S1 {{unc_S_weight_threshold}}, ST {{unc_ST_weight_threshold}}). Everything else is unresolved at this sample size ({{unc_at_noise_floor}}), including the external drive, whose S1 of {{unc_S_drive}} sits below the largest observed negative-index magnitude of {{unc_noise_floor}} on an interval of {{unc_CI_drive}}. This finite-sample diagnostic is set by {{unc_noise_floor_factor}}, not by the declared null factor `lif_seed` at {{unc_null_factor_S}}; the previous version of this section quoted the null factor and then ranked a factor inside the resulting noise. The residual `1 − sum(S1)` is {{unc_interaction_share}} raw and {{unc_interaction_share_clipped}} with negatives clipped; estimator error contributes to this difference, so neither is a precise interaction fraction. On the analytic benchmark the estimator returns {{ish_estimated}} against the exact {{ish_analytic}} at n_base = {{ish_n_base}} (max error {{ish_max_abs_error}}). -The top factor is **{{voi_rank1_factor}}** ({{voi_rank1_var}} of {{voi_var_total}} Hz²), resolved by *{{voi_rank1_experiment}}* — reading the *shape* of the response-versus-engagement curve at a known synapse, which is precisely the assumption §3.4 shows the suppression conclusion turns on. The second, **{{voi_rank2_factor}}** ({{voi_rank2_var}} Hz²), needs **no experiment at all**: it is resolved by re-analysing synapse-confidence strata of data already held. +The top factor is **{{voi_rank1_factor}}** ({{voi_rank1_var}} of {{voi_var_total}} squared rate-model units), resolved by *{{voi_rank1_experiment}}*; reading the *shape* of the response-versus-engagement curve at a known synapse, which is precisely the assumption §3.4 shows the suppression conclusion turns on. The second, **{{voi_rank2_factor}}** ({{voi_rank2_var}} squared rate-model units), motivates re-analysis of synapse-confidence strata. Such analysis measures sensitivity to extraction choices; it does not resolve missing connectivity or guarantee the stated variance reduction in practice. ### 3.6 What the refusal rule actually does (Fig. F11, Tables T10, T20) -The type system changes results rather than documentation. Of {{ev_rows}} library rows, {{ev_rows_not_modelled}} carry no admissible evidence and return N/A instead of a small plausible number, and the admissible transformation is decided by the (parameter type, distance) pair rather than by the author: {{ev_model_binding_occupancy}} row earns a physical binding model, {{ev_model_binding_engagement_proxy}} a labelled binding proxy, {{ev_model_functional_engagement}} a direct functional engagement and {{ev_model_functional_engagement_proxy}} a transferred one. The single physical row is {{ev_binding_row_names}} [77]; the library's previous binding constant, {{ev_binding_proxy_row_names}}, is demoted to a cross-species proxy — the same number under a name that says it was transferred. Removing that one row would leave a branch of the type system with no instances, which is the kind of fact a lookup table hides and an exhaustively checked invariant surfaces. The full census, the domain readouts the engagements feed, and the literature-concordance analysis are in §S6 and §S10. +Evidence gating changes numerical outputs and their interpretation. Of {{ev_rows}} library rows, {{ev_rows_not_modelled}} carry no admissible evidence and return N/A instead of a small plausible number. Dispatch on the (parameter type, distance) pair assigns {{ev_model_binding_occupancy}} row a physical binding-occupancy model, {{ev_model_binding_engagement_proxy}} a binding proxy, {{ev_model_functional_engagement}} a direct functional engagement and {{ev_model_functional_engagement_proxy}} a transferred one. The physical row is {{ev_binding_row_names}} [76]; the previous binding constant, {{ev_binding_proxy_row_names}}, is retained numerically but labelled as a cross-species proxy. Without the new exact-target row, the shipped library would have no example exercising the physical-occupancy branch. The finite audit makes that coverage visible; it does not introduce new Hill algebra. The full census, downstream domain readouts and literature-concordance analysis are in §S6 and §S10. -### 3.7 Reproduction, determinism and cross-backend parity (Fig. F1) +### 3.7 Availability and reproducibility (engineering support; Fig. F1) -One command regenerates every figure, table and number in this paper from committed artifacts, with no download and no access to the 1.1 GB weight matrix: +The standard command recomputes the core analyses on committed cuts and renders the paper. The scale table is an exception: it imports `papers/scale_study.json`, a committed earlier result, rather than recomputing the large-rung profiles. Only the smallest scale cut is committed; a fresh scale study requires the missing cuts and their upstream data. Re-rendering T27 is therefore not an independent reproduction of that experiment: ```bash python -m pip install -e ".[dev,viz]" @@ -191,37 +200,45 @@ python scripts/reproduce_paper.py --fast # reduced permutation and replicate c python scripts/reproduce_paper.py --only dependence --outdir /tmp/x ``` -Beyond the paper, `flylab run` turns a declarative spec into a self-describing run directory whose manifest records the spec, code and library hashes, the seeds and two hashes per artifact; the same spec and seed reproduce it byte for byte (`docs/WORKFLOW.md`). `papers/results.json` carries every scalar quoted above with its value, rendered text, unit and producing step; two runs at one commit and seed produce identical values, runtimes excepted, and a test asserts it. Every numeric span in the manuscript is a substitution from that record, so a number cannot enter the prose unless the pipeline produced it, and the test suite asserts that every key either document quotes exists in the record, that the record did not come from a reduced-effort run, and — new here, because the previous record's shuffle count did not match the shipped default and nothing caught it — that **every statistical parameter in the record equals the shipped default**. +This section records engineering support rather than a fourth research question. `flylab run` produces a manifest containing spec, code and library hashes, seeds and two hashes per artifact; a test reproduces that artifact byte for byte (`docs/WORKFLOW.md`). The paper determinism test covers only the reduced provenance-and-census path, not every full analysis. `papers/results.json` records each quoted value, unit and producing step, and guards reject reduced-effort records or mismatched declared defaults. + +The same wheel runs on Pyodide. `tests/test_browser_bridge.py` asserts **endpoint parity** for a representative call to each route, including JSON readouts, warnings and provenance; it does not establish whole-analysis equality. An archival DOI and release tag remain outstanding. -The same wheel runs in a browser tab on Pyodide and is published as a static site, so these assays can be re-run by opening one link with nothing installed; `tests/test_browser_bridge.py` asserts **endpoint parity** — equal JSON, every readout, warning and provenance field, for a representative call to each route, which is route-level equality rather than a whole-analysis equality claim. MaleCNS v1.0 and FlyWire are CC-BY 4.0 from their consortia [5], [6] and must be cited before this software; licences are listed under Statements, and an archival Zenodo deposit will carry the DOI **[to be minted on acceptance]**. +The latest attempted full numerical verification was interrupted during dependence analysis. This revision preserves the recorded numerical source and edits its prose; it is not presented as a completed independent rerun. ## 4. Discussion -**What transfers.** *First*, the dependence verdict is a joint property of a prediction and the extract it is measured on, and the extract's recurrence predicts which way it goes. The same validated instrument, with no parameter changed, classifies {{dep_land_composition_dominated}} of {{dep_land_cells}} cells composition-dominated on a sparse extract (mean degree {{named_mean_degree}}) and {{dep_land_taste_composition_dominated}} of {{dep_land_taste_cells}} on a denser one (mean degree {{taste_motor_mean_degree}}), and across {{scale_study_n_cuts}} extracts the composition-dominated verdict survives on {{scale_study_n_composition_dominated}} — the in-star — while an extract with *fewer* nodes and twenty times the mean degree is already topology-dependent. The practical rule generalises to any pipeline of this shape: **a dependence result must be reported with the structural statistics of the extract it was measured on, and a negative measured on a sparse extract is not evidence about the network it was cut from.** Because the structural statistic that predicts the flip is computable before any permutation is run (T19, T27), this is a cheap check, not a research programme. +**What transfers.** Dependence labels change across extracts: {{dep_land_composition_dominated}} of {{dep_land_cells}} cells are composition-dominated on `named`, versus {{dep_land_taste_composition_dominated}} of {{dep_land_taste_cells}} on `taste_motor`. Only one extract has that label, and several structural properties differ at once, so recurrence is not isolated. Evidence typing also changes outputs by refusing unsupported numbers and labelling transferred constants. Finally, the biased null and disconnected specification context show why reproducibility alone is insufficient: both broken comparisons were deterministic and both flattered the conclusions. The gain transformation is consequently both the source of the direction reversal and the largest resolved variance factor. -*Second*, typing the evidence changes results rather than documentation: refusing inadmissible rows removed a class of small, plausible, invented numbers no reader could have told apart from measurements, and grading a source's distance demoted a cross-species constant to a labelled proxy without changing its value. Stated as an invariant and checked exhaustively, this is a property of the pipeline rather than a convention its authors intend to follow. +Other literature-parameterised network domains can apply the same audit without adopting this biological model: -*Third*, two of the instruments were themselves broken in ways that flattered the conclusions, and both were found by pointing the instruments inward: a null that moved the quantity it claimed to preserve exactly, and a specification family whose alternatives never reached the engine its predicates ran on, so a uniform robustness rate was guaranteed by construction. Both failures share a signature worth naming — **a test whose result was determined by its own implementation rather than by the data** — and both were invisible to determinism checks, because a record can be perfectly reproducible and still be produced by a test that could not have returned another answer. The specification family and the variance decomposition then converge on one weak point: the direction of the headline effect is set by an asserted functional form, which is also the largest resolved factor in the variance budget and the top-ranked measurement. +1. type each external parameter and reject transformations its provenance does not license; +2. state the extract, readout, runtime normalisation and statistical effort with every dependence result; +3. design each null around the quantity it must preserve, then measure whether it actually preserves it; +4. trace every alternative specification into the engine used by the test, including null-generation paths; +5. compare extracts and runtime conventions before generalising, and use variance attribution to select the next measurement. -**Encoded assumptions versus generated observations.** The separation is itself generated rather than asserted: an audit of the dependency chain behind one readout returns {{claims_chain_links}} links (T18), of which exactly {{claims_observed_phrase}} is a measurement of the system being simulated, {{claims_literature_derived}} reuses a measurement of something else, {{claims_model_assumption}} are asserted modelling choices and {{claims_computed}} are arithmetic on the links above. A companion decomposition — a different decomposition, not a repartition of those links — reports {{claims_facts_phrase}}, {{claims_inference_phrase}} and {{claims_unknown_phrase}}. That one link in nine is a measurement of the simulated system is worth stating plainly for any pipeline of this shape. +These are proposed audit steps, not evidence that FlyLab has been implemented in power, metabolic or epidemic networks. -**Relation to adjacent work.** Our version of the [69], [70] result survives neither a change of extract (§3.2) nor a change of the runtime's normalisation (§3.3), and both models normalise or regionally average, so the parallel is at least as consistent with a shared modelling artefact as with a property of receptor-weighted network perturbation. That is itself a methodological observation: a result reproduced across two pipelines that share a normalisation convention has not been independently confirmed. +**Encoded assumptions versus generated observations.** The audit returns {{claims_chain_links}} links (T18): {{claims_observed_phrase}} measures the simulated system, {{claims_literature_derived}} measures something else, {{claims_model_assumption}} are asserted and {{claims_computed}} are computed. A separate decomposition, not a repartition of those links, reports {{claims_facts_phrase}}, {{claims_inference_phrase}} and {{claims_unknown_phrase}}. -**What the pipeline cannot support**, and these limits are themselves computing results about what such a pipeline can establish. It cannot predict a dose–response in an animal: with no pharmacokinetics available, the free concentration is an *input*, not a derived quantity. It cannot support a cross-species safety claim, nor a subunit-level claim where the sources used hybrid constructs. It cannot support a general statement about whether perturbation on networks of this kind needs topology — only a statement per extract with that extract's structure reported. And it cannot produce a live-animal result, which no code path may write. +**Relation to adjacent work.** Our version of the [68], [69] result survives neither a change of extract (§3.2) nor runtime normalisation (§3.3). Because both models normalise or regionally average, the parallel may reflect a shared convention rather than independent confirmation. -**Five model-specific caveats** are given in full in §S9: an asserted sign table exempt from the specification family and unbounded by the label-permuting null; the load-bearing row normalisation; perturbations applied across transmitter-defined rather than receptor-defined synapses, which the expression layer bounds rather than corrects; hops-limited extracts, whose missing inhibition can flip the sign of any readout and whose cost §3.2 now quantifies; and agreement of direction only between the two runtimes. +**What the pipeline cannot support.** Without pharmacokinetics, free concentration is an input rather than an animal dose. The model supports neither cross-species safety nor subunit claims from hybrid constructs, and it cannot generalise topology dependence beyond the reported extract. No code path may write a live-animal result. -**Prospective, not pre-registered.** The hypotheses in §S5 are written before any experiment and labelled as software predictions with no live result; a mutable repository is not a registry, so they become pre-registered only once a tagged release is externally archived. Two of them predict a readout whose *sign* is a modelling choice over this paper's own specification family, and the supplement now says so beside them. +**Five model-specific caveats** are given in full in §S9: an asserted sign table exempt from the specification family and unbounded by the label-permuting null; the load-bearing row normalisation; perturbations applied across transmitter-defined rather than receptor-defined synapses, which the expression layer bounds rather than corrects; hops-limited extracts with unobserved boundary inputs and extract-sensitive results; and agreement of direction only between the two runtimes. Expression mapping coverage is {{expr_coverage_overall}}, and the principal MN9 readout remains unresolved. + +**Prospective, not pre-registered.** The hypotheses in §S5 are written before any experiment and labelled as software predictions with no live result; a mutable repository is not a registry, and an archive DOI alone does not establish preregistration. A dated, frozen experimental protocol specifying hypotheses, outcomes and analyses before collection is also needed. Two of them predict a readout whose *sign* is a modelling choice over this paper's own specification family, and the supplement now says so beside them. ## 5. Conclusion -We set out to make a network simulation say which of its inputs its predictions depend on, and to evaluate that machinery rather than assert it. The method is four instruments — typed evidence propagation under a checked soundness invariant, input-dependence testing against an ordered ladder with an explicit equivalence margin and multiplicity control, an ablation ladder scored against a matched reference, and specification-family robustness with variance attribution — none specific to the domain we demonstrate them on. Pointed at its own results, the method establishes that the headline effect's direction is a property of an asserted functional form; that a layer-agreement score near 1 can be an algebraic consequence of the pipeline's own normalisation; and — the result we would most want another group to break — that **a dependence verdict obtained on one extract of a network can invert on another extract of the same network, with recurrence rather than size predicting which way**. A "this prediction does not need the structure" result is only as good as the extract it was measured on, and the statistic that says so is cheap to compute in advance. +We set out to make a network simulation say which of its inputs its predictions depend on, and to evaluate that machinery rather than assert it. The method combines four components: typed evidence propagation under an enforced transformation rule and finite-library invariant audit, input-dependence diagnostics with an explicit point-gap tolerance and component-test multiplicity adjustment, an ablation ladder scored against a matched reference, and specification-family robustness with variance attribution; none specific to the domain we demonstrate them on. Pointed at its own results, the method establishes that the headline effect's direction is a property of an asserted functional form; that layer-agreement scores depend on shared gains and runtime normalisation; and that **dependence labels can change across extracts of the same network**. The present study does not isolate recurrence from correlated extraction properties. A "this prediction does not need the structure" result is only as good as the extract it was measured on, and structural summaries can flag the need for additional checks; they are not yet validated substitutes for those checks. What would change the picture is short and specific, and the variance budget rather than taste names it: a calibration of the parameter-to-gain transformation at a known synapse; a reconstruction-confidence re-analysis, which needs no new measurement; a full landscape on an extract that is neither an in-star nor hops-limited; and one live assay. Until then this is a description of a method, and of a tool that tests itself. ## Statements -**AI assistance.** The author defined the research questions, methods and analyses; verified the source literature, checking each cited publication against the value attributed to it; executed and reviewed every analysis; and accepts full responsibility for the scientific content. AI assistance was used for software implementation and for drafting and editing prose. No data were generated, imputed, estimated or extrapolated by a language model; every number in the results, tables and figures was produced by the reproduction pipeline from committed artifacts and is keyed in `papers/results.json`. +**AI assistance.** AI tools assisted implementation, statistical checking, literature lookup, peer review and drafting. The reported simulation outputs were computed by the analysis code. The author is responsible for verifying the source literature, analysis choices and final scientific claims before submission; automated checks do not establish that this author verification has occurred. **Animal research.** No animal experiments were performed and no live-animal data are reported. The `live_lab` field of every notebook produced by this pipeline is `null`. The protocols in `protocols/` are unexecuted drafts. @@ -233,157 +250,155 @@ What would change the picture is short and specific, and the variance budget rat [1] J. Dempster, *RatCVS*, University of Strathclyde, Glasgow, UK. Simulated pithed-rat cardiovascular pharmacology teaching package. -[2] S. Okabe *et al.*, "A web-based simulation of cardiovascular pharmacology for undergraduate teaching," *Eur. J. Pharmacol.*, vol. 997, art. 177618, 2025. - -[3] *PharmVR*: immersive virtual-reality pharmacology training, preprint, 2025. +[2] S. Okabe, T. Goto, D. Hirayama and Y. Nishimura, "Virtual rat web: A versatile simulation tool for pharmacology education in a variety of settings," *Eur. J. Pharmacol.*, vol. 997, art. 177618, 2025. DOI 10.1016/j.ejphar.2025.177618. PMID 40222443. -[4] "Virtual simulation in pharmacology and physiology education: an evaluation," *J. Educ. Eval. Health Prof.*, 2020. +[3] K. Thiyagarajan, "PharmVR, Immersive Virtual Reality based Animal Simulator Software," *Preprints.org*, concept paper, 2025. DOI 10.20944/preprints202507.0959.v1. Not peer reviewed. -[5] S. Dorkenwald *et al.*, "Neuronal wiring diagram of an adult brain," *Nature*, vol. 634, pp. 124–138, 2024. DOI 10.1038/s41586-024-07558-y. +[4] S. Dorkenwald *et al.*, "Neuronal wiring diagram of an adult brain," *Nature*, vol. 634, pp. 124–138, 2024. DOI 10.1038/s41586-024-07558-y. -[6] HHMI Janelia FlyEM, University of Cambridge, MRC LMB and Google Research, *MaleCNS v1.0: a synapse-resolution connectome of the adult male* Drosophila *central nervous system*, 2026. CC-BY 4.0. +[5] HHMI Janelia FlyEM, University of Cambridge, MRC LMB and Google Research, *MaleCNS v1.0: a synapse-resolution connectome of the adult male* Drosophila *central nervous system*, 2026. CC-BY 4.0. -[7] P. K. Shiu, G. R. Sterne, N. Spiller *et al.*, "A leaky integrate-and-fire computational model based on the connectome of the entire adult *Drosophila* brain reveals insights into sensorimotor processing," *Nature*, vol. 634, 2024. +[6] P. K. Shiu, G. R. Sterne, N. Spiller *et al.*, "A leaky integrate-and-fire computational model based on the connectome of the entire adult *Drosophila* brain reveals insights into sensorimotor processing," *Nature*, vol. 634, 2024. -[8] C. Bass *et al.*, "Mutation analysis and functional characterisation of insect nicotinic acetylcholine receptors," *Insect Biochem. Mol. Biol.*, vol. 36, pp. 86–96, 2006. +[7] C. Bass *et al.*, "Mutation analysis and functional characterisation of insect nicotinic acetylcholine receptors," *Insect Biochem. Mol. Biol.*, vol. 36, pp. 86–96, 2006. -[9] H. Dederer *et al.*, "Selectivity of neonicotinoids at hybrid insect nicotinic acetylcholine receptors," *Insect Biochem. Mol. Biol.*, vol. 41, p. 51, 2011. +[8] H. Dederer *et al.*, "Selectivity of neonicotinoids at hybrid insect nicotinic acetylcholine receptors," *Insect Biochem. Mol. Biol.*, vol. 41, p. 51, 2011. -[10] FlyConnectome, *Neurotransmitter ground-truth repository for* Drosophila *connectomes*, 2024. +[9] N. Eckstein *et al.*, "Neurotransmitter classification from electron microscopy images at synaptic sites in *Drosophila melanogaster*," *Cell*, vol. 187, pp. 2574–2594.e23, 2024. DOI 10.1016/j.cell.2024.03.016. PMID 38729112. PMCID PMC11106717. -[11] P. Mineault, "Deconstructing viral fly simulations," technical note, 2026. +[10] P. Mineault, "Deconstructing viral fly sims," *The NeuroAI archive*, 14 September 2026. https://www.neuroai.science/p/are-flies-playing-beat-saber. Newsletter essay, not peer reviewed. -[12] Y. Deng *et al.*, "Chemoconnectomics: mapping chemical transmission in *Drosophila*," *Neuron*, vol. 101, pp. 876–893, 2019. +[11] Y. Deng *et al.*, "Chemoconnectomics: mapping chemical transmission in *Drosophila*," *Neuron*, vol. 101, pp. 876–893, 2019. -[13] C. Bass, A. M. Puinean, M. Andrews *et al.*, "Mutation of a nicotinic acetylcholine receptor β subunit is associated with resistance to neonicotinoid insecticides in the aphid *Myzus persicae*," *BMC Neurosci.*, vol. 12, art. 51, 2011. DOI 10.1186/1471-2202-12-51. PMID 21627790. +[12] C. Bass, A. M. Puinean, M. Andrews *et al.*, "Mutation of a nicotinic acetylcholine receptor β subunit is associated with resistance to neonicotinoid insecticides in the aphid *Myzus persicae*," *BMC Neurosci.*, vol. 12, art. 51, 2011. DOI 10.1186/1471-2202-12-51. PMID 21627790. -[14] R. H. ffrench-Constant, T. A. Rocheleau, J. C. Steichen and A. E. Chalmers, "A point mutation in a *Drosophila* GABA receptor confers insecticide resistance," *Nature*, vol. 363, pp. 449–451, 1993. DOI 10.1038/363449a0. +[13] R. H. ffrench-Constant, T. A. Rocheleau, J. C. Steichen and A. E. Chalmers, "A point mutation in a *Drosophila* GABA receptor confers insecticide resistance," *Nature*, vol. 363, pp. 449–451, 1993. DOI 10.1038/363449a0. -[15] W. T. Garrood, C. T. Zimmer, O. Gutbrod *et al.*, "Influence of the RDL A301S mutation in the brown planthopper *Nilaparvata lugens* on the activity of phenylpyrazole insecticides," *Pestic. Biochem. Physiol.*, vol. 142, pp. 1–8, 2017. DOI 10.1016/j.pestbp.2017.01.007. PMID 29107231. +[14] W. T. Garrood, C. T. Zimmer, O. Gutbrod *et al.*, "Influence of the RDL A301S mutation in the brown planthopper *Nilaparvata lugens* on the activity of phenylpyrazole insecticides," *Pestic. Biochem. Physiol.*, vol. 142, pp. 1–8, 2017. DOI 10.1016/j.pestbp.2017.01.007. PMID 29107231. -[16] A. M. Hosie, H. A. Baylis, S. D. Buckingham and D. B. Sattelle, "Actions of the insecticide fipronil on dieldrin-sensitive and -resistant GABA receptors of *Drosophila melanogaster*," *Br. J. Pharmacol.*, vol. 115, no. 6, pp. 909–912, 1995. DOI 10.1111/j.1476-5381.1995.tb15896.x. PMID 7582519. +[15] A. M. Hosie, H. A. Baylis, S. D. Buckingham and D. B. Sattelle, "Actions of the insecticide fipronil on dieldrin-sensitive and -resistant GABA receptors of *Drosophila melanogaster*," *Br. J. Pharmacol.*, vol. 115, no. 6, pp. 909–912, 1995. DOI 10.1111/j.1476-5381.1995.tb15896.x. PMID 7582519. -[17] A. M. Hosie and D. B. Sattelle, "Allosteric modulation of an expressed homo-oligomeric GABA-gated chloride channel of *Drosophila melanogaster*," *Br. J. Pharmacol.*, vol. 117, no. 6, pp. 1229–1237, 1996. DOI 10.1111/j.1476-5381.1996.tb16720.x. PMID 8882620. +[16] A. M. Hosie and D. B. Sattelle, "Allosteric modulation of an expressed homo-oligomeric GABA-gated chloride channel of *Drosophila melanogaster*," *Br. J. Pharmacol.*, vol. 117, no. 6, pp. 1229–1237, 1996. DOI 10.1111/j.1476-5381.1996.tb16720.x. PMID 8882620. -[18] D. H. Kwon, K. S. Yoon, J. M. Clark and S. H. Lee, "A point mutation in a glutamate-gated chloride channel confers abamectin resistance in the two-spotted spider mite, *Tetranychus urticae* Koch," *Insect Mol. Biol.*, vol. 19, pp. 583–591, 2010. DOI 10.1111/j.1365-2583.2010.01017.x. PMID 20522121. +[17] D. H. Kwon, K. S. Yoon, J. M. Clark and S. H. Lee, "A point mutation in a glutamate-gated chloride channel confers abamectin resistance in the two-spotted spider mite, *Tetranychus urticae* Koch," *Insect Mol. Biol.*, vol. 19, pp. 583–591, 2010. DOI 10.1111/j.1365-2583.2010.01017.x. PMID 20522121. -[19] G. Le Goff, A. Hamon, J.-B. Bergé and M. Amichot, "Resistance to fipronil in *Drosophila simulans*: influence of two point mutations in the RDL GABA receptor subunit," *J. Neurochem.*, vol. 92, pp. 1295–1305, 2005. DOI 10.1111/j.1471-4159.2004.02922.x. PMID 15748149. +[18] G. Le Goff, A. Hamon, J.-B. Bergé and M. Amichot, "Resistance to fipronil in *Drosophila simulans*: influence of two point mutations in the RDL GABA receptor subunit," *J. Neurochem.*, vol. 92, pp. 1295–1305, 2005. DOI 10.1111/j.1471-4159.2004.02922.x. PMID 15748149. -[20] P. Menozzi, M. A. Shi, A. Lougarre, Z. H. Tang and D. Fournier, "Mutations of acetylcholinesterase which confer insecticide resistance in *Drosophila melanogaster* populations," *BMC Evol. Biol.*, vol. 4, art. 4, 2004. DOI 10.1186/1471-2148-4-4. PMID 15018651. +[19] P. Menozzi, M. A. Shi, A. Lougarre, Z. H. Tang and D. Fournier, "Mutations of acetylcholinesterase which confer insecticide resistance in *Drosophila melanogaster* populations," *BMC Evol. Biol.*, vol. 4, art. 4, 2004. DOI 10.1186/1471-2148-4-4. PMID 15018651. -[21] A. Mutero, M. Pralavorio, J. M. Bride and D. Fournier, "Resistance-associated point mutations in insecticide-insensitive acetylcholinesterase," *Proc. Natl. Acad. Sci. USA*, vol. 91, no. 13, pp. 5922–5926, 1994. DOI 10.1073/pnas.91.13.5922. PMID 8016090. +[20] A. Mutero, M. Pralavorio, J. M. Bride and D. Fournier, "Resistance-associated point mutations in insecticide-insensitive acetylcholinesterase," *Proc. Natl. Acad. Sci. USA*, vol. 91, no. 13, pp. 5922–5926, 1994. DOI 10.1073/pnas.91.13.5922. PMID 8016090. -[22] P. N. R. Usherwood, H. Vais, B. P. S. Khambay, T. G. E. Davies and M. S. Williamson, "Sensitivity of the *Drosophila para* sodium channel to DDT is not lowered by the super-kdr mutation M918T on the IIS4–S5 linker that profoundly reduces sensitivity to permethrin and deltamethrin," *FEBS Lett.*, vol. 579, no. 28, pp. 6317–6325, 2005. DOI 10.1016/j.febslet.2005.09.096. PMID 16263118. +[21] P. N. R. Usherwood, H. Vais, B. P. S. Khambay, T. G. E. Davies and M. S. Williamson, "Sensitivity of the *Drosophila para* sodium channel to DDT is not lowered by the super-kdr mutation M918T on the IIS4–S5 linker that profoundly reduces sensitivity to permethrin and deltamethrin," *FEBS Lett.*, vol. 579, no. 28, pp. 6317–6325, 2005. DOI 10.1016/j.febslet.2005.09.096. PMID 16263118. -[23] H. Vais, M. S. Williamson, S. J. Goodson *et al.*, "Activation of *Drosophila* sodium channels promotes modification by deltamethrin: reductions in affinity caused by knock-down resistance mutations," *J. Gen. Physiol.*, vol. 115, no. 3, pp. 305–318, 2000. DOI 10.1085/jgp.115.3.305. PMID 10694259. +[22] H. Vais, M. S. Williamson, S. J. Goodson *et al.*, "Activation of *Drosophila* sodium channels promotes modification by deltamethrin: reductions in affinity caused by knock-down resistance mutations," *J. Gen. Physiol.*, vol. 115, no. 3, pp. 305–318, 2000. DOI 10.1085/jgp.115.3.305. PMID 10694259. -[24] X. Wang, A. M. Puinean, A. O. O'Reilly *et al.*, "Mutations on M3 helix of *Plutella xylostella* glutamate-gated chloride channel confer unequal resistance to abamectin by two different mechanisms," *Insect Biochem. Mol. Biol.*, vol. 86, pp. 50–57, 2017. DOI 10.1016/j.ibmb.2017.05.006. PMID 28576654. +[23] X. Wang, A. M. Puinean, A. O. O'Reilly *et al.*, "Mutations on M3 helix of *Plutella xylostella* glutamate-gated chloride channel confer unequal resistance to abamectin by two different mechanisms," *Insect Biochem. Mol. Biol.*, vol. 86, pp. 50–57, 2017. DOI 10.1016/j.ibmb.2017.05.006. PMID 28576654. -[25] H. G. Zhang, R. H. ffrench-Constant and M. B. Jackson, "A unique amino acid of the *Drosophila* GABA receptor with influence on drug sensitivity by two mechanisms," *J. Physiol.*, vol. 479, no. 1, pp. 65–75, 1994. DOI 10.1113/jphysiol.1994.sp020278. PMID 7527461. +[24] H. G. Zhang, R. H. ffrench-Constant and M. B. Jackson, "A unique amino acid of the *Drosophila* GABA receptor with influence on drug sensitivity by two mechanisms," *J. Physiol.*, vol. 479, no. 1, pp. 65–75, 1994. DOI 10.1113/jphysiol.1994.sp020278. PMID 7527461. -[26] A. Cartereau, C. Martin and S. H. Thany, "Neonicotinoid insecticides differently modulate acetylcholine-induced currents on mammalian α7 nicotinic acetylcholine receptors," *Br. J. Pharmacol.*, vol. 175, no. 11, pp. 1987–1998, 2018. DOI 10.1111/bph.14018. +[25] A. Cartereau, C. Martin and S. H. Thany, "Neonicotinoid insecticides differently modulate acetylcholine-induced currents on mammalian α7 nicotinic acetylcholine receptors," *Br. J. Pharmacol.*, vol. 175, no. 11, pp. 1987–1998, 2018. DOI 10.1111/bph.14018. -[27] L. M. Cole, R. A. Nicholson and J. E. Casida, "Action of phenylpyrazole insecticides at the GABA-gated chloride channel," *Pestic. Biochem. Physiol.*, vol. 46, no. 1, pp. 47–54, 1993. DOI 10.1006/pest.1993.1035. +[26] L. M. Cole, R. A. Nicholson and J. E. Casida, "Action of phenylpyrazole insecticides at the GABA-gated chloride channel," *Pestic. Biochem. Physiol.*, vol. 46, no. 1, pp. 47–54, 1993. DOI 10.1006/pest.1993.1035. -[28] G. R. Dawson, K. A. Wafford, A. Smith *et al.*, "Anticonvulsant and adverse effects of avermectin analogs in mice are mediated through the γ-aminobutyric acid A receptor," *J. Pharmacol. Exp. Ther.*, vol. 295, no. 3, pp. 1051–1060, 2000. PMID 11082440. +[27] G. R. Dawson, K. A. Wafford, A. Smith *et al.*, "Anticonvulsant and adverse effects of avermectin analogs in mice are mediated through the γ-aminobutyric acid A receptor," *J. Pharmacol. Exp. Ther.*, vol. 295, no. 3, pp. 1051–1060, 2000. PMID 11082440. -[29] K. Lees, M. Musgaard, S. Suwanmanee *et al.*, "Actions of agonists, fipronil and ivermectin on the predominant in vivo splice and edit variant (RDLbd, I/V) of the *Drosophila* GABA receptor expressed in *Xenopus laevis* oocytes," *PLoS ONE*, vol. 9, no. 5, art. e97468, 2014. DOI 10.1371/journal.pone.0097468. +[28] K. Lees, M. Musgaard, S. Suwanmanee *et al.*, "Actions of agonists, fipronil and ivermectin on the predominant in vivo splice and edit variant (RDLbd, I/V) of the *Drosophila* GABA receptor expressed in *Xenopus laevis* oocytes," *PLoS ONE*, vol. 9, no. 5, art. e97468, 2014. DOI 10.1371/journal.pone.0097468. -[30] T. Perry, D. G. Heckel, J. A. McKenzie and P. Batterham, "Effects of mutations in *Drosophila* nicotinic acetylcholine receptor subunits on sensitivity to insecticides targeting nicotinic acetylcholine receptors," *Pestic. Biochem. Physiol.*, vol. 102, no. 1, pp. 56–60, 2012. DOI 10.1016/j.pestbp.2011.10.010. +[29] T. Perry, D. G. Heckel, J. A. McKenzie and P. Batterham, "Effects of mutations in *Drosophila* nicotinic acetylcholine receptor subunits on sensitivity to insecticides targeting nicotinic acetylcholine receptors," *Pestic. Biochem. Physiol.*, vol. 102, no. 1, pp. 56–60, 2012. DOI 10.1016/j.pestbp.2011.10.010. -[31] G. S. Ratra and J. E. Casida, "GABA receptor subunit composition relative to insecticide potency and selectivity," *Toxicol. Lett.*, vol. 122, no. 3, pp. 215–222, 2001. DOI 10.1016/s0378-4274(01)00366-6. PMID 11489356. +[30] G. S. Ratra and J. E. Casida, "GABA receptor subunit composition relative to insecticide potency and selectivity," *Toxicol. Lett.*, vol. 122, no. 3, pp. 215–222, 2001. DOI 10.1016/s0378-4274(01)00366-6. PMID 11489356. -[32] G. S. Ratra, S. G. Kamita and J. E. Casida, "Role of human GABA_A receptor β3 subunit in insecticide toxicity," *Toxicol. Appl. Pharmacol.*, vol. 172, no. 3, pp. 233–240, 2001. DOI 10.1006/taap.2001.9154. PMID 11312652. +[31] G. S. Ratra, S. G. Kamita and J. E. Casida, "Role of human GABA_A receptor β3 subunit in insecticide toxicity," *Toxicol. Appl. Pharmacol.*, vol. 172, no. 3, pp. 233–240, 2001. DOI 10.1006/taap.2001.9154. PMID 11312652. -[33] Q. Shan, J. L. Haddrill and J. W. Lynch, "Ivermectin, an unconventional agonist of the glycine receptor chloride channel," *J. Biol. Chem.*, vol. 276, no. 16, pp. 12556–12564, 2001. DOI 10.1074/jbc.M011264200. +[32] Q. Shan, J. L. Haddrill and J. W. Lynch, "Ivermectin, an unconventional agonist of the glycine receptor chloride channel," *J. Biol. Chem.*, vol. 276, no. 16, pp. 12556–12564, 2001. DOI 10.1074/jbc.M011264200. -[34] K. Tasman, S. A. Rands and J. J. L. Hodge, "The power of *Drosophila melanogaster* for modeling neonicotinoid effects on pollinators and identifying novel mechanisms," *Front. Physiol.*, vol. 12, art. 659440, 2021. DOI 10.3389/fphys.2021.659440. +[33] K. Tasman, S. A. Rands and J. J. L. Hodge, "The power of *Drosophila melanogaster* for modeling neonicotinoid effects on pollinators and identifying novel mechanisms," *Front. Physiol.*, vol. 12, art. 659440, 2021. DOI 10.3389/fphys.2021.659440. -[35] M. Tomizawa and J. E. Casida, "Selective toxicity of neonicotinoids attributable to specificity of insect and mammalian nicotinic receptors," *Annu. Rev. Entomol.*, vol. 48, pp. 339–364, 2003. DOI 10.1146/annurev.ento.48.091801.112731. PMID 12208819. +[34] M. Tomizawa and J. E. Casida, "Selective toxicity of neonicotinoids attributable to specificity of insect and mammalian nicotinic receptors," *Annu. Rev. Entomol.*, vol. 48, pp. 339–364, 2003. DOI 10.1146/annurev.ento.48.091801.112731. PMID 12208819. -[36] M. Tomizawa and J. E. Casida, "Neonicotinoid insecticide toxicology: mechanisms of selective action," *Annu. Rev. Pharmacol. Toxicol.*, vol. 45, pp. 247–268, 2005. DOI 10.1146/annurev.pharmtox.45.120403.095930. +[35] M. Tomizawa and J. E. Casida, "Neonicotinoid insecticide toxicology: mechanisms of selective action," *Annu. Rev. Pharmacol. Toxicol.*, vol. 45, pp. 247–268, 2005. DOI 10.1146/annurev.pharmtox.45.120403.095930. -[37] M. Tomizawa, D. L. Lee and J. E. Casida, "Neonicotinoid insecticides: molecular features conferring selectivity for insect versus mammalian nicotinic receptors," *J. Agric. Food Chem.*, vol. 48, no. 12, pp. 6016–6024, 2000. DOI 10.1021/jf000873c. PMID 11312774. +[36] M. Tomizawa, D. L. Lee and J. E. Casida, "Neonicotinoid insecticides: molecular features conferring selectivity for insect versus mammalian nicotinic receptors," *J. Agric. Food Chem.*, vol. 48, no. 12, pp. 6016–6024, 2000. DOI 10.1021/jf000873c. PMID 11312774. -[38] A. J. Wolstenholme, "Glutamate-gated chloride channels," *J. Biol. Chem.*, vol. 287, no. 48, pp. 40232–40238, 2012. DOI 10.1074/jbc.R112.406280. +[37] A. J. Wolstenholme, "Glutamate-gated chloride channels," *J. Biol. Chem.*, vol. 287, no. 48, pp. 40232–40238, 2012. DOI 10.1074/jbc.R112.406280. -[39] A. M. Allen, M. C. Neville, S. Birtles *et al.*, "A single-cell transcriptomic atlas of the adult *Drosophila* ventral nerve cord," *eLife*, vol. 9, art. e54074, 2020. DOI 10.7554/eLife.54074. PMID 32314735. +[38] A. M. Allen, M. C. Neville, S. Birtles *et al.*, "A single-cell transcriptomic atlas of the adult *Drosophila* ventral nerve cord," *eLife*, vol. 9, art. e54074, 2020. DOI 10.7554/eLife.54074. PMID 32314735. -[40] V. Croset, C. D. Treiber and S. Waddell, "Cellular diversity in the *Drosophila* midbrain revealed by single-cell transcriptomics," *eLife*, vol. 7, art. e34550, 2018. DOI 10.7554/eLife.34550. PMID 29671739. +[39] V. Croset, C. D. Treiber and S. Waddell, "Cellular diversity in the *Drosophila* midbrain revealed by single-cell transcriptomics," *eLife*, vol. 7, art. e34550, 2018. DOI 10.7554/eLife.34550. PMID 29671739. -[41] K. Davie, J. Janssens, D. Koldere *et al.*, "A single-cell transcriptome atlas of the aging *Drosophila* brain," *Cell*, vol. 174, no. 4, pp. 982–998.e20, 2018. DOI 10.1016/j.cell.2018.05.057. PMID 29909982. +[40] K. Davie, J. Janssens, D. Koldere *et al.*, "A single-cell transcriptome atlas of the aging *Drosophila* brain," *Cell*, vol. 174, no. 4, pp. 982–998.e20, 2018. DOI 10.1016/j.cell.2018.05.057. PMID 29909982. -[42] L. Enell, Y. Hamasaka, A. Kolodziejczyk and D. R. Nässel, "γ-Aminobutyric acid (GABA) signaling components in *Drosophila*: immunocytochemical localization of GABA_B receptors in relation to the GABA_A receptor subunit RDL and a vesicular GABA transporter," *J. Comp. Neurol.*, vol. 505, no. 1, pp. 18–31, 2007. DOI 10.1002/cne.21472. PMID 17729251. +[41] L. Enell, Y. Hamasaka, A. Kolodziejczyk and D. R. Nässel, "γ-Aminobutyric acid (GABA) signaling components in *Drosophila*: immunocytochemical localization of GABA_B receptors in relation to the GABA_A receptor subunit RDL and a vesicular GABA transporter," *J. Comp. Neurol.*, vol. 505, no. 1, pp. 18–31, 2007. DOI 10.1002/cne.21472. PMID 17729251. -[43] N. Konstantinides, K. Kapuralin, C. Fadil *et al.*, "Phenotypic convergence: distinct transcription factors regulate common terminal features," *Cell*, vol. 174, no. 3, pp. 622–635.e13, 2018. DOI 10.1016/j.cell.2018.05.021. PMID 29909983. +[42] N. Konstantinides, K. Kapuralin, C. Fadil *et al.*, "Phenotypic convergence: distinct transcription factors regulate common terminal features," *Cell*, vol. 174, no. 3, pp. 622–635.e13, 2018. DOI 10.1016/j.cell.2018.05.021. PMID 29909983. -[44] H. Li, J. Janssens, M. De Waegeneer *et al.*, "Fly Cell Atlas: a single-nucleus transcriptomic atlas of the adult fruit fly," *Science*, vol. 375, no. 6584, art. eabk2432, 2022. DOI 10.1126/science.abk2432. PMID 35239393. +[43] H. Li, J. Janssens, M. De Waegeneer *et al.*, "Fly Cell Atlas: a single-nucleus transcriptomic atlas of the adult fruit fly," *Science*, vol. 375, no. 6584, art. eabk2432, 2022. DOI 10.1126/science.abk2432. PMID 35239393. -[45] C. I. Bliss, "The toxicity of poisons applied jointly," *Ann. Appl. Biol.*, vol. 26, no. 3, pp. 585–615, 1939. DOI 10.1111/j.1744-7348.1939.tb06990.x. +[44] C. I. Bliss, "The toxicity of poisons applied jointly," *Ann. Appl. Biol.*, vol. 26, no. 3, pp. 585–615, 1939. DOI 10.1111/j.1744-7348.1939.tb06990.x. -[46] W. R. Greco, G. Bravo and J. C. Parsons, "The search for synergy: a critical review from a response surface perspective," *Pharmacol. Rev.*, vol. 47, no. 2, pp. 331–385, 1995. PMID 7568331. +[45] W. R. Greco, G. Bravo and J. C. Parsons, "The search for synergy: a critical review from a response surface perspective," *Pharmacol. Rev.*, vol. 47, no. 2, pp. 331–385, 1995. PMID 7568331. -[47] Q. Huang, Y. Deng, T. Zhan and Y. He, "Synergistic and antagonistic effects of piperonyl butoxide in fipronil-susceptible and resistant rice stem borers, *Chilo suppressalis*," *J. Insect Sci.*, vol. 10, art. 182, 2010. DOI 10.1673/031.010.14142. PMID 21062143. +[46] Q. Huang, Y. Deng, T. Zhan and Y. He, "Synergistic and antagonistic effects of piperonyl butoxide in fipronil-susceptible and resistant rice stem borers, *Chilo suppressalis*," *J. Insect Sci.*, vol. 10, art. 182, 2010. DOI 10.1673/031.010.14142. PMID 21062143. -[48] S. Loewe and H. Muischnek, "Über Kombinationswirkungen: 1. Mitteilung: Hilfsmittel der Fragestellung," *Naunyn-Schmiedebergs Arch. Exp. Pathol. Pharmakol.*, vol. 114, pp. 313–326, 1926. DOI 10.1007/BF01952257. +[47] S. Loewe and H. Muischnek, "Über Kombinationswirkungen: 1. Mitteilung: Hilfsmittel der Fragestellung," *Naunyn-Schmiedebergs Arch. Exp. Pathol. Pharmakol.*, vol. 114, pp. 313–326, 1926. DOI 10.1007/BF01952257. -[49] S. Loewe, "The problem of synergism and antagonism of combined drugs," *Arzneimittelforschung*, vol. 3, no. 6, pp. 285–290, 1953. PMID 13081480. +[48] S. Loewe, "The problem of synergism and antagonism of combined drugs," *Arzneimittelforschung*, vol. 3, no. 6, pp. 285–290, 1953. PMID 13081480. -[50] D. Philippou, V. Borzatta, E. Capparella *et al.*, "The use of substituted alkynyl phenoxy derivatives of piperonyl butoxide to control insecticide-resistant pests," *Pest Manag. Sci.*, vol. 72, no. 10, pp. 1946–1950, 2016. DOI 10.1002/ps.4234. PMID 26800141. +[49] D. Philippou, V. Borzatta, E. Capparella *et al.*, "The use of substituted alkynyl phenoxy derivatives of piperonyl butoxide to control insecticide-resistant pests," *Pest Manag. Sci.*, vol. 72, no. 10, pp. 1946–1950, 2016. DOI 10.1002/ps.4234. PMID 26800141. -[51] Y. C. Zhu, J. Yao, J. Adamczyk and R. Luttrell, "Synergistic toxicity and physiological impact of imidacloprid alone and binary mixtures with seven representative pesticides on honey bee (*Apis mellifera*)," *PLoS ONE*, vol. 12, no. 5, art. e0176837, 2017. DOI 10.1371/journal.pone.0176837. PMID 28467462. +[50] Y. C. Zhu, J. Yao, J. Adamczyk and R. Luttrell, "Synergistic toxicity and physiological impact of imidacloprid alone and binary mixtures with seven representative pesticides on honey bee (*Apis mellifera*)," *PLoS ONE*, vol. 12, no. 5, art. e0176837, 2017. DOI 10.1371/journal.pone.0176837. PMID 28467462. -[52] P. J. Daborn, J. L. Yen, M. R. Bogwitz *et al.*, "A single P450 allele associated with insecticide resistance in *Drosophila*," *Science*, vol. 297, no. 5590, pp. 2253–2256, 2002. DOI 10.1126/science.1074170. PMID 12351787. +[51] P. J. Daborn, J. L. Yen, M. R. Bogwitz *et al.*, "A single P450 allele associated with insecticide resistance in *Drosophila*," *Science*, vol. 297, no. 5590, pp. 2253–2256, 2002. DOI 10.1126/science.1074170. PMID 12351787. -[53] D. G. Folk, C. Han and T. J. Bradley, "Water acquisition and partitioning in *Drosophila melanogaster*: effects of selection for desiccation-resistance," *J. Exp. Biol.*, vol. 204, no. 19, pp. 3323–3331, 2001. DOI 10.1242/jeb.204.19.3323. PMID 11606606. +[52] D. G. Folk, C. Han and T. J. Bradley, "Water acquisition and partitioning in *Drosophila melanogaster*: effects of selection for desiccation-resistance," *J. Exp. Biol.*, vol. 204, no. 19, pp. 3323–3331, 2001. DOI 10.1242/jeb.204.19.3323. PMID 11606606. -[54] R. Fusetto, S. Denecke, T. Perry, R. A. J. O'Hair and P. Batterham, "Partitioning the roles of CYP6G1 and gut microbes in the metabolism of the insecticide imidacloprid in *Drosophila melanogaster*," *Sci. Rep.*, vol. 7, art. 11339, 2017. DOI 10.1038/s41598-017-09800-2. PMID 28900131. +[53] R. Fusetto, S. Denecke, T. Perry, R. A. J. O'Hair and P. Batterham, "Partitioning the roles of CYP6G1 and gut microbes in the metabolism of the insecticide imidacloprid in *Drosophila melanogaster*," *Sci. Rep.*, vol. 7, art. 11339, 2017. DOI 10.1038/s41598-017-09800-2. PMID 28900131. -[55] N. Joußen, D. G. Heckel, M. Haas, I. Schuphan and B. Schmidt, "Metabolism of imidacloprid and DDT by P450 CYP6G1 expressed in cell cultures of *Nicotiana tabacum*," *Pest Manag. Sci.*, vol. 64, no. 1, pp. 65–73, 2008. DOI 10.1002/ps.1472. PMID 17912692. +[54] N. Joußen, D. G. Heckel, M. Haas, I. Schuphan and B. Schmidt, "Metabolism of imidacloprid and DDT by P450 CYP6G1 expressed in cell cultures of *Nicotiana tabacum*," *Pest Manag. Sci.*, vol. 64, no. 1, pp. 65–73, 2008. DOI 10.1002/ps.1472. PMID 17912692. -[56] S. C. Piyankarage, H. Augustin, Y. Grosjean, D. E. Featherstone and S. A. Shippy, "Hemolymph amino acid analysis of individual *Drosophila* larvae," *Anal. Chem.*, vol. 80, no. 4, pp. 1201–1207, 2008. DOI 10.1021/ac701785z. +[55] S. C. Piyankarage, H. Augustin, Y. Grosjean, D. E. Featherstone and S. A. Shippy, "Hemolymph amino acid analysis of individual *Drosophila* larvae," *Anal. Chem.*, vol. 80, no. 4, pp. 1201–1207, 2008. DOI 10.1021/ac701785z. -[57] S. C. Piyankarage, D. E. Featherstone and S. A. Shippy, "Nanoliter hemolymph sampling and analysis of individual adult *Drosophila melanogaster*," *Anal. Chem.*, vol. 84, no. 10, pp. 4460–4466, 2012. DOI 10.1021/ac3002319. +[56] S. C. Piyankarage, D. E. Featherstone and S. A. Shippy, "Nanoliter hemolymph sampling and analysis of individual adult *Drosophila melanogaster*," *Anal. Chem.*, vol. 84, no. 10, pp. 4460–4466, 2012. DOI 10.1021/ac3002319. -[58] S. Suchail, L. Debrauwer and L. P. Belzunces, "Metabolism of imidacloprid in *Apis mellifera*," *Pest Manag. Sci.*, vol. 60, no. 3, pp. 291–296, 2004. DOI 10.1002/ps.772. PMID 15025241. +[57] S. Suchail, L. Debrauwer and L. P. Belzunces, "Metabolism of imidacloprid in *Apis mellifera*," *Pest Manag. Sci.*, vol. 60, no. 3, pp. 291–296, 2004. DOI 10.1002/ps.772. PMID 15025241. -[59] S. Suchail, G. De Sousa, R. Rahmani and L. P. Belzunces, "In vivo distribution and metabolisation of ¹⁴C-imidacloprid in different compartments of *Apis mellifera* L.," *Pest Manag. Sci.*, vol. 60, no. 11, pp. 1056–1062, 2004. DOI 10.1002/ps.895. +[58] S. Suchail, G. De Sousa, R. Rahmani and L. P. Belzunces, "In vivo distribution and metabolisation of ¹⁴C-imidacloprid in different compartments of *Apis mellifera* L.," *Pest Manag. Sci.*, vol. 60, no. 11, pp. 1056–1062, 2004. DOI 10.1002/ps.895. -[60] W. Cao, L. Song, J. Cheng *et al.*, "An automated rapid iterative negative geotaxis assay for analyzing adult climbing behavior in a *Drosophila* model of neurodegeneration," *J. Vis. Exp.*, no. 127, art. 56507, 2017. DOI 10.3791/56507. PMID 28931001. +[59] W. Cao, L. Song, J. Cheng *et al.*, "An automated rapid iterative negative geotaxis assay for analyzing adult climbing behavior in a *Drosophila* model of neurodegeneration," *J. Vis. Exp.*, no. 127, art. 56507, 2017. DOI 10.3791/56507. PMID 28931001. -[61] M. Ö. Çevik and A. Erden, "The course of habituation of the proboscis extension reflex can be predicted by sucrose responsiveness in *Drosophila*," *PLoS ONE*, vol. 7, no. 6, art. e39863, 2012. DOI 10.1371/journal.pone.0039863. PMID 22761915. +[60] M. Ö. Çevik and A. Erden, "The course of habituation of the proboscis extension reflex can be predicted by sucrose responsiveness in *Drosophila*," *PLoS ONE*, vol. 7, no. 6, art. e39863, 2012. DOI 10.1371/journal.pone.0039863. PMID 22761915. -[62] W. Chen, X. Gu, Y. T. Yang, P. Batterham and T. Perry, "Dual nicotinic acetylcholine receptor subunit gene knockouts reveal limits to functional redundancy," *Pestic. Biochem. Physiol.*, vol. 184, art. 105118, 2022. DOI 10.1016/j.pestbp.2022.105118. PMID 35715057. +[61] W. Chen, X. Gu, Y. T. Yang, P. Batterham and T. Perry, "Dual nicotinic acetylcholine receptor subunit gene knockouts reveal limits to functional redundancy," *Pestic. Biochem. Physiol.*, vol. 184, art. 105118, 2022. DOI 10.1016/j.pestbp.2022.105118. PMID 35715057. -[63] A. S. French, M. J. Sellier, M. Ali Agha *et al.*, "Dual mechanism for bitter avoidance in *Drosophila*," *J. Neurosci.*, vol. 35, no. 9, pp. 3990–4004, 2015. DOI 10.1523/JNEUROSCI.1312-14.2015. PMID 25740527. +[62] A. S. French, M. J. Sellier, M. Ali Agha *et al.*, "Dual mechanism for bitter avoidance in *Drosophila*," *J. Neurosci.*, vol. 35, no. 9, pp. 3990–4004, 2015. DOI 10.1523/JNEUROSCI.1312-14.2015. PMID 25740527. -[64] J. W. Gargano, I. Martin, P. Bhandari and M. S. Grotewiel, "Rapid iterative negative geotaxis (RING): a new method for assessing age-related locomotor decline in *Drosophila*," *Exp. Gerontol.*, vol. 40, no. 5, pp. 386–395, 2005. DOI 10.1016/j.exger.2005.02.005. PMID 15919590. +[63] J. W. Gargano, I. Martin, P. Bhandari and M. S. Grotewiel, "Rapid iterative negative geotaxis (RING): a new method for assessing age-related locomotor decline in *Drosophila*," *Exp. Gerontol.*, vol. 40, no. 5, pp. 386–395, 2005. DOI 10.1016/j.exger.2005.02.005. PMID 15919590. -[65] H. Y. Lee, L. Zhou, S. Ghanta *et al.*, "Mechanisms of naturally evolved ethanol resistance in *Drosophila melanogaster*," *J. Exp. Biol.*, 2014. DOI 10.1242/jeb.110510. +[64] H. Y. Lee, L. Zhou, S. Ghanta *et al.*, "Mechanisms of naturally evolved ethanol resistance in *Drosophila melanogaster*," *J. Exp. Biol.*, 2014. DOI 10.1242/jeb.110510. -[66] F. Martelli, Z. Zhongyuan, J. Wang *et al.*, "Low doses of the neonicotinoid insecticide imidacloprid induce ROS triggering neurological and metabolic impairments in *Drosophila*," *Proc. Natl. Acad. Sci. USA*, vol. 117, no. 41, pp. 25840–25850, 2020. DOI 10.1073/pnas.2011828117. PMID 32989137. +[65] F. Martelli, Z. Zhongyuan, J. Wang *et al.*, "Low doses of the neonicotinoid insecticide imidacloprid induce ROS triggering neurological and metabolic impairments in *Drosophila*," *Proc. Natl. Acad. Sci. USA*, vol. 117, no. 41, pp. 25840–25850, 2020. DOI 10.1073/pnas.2011828117. PMID 32989137. -[67] M. J. Burton, I. R. Mellor, I. R. Duce *et al.*, "Differential resistance of insect sodium channels with kdr mutations to deltamethrin, permethrin and DDT," *Insect Biochem. Mol. Biol.*, vol. 41, no. 9, pp. 723–732, 2011. DOI 10.1016/j.ibmb.2011.05.004. +[66] M. J. Burton, I. R. Mellor, I. R. Duce *et al.*, "Differential resistance of insect sodium channels with kdr mutations to deltamethrin, permethrin and DDT," *Insect Biochem. Mol. Biol.*, vol. 41, no. 9, pp. 723–732, 2011. DOI 10.1016/j.ibmb.2011.05.004. -[68] A. A. Lazar, T. Liu, M. K. Turkcan and Y. Zhou, "Accelerating with FlyBrainLab the discovery of the functional logic of the *Drosophila* brain in the connectomic and synaptomic era," *eLife*, vol. 10, art. e62362, 2021. DOI 10.7554/eLife.62362. PMID 33616035. +[67] A. A. Lazar, T. Liu, M. K. Turkcan and Y. Zhou, "Accelerating with FlyBrainLab the discovery of the functional logic of the *Drosophila* brain in the connectomic and synaptomic era," *eLife*, vol. 10, art. e62362, 2021. DOI 10.7554/eLife.62362. PMID 33616035. -[69] I. Mindlin, R. Herzog, L. Belloli *et al.*, "Whole brain modelling for simulating pharmacological interventions on patients with disorders of consciousness," *Commun. Biol.*, vol. 7, art. 1176, 2024. DOI 10.1038/s42003-024-06852-9. PMID 39300281. +[68] I. Mindlin, R. Herzog, L. Belloli *et al.*, "Whole brain modelling for simulating pharmacological interventions on patients with disorders of consciousness," *Commun. Biol.*, vol. 7, art. 1176, 2024. DOI 10.1038/s42003-024-06852-9. PMID 39300281. -[70] G. Deco, J. Cruzat, J. Cabral *et al.*, "Whole-brain multimodal neuroimaging model using serotonin receptor maps explains non-linear functional effects of LSD," *Curr. Biol.*, vol. 28, no. 19, pp. 3065–3074.e6, 2018. DOI 10.1016/j.cub.2018.07.083. PMID 30270185. +[69] G. Deco, J. Cruzat, J. Cabral *et al.*, "Whole-brain multimodal neuroimaging model using serotonin receptor maps explains non-linear functional effects of LSD," *Curr. Biol.*, vol. 28, no. 19, pp. 3065–3074.e6, 2018. DOI 10.1016/j.cub.2018.07.083. PMID 30270185. -[71] A. Saltelli, P. Annoni, I. Azzini, F. Campolongo, M. Ratto and S. Tarantola, "Variance based sensitivity analysis of model output: design and estimator for the total sensitivity index," *Comput. Phys. Commun.*, vol. 181, no. 2, pp. 259–270, 2010. DOI 10.1016/j.cpc.2009.09.018. +[70] A. Saltelli, P. Annoni, I. Azzini, F. Campolongo, M. Ratto and S. Tarantola, "Variance based sensitivity analysis of model output: design and estimator for the total sensitivity index," *Comput. Phys. Commun.*, vol. 181, no. 2, pp. 259–270, 2010. DOI 10.1016/j.cpc.2009.09.018. -[72] M. J. W. Jansen, "Analysis of variance designs for model output," *Comput. Phys. Commun.*, vol. 117, no. 1–2, pp. 35–43, 1999. DOI 10.1016/S0010-4655(98)00154-4. +[71] M. J. W. Jansen, "Analysis of variance designs for model output," *Comput. Phys. Commun.*, vol. 117, no. 1–2, pp. 35–43, 1999. DOI 10.1016/S0010-4655(98)00154-4. -[73] T. Ishigami and T. Homma, "An importance quantification technique in uncertainty analysis for computer models," in *Proc. 1st Int. Symp. Uncertainty Modeling and Analysis (ISUMA)*, 1990, pp. 398–403. DOI 10.1109/ISUMA.1990.151285. +[72] T. Ishigami and T. Homma, "An importance quantification technique in uncertainty analysis for computer models," in *Proc. 1st Int. Symp. Uncertainty Modeling and Analysis (ISUMA)*, 1990, pp. 398–403. DOI 10.1109/ISUMA.1990.151285. -[74] S. Steegen, F. Tuerlinckx, A. Gelman and W. Vanpaemel, "Increasing transparency through a multiverse analysis," *Perspect. Psychol. Sci.*, vol. 11, no. 5, pp. 702–712, 2016. DOI 10.1177/1745691616658637. PMID 27694465. +[73] S. Steegen, F. Tuerlinckx, A. Gelman and W. Vanpaemel, "Increasing transparency through a multiverse analysis," *Perspect. Psychol. Sci.*, vol. 11, no. 5, pp. 702–712, 2016. DOI 10.1177/1745691616658637. PMID 27694465. -[75] U. Simonsohn, J. P. Simmons and L. D. Nelson, "Specification curve analysis," *Nat. Hum. Behav.*, vol. 4, pp. 1208–1214, 2020. DOI 10.1038/s41562-020-0912-z. PMID 32719546. +[74] U. Simonsohn, J. P. Simmons and L. D. Nelson, "Specification curve analysis," *Nat. Hum. Behav.*, vol. 4, pp. 1208–1214, 2020. DOI 10.1038/s41562-020-0912-z. PMID 32719546. -[76] R. A. Howard, "Information value theory," *IEEE Trans. Syst. Sci. Cybern.*, vol. 2, no. 1, pp. 22–26, 1966. DOI 10.1109/TSSC.1966.300074. +[75] R. A. Howard, "Information value theory," *IEEE Trans. Syst. Sci. Cybern.*, vol. 2, no. 1, pp. 22–26, 1966. DOI 10.1109/TSSC.1966.300074. -[77] M. Tomizawa, B. Latli and J. E. Casida, "Novel neonicotinoid-agarose affinity column for *Drosophila* and *Musca* nicotinic acetylcholine receptors," *J. Neurochem.*, vol. 67, no. 4, pp. 1669–1676, 1996. DOI 10.1046/j.1471-4159.1996.67041669.x. PMID 8858952. +[76] M. Tomizawa, B. Latli and J. E. Casida, "Novel neonicotinoid-agarose affinity column for *Drosophila* and *Musca* nicotinic acetylcholine receptors," *J. Neurochem.*, vol. 67, no. 4, pp. 1669–1676, 1996. DOI 10.1046/j.1471-4159.1996.67041669.x. PMID 8858952. --- diff --git a/papers/STATUS.md b/papers/STATUS.md index 4039b52..16396fd 100644 --- a/papers/STATUS.md +++ b/papers/STATUS.md @@ -3,12 +3,16 @@ **Files:** `papers/IJRC_FlyLab_draft.md` and `papers/SUPPLEMENT.md` — **both generated**. Edit `papers/IJRC_FlyLab_draft.md.in` / `papers/SUPPLEMENT.md.in` and re-render; never hand-edit a number into a rendered document. **Venue:** IJRC (ijrcom.org), as a **Research Article** (Introduction, Methods, Results, Discussion, Conclusion, References). -**State:** v0.6.1 revision complete, responding to the **second-round** referee report (`papers/reviews/round2_referee_report.md`). Not submitted. Point-by-point response, by the report's own item numbers: `papers/SUBMISSION_CHECKLIST.md` §R. +**State:** additional internal-review and readability revision of main `38cad67`. Not submitted. Review decisions: `papers/reviews/current_revision_decisions.md` and `papers/reviews/round4_editorial_decisions.md`. Author details are supplied; an editable Word preparation copy is in `papers/submission/`. -**The paper is framed for a computing venue, and the central empirical claim reversed in this round.** IJRC is a computing journal, so the manuscript leads with the general problem — a network simulation cannot say which of its inputs a prediction depends on — and treats the fly pharmacology as the domain the method is demonstrated on. The headline is methodological: a dependence verdict obtained on one extract of a network can invert on another extract of the same network, and the extract's **recurrence, not its size**, predicts which way. "Most predictions do not need the connectome" and "exactly the chloride-channel blockers" are withdrawn. +**Current internal-review update:** the denser-cut study, synthetic recovery cases, corrected evidence propagation and specification fixes are retained. The current revision narrows the remaining claims: extract sensitivity is observed, recurrence is not isolated causally, the legacy equivalence label is only a median-gap diagnostic, and the small synthetic detection grid does not establish endpoint-specific power. Review reports and decisions are in `papers/reviews/`. + +**This review's verification:** the existing default-effort numerical record is retained. The attempted full rerun was interrupted during dependence analysis. `scripts/paper_review_metadata.py` refreshes interpretation and captions while asserting that numerical value payloads and table CSV bytes are unchanged; `results.json.review_metadata` distinguishes this replay from a completed rerun. See `papers/reviews/current_verification.md` for completed checks. + +**Reproduction boundary:** the standard pipeline recomputes core committed-cut analyses but imports the previously computed `papers/scale_study.json`. Only `scale_1k` is committed; large-rung results require upstream cuts to recompute and the saved study lacks source/input hashes. Re-rendering its table is not an independent rerun. ```bash -python scripts/reproduce_paper.py # regenerate every figure, table and number (~35-45 min) +python scripts/reproduce_paper.py # recompute core analyses and replay the cached scale table python scripts/reproduce_paper.py --only paper # re-render both documents from results.json python -m pytest -q # must be green ``` @@ -18,47 +22,46 @@ python -m pytest -q # must be green | Part | State | |---|---| | Abstract (architecture + one topology result + one reproducibility result + the central limitation; individual numbers stripped) | done | -| §1 Introduction — the four research questions, related work, and the narrow novelty claim with a comparison table | done | +| §1 Introduction: three research questions, related work, and the narrow novelty claim with a comparison table | done; reproduction is supporting evidence rather than RQ4 | | §2 Methods — typed evidence, the circuit substrate, connectome-dependence analysis, the ablation ladder and specification family, the evaluation taxonomy, global uncertainty and VOI | done (6 subsections) | | §3 Results — the instruments evaluated first (ground truth and power, then the inversion and the scale ladder), with the domain readouts as supporting material; every number from `results.json` | done (7 subsections) | | §4 **Discussion** — what was learned computationally (including the substrate dependence of the method's own verdict, and two instruments that were broken in ways that flattered the conclusions); model-generated vs encoded; adjacent work; **what the software cannot support**; five model-specific caveats, at length in §S9; prospective-not-pre-registered | done | | §5 Conclusion | done | | Statements — precise AI declaration, animal research, data/code availability, competing interests, funding | done | -| References — 77 entries with DOIs/PMIDs, including FlyBrainLab [68], the receptor-map whole-brain models [69], [70] and the *Drosophila* binding constant [77] | done | +| References — 76 entries with DOIs/PMIDs, including FlyBrainLab [67], the receptor-map whole-brain models [68], [69] and the *Drosophila* binding constant [76] | done | | Supplement — mechanism rationale and the level-C rule (S1), the specification family and why the previous matrix was vacuous (S1.1), null-model definitions with the weight-matched transmitter null (S2), the three-way verdict and the equivalence margin (S2.1), effect floor and multiplicity (S2.2), ground-truth recovery and power (S2.3), the scale ladder and where the verdict settles (S2.4), runtimes and the row normalisation (S3), supporting capabilities (S4), prospective predictions with the H2/H3 caveat (S5), evidence-typing rules and the soundness invariant (S6), reproduction knobs (S7), notebook warnings (S8), model-specific caveats in full (S9), and the domain readouts, selectivity and literature concordance moved out of the body (S10) | done | | 16 figures (F16 is the instrument validation) + T0–T25, regenerated at 300 dpi | done | -**Body word count is generated, not estimated**, and a test fails if it leaves the 5000–7000 target: `papers/results.json → values.paper_words_body`. +**Body word count is generated, not estimated**: `papers/results.json → values.paper_words_body`. The editorial test guard is 5000–8000. The concise revision is below the journal's typical Research Articles 6000–8000 range but within its general 5000–10000 guidance. Confirm the length rather than pad the text. Final format and blind-review packaging remain author checks. -## True in the repo today (all reproducible with committed artifacts) +## Implemented capabilities and committed artifacts -- Typed evidence: 102 rows / 21 compounds / 15 receptor keys; 34 EC50, 10 IC50, **2 Kd**, 56 not modelled and returning N/A, enforced by the type system and by a soundness invariant checked over every row and every entry point. Evidence distance E0–E4: 10 / 15 / 10 / 11 / 56. +- Typed evidence: 102 rows / 21 compounds / 15 receptor keys; 34 EC50, 10 IC50, **2 Kd**, 56 not modelled and returning N/A, enforced by a tagged-record dispatch discipline and regression checks across all shipped rows and enumerated pipeline paths. Evidence distance E0–E4: 10 / 15 / 10 / 11 / 56. - Two committed MaleCNS v1.0 cuts (1126/1360 and 1841/19 066) plus a 165 122-cell census fallback; all 16 labellar `LB*` types proven present. -- Connectome-dependence analysis at n = 1000 with three-way equivalence verdicts, a weight-matched transmitter null, a permutation-count sweep, ground-truth validation of the ladder, and an FDR-corrected landscape over 21 compounds × 4 concentrations **on both committed cuts**. +- Connectome-dependence analysis at n = 1000 with three-way descriptive tolerance diagnostics, a weight-matched transmitter null, a permutation-count sweep, synthetic positive-control checks of the ladder, and a landscape classified using BH-adjusted component tests over 21 compounds × 4 concentrations **on both committed cuts**. - Ablation ladder, conclusion-stability matrix over 25 gain specifications, a 3×3 threshold grid, a Sobol' budget (11 264 evaluations) with Ishigami validation, and a VOI ranking. - Map-extracted GRN → MN9 path with a complete bitter veto in vehicle and partial relief under fipronil, on both engines. -- Everything above regenerated by one command in about 35–45 minutes with no downloads, and a test asserts the committed record was produced at the shipped statistical effort. - -## Honest negatives the draft reports as results +- Core committed-cut analyses regenerated by the standard command; scale profiles are imported from the saved study, and a test asserts the committed record was produced at the shipped statistical effort. -These are in the manuscript on purpose and must not be removed to make it look better: +## Interpretation limits -- **The central RQ2 claim reversed, and the scaling study says what it tracks.** On the in-star `named` cut the landscape gives 11 topology-dependent / 42 composition-dominated / 31 no-effect of 84; repeated unchanged on `taste_motor` it gives 54 / 0 / 30. Across a ladder of 7 extracts (`scale_1k` … `scale_50k` plus the two committed cuts) the composition-dominated verdict survives on **exactly one** — the in-star — and `scale_1k` has *fewer* nodes (1000 vs 1126) with mean degree 22.9 vs 1.21 and is already topology-dependent. **Recurrence, not size, predicts the inversion.** At the denser middle rungs (5k, 10k) imidacloprid is distinguishable from *every* null. The top rungs run at n = 20 (resolution 0.048), so they confirm rather than establish the settled verdict. -- Imidacloprid's `named` mean-rate effect is **composition-dominated** at n = 1000 — equivalent within tolerance for rewire_degree_preserving, sign_permute_weight_matched, weight_permute and merely indeterminate for sign_permute against a margin of 0.332 Hz. Fipronil is topology-dependent there (p = 0.006) but does **not** survive correction inside the exploratory landscape (q = 0.050). -- The fipronil bitter-veto arm is a **properly powered negative** at n = 300 (all p > 0.05, |z| < 0.25). -- **The suppression conclusion is specification-dependent**: retained by 15/25 specifications, reversed by all 10 monotone ones. And the previous 25/25 topology result was **vacuous** — 6 shuffles could not reject, and the specification never reached the null engine. -- The composition-only baseline tracks the full model's ordering (ρ 0.988 at 1 µM), but pharmacology-free pseudo-compounds reach 0.847 and a degree-corrected engine gives 0.666; the direction-aware topology-only level reaches 0.961, against 0.242 for the sign-broken rule whose conclusion is withdrawn. -- The amplify/buffer split is **not** stable across the threshold grid. -- At a saturating dose the model is blind to potency and Hill *n*; only gain_transform, weight_threshold have a first-order interval excluding zero, and the estimator's noise floor is 0.046 (set by gain_coef, not by the null factor). -- There is **no independent out-of-sample validation** of the circuit model, and the rank comparisons that share a source with the library are labelled as such. -- The rate and LIF engines agree on **direction only** (≈0.29 Hz vs ≈105 Hz). +- Dependence labels change between extracts. Current evidence does not select recurrence over density, extraction recipe or correlated structural properties. +- `equivalent_within_tolerance` is a legacy name for non-rejection plus a small point gap from the null median, not a confidence-bounded equivalence test. +- BH-adjusted values are reported; shared shuffles alone do not prove the required dependence assumptions. +- The taste-motor veto-ratio non-rejection has no endpoint-specific power analysis. +- The six-network-per-setting synthetic grid is an initial recovery check with imprecise detection rates. +- Top scale rungs use 20 permutations and report probabilities at the resolution floor. Matching labels are not proof of convergence. +- High ablation correlation depends on shared assumptions; its observed value also exceeds the reported reference 95th percentile, so the reference does not explain it away entirely. +- Nicotinic suppression reverses under monotone gain rules. +- Rate-model units and LIF spikes/s are not interchangeable measured firing rates. +- No independent circuit-level or live-animal validation exists. ## Still missing for a **results** paper 1. One live PER or climbing table in `live_lab` (climbing first — the control statistics exist; the PER baseline does not). 2. The one allowed fit — and the uncertainty budget now names it without argument: the engagement→gain transformation, not a potency value. 3. Expression coverage above 0.205, with adult motor neurons resolved. -4. An externally archived, timestamped release, without which the predictions stay *prospective* rather than pre-registered. +4. A source-pinned archival release and, for preregistration, a frozen experimental protocol dated before data collection. ## Submit when diff --git a/papers/SUBMISSION_CHECKLIST.md b/papers/SUBMISSION_CHECKLIST.md index 45bf195..ec96f21 100644 --- a/papers/SUBMISSION_CHECKLIST.md +++ b/papers/SUBMISSION_CHECKLIST.md @@ -12,9 +12,9 @@ python -m pytest -q # must be green ## Framing -IJRC is a computing research journal and the manuscript is written as one. The lead is the general problem — *a simulation built on a measured network and literature parameters cannot say which of its inputs its predictions depend on* — and the contribution is four transferable instruments: typed evidence propagation under a checked soundness invariant; input-dependence testing against an ordered ladder with an equivalence margin and multiplicity control; an ablation ladder scored against a matched reference; and specification-family robustness with variance attribution. §3 evaluates **the instruments**. The fly pharmacology is the demonstration domain, with its census, readouts and literature concordance as supporting material in §S6 and §S10. +IJRC is a computing research journal and the manuscript is written as one. The lead is the general problem — *a simulation built on a measured network and literature parameters cannot say which of its inputs its predictions depend on* — and the contribution is four transferable instruments: typed evidence propagation under a checked soundness invariant; input-dependence testing against an ordered ladder with a descriptive median-gap tolerance and multiplicity adjustment; an ablation ladder scored against a matched reference; and specification-family robustness with variance attribution. §3 evaluates **the instruments**. The fly pharmacology is the demonstration domain, with its census, readouts and literature concordance as supporting material in §S6 and §S10. -The headline is methodological: **a dependence verdict obtained on one extract of a network can invert on another extract of the same network, and the extract's recurrence — not its size — predicts which way** (§3.2, T27). +The headline is extract sensitivity: dependence labels change across extracts. The current evidence does not isolate recurrence as the cause. A point-gap tolerance is descriptive, not a formal equivalence test. Current review decisions are in `papers/reviews/current_revision_decisions.md`. Internal review history lives in `papers/reviews/`. It is archived critique, not journal correspondence: the corrections it forced are in the manuscript and the code, and a formal point-by-point response belongs to an actual resubmission against real referees' numbered comments. @@ -22,19 +22,19 @@ Internal review history lives in `papers/reviews/`. It is archived critique, not | Item | State | Where | |---|---|---| -| Title, author block, affiliations | **missing** — authors are "FlyLab contributors" | needs real names, ORCIDs, affiliations, a corresponding author and an email | +| Title, author block, affiliations | supplied by author on 22 September 2026 | Michél Nguyen; University of the People; michel_ng@icloud.com; ORCID 0000-0001-6834-4422 | | Abstract | done — computing problem first, the methodological headline, the central limitation | draft §Abstract | | Index terms / keywords | done | draft §Index terms | -| Numbered IEEE-style references with DOIs | done — 77 entries | draft §References | +| Numbered IEEE-style references with DOIs | 76 entries after source audit; final author verification pending | draft §References | | In-text citation of every reference | check with the grep in §D | — | | Figures, 300 dpi, captions | done — 16 figures (F16 is the instrument validation) | `papers/figures/`, `figures/captions.md` | | Tables | done — T0–T27 as CSV and Markdown | `papers/tables/` | -| Word count | done — inside the 5000–7000 target, generated and test-enforced | `values.paper_words_body` | +| Word count | concise draft below typical Research Articles 6000–8000, within general 5000–10000; confirm with editor | `values.paper_words_body`; editorial test guard 5000–8000 | | Required section structure | done | draft §1–§5 | | Supplementary material | done — S1–S10 | `papers/SUPPLEMENT.md` | | Data / code / ethics / AI statements | done | draft §Statements | | Archival DOI | **placeholder** — Zenodo deposit not created | draft §3.7, `CITATION.cff` | -| Submission format (LaTeX/Word template) | **not done** — the draft is Markdown | convert once the venue's template is confirmed | +| Submission format (LaTeX/Word template) | editable Word preparation copy; final template/citation check pending | `papers/submission/FlyLab_manuscript.docx` and preparation notes | | Cover letter | **not written** | — | | Suggested reviewers | **not chosen** | — | @@ -42,20 +42,18 @@ Internal review history lives in `papers/reviews/`. It is archived critique, not 1. **One live *Drosophila* assay.** Climbing first: a complete published protocol and control statistics exist. Until a real table is imported by hand into `live_lab`, H1–H7 stay prospective software predictions. 2. **The one fitted parameter.** The variance budget names it: the engagement→gain transformation, not a potency value, which the model cannot see at a saturating dose. -3. **Better power at the top of the scale ladder.** The ladder settles the verdict, but its 25k/50k rungs run at n = 20 (resolution 0.048), where a rejection is the smallest the test can express. More permutations there, or a cheaper null, would turn confirmation into evidence. +3. **Better power at the top of the scale ladder.** The ladder reports matching labels, but its 25k/50k rungs run at n = 20 (resolution 0.048), where a rejection is the smallest the test can express. More permutations there, or a cheaper null, would provide finer-resolution estimates. 4. **Expression coverage above its current value**, with adult motor-neuron receptor expression sourced or the bound declared permanent. -5. **An externally archived release**, without which "prospective" may not become "pre-registered". +5. **An externally archived release**, plus a frozen experimental protocol before using the term "pre-registered". -## C. Known weaknesses a reviewer will find first +## C. Remaining methodological limits -1. **The central RQ2 result reverses between the two cuts.** Answer: that *is* the result, it is measured rather than argued, the instrument is validated against planted ground truth (F16, T22) so the reversal is not an artefact of power, the structural difference is reported in advance (T19), and the scaling study over seven extracts (T27) identifies recurrence rather than size as what predicts it. -2. **The top rungs of the scale ladder run at n = 20.** Answer: stated in §3.2 and §S2.4 with the resolution (0.048 against α = 0.05) — those rungs confirm a verdict already settled at better-powered rungs, and the paper says they do not establish it. -3. **The composition-versus-full correlation is nearly an algebraic identity.** Answer: stated as such, with a matched reference distribution (T24) and a normalisation sweep (T25); the generalisation drawn from it is withdrawn. -4. **The headline circuit result's direction depends on the gain rule.** Answer: a first-class result (§3.4), quantified over the prespecified specification family, and nominated by the variance budget as the measurement to make next. -5. **The gain rules and five transmitter signs are asserted coefficients.** Answer: published (T2, T21) and named as such; the paper says which of its own claims that costs it. -6. **No independent out-of-sample validation.** Answer: stated in §2.5 and §S10, with shared-source rank comparisons flagged by the pipeline. The ground-truth experiment validates the instrument, not the domain model, and the paper says so. -7. **Two of the instruments were broken.** Answer: found by us, published in §3.4, §S1.1 and §S2, and both fixed with the corrected numbers in the text. Neither failure was detectable by a determinism check, which is itself reported as a finding. -8. **`--fast` is not the paper.** Answer: one test refuses a committed record from a fast run, another asserts the record's statistical knobs equal the shipped defaults. +1. A controlled intervention or better matched extract comparison is needed to isolate recurrence from other structural differences. +2. A confidence-bounded equivalence analysis is needed before interpreting a within-margin median as statistical equivalence. +3. The small synthetic grid does not establish general calibration or power for the taste-motor ratio endpoint. +4. The BH dependence condition has not been established by merely reusing shuffle streams. +5. Large scale profiles are cached results with missing cut/source hashes. Archive those inputs and rerun with adequate permutation effort before calling them independently reproduced. +6. Biological validation, receptor expression and gain calibration remain unresolved. ## D. Pre-submission mechanical checks @@ -100,7 +98,7 @@ python -m pytest -q tests/test_reproduce.py ## E. Final pass before sending - [ ] Real authors, affiliations, ORCIDs, corresponding author. -- [ ] Zenodo deposit created; DOI substituted for the placeholder in the draft and in `CITATION.cff`; only then may the predictions be called pre-registered. +- [ ] Zenodo deposit created; DOI substituted for the placeholder in the draft and in `CITATION.cff`; preregistration additionally requires a frozen, dated experimental protocol before collection. - [ ] Tag the release; confirm `git_sha` in `papers/results.json` matches the tag. - [ ] Re-run `scripts/reproduce_paper.py` on a clean checkout and confirm the values are unchanged. - [ ] `python -m pytest -q` green; `pytest -m slow` green. diff --git a/papers/SUPPLEMENT.md b/papers/SUPPLEMENT.md index 086be80..9b69689 100644 --- a/papers/SUPPLEMENT.md +++ b/papers/SUPPLEMENT.md @@ -8,20 +8,20 @@ FlyLab 0.6.0, commit `f95fd88a2165`, library SHA-256 `594b8d32c5e6`, map `male-c `flylab/pharm/mechanisms.py` is the single source of truth mapping engagement rows onto circuit gains; no assay carries its own copy, which is what makes Table T2 a truthful description of the software. Gains are dimensionless multipliers on the synaptic weight of a transmitter class (or, for `g_nav`, on node excitability); 1.0 is vehicle, and every gain is floored at 0.05 because a network with a gain of exactly zero is not informative. The 15 rules are listed in full in T2. The four shipped since v0.4 may not be changed silently: -* **Insect nAChR, activating:** `g_ach = max(0.05, 1 + 0.4θ − 1.6θ²)`. Rises at low engagement and falls at high engagement, encoding excitation followed by desensitisation and depolarisation block. At θ = 0.99 it is at the floor, which is why a saturating dose of a nicotinic agonist silences cholinergic transmission in the model rather than merely reducing it — and, as §3.4 of the manuscript shows, why the suppression conclusion is a property of *this* rule. +* **Insect nAChR, activating:** `g_ach = max(0.05, 1 + 0.4θ − 1.6θ²)`. Rises at low engagement and falls at high engagement, encoding excitation followed by desensitisation and depolarisation block. At θ = 0.99 it is at the floor, which is why a saturating dose of a nicotinic agonist silences cholinergic transmission in the model rather than merely reducing it; and, as §3.4 of the manuscript shows, why the suppression conclusion is a property of *this* rule. * **Insect nAChR, blocking:** `g_ach = max(0.05, 1 − θ)`. * **Insect RDL, blocking:** `g_gaba = max(0.05, 1 − θ)`. Blocking the GABA-gated chloride channel removes inhibition. * **Insect RDL, activating:** `g_gaba = max(0.05, 1 + 0.4θ)`. -Four families added in v0.5 share the shape and state a rationale each: GluCl activation uses a larger coefficient (`1 + 0.8θ`) than RDL because avermectin opening is effectively irreversible [38]; AChE inhibition raises an ambient cholinergic tone (`ach_tone = 1 + 2.0θ`) multiplying `g_ach` together with the agonist curve, reproducing the organophosphate excitation-then-block picture; Nav modulation scales node excitability globally (`g_nav = 1 + 1.5θ`), since pyrethroids and DDT act on every cell regardless of transmitter [23], [67]; octopamine agonism uses the smallest coefficient (`1 + 0.5θ`) because it is modulatory. +Four families added in v0.5 share the shape and state a rationale each: GluCl activation uses a larger coefficient (`1 + 0.8θ`) than RDL because avermectin opening is effectively irreversible [37]; AChE inhibition raises an ambient cholinergic tone (`ach_tone = 1 + 2.0θ`) multiplying `g_ach` together with the agonist curve, reproducing the organophosphate excitation-then-block picture; Nav modulation scales node excitability globally (`g_nav = 1 + 1.5θ`), since pyrethroids and DDT act on every cell regardless of transmitter [22], [66]; octopamine agonism uses the smallest coefficient (`1 + 0.5θ`) because it is modulatory. -**The level-C generic rule.** The ablation ladder's topology-only level is given one number per compound instead of a mechanism-specific gain vector. Until v0.6.1 that number was `g_all = max(0.05, 1 - theta_max)`, applied to every transmitter alike — a rule that can only depress, so every one of its 21 × 4 entries was at most zero and the level could not express disinhibition at all. It is now `g_all = max(0.05, 1 - theta_max) if the compound's dominant mechanism lowers net synaptic drive, else 1 + theta_max`: the same generic magnitude, with one bit of sign read off the mechanism table. The historical rule: max(0.05, 1 - theta_max) on every transmitter. Structurally incapable of a positive effect, so it cannot order a library containing RDL/GluCl blockers. Reported as a FLOOR (what a depression-only connectome model achieves), never as a competitor. Both are reported, the old one as `C_topology_only_floor` in T13 and T13b. +**The level-C generic rule.** The ablation ladder's topology-only level is given one number per compound instead of a mechanism-specific gain vector. Until v0.6.1 that number was `g_all = max(0.05, 1 - theta_max)`, applied to every transmitter alike; a rule that can only depress, so every one of its 21 × 4 entries was at most zero and the level could not express disinhibition at all. It is now `g_all = max(0.05, 1 - theta_max) if the compound's dominant mechanism lowers net synaptic drive, else 1 + theta_max`: the same generic magnitude, with one bit of sign read off the mechanism table. The historical rule: max(0.05, 1 - theta_max) on every transmitter. Structurally incapable of a positive effect, so it cannot order a library containing RDL/GluCl blockers. Reported as a FLOOR (what a depression-only connectome model achieves), never as a competitor. Both are reported, the old one as `C_topology_only_floor` in T13 and T13b. **These coefficients and shapes are modelling choices, not measurements.** That is why the manuscript prespecifies a family of 25 alternatives (§2.4) and re-derives every conclusion under each, and why the variance analysis nominates this transformation as the assumption worth measuring first. The **transmitter sign table** the rate engine iterates is a second set of asserted coefficients and is published as T21: 5 of the seven magnitudes (glutamate -0.4, histamine -0.5, dopamine +0.2, serotonin +0.2, octopamine +0.2) carry no measurement, are members of no specification family, and are not bounded by the label-permuting null, which permutes assignments and leaves the table itself untouched. ### S1.1 The specification family -Five shapes — monotone-linear, saturating, weakly biphasic, the shipped biphasic rule and a strongly biphasic rule — at five coefficient scales (0.50×, 0.75×, 1.00×, 1.25×, 1.50×), giving 25 specifications, of which `flylab_biphasic@1.00x` is the shipped one. Installation is process-global, so specifications are evaluated serially. Full matrix: T14; per-specification diagnostics: T14b. +Five shapes; monotone-linear, saturating, weakly biphasic, the shipped biphasic rule and a strongly biphasic rule; at five coefficient scales (0.50×, 0.75×, 1.00×, 1.25×, 1.50×), giving 25 specifications, of which `flylab_biphasic@1.00x` is the shipped one. Installation is process-global, so specifications are evaluated serially. Full matrix: T14; per-specification diagnostics: T14b. | Conclusion | Claim | Retained | Reversed | Undecidable | |---|---|---|---|---| @@ -38,7 +38,7 @@ Five shapes — monotone-linear, saturating, weakly biphasic, the shipped biphas 1. *The shuffle count could not reject.* The topology predicates were evaluated at 6 shuffles per specification, where the smallest attainable empirical permutation probability is 1/7 = 0.143. `p ≤ 0.05` was arithmetically unattainable, so the predicate fell back on a Gaussian `|z| < 2` and C3's 25/25 was a non-rejection under a test that could not reject. The run behind this document uses 100 shuffles per specification (resolution 0.010), which supports α = 0.050 with room to spare. 2. *The specification never reached the null engine.* `robustness.mechanism_spec` rebinds `compute_gains` on the assay modules, but the permutation path resolves gains through `flylab.circuit.rate.compute_gains`, which it did not rebind. Every specification therefore fed the **default** gains to its nulls and produced identical effects, so a uniform retention rate across the family was guaranteed by construction. The dependence engine now accepts the gain patch explicitly (`gains_by_compound`), and the per-specification topology verdicts in T14 are computed under the specification they are labelled with. -Both topology conclusions survive the correction (25/25 and 25/25 after Benjamini-Hochberg across the 100 structural tests of the run; 25 and 25 uncorrected), but a retention rate for a conclusion shaped as a *failure to reject* is only as strong as the equivalence behind it. Of C3's 25, **17 reach equivalence within the prespecified margin and 8 are merely indeterminate**; the corresponding split for C4 is 0 / 0. Every specification's topology cell is run on the same shuffle stream, so the structural tests are positively dependent rather than independent; Benjamini-Hochberg controls the FDR under positive dependence, but the fractions describe this shuffle stream and are not 25 independent replications. +C4 remains topology-dependent in 25/25 specifications after Benjamini-Hochberg adjustment across the 100 component structural tests of the run (25 unadjusted): it is retained when at least one adjusted structural test rejects. C3 instead has no structural rejection in 25/25, unchanged from 25 unadjusted retentions. Adjustment cannot strengthen a conclusion shaped as a failure to reject or establish absence of dependence. Of C3's 25, **17 pass the configured descriptive median-gap criterion and 8 are merely indeterminate**. The adjustment targets the component tests; because a topology label is an OR over two modes, it does not establish FDR control for specification-level labels or conclusion counts. Every specification's topology cell is run on the same shuffle stream, so the tests are dependent, and shared draws do not by themselves establish the positive-regression dependence condition required for the usual Benjamini-Hochberg guarantee. Adjusted values are therefore reported as a sensitivity analysis, not an unconditional FDR guarantee; the fractions are not 25 independent replications. Specifications reversing C1: linear@0.50x, linear@0.75x, linear@1.00x, linear@1.25x, linear@1.50x, saturating@0.50x, saturating@0.75x, saturating@1.00x, saturating@1.25x, saturating@1.50x. Mean SI gap over the nicotinic set under the shipped rule: -0.884 log₁₀ units; averaged over the 6 reversing specifications for which the index is defined: -0.893. **The v0.6 supplement said the monotone rules "buffer at least as much, never less". That is false.** The per-specification range runs -0.199 to -1.734, 3 of them buffer *less* than the shipped rule (linear@1.50x, saturating@1.25x, saturating@1.50x), and the aggregate is carried by the extreme of the range. C5 survives the specifications that reverse C1, but not for the reason the previous version gave. @@ -46,57 +46,57 @@ Specifications reversing C1: linear@0.50x, linear@0.75x, linear@1.00x, linear@1. `flylab/analysis/nullmodels.py` and `flylab/analysis/dependence.py` degrade the same committed cut, in increasing order of destruction: -* **`sign_permute_weight_matched`** — a constrained shuffle of consensus transmitter labels: random pairs of differently-labelled nodes are swapped, and a swap is kept only if every tracked transmitter's share of *total outgoing synaptic weight* stays within 0.002 of the real graph's. The node count of each label is preserved exactly, the weighted share to within the tolerance. This is the ladder's rank-3 rung. -* **`sign_permute`** — the plain permutation of the node label array. It is a **joint target-set-and-sign null**, not a rung: see below. -* **`weight_permute`** — synapse counts are permuted across edges. Topology and transmitters are kept; the pairing between wiring and strength is destroyed. -* **`rewire_degree_preserving`** — double-edge swaps, roughly 10 × |E| accepted swaps. Every node keeps its in-degree, out-degree and transmitter; the wiring pattern is destroyed. -* **`erdos_renyi`** — the same node and edge counts, endpoints drawn uniformly, weights drawn from the empirical weight list. Only the size of the graph and its census survive. +* **`sign_permute_weight_matched`**; a constrained shuffle of consensus transmitter labels: random pairs of differently-labelled nodes are swapped, and a swap is kept only if every tracked transmitter's share of *total outgoing synaptic weight* stays within 0.002 of the real graph's. The node count of each label is preserved exactly, the weighted share to within the tolerance. This is the ladder's rank-3 rung. +* **`sign_permute`**; the plain permutation of the node label array. It is a **joint target-set-and-sign null**, not a rung: see below. +* **`weight_permute`**; synapse counts are permuted across edges. Topology and transmitters are kept; the pairing between wiring and strength is destroyed. +* **`rewire_degree_preserving`**; double-edge swaps, roughly 10 × |E| accepted swaps. Every node keeps its in-degree, out-degree and transmitter; the wiring pattern is destroyed. +* **`erdos_renyi`**; the same node and edge counts, endpoints drawn uniformly, weights drawn from the empirical weight list. Only the size of the graph and its census survive. -**Why `sign_permute` is not a transmitter-identity null.** The v0.6 supplement stated that permuting labels preserves the transmitter histogram "and therefore the global excitation/inhibition balance, preserved *exactly*". That is wrong, and the error was load-bearing: out-strength is heterogeneous across nodes, so the *weighted* balance moves. Measured (T23), the cholinergic share of total outgoing synaptic weight on the `named` cut is 0.619 on the real graph against 0.544 ± 0.027 across 1000 permutations, range 0.461–0.616, putting the real graph at the 100th percentile of its own null. Because a gain patch acts *through* transmitter identity, the mode moves the drug's target set and the sign matrix at once, and the extra variance makes it conservative. The weight-matched null holds every tracked share to within 0.002. Substituting it does not change imidacloprid's permutation probability but it shrinks the equivalence gap from 2.652 Hz to 0.236 Hz against a margin of 0.332 Hz. +**Why `sign_permute` is not a transmitter-identity null.** The v0.6 supplement stated that permuting labels preserves the transmitter histogram "and therefore the global excitation/inhibition balance, preserved *exactly*". That is wrong, and the error was load-bearing: out-strength is heterogeneous across nodes, so the *weighted* balance moves. Measured (T23), the cholinergic share of total outgoing synaptic weight on the `named` cut is 0.619 on the real graph against 0.544 ± 0.027 across 1000 permutations, range 0.461–0.616, putting the real graph at the 100th percentile of its own null. Because a gain patch acts *through* transmitter identity, the mode moves the drug's target set and the sign matrix at once, and its changed variance can affect rejection rates; conservativeness is not established in general. The weight-matched null holds every tracked share to within 0.002. Substituting it changes the null ensemble and its probability, while shrinking the median gap from 2.652 rate-model units to 0.236 rate-model units against a margin of 0.332 rate-model units. Every mode preserves the node set, the node order and the seed block, and never creates self-loops or duplicate edges. The effect is always `treated − vehicle` with **both arms on the same graph**. The reported probability is the empirical two-sided permutation *p*, `(k+1)/(n+1)`, with resolution `1/(n+1)`; on a landscape the same `n` shuffled graphs are reused by every cell (a paired design), and shuffle *i* depends only on `(seed, i)`, so each cell's numbers are identical to running its profile alone. **Why `z` is secondary.** A null distribution built on degraded graphs is not normal, and for the Erdős–Rényi mode its variance can collapse, producing numerically enormous *z* values that mean only "decisively different". -### S2.1 The three-way verdict and the equivalence margin +### S2.1 The three-way diagnostic and median-gap tolerance -Failing to reject a null is not evidence of equivalence, and none of the software's return values says it is. Each mode gets exactly one of `distinguishable`, `equivalent_within_tolerance` or `indeterminate`, against a margin δ prespecified as 5% of the **vehicle readout** of the real graph. The vehicle readout is the right scale because it is in the units of the readout, does not depend on the drug, on the null ensemble or on the number of shuffles, and is the same for every mode of a cell, so an equivalence claim made against it is comparable across the ladder. It is one tenth of the 50 % relative change the selectivity module treats as a circuit response. A cell whose baseline is undefined or zero can form no margin; its non-rejected modes are then `indeterminate`, which is the correct answer rather than a missing one. +Failing to reject a null is not evidence of equivalence. The implementation first assesses the empirical permutation probability. For a non-rejected mode, it compares the absolute real-effect to null-median gap against a configured tolerance δ = 5% times the real graph's vehicle readout. Passing this point-gap check returns the legacy API string `equivalent_within_tolerance`; throughout this paper this is a **descriptive within-margin diagnostic**, not inferential equivalence. No confidence interval for the median gap, TOST procedure, distributional closeness criterion or error-controlled equivalence decision is computed. A noisy or broad null can pass the criterion. A cell with an undefined or zero vehicle value cannot form a positive margin and remains `indeterminate`. The vehicle-based tolerance is a modelling convention, not a measured biological relevance bound or choice registered before inspecting results. -**The confirmatory profile in full.** For imidacloprid at 1 µM on `named`, at 1000 permutations against a margin of 0.332 Hz, the per-mode probabilities and verdicts are: weight permutation *p* = 0.586, equivalent_within_tolerance; degree-preserving rewiring *p* = 0.472, equivalent_within_tolerance; the weight-matched transmitter null *p* = 0.389, equivalent_within_tolerance; the joint label-permutation null *p* = 0.275, indeterminate; and the Erdős–Rényi control *p* = 9.99e-04, distinguishable. The resulting necessary level is `degree_sequence` with verdict equivalent_within_tolerance. +**The reference profile in full.** For imidacloprid at 1 µM on `named`, at 1000 permutations against a margin of 0.332 rate-model units, the per-mode probabilities and verdicts are: weight permutation *p* = 0.586, equivalent_within_tolerance; degree-preserving rewiring *p* = 0.472, equivalent_within_tolerance; the weight-matched transmitter null *p* = 0.389, equivalent_within_tolerance; the joint label-permutation null *p* = 0.275, indeterminate; and the Erdős–Rényi control *p* = 9.99e-04, distinguishable. The resulting necessary level is `degree_sequence` with verdict equivalent_within_tolerance. -The **necessary information level** is the weakest rung the test could not distinguish from the real cut, reported with its verdict. Ranks 2 and 3 are not strictly nested — `weight_permute` keeps transmitter identity and loses the weight-topology pairing, the transmitter nulls the reverse — so a richer model can be distinguishable where a weaker one is not. Such cells are **non-monotone** and are flagged: 11 of 84 on the `named` cut (chlorpyrifos_oxon 10 nM, dieldrin 10 nM, dieldrin 100 nM, dieldrin 1 µM, fipronil 10 nM, fipronil 100 nM, ivermectin 1 µM, ivermectin 10 µM, picrotoxin 100 nM, picrotoxin 1 µM, picrotoxin 10 µM) and 0 of 84 on `taste_motor`. Their level must be read as "the cheapest graph model this test cannot tell apart from the real one", never as "everything above it is indistinguishable too". +The **necessary information level** is the weakest rung the test could not distinguish from the real cut, reported with its verdict. Ranks 2 and 3 are not strictly nested; `weight_permute` keeps transmitter identity and loses the weight-topology pairing, the transmitter nulls the reverse; so a richer model can be distinguishable where a weaker one is not. Such cells are **non-monotone** and are flagged: 11 of 84 on the `named` cut (chlorpyrifos_oxon 10 nM, dieldrin 10 nM, dieldrin 100 nM, dieldrin 1 µM, fipronil 10 nM, fipronil 100 nM, ivermectin 1 µM, ivermectin 10 µM, picrotoxin 100 nM, picrotoxin 1 µM, picrotoxin 10 µM) and 0 of 84 on `taste_motor`. Their level must be read as "the cheapest graph model this test cannot tell apart from the real one", never as "everything above it is indistinguishable too". -### S2.2 Effect floor, multiplicity, and confirmatory versus exploratory +### S2.2 Effect floor and multiplicity -**Effect floor.** An absolute floor of 1e-6 Hz against a ~6.6 Hz baseline admits effects at numerical-noise level, so the effective floor is `max(absolute, 1 % × |vehicle readout|)`. Landscapes report class counts under both. On the `named` cut the relative floor moves 18 cells into `no-effect` (acetylcholine 10 nM (0.000761 Hz), acetylcholine 100 nM (0.0118 Hz), ddt 10 nM (0.00309 Hz), ddt 100 nM (0.0309 Hz), gaba 100 nM (-2.74e-05 Hz), gaba 1 µM (-0.000864 Hz), gaba 10 µM (-0.025 Hz), ivermectin 10 nM (-0.0156 Hz), ivermectin 100 nM (-0.0587 Hz), picrotoxin 10 nM (0.0076 Hz), spinosad 10 nM (0.00601 Hz), spinosad 100 nM (0.0549 Hz), sulfoxaflor 10 nM (0.000202 Hz), sulfoxaflor 100 nM (0.00202 Hz), sulfoxaflor 1 µM (0.0196 Hz), thiamethoxam 10 nM (0.000303 Hz), thiamethoxam 100 nM (0.00302 Hz), thiamethoxam 1 µM (0.0289 Hz)), changing the counts from 16 / 55 / 13 to 11 / 42 / 31. +**Effect floor.** An absolute floor of 1e-6 rate-model units against a ~6.6 rate-model units baseline admits effects at numerical-noise level, so the effective floor is `max(absolute, 1 % × |vehicle readout|)`. Landscapes report class counts under both. On the `named` cut the relative floor moves 18 cells into `no-effect` (acetylcholine 10 nM (0.000761 Hz), acetylcholine 100 nM (0.0118 Hz), ddt 10 nM (0.00309 Hz), ddt 100 nM (0.0309 Hz), gaba 100 nM (-2.74e-05 Hz), gaba 1 µM (-0.000864 Hz), gaba 10 µM (-0.025 Hz), ivermectin 10 nM (-0.0156 Hz), ivermectin 100 nM (-0.0587 Hz), picrotoxin 10 nM (0.0076 Hz), spinosad 10 nM (0.00601 Hz), spinosad 100 nM (0.0549 Hz), sulfoxaflor 10 nM (0.000202 Hz), sulfoxaflor 100 nM (0.00202 Hz), sulfoxaflor 1 µM (0.0196 Hz), thiamethoxam 10 nM (0.000303 Hz), thiamethoxam 100 nM (0.00302 Hz), thiamethoxam 1 µM (0.0289 Hz)), changing the counts from 16 / 55 / 13 to 11 / 42 / 31. -**Multiplicity.** A landscape is a screen. Benjamini-Hochberg is applied across the structural tests of the run (168 on `named`, 168 on `taste_motor`), the class is decided on the adjusted probabilities, and both counts are reported. Because an empirical permutation probability cannot fall below `1/(n+1)`, the result also reports how many tests must sit at that floor before anything can be rejected, so that a corrected count of zero is never mistaken for a finding (`fdr_can_reject` = yes here). +**Multiplicity.** A landscape is a screen. Benjamini-Hochberg is applied across the component structural tests of the run (168 on `named`, 168 on `taste_motor`), the class is decided from those adjusted probabilities, and both adjusted-test and unadjusted counts are reported. A topology-dependent cell is an OR decision over its two structural modes, so test-level adjustment does not by itself establish FDR control for the resulting cell labels or counts. The shared shuffle draws make the tests dependent; reuse alone does not establish the positive-regression condition required for the usual Benjamini-Hochberg guarantee. The adjusted values are therefore a multiplicity sensitivity analysis. Because an empirical permutation probability cannot fall below `1/(n+1)`, the result also reports how many tests must sit at that floor before anything can be rejected, so that an adjusted-test count of zero is never mistaken for a finding (`fdr_can_reject` = yes here). -**Confirmatory versus exploratory.** The imidacloprid and fipronil profiles at *n* ≥ 1000 are prespecified planned comparisons and carry no correction; everything else is exploratory and says so in its warnings. Fipronil at 1 µM is the clean illustration: confirmatory *p* = 0.006, but inside the exploratory landscape its adjusted probability is 0.050 and the cell does not survive correction (composition-dominated corrected against topology-dependent raw). +**Reference profiles versus exploratory screens.** The imidacloprid and fipronil profiles at *n* ≥ 1000 receive higher computational effort and are reported with nominal probabilities. They were refined during iterative development and are not pre-registered confirmations. Fipronil at 1 µM has reference-profile *p* = 0.006; inside the library screen its adjusted probability is 0.050, with composition-dominated corrected versus topology-dependent raw labels. These differ in both permutation effort and adjustment, so the change cannot be attributed to adjustment alone. Labels based on non-rejection do not establish equivalence, including after adjustment. ### S2.3 Does the instrument work? Ground-truth recovery and power Without this, a negative from the ladder is indistinguishable from an under-powered test. A synthetic cut of 250 nodes and about 1500 edges is built with log-normal weights and a transmitter composition close to the real cut's, and a directed cycle of cholinergic nodes passing through the driven seeds is planted in it at a known strength. The cycle is recurrent cholinergic gain, so a patch that collapses `g_ach` removes an amplification that exists only while the cycle is intact: the drug *contrast*, not merely the rate, depends on the wiring. Strength 0 plants nothing and is the negative control. Nothing here touches the connectome or the compound library. -the ladder recovered 2 of 3 planted topology-dependent effects at n = 200 and returned no false positive on the unplanted control. Over the power grid (6 independently generated graphs per cell, planted strengths 0, 0.25, 0.5, 1, permutation counts 50, 200, 1000) the detection rate by strength is strength 0.25: 0.00-0.17; strength 0.5: 0.67-0.83; strength 1: 1.00-1.00 and by permutation count n=50: 0.61, n=200: 0.61, n=1000: 0.67; the empirical false-positive rate on the unplanted control is 0 of 18 (0.000). Effect size dominates: raising the permutation count buys far less than raising the planted strength. Full grid: T22, Fig. F16. +the ladder recovered 2 of 3 planted topology-dependent effects at n = 200 and returned no false positive on the unplanted control. Over the power grid (6 independently generated graphs per cell, planted strengths 0, 0.25, 0.5, 1, permutation counts 50, 200, 1000) the detection rate by strength is strength 0.25: 0.00-0.17; strength 0.5: 0.67-0.83; strength 1: 1.00-1.00 and by permutation count n=50: 0.61, n=200: 0.61, n=1000: 0.67; the empirical false-positive rate on the unplanted control is 0 of 18 (0.000). The grid is small, and changes with strength and permutation count are descriptive. Six detections from six networks would still have a two-sided Wilson 95% interval of approximately 0.61–1.00; zero false positives in eighteen independent trials would allow an upper bound near 0.18. Dependence between reused networks/settings further limits pooled interpretation. The unplanted case also omits cycle edges and associated transmitter relabelling, so it is not a perfectly matched zero-strength intervention. Full grid: T22, Fig. F16. **Permutation counts in this document.** 1000 per mode for the `named` profiles (resolution 0.0010), 300 for the `taste_motor` path arm, 1000 for each of the 84 `named` landscape cells and 300 for each of the 84 `taste_motor` cells. The sweep (T11) shows the verdict at α = 0.05 settling from *n* ≈ 50–400, and a mid-range *p* reaching ±0.02 of its final value between *n* = 25 and 400. -### S2.4 Where does the verdict settle with the size of the cut? +### S2.4 How do labels vary across the examined cut ladder? -The verdict reversed between the two committed cuts, and two cuts are not a trend. `flylab.analysis.scale.dependence_vs_scale` runs the *same* dependence profile (`dependence_profile`, unchanged) across a ladder of nested cuts of 1k, 5k, 10k, 25k, 50k cells grown from the taste-motor seed set by one fixed rule, so a verdict can be compared across size without the construction changing underneath it. The two committed cuts are built by a different recipe, so the clean scale axis is the `scale_*` rungs alone; a sequence that interleaves the committed cuts mixes a size effect with a construction effect, and `verdict_stability(recipe="scale_ladder")` reports each separately. A verdict counts as *settled* only if it is the same on the two largest cuts run with no change above that point — one rung of agreement is n = 1 and is not stability. +The label reversed between the two committed cuts, and two cuts are not a trend. `flylab.analysis.scale.dependence_vs_scale` repeats the same profile across nested cuts of 1k, 5k, 10k, 25k, 50k cells grown from the taste-motor seeds by one fixed rule. Holding that recipe fixed removes one construction difference, but size, edge count, density, composition, paths, readout support and permutation effort still co-vary. The two committed cuts use another recipe, so interleaving them with `scale_*` adds a construction confound. The legacy *settled* field is only a matching-label indicator: it is true when the two largest cuts run have the same label with no later change; it is not statistical stability or convergence. -Cost is the constraint, not correctness. From the per-shuffle and per-edge constants measured on this machine (T26), a profile at *n* = 1000 on the 50k rung is 21.1 hours on one core and the landscape over the specification family is 7 days per specification; a whole-CNS profile at the weight-5 floor is 59 hours and the dense LIF matrix stops fitting in memory well before it. The permutation count must therefore fall with the cut, and `permutation_budget` refuses to go below the resolution floor `1/α − 1`, so every rung records the *n* it was run at. +Computational cost restricts permutation resolution and therefore the strength of the statistical conclusions. From the per-shuffle and per-edge constants measured on this machine (T26), a profile at *n* = 1000 on the 50k rung is 21.1 hours on one core and the landscape over the specification family is 7 days per specification; a whole-CNS profile at the weight-5 floor is 59 hours and the dense LIF matrix stops fitting in memory well before it. The permutation count must therefore fall with the cut, and `permutation_budget` refuses to go below the resolution floor `1/α − 1`, so every rung records the *n* it was run at. -The ladder's smallest rung, `scale_1k`, is already the decisive contrast with the two cuts of T19: 1000 nodes, 22 857 edges, mean degree 22.9, in-star no — *fewer* nodes than `named` and a densely recurrent cut nothing like it. +The ladder's smallest rung, `scale_1k`, is a descriptive contrast with the two cuts of T19: 1000 nodes, 22 857 edges, mean degree 22.9, in-star no; fewer nodes than `named`, but a much denser non-in-star extract. -**Result** (T27; the study record is read from `scale_study.json` beside the paper artifacts, or from `data/derived/` when it is promoted there). Over the 7 cuts run (scale_1k, named, taste_motor, scale_5k, scale_10k, scale_25k, scale_50k), the composition-dominated verdict survives on 1 — named — so the inversion does not track the number of cells: `scale_1k` has 1000 nodes against `named`'s 1126 and a mean degree of 22.9 against 1.21, and it is already topology-dependent. Per-rung sequences: scale_1k (1000 nodes, mean degree 22.9): topology-dependent; named (1126 nodes, mean degree 1.2): composition-dominated; taste_motor (1841 nodes, mean degree 10.4): topology-dependent; scale_5k (5000 nodes, mean degree 56.0): topology-dependent; scale_10k (10000 nodes, mean degree 62.2): topology-dependent; scale_25k (25000 nodes, mean degree 54.6): topology-dependent; scale_50k (50000 nodes, mean degree 44.3): topology-dependent. Fipronil is topology-dependent on every rung. In summary: imidacloprid: the verdict is 'topology-dependent' on the largest 5 of 7 cuts tested, first reached at 'taste_motor'. fipronil: the verdict is 'topology-dependent' on the largest 7 of 7 cuts tested, first reached at 'scale_1k'. On the clean `scale_*` ladder the verdicts settle: yes. At scale_5k, scale_10k the necessary information level is the real network itself, i.e. the prediction is distinguishable from *every* rung of the ladder. Imidacloprid's relative effect stays within 0.52-0.93 of the vehicle readout across the ladder, so the verdict changes while the quantity being measured does not. +**Result** (T27; the study record is read from `scale_study.json` beside the paper artifacts, or from `data/derived/` when it is promoted there). Across the 7 cuts examined (scale_1k, named, taste_motor, scale_5k, scale_10k, scale_25k, scale_50k), the composition-dominated label occurs on 1 cut, named. Because `scale_1k` has fewer nodes but a different label, node count alone does not order these observations. The imidacloprid sequence is scale_1k (1000 nodes, mean degree 22.9): topology-dependent; named (1126 nodes, mean degree 1.2): composition-dominated; taste_motor (1841 nodes, mean degree 10.4): topology-dependent; scale_5k (5000 nodes, mean degree 56.0): topology-dependent; scale_10k (10000 nodes, mean degree 62.2): topology-dependent; scale_25k (25000 nodes, mean degree 54.6): topology-dependent; scale_50k (50000 nodes, mean degree 44.3): topology-dependent; fipronil is labelled topology-dependent on every examined rung. The legacy matching-label indicator on the `scale_*` subset is yes; this records final-label agreement, not convergence. At scale_5k, scale_10k all tested null modes are rejected at the rung's stated effort. Imidacloprid's relative effect stays within 0.52-0.93 of the vehicle readout, but similar aggregate magnitude does not make the cell populations, drive support or other extract properties identical. -**Two caveats, both reported rather than absorbed.** The permutation budget falls with the edge count (20-1000 draws, resolution 0.0010-0.0476): at scale_10k, scale_25k, scale_50k the permutation budget falls to n = 20, where the smallest attainable probability is 0.048 against alpha = 0.05: a rejection there is the smallest the test can express, so those rungs confirm a verdict already settled lower down rather than establishing it. And `named` and `taste_motor` are hops-limited neighbourhoods while the `scale_*` rungs are nested ranked-BFS cuts, so a difference *between* the two families confounds scale with construction; differences *within* the `scale_*` rungs do not. +**Two caveats, both reported rather than absorbed.** The permutation budget falls with edge count (20-1000 draws, resolution 0.0010-0.0476): at scale_10k, scale_25k, scale_50k the permutation budget falls to n = 20, where the smallest attainable probability is 0.048 against alpha = 0.05: a rejection there is the smallest the test can express, so the top-rung labels are resolution-limited observations, not independent confirmations. And `named` and `taste_motor` are hops-limited neighbourhoods while `scale_*` uses nested ranked-BFS cuts. Comparisons between those families confound recipe with structure; comparisons within `scale_*` hold the recipe fixed but do not isolate size, density, recurrence or the other properties that change together. ## S3. Runtimes and their calibration -**Rate runtime.** `flylab/circuit/rate.py` holds a dense signed weight matrix `W[j,i] = sign(nt_i) × synapses(i→j)`, with the signs of Table T21. **That matrix is then row-normalised**: `W_signed = W / max(Σ_i |W[j,i]|, 1)`, one denominator per postsynaptic cell. A gain patch afterwards scales the columns by the presynaptic transmitter's gain, so `W_eff = W_signed · diag(g)` and **each cell's recurrent input is a composition-weighted average of its presynaptic gains, normalised by its own total absolute input**. The update is a clipped leaky iteration `r ← clip((1 − α)·r + α·g_nav·(d + W_eff·r), 0, r_max)` with α = 0.3, 80 steps, `r_max` = 300 Hz and *d* the per-node external drive in Hz. It is deterministic, and one evaluation of the 1126-node graph costs about 50 ms with single-threaded BLAS — which is what makes 1000-permutation profiles and an 84-cell landscape affordable at all. +**Rate runtime.** `flylab/circuit/rate.py` holds a dense signed weight matrix `W[j,i] = sign(nt_i) × synapses(i→j)`, with the signs of Table T21. **That matrix is then row-normalised**: `W_signed = W / max(Σ_i |W[j,i]|, 1)`, one denominator per postsynaptic cell. A gain patch afterwards scales the columns by the presynaptic transmitter's gain, so `W_eff = W_signed · diag(g)` and **each cell's recurrent input is a composition-weighted average of its presynaptic gains, normalised by its own total absolute input**. The update is a clipped leaky iteration `r ← clip((1 − α)·r + α·g_nav·(d + W_eff·r), 0, r_max)` with α = 0.3, 80 steps, `r_max` = 300 rate-model units and *d* the per-node external drive in rate-model units. It is deterministic, and one evaluation of the 1126-node graph costs about 50 ms with single-threaded BLAS; which is what makes 1000-permutation profiles and an 84-cell landscape affordable at all. The normalisation is an **asserted modelling choice**, not a measurement, and it is consequential for RQ2: a readout in which each cell averages its presynaptic gains builds part of "composition-dominated" into itself. Three modes are implemented and swept in §3.3 (T25): @@ -106,37 +106,37 @@ The normalisation is an **asserted modelling choice**, not a measurement, and it | `none` | 1 | denom[j] = 1: raw signed synapse counts. The recurrent operator is supercritical (spectral radius ~150 on the named cut) and the r_max clip, not the wiring, bounds the rates. | | `degree` | `max(in-degree, 1)` | denom[j] = max(in-degree[j], 1): degree-corrected. Relative input strengths survive (unlike row_abs) but hubs are discounted by fan-in. | -The composition-versus-full rank correlation at 1 µM is 0.988 under `row_abs`, 0.906 under `none` and 0.666 under `degree` (-0.372 at 10 nM): the composition verdict survives removing the normalisation but not replacing it with a degree-corrected one. Levels are not comparable across modes — the unnormalised operator is supercritical on this cut, so the `r_max` clip rather than the wiring bounds the rates — only the orderings are. +The composition-versus-full rank correlation at 1 µM is 0.988 under `row_abs`, 0.906 under `none` and 0.666 under `degree` (-0.372 at 10 nM): the composition verdict survives removing the normalisation but not replacing it with a degree-corrected one. Levels are not comparable across modes; the unnormalised operator is supercritical on this cut, so the `r_max` clip rather than the wiring bounds the rates; only the orderings are. -**LIF runtime.** `flylab/circuit/lif.py` is a current-based leaky integrate-and-fire network with the parameters of Shiu *et al.* [7]: τ_m = 20 ms, V_rest = −52 mV, V_th = −45 mV, V_reset = −52 mV, refractory 2.2 ms, exponential synaptic current with τ_s = 5 ms, forward Euler at dt = 0.1 ms. A presynaptic spike on edge *i*→*j* injects `w_scale × synapses × sign(nt_i) × gain(nt_i)` mV into *j*. +**LIF runtime.** `flylab/circuit/lif.py` is a current-based leaky integrate-and-fire network with the parameters of Shiu *et al.* [6]: τ_m = 20 ms, V_rest = −52 mV, V_th = −45 mV, V_reset = −52 mV, refractory 2.2 ms, exponential synaptic current with τ_s = 5 ms, forward Euler at dt = 0.1 ms. A presynaptic spike on edge *i*→*j* injects `w_scale × synapses × sign(nt_i) × gain(nt_i)` mV into *j*. -Two choices are modelling decisions, not literature values. *The synaptic scale is a one-time calibration scan*: `w_scale` = 0.05 mV per synapse and a drive weight of 0.5 mV per event were chosen once, by scanning, so that vehicle MN9 cells fire in a plausible band under the standard 150 Hz seed drive. No recording was used and no error was minimised against data, so this does not consume the one fitted parameter the project allows — but it is a tuned constant. *A background Poisson drive is required, and it is a choice*: excising a 1-hop neighbourhood removes every input from outside it, so every non-seed cell receives an independent Poisson drive at 0.65 × the seed rate, standing in for the amputated inputs. This is the most consequential free choice in the LIF arm; it sets the operating point and it is why the two engines report sugar-driven MN9 at 0.293 Hz (rate) against 105 Hz (LIF). +Two choices are modelling decisions, not literature values. *The synaptic scale is a one-time calibration scan*: `w_scale` = 0.05 mV per synapse and a drive weight of 0.5 mV per event were chosen once, by scanning, so that vehicle MN9 cells fire in a plausible band under the standard 150 Hz input drive seed drive. No recording was used and no error was minimised against data, so this does not consume the one fitted parameter the project allows; but it is a tuned constant. *A background Poisson drive is required, and it is a choice*: excising a 1-hop neighbourhood removes every input from outside it, so every non-seed cell receives an independent Poisson drive at 0.65 × the seed rate, standing in for the amputated inputs. This is the most consequential free choice in the LIF arm; it sets the operating point and it is why the two engines report sugar-driven MN9 at 0.293 rate-model units (rate) against 105 LIF spikes/s (LIF). ## S4. Supporting capabilities These are implemented, tested and used, but are supporting capabilities rather than contributions of this paper. -**Exposure.** `flylab/pharm/exposure.py` maps a dose and route onto *C(t)* and hence engagement(t) in one compartment. Its constants are class placeholders chosen so that the three routes differ in the direction the literature describes, and are fitted to no *Drosophila* ADME dataset because that dataset does not exist: of the seven quantities such a model needs, `data/literature/fly_pharmacokinetics.yaml` records only metabolite identity as well established [52], [54], [55], [58]. Adult haemolymph volume, adult body mass, cuticular penetration rate and any *Drosophila* elimination half-life are `null`; what is recorded is 25 nL recoverable per adult [57], 174 ± 81 nL in third-instar larvae [56], a water-partitioning measurement [53], an ethanol entry that could not supply a rate [65], and an imidacloprid half-life of 4.5–5 h in *honeybee* [58] — the only verified insect half-life for a library compound. Two structural warnings travel with every profile: haemolymph is the lowest-radioactivity compartment for imidacloprid, so a one-compartment haemolymph model under-estimates CNS exposure [59]; and a substantial metabolic route runs through the gut microbiome, so a single-enzyme clearance term is wrong for this compound [54]. +**Exposure.** `flylab/pharm/exposure.py` maps a dose and route onto *C(t)* and hence engagement(t) in one compartment. Its constants are class placeholders chosen so that the three routes differ in the direction the literature describes, and are fitted to no *Drosophila* ADME dataset because that dataset does not exist: of the seven quantities such a model needs, `data/literature/fly_pharmacokinetics.yaml` records only metabolite identity as well established [51], [53], [54], [57]. Adult haemolymph volume, adult body mass, cuticular penetration rate and any *Drosophila* elimination half-life are `null`; what is recorded is 25 nL recoverable per adult [56], 174 ± 81 nL in third-instar larvae [55], a water-partitioning measurement [52], an ethanol entry that could not supply a rate [64], and an imidacloprid half-life of 4.5–5 h in *honeybee* [57]; the only verified insect half-life for a library compound. Two structural warnings travel with every profile: haemolymph is the lowest-radioactivity compartment for imidacloprid, so a one-compartment haemolymph model under-estimates CNS exposure [58]; and a substantial metabolic route runs through the gut microbiome, so a single-enzyme clearance term is wrong for this compound [53]. -**Genotype.** `flylab/pharm/genotype.py` applies published target-site resistance alleles as multiplicative potency shifts from `data/literature/resistance_alleles.yaml` [13]–[25], enforcing one rule: **a shift applies per compound, never per receptor.** Two published findings make the naive implementation wrong — *Rdl* A301S shifts GABA potency 33.9-fold but fipronil only 1.1-fold on the expressed planthopper receptor [15], and *para* M918T shifts deltamethrin 100-fold while leaving DDT at 1-fold [22]. An allele/compound pair with no sourced number is left unchanged and listed under `unshifted`, never extrapolated. One row is deliberately ugly: *Rdl* A301G in *D. simulans* carries a 20000-fold shift derived from a whole-animal resistance ratio, which drives fipronil engagement to 4.63e-04 and is almost certainly an overstatement of the target-site change [19]. The notebook says so. Panel: T9, F9. +**Genotype.** `flylab/pharm/genotype.py` applies published target-site resistance alleles as multiplicative potency shifts from `data/literature/resistance_alleles.yaml` [12]–[24], enforcing one rule: **a shift applies per compound, never per receptor.** Two published findings make the naive implementation wrong; *Rdl* A301S shifts GABA potency 33.9-fold but fipronil only 1.1-fold on the expressed planthopper receptor [14], and *para* M918T shifts deltamethrin 100-fold while leaving DDT at 1-fold [21]. An allele/compound pair with no sourced number is left unchanged and listed under `unshifted`, never extrapolated. One row is deliberately ugly: *Rdl* A301G in *D. simulans* carries a 20000-fold shift derived from a whole-animal resistance ratio, which drives fipronil engagement to 4.63e-04 and is almost certainly an overstatement of the target-site change [18]. The notebook says so. Panel: T9, F9. -**Mixtures.** `flylab/pharm/mixtures.py` implements Bliss independence [45] and Loewe additivity [48], [49] as explicit null models, following the response-surface framing of Greco *et al.* [46], and selects between them per pair: Loewe where the components share a receptor key, Bliss where their targets are disjoint. The three-case comparison against the honeybee dataset [51] is a qualitative concordance check and not validation. Table T5. +**Mixtures.** `flylab/pharm/mixtures.py` implements Bliss independence [44] and Loewe additivity [47], [48] as explicit null models, following the response-surface framing of Greco *et al.* [45], and selects between them per pair: Loewe where the components share a receptor key, Bliss where their targets are disjoint. The three-case comparison against the honeybee dataset [50] is a qualitative concordance check and not validation. Table T5. -**Expression weighting.** `flylab/pharm/expression.py` relaxes the assumption that every cell expresses the receptor, using class-level statements from single-cell atlases [39]–[44] mapped by hand onto MaleCNS superclasses, and interpolates each node's gain between 1.0 and the full scalar gain. It is a **sensitivity layer and never the primary model**: none of the cited atlases uses MaleCNS superclass labels, so every mapping row is tagged `cross_atlas_inference`; coverage is 0.205 of cells; and receptor expression in identified adult motor neurons is a confirmed gap, so MN9 carries the `unknown` default. Weighting shifts MN9 by -2.622 Hz at 1 µM imidacloprid, which we report as the size of the uncertainty rather than as a correction. +**Expression weighting.** `flylab/pharm/expression.py` relaxes the assumption that every cell expresses the receptor, using class-level statements from single-cell atlases [38]–[43] mapped by hand onto MaleCNS superclasses, and interpolates each node's gain between 1.0 and the full scalar gain. It is a **sensitivity layer and never the primary model**: none of the cited atlases uses MaleCNS superclass labels, so every mapping row is tagged `cross_atlas_inference`; 0.205 is the fraction of cell–receptor slots with a non-unknown ordinal class mapping, not a fraction of cells with measured expression; and receptor expression in identified adult motor neurons is a confirmed gap, so MN9 carries the `unknown` default. Weighting shifts MN9 by -2.622 rate-model units at 1 µM imidacloprid, which we report as the size of the uncertainty rather than as a correction. -**Model dose-response.** `flylab/assays/ensemble.py` builds the model's own concentration-response curve, fits a four-parameter Hill function by Nelder–Mead in log-concentration space and bootstraps it. For imidacloprid on the `named` cut the circuit midpoint is 34.7 nM with a bootstrap 95 % CI of 35.2 nM to 71.5 nM (r² = 0.9999). The interval does not bracket the point estimate, and that is reported rather than hidden: the bootstrap resamples replicates within each concentration, the lower plateau of the ladder is short, and the midpoint is therefore poorly identified. This is in any case the **model's** circuit midpoint, not a receptor or an animal IC50. The one-at-a-time tornado at 1 µM finds exactly zero span for ec50, hill_n — the engagement is saturated — which the global analysis (§3.5) confirms and quantifies. +**Model dose-response.** `flylab/assays/ensemble.py` builds the model's own concentration-response curve, fits a four-parameter Hill function by Nelder–Mead in log-concentration space and bootstraps it. For imidacloprid on the `named` cut the circuit midpoint is 34.7 nM with a bootstrap 95 % CI of 35.2 nM to 71.5 nM (r² = 0.9999). The interval does not bracket the point estimate, and that is reported rather than hidden: the bootstrap resamples replicates within each concentration, the lower plateau of the ladder is short, and the midpoint is therefore poorly identified. This is in any case the **model's** circuit midpoint, not a receptor or an animal IC50. The one-at-a-time tornado at 1 µM finds exactly zero span for ec50, hill_n; the engagement is saturated; which the global analysis (§3.5) confirms and quantifies. ## S5. Prospective predictions The following 7 hypotheses are written before any experiment and labelled `software_prediction`; `live_result` is `null` in every row and stays null until a real table is imported by hand. They are **prospective, not pre-registered**: a mutable repository is not a registry, and the wording will change only if a tagged release is externally archived with a timestamp. Full table: T3. * **H1** (receptor engagement). At 1 µM, imidacloprid insect nicotinic engagement ≥ 0.9 while vertebrate α4β2 stays ≤ 0.2. Predicted gap 0.854. Checkable against published expressed-receptor work rather than against flies. -* **H2** (PER, `protocols/per.md`). The same dose lowers cholinergic drive in the MN9/DNp01 neighbourhood: predicted change in network mean rate -6.145 Hz. **Specification-dependence caveat.** The *sign* of this prediction is a modelling choice, not a model output: §3.4 shows the same compound at the same engagement takes the network from 6.631 Hz to 0.458 Hz under the shipped biphasic rule and from 6.631 Hz to 10.1 Hz under a monotone rule of the same admissible family, and only 15 of 25 specifications retain the suppression direction. An experimenter should treat H2 as a signed interval over the family rather than as a point prediction, and a live result in either direction falsifies a specification rather than the integration. The control rate this needs could not be sourced; the nearest substitute is 40 % responders at 1 h food deprivation rising to 68 % at 2 h in 463 wild-type flies [61], which also shows that deprivation time is a larger nuisance variable than most effects FlyLab would try to detect. -* **H3** (climbing, `protocols/climbing.md`). Fipronil at 1 µM disinhibits the neighbourhood: predicted change 0.896 Hz, opposite in sign to H2. **The same caveat applies transitively**: H3 is defined relative to H2, so its contrast inherits H2's specification dependence even though the fipronil arm itself is retained by 25 of 25 specifications. Negative geotaxis is the best-served endpoint [60], [64], [66]. -* **H4** (negative control). Diazepam engages vertebrate GABA-A and not insect RDL in this library, so it moves no fly readout: predicted effect 0.000. A non-zero live effect would falsify either the library row [17] or the assumption that the compound reaches the CNS unmetabolised. -* **H5** (within-animal control). Bitter input vetoes sugar-driven MN9 on the map-extracted path: predicted change -0.296 Hz relative to sugar alone, matching the published direction [7], [63]. -* **H6**. Fipronil relieves the bitter veto on the map path: predicted rise in the veto ratio 0.413. The dependence analysis shows this readout is not distinguishable from its shuffled controls, so a live confirmation would support the pharmacology and **not** the wiring. -* **H7** (selectivity negative control). Picrotoxin perturbs the fly neighbourhood while its receptor selectivity index is ≈ 0: predicted change 0.340 Hz. Any insect-selectivity claim the tool makes must fail for picrotoxin, and picrotoxin's lack of insect-over-vertebrate selectivity is independently attested [31]. +* **H2** (PER, `protocols/per.md`). The same dose lowers cholinergic drive in the MN9/DNp01 neighbourhood: predicted change in network mean rate -6.145 rate-model units. **Specification-dependence caveat.** The *sign* of this prediction is a modelling choice, not a model output: §3.4 shows the same compound at the same engagement takes the network from 6.631 rate-model units to 0.458 rate-model units under the shipped biphasic rule and from 6.631 rate-model units to 10.1 rate-model units under a monotone rule of the same admissible family, and only 15 of 25 specifications retain the suppression direction. An experimenter should treat H2 as a signed interval over the family rather than as a point prediction, and a live result in either direction falsifies a specification rather than the integration. The control rate this needs could not be sourced; the nearest substitute is 40 % responders at 1 h food deprivation rising to 68 % at 2 h in 463 wild-type flies [60], which also shows that deprivation time is a larger nuisance variable than most effects FlyLab would try to detect. +* **H3** (climbing, `protocols/climbing.md`). Fipronil at 1 µM disinhibits the neighbourhood: predicted change 0.896 rate-model units, opposite in sign to H2. **The same caveat applies transitively**: H3 is defined relative to H2, so its contrast inherits H2's specification dependence even though the fipronil arm itself is retained by 25 of 25 specifications. Negative geotaxis is the best-served endpoint [59], [63], [65]. +* **H4** (negative control). Diazepam engages vertebrate GABA-A and not insect RDL in this library, so it moves no fly readout: predicted effect 0.000. A non-zero live effect could reveal an omitted target or model failure rather than falsifying the absence of evidence in the library row [16] or the assumption that the compound reaches the CNS unmetabolised. +* **H5** (within-animal control). Bitter input vetoes sugar-driven MN9 on the map-extracted path: predicted change -0.296 rate-model units relative to sugar alone, matching the published direction [6], [62]. +* **H6**. Fipronil relieves the bitter veto on the map path: predicted rise in the veto ratio 0.413. The dependence analysis shows this readout is not distinguishable from its shuffled controls, so a matching live response alone would not discriminate the tested wiring alternatives. This non-rejection has no endpoint-specific power calibration. +* **H7** (selectivity negative control). Picrotoxin perturbs the fly neighbourhood while its receptor selectivity index is ≈ 0: predicted change 0.340 rate-model units. Any insect-selectivity claim the tool makes must fail for picrotoxin, and picrotoxin's lack of insect-over-vertebrate selectivity is independently attested [30]. Rows where the model's own output disagrees with the prospective direction: none. @@ -144,7 +144,7 @@ Rows where the model's own output disagrees with the prospective direction: none ## S6. Evidence typing: the full rules -The permitted transformation is a function of two facts — the **parameter type** the source measured and the **evidence distance** between that source and this compound at this receptor in this species. The distance grades are E0 (on-target); E1 (cross-species); E2 (related receptor); E3 (class extrapolation); E4 (unsupported). The full (parameter type × distance) → model table is generated as T20; its shape is: +The permitted transformation is a function of two facts; the **parameter type** the source measured and the **evidence distance** between that source and this compound at this receptor in this species. The distance grades are E0 (on-target); E1 (cross-species); E2 (related receptor); E3 (class extrapolation); E4 (unsupported). A Kd/Ki earns physical `binding_occupancy` only at E0 (the exact compound, receptor and species); at E1–E3 it is explicitly a binding-derived engagement proxy, and at E4 it is not modelled. The full (parameter type × distance) → model table is generated as T20; its shape is: | Parameter type | E0 | E1–E2 | E3 | E4 | |---|---|---|---|---| @@ -152,11 +152,11 @@ The permitted transformation is a function of two facts — the **parameter type | `EC50`, `IC50`, `Kb` | `functional_engagement` | `functional_engagement_proxy` | `functional_engagement_proxy` | `not_modelled` | | `relative_potency`, `class_order`, `unknown` | `not_modelled` | `not_modelled` | `not_modelled` | `not_modelled` | -A proxy computes the same Hill expression as its non-proxy sibling but reports itself as a transfer and carries a provenance warning naming what was transferred — the species, the preparation or the chemical class. Asking for a binding occupancy from an EC50, or for any numeric engagement from a placeholder row, raises `EvidenceTypeError`; the library validator rejects a placeholder that carries a value and a non-placeholder that does not. The library currently reports none validation problems. +A proxy computes the same Hill expression as its non-proxy sibling but reports itself as a transfer and carries a provenance warning naming what was transferred; the species, the preparation or the chemical class. Asking for a binding occupancy from an EC50, or for any numeric engagement from a placeholder or unsupported row, raises/refuses via `EvidenceTypeError`; the library validator rejects a placeholder that carries a value and a non-placeholder that does not. The library currently reports none validation problems. -**The soundness invariant.** *No numeric engagement reaching a circuit readout may originate from a row whose (parameter type, evidence distance) pair does not admit one.* `tests/test_pharm_evidence.py` checks it exhaustively rather than at a call site: over every row of the shipped library and every entry point that can produce an engagement, and with a companion check that dropping the inadmissible rows changes no gain vector. Mutating the rule table makes the test fail with a message naming the row and the pair. This, rather than the enum, is what the manuscript claims as a computing contribution. +**The soundness invariant.** *No numeric engagement reaching a circuit readout may originate from a row whose (parameter type, evidence distance) pair does not admit one.* `tests/test_pharm_evidence.py` performs a finite regression audit over every row of the shipped library through the enumerated comparison, dose-curve, uncertainty, subgraph, taste-assay and dashboard paths. Companion checks show that dropping inadmissible rows changes no gain vector, and deliberate invalid-transformation requests raise with the offending evidence pair. These tests cover the shipped rows and enumerated paths; they do not prove coverage of an unenumerated future entry point, nor do they mutate the rule table end to end. This enforced refusal plus finite audit, rather than the enum alone, is what the manuscript claims as a computing contribution; legacy untyped rows are normalised as class-extrapolated EC50s for backwards compatibility and are not evidence of a measured target. -**Census.** 102 rows over 21 compounds and 15 receptor keys (Table T10, Figure F11). Parameter types: 34 EC50, 10 IC50, 2 Kd, 56 unknown. Evidence distances: 10 E0, 15 E1, 10 E2, 11 E3, 56 E4. Engagement models: 1 `binding_occupancy`, 1 `binding_engagement_proxy`, 9 `functional_engagement`, 35 `functional_engagement_proxy`, 56 `not_modelled`. Rows entitled to a binding-occupancy model: imidacloprid at insect_nAChR_native_dmel (Kd 2e-09 M) [77]. Rows demoted to a binding proxy by distance: imidacloprid at insect_nAChR_beta1 (Kd 8.3e-11 M). +**Census.** 102 rows over 21 compounds and 15 receptor keys (Table T10, Figure F11). Parameter types: 34 EC50, 10 IC50, 2 Kd, 56 unknown. Evidence distances: 10 E0, 15 E1, 10 E2, 11 E3, 56 E4. Engagement models: 1 `binding_occupancy`, 1 `binding_engagement_proxy`, 9 `functional_engagement`, 35 `functional_engagement_proxy`, 56 `not_modelled`. Rows entitled to a binding-occupancy model: imidacloprid at insect_nAChR_native_dmel (Kd 2e-09 M) [76]. Rows demoted to a binding proxy by distance: imidacloprid at insect_nAChR_beta1 (Kd 8.3e-11 M). ## S7. Reproduction details @@ -177,7 +177,7 @@ A proxy computes the same Hill expression as its non-proxy sibling but reports i | dose-response bootstrap | 200 resamples | 40 | | prediction replicates | 6 | 2 | -The middle column is read from `results.json`, so it describes the run that rendered this document rather than an aspiration. `--fast` changes statistical effort only, never the model: point estimates that do not depend on replicate or permutation count are identical, and a `--fast` *p* is coarser. Two tests guard the committed record: it must not come from a `--fast` run, and — new in this revision — **every statistical parameter it records must equal the shipped default**, because the v0.6 record was generated at a shuffle count the code no longer used and nothing detected it. +The middle column is read from `results.json`, so it describes the run that rendered this document rather than an aspiration. `--fast` changes statistical effort only, never the model: point estimates that do not depend on replicate or permutation count are identical, and a `--fast` *p* is coarser. Two tests guard the committed record: it must not come from a `--fast` run, and; new in this revision; **every statistical parameter it records must equal the shipped default**, because the v0.6 record was generated at a shuffle count the code no longer used and nothing detected it. Two of the analyses are started in worker processes when the run begins and joined at their own step, because they dominate the wall clock and depend on nothing the other steps produce. `--jobs` controls that and the permutation workers; results are identical at any value. @@ -189,13 +189,13 @@ Every assay returns a notebook (schema 0.3) whose `warnings` list ships with the These are summarised in §4 and given here at length. -**(i) The transmitter sign table is mostly asserted.** Acetylcholine at +1.0 is the normalisation every other entry is scaled against, and GABA at −1.0 is the convention that inhibition is as strong per synapse as excitation. The remaining five — glutamate -0.4, histamine -0.5, dopamine +0.2, serotonin +0.2, octopamine +0.2 — are asserted magnitudes with no measurement behind them (T21). Glutamate in particular is signed inhibitory at 40 % of the stated strength, which is a hedge against the cells whose glutamatergic output is excitatory rather than a measurement of anything; any such cell is mis-signed in this model. None of these five is a member of the specification family, and the transmitter nulls do not bound them, because they permute *assignments* and leave the table untouched. Separately, the ablation ladder's level B inherits a whole-CNS excitation index that scores glutamate as partly excitatory while level D signs it inhibitory; the two levels of one ladder therefore disagree about a transmitter sign, and reconciling them moves the composition-versus-full rank correlation by at most 0.057. +**(i) The transmitter sign table is mostly asserted.** Acetylcholine at +1.0 is the normalisation every other entry is scaled against, and GABA at −1.0 is the convention that inhibition is as strong per synapse as excitation. The remaining five; glutamate -0.4, histamine -0.5, dopamine +0.2, serotonin +0.2, octopamine +0.2; are asserted magnitudes with no measurement behind them (T21). Glutamate in particular is signed inhibitory at 40 % of the stated strength, which is a hedge against the cells whose glutamatergic output is excitatory rather than a measurement of anything; any such cell is mis-signed in this model. None of these five is a member of the specification family, and the transmitter nulls do not bound them, because they permute *assignments* and leave the table untouched. Separately, the ablation ladder's level B inherits a whole-CNS excitation index that scores glutamate as partly excitatory while level D signs it inhibitory; the two levels of one ladder therefore disagree about a transmitter sign, and reconciling them moves the composition-versus-full rank correlation by at most 0.057. **(ii) The row normalisation is load-bearing** (§S3, §3.3). -**(iii) Receptor effects are applied across transmitter-defined synapses.** A gain on `g_ach` scales every cholinergic synapse, whether or not that synapse expresses the receptor the compound binds. This is the assumption the expression layer relaxes, and it is reported as a bound rather than an improvement: coverage is 0.205 of cells, every mapping row is a cross-atlas inference, and receptor expression in identified adult motor neurons — MN9's class — is a confirmed gap in the atlas literature, so the paper's principal readout carries the default weight. Weighting shifts MN9 by -2.622 Hz at 1 µM imidacloprid. +**(iii) Receptor effects are applied across transmitter-defined synapses.** A gain on `g_ach` scales every cholinergic synapse, whether or not that synapse expresses the receptor the compound binds. This is the assumption the expression layer relaxes, and it is reported as a bound rather than an improvement: 0.205 is the fraction of cell–receptor slots with a non-unknown ordinal class mapping, not a fraction of cells with measured expression; every mapping row is a cross-atlas inference; and receptor expression in identified adult motor neurons; MN9's class; is a confirmed gap in the atlas literature, so the paper's principal readout carries the default weight. Weighting shifts MN9 by -2.622 rate-model units at 1 µM imidacloprid. -**(iv) Hops-limited cuts.** A 1-hop neighbourhood with a weight floor is a fraction of MN9's real input; missing inhibition can flip the sign of any readout, and the disinhibition readout is bounded by the GABA share that survived the cut. Growing the cut is deliberately constrained: past about 1 MB of JSON the weight floor is raised rather than hops added, because a 2-hop cut of a motor neuron in a 25-million-edge CNS is most of the animal. §3.2 is the quantitative demonstration of how much the choice of cut matters — the dependence verdict itself reverses between the two committed cuts. +**(iv) Hops-limited cuts.** A 1-hop neighbourhood with a weight floor is a fraction of MN9's real input; missing inhibition can flip the sign of any readout, and the disinhibition readout is bounded by the GABA share that survived the cut. Growing the cut is deliberately constrained: past about 1 MB of JSON the weight floor is raised rather than hops added, because a 2-hop cut of a motor neuron in a 25-million-edge CNS is most of the animal. §3.2 is the quantitative demonstration of how much the choice of cut matters; the dependence verdict itself reverses between the two committed cuts. **(v) The two cuts differ in what they can express.** The `named` cut carries 612 cholinergic, 295 GABAergic and 56 glutamatergic cells, and contains 0 octopaminergic cells and 0 on `taste_motor`, so a compound whose only insect mechanism acts through `g_oct` cannot move either readout for substrate reasons rather than pharmacological ones; 14% of `named` cells carry no consensus transmitter prediction and are presynaptically inert. Full structural census: T19. @@ -207,9 +207,9 @@ These are the domain-side results the method is demonstrated on. They are suppor **The evidence census in full.** Of 102 rows over 21 compounds and 15 receptor keys, 34 carry an EC50, 10 an IC50 and 2 a binding constant. By evidence distance the split is 10 / 15 / 10 / 11 / 56 rows at E0–E4. At the 1 µM working concentration the two headline compounds engage their insect targets at 0.991 and 0.985 against 0.091 and 0.476 at the vertebrate counterparts (500-fold and 37-fold; 2.70 log₁₀ units of window for the first). Placeholder rows contribute nothing: the negative-control row reports "not modelled (N/A)". -**Sensory-veto readout.** Driving the 34 sweet GRN body IDs at 150 Hz and reading MN9 reproduces the published direction of a bitter veto [7], [63] on both runtimes (0.000 rate, 0.000 LIF in vehicle; 0.413 and 0.159 under fipronil). The two runtimes agree on **direction only** — sugar-driven MN9 is 0.293 Hz on the rate engine against 105 Hz on the LIF engine, a difference following from the LIF background drive (§S3) — and we do not present that agreement as cross-model validation. +**Sensory-veto readout.** Driving the 34 sweet GRN body IDs at 150 Hz input drive and reading MN9 reproduces the published direction of a bitter veto [6], [62] on both runtimes (0.000 rate, 0.000 LIF in vehicle; 0.413 and 0.159 under fipronil). The `LB3b/c` sweet and `LB1a-d` bitter labels are working hypotheses from published gustatory nomenclature, not MaleCNS modality annotations. The two runtimes agree on **direction only**; sugar-driven MN9 is 0.293 rate-model units on the rate engine against 105 LIF spikes/s on the LIF engine, a difference following from the LIF background drive (§S3); and we do not present that agreement as cross-model validation. -**A properly powered negative.** At 300 permutations the veto ratio on the extracted sensory path is not distinguishable from any shuffled control (max |z| = 0.220), equivalent within tolerance for none and indeterminate for erdos_renyi, rewire_degree_preserving, sign_permute, sign_permute_weight_matched, weight_permute. Because §3.1 characterises the instrument's power, this is reportable as a negative rather than as an absence of evidence. The comparator compound has no veto ratio at this dose at all — the readout is *undefined*, not zero, and is reported as undefined. +**Non-rejection with unquantified endpoint-specific power.** At 300 permutations the sensory-veto ratio is not distinguished from the tested controls (max |z| = 0.220). The descriptive median-gap criterion passes for none and is indeterminate for erdos_renyi, rewire_degree_preserving, sign_permute, sign_permute_weight_matched, weight_permute. The synthetic mean-rate study in §3.1 does not establish power for this different graph and ratio endpoint, so the result cannot rule out a wiring effect. The comparator's ratio is undefined at this dose because its denominator is silenced; it is not zero. **Selectivity landscape.** The split is by mechanism rather than potency: chlorpyrifos_oxon, ddt, deltamethrin, permethrin are amplified by the circuit and acetamiprid, acetylcholine, clothianidin, imidacloprid, nicotine, nitenpyram buffered, subject to the threshold caveat of §3.4. Compounds never reaching the circuit threshold on either cut (chlordimeform, dieldrin, fipronil, gaba, ivermectin, picrotoxin, spinosad, sulfoxaflor, thiamethoxam) are reported **unscored** rather than buffered, and one of them is unscored for a substrate reason rather than a pharmacological one: chlordimeform's only insect mechanism acts through octopamine, and neither cut contains an octopaminergic cell. diff --git a/papers/SUPPLEMENT.md.in b/papers/SUPPLEMENT.md.in index e0255dd..d0ed5c6 100644 --- a/papers/SUPPLEMENT.md.in +++ b/papers/SUPPLEMENT.md.in @@ -15,20 +15,20 @@ FlyLab {{flylab_version}}, commit `{{git_sha}}`, library SHA-256 `{{library_sha2 `flylab/pharm/mechanisms.py` is the single source of truth mapping engagement rows onto circuit gains; no assay carries its own copy, which is what makes Table T2 a truthful description of the software. Gains are dimensionless multipliers on the synaptic weight of a transmitter class (or, for `g_nav`, on node excitability); 1.0 is vehicle, and every gain is floored at 0.05 because a network with a gain of exactly zero is not informative. The {{mechanism_rules}} rules are listed in full in T2. The four shipped since v0.4 may not be changed silently: -* **Insect nAChR, activating:** `g_ach = max(0.05, 1 + 0.4θ − 1.6θ²)`. Rises at low engagement and falls at high engagement, encoding excitation followed by desensitisation and depolarisation block. At θ = 0.99 it is at the floor, which is why a saturating dose of a nicotinic agonist silences cholinergic transmission in the model rather than merely reducing it — and, as §3.4 of the manuscript shows, why the suppression conclusion is a property of *this* rule. +* **Insect nAChR, activating:** `g_ach = max(0.05, 1 + 0.4θ − 1.6θ²)`. Rises at low engagement and falls at high engagement, encoding excitation followed by desensitisation and depolarisation block. At θ = 0.99 it is at the floor, which is why a saturating dose of a nicotinic agonist silences cholinergic transmission in the model rather than merely reducing it; and, as §3.4 of the manuscript shows, why the suppression conclusion is a property of *this* rule. * **Insect nAChR, blocking:** `g_ach = max(0.05, 1 − θ)`. * **Insect RDL, blocking:** `g_gaba = max(0.05, 1 − θ)`. Blocking the GABA-gated chloride channel removes inhibition. * **Insect RDL, activating:** `g_gaba = max(0.05, 1 + 0.4θ)`. -Four families added in v0.5 share the shape and state a rationale each: GluCl activation uses a larger coefficient (`1 + 0.8θ`) than RDL because avermectin opening is effectively irreversible [38]; AChE inhibition raises an ambient cholinergic tone (`ach_tone = 1 + 2.0θ`) multiplying `g_ach` together with the agonist curve, reproducing the organophosphate excitation-then-block picture; Nav modulation scales node excitability globally (`g_nav = 1 + 1.5θ`), since pyrethroids and DDT act on every cell regardless of transmitter [23], [67]; octopamine agonism uses the smallest coefficient (`1 + 0.5θ`) because it is modulatory. +Four families added in v0.5 share the shape and state a rationale each: GluCl activation uses a larger coefficient (`1 + 0.8θ`) than RDL because avermectin opening is effectively irreversible [37]; AChE inhibition raises an ambient cholinergic tone (`ach_tone = 1 + 2.0θ`) multiplying `g_ach` together with the agonist curve, reproducing the organophosphate excitation-then-block picture; Nav modulation scales node excitability globally (`g_nav = 1 + 1.5θ`), since pyrethroids and DDT act on every cell regardless of transmitter [22], [66]; octopamine agonism uses the smallest coefficient (`1 + 0.5θ`) because it is modulatory. -**The level-C generic rule.** The ablation ladder's topology-only level is given one number per compound instead of a mechanism-specific gain vector. Until v0.6.1 that number was `{{abl_generic_floor_rule}}`, applied to every transmitter alike — a rule that can only depress, so every one of its {{abl_compounds}} × {{abl_concs}} entries was at most zero and the level could not express disinhibition at all. It is now `{{abl_generic_rule}}`: the same generic magnitude, with one bit of sign read off the mechanism table. {{abl_generic_floor_note}} Both are reported, the old one as `C_topology_only_floor` in T13 and T13b. +**The level-C generic rule.** The ablation ladder's topology-only level is given one number per compound instead of a mechanism-specific gain vector. Until v0.6.1 that number was `{{abl_generic_floor_rule}}`, applied to every transmitter alike; a rule that can only depress, so every one of its {{abl_compounds}} × {{abl_concs}} entries was at most zero and the level could not express disinhibition at all. It is now `{{abl_generic_rule}}`: the same generic magnitude, with one bit of sign read off the mechanism table. {{abl_generic_floor_note}} Both are reported, the old one as `C_topology_only_floor` in T13 and T13b. **These coefficients and shapes are modelling choices, not measurements.** That is why the manuscript prespecifies a family of {{stab_family_size}} alternatives (§2.4) and re-derives every conclusion under each, and why the variance analysis nominates this transformation as the assumption worth measuring first. The **transmitter sign table** the rate engine iterates is a second set of asserted coefficients and is published as T21: {{sign_asserted}} of the seven magnitudes ({{sign_asserted_transmitters}}) carry no measurement, are members of no specification family, and are not bounded by the label-permuting null, which permutes assignments and leaves the table itself untouched. ### S1.1 The specification family -Five shapes — monotone-linear, saturating, weakly biphasic, the shipped biphasic rule and a strongly biphasic rule — at five coefficient scales (0.50×, 0.75×, 1.00×, 1.25×, 1.50×), giving {{stab_family_size}} specifications, of which `{{stab_default_spec}}` is the shipped one. Installation is process-global, so specifications are evaluated serially. Full matrix: T14; per-specification diagnostics: T14b. +Five shapes; monotone-linear, saturating, weakly biphasic, the shipped biphasic rule and a strongly biphasic rule; at five coefficient scales (0.50×, 0.75×, 1.00×, 1.25×, 1.50×), giving {{stab_family_size}} specifications, of which `{{stab_default_spec}}` is the shipped one. Installation is process-global, so specifications are evaluated serially. Full matrix: T14; per-specification diagnostics: T14b. | Conclusion | Claim | Retained | Reversed | Undecidable | |---|---|---|---|---| @@ -45,7 +45,7 @@ Five shapes — monotone-linear, saturating, weakly biphasic, the shipped biphas 1. *The shuffle count could not reject.* The topology predicates were evaluated at 6 shuffles per specification, where the smallest attainable empirical permutation probability is 1/7 = 0.143. `p ≤ 0.05` was arithmetically unattainable, so the predicate fell back on a Gaussian `|z| < 2` and C3's 25/25 was a non-rejection under a test that could not reject. The run behind this document uses {{stab_n_shuffles}} shuffles per specification (resolution {{stab_shuffle_resolution}}), which supports α = {{stab_topology_alpha}} with room to spare. 2. *The specification never reached the null engine.* `robustness.mechanism_spec` rebinds `compute_gains` on the assay modules, but the permutation path resolves gains through `flylab.circuit.rate.compute_gains`, which it did not rebind. Every specification therefore fed the **default** gains to its nulls and produced identical effects, so a uniform retention rate across the family was guaranteed by construction. The dependence engine now accepts the gain patch explicitly (`gains_by_compound`), and the per-specification topology verdicts in T14 are computed under the specification they are labelled with. -Both topology conclusions survive the correction ({{stab_C3_retained}}/{{stab_family_size}} and {{stab_C4_retained}}/{{stab_family_size}} after Benjamini-Hochberg across the {{stab_n_structural_tests}} structural tests of the run; {{stab_C3_retained_uncorrected}} and {{stab_C4_retained_uncorrected}} uncorrected), but a retention rate for a conclusion shaped as a *failure to reject* is only as strong as the equivalence behind it. Of C3's {{stab_C3_retained}}, **{{stab_C3_equivalent}} reach equivalence within the prespecified margin and {{stab_C3_indeterminate}} are merely indeterminate**; the corresponding split for C4 is {{stab_C4_equivalent}} / {{stab_C4_indeterminate}}. Every specification's topology cell is run on the same shuffle stream, so the structural tests are positively dependent rather than independent; Benjamini-Hochberg controls the FDR under positive dependence, but the fractions describe this shuffle stream and are not {{stab_family_size}} independent replications. +C4 remains topology-dependent in {{stab_C4_retained}}/{{stab_family_size}} specifications after Benjamini-Hochberg adjustment across the {{stab_n_structural_tests}} component structural tests of the run ({{stab_C4_retained_uncorrected}} unadjusted): it is retained when at least one adjusted structural test rejects. C3 instead has no structural rejection in {{stab_C3_retained}}/{{stab_family_size}}, unchanged from {{stab_C3_retained_uncorrected}} unadjusted retentions. Adjustment cannot strengthen a conclusion shaped as a failure to reject or establish absence of dependence. Of C3's {{stab_C3_retained}}, **{{stab_C3_equivalent}} pass the configured descriptive median-gap criterion and {{stab_C3_indeterminate}} are merely indeterminate**. The adjustment targets the component tests; because a topology label is an OR over two modes, it does not establish FDR control for specification-level labels or conclusion counts. Every specification's topology cell is run on the same shuffle stream, so the tests are dependent, and shared draws do not by themselves establish the positive-regression dependence condition required for the usual Benjamini-Hochberg guarantee. Adjusted values are therefore reported as a sensitivity analysis, not an unconditional FDR guarantee; the fractions are not {{stab_family_size}} independent replications. Specifications reversing C1: {{stab_C1_failing}}. Mean SI gap over the nicotinic set under the shipped rule: {{stab_mean_gap_nicotinic_default}} log₁₀ units; averaged over the {{stab_mean_gap_nicotinic_monotone_n}} reversing specifications for which the index is defined: {{stab_mean_gap_nicotinic_monotone}}. **The v0.6 supplement said the monotone rules "buffer at least as much, never less". That is false.** The per-specification range runs {{stab_mean_gap_nicotinic_monotone_max}} to {{stab_mean_gap_nicotinic_monotone_min}}, {{stab_monotone_buffer_less_n}} of them buffer *less* than the shipped rule ({{stab_monotone_buffer_less}}), and the aggregate is carried by the extreme of the range. C5 survives the specifications that reverse C1, but not for the reason the previous version gave. @@ -53,57 +53,57 @@ Specifications reversing C1: {{stab_C1_failing}}. Mean SI gap over the nicotinic `flylab/analysis/nullmodels.py` and `flylab/analysis/dependence.py` degrade the same committed cut, in increasing order of destruction: -* **`sign_permute_weight_matched`** — a constrained shuffle of consensus transmitter labels: random pairs of differently-labelled nodes are swapped, and a swap is kept only if every tracked transmitter's share of *total outgoing synaptic weight* stays within {{bal_tol}} of the real graph's. The node count of each label is preserved exactly, the weighted share to within the tolerance. This is the ladder's rank-3 rung. -* **`sign_permute`** — the plain permutation of the node label array. It is a **joint target-set-and-sign null**, not a rung: see below. -* **`weight_permute`** — synapse counts are permuted across edges. Topology and transmitters are kept; the pairing between wiring and strength is destroyed. -* **`rewire_degree_preserving`** — double-edge swaps, roughly 10 × |E| accepted swaps. Every node keeps its in-degree, out-degree and transmitter; the wiring pattern is destroyed. -* **`erdos_renyi`** — the same node and edge counts, endpoints drawn uniformly, weights drawn from the empirical weight list. Only the size of the graph and its census survive. +* **`sign_permute_weight_matched`**; a constrained shuffle of consensus transmitter labels: random pairs of differently-labelled nodes are swapped, and a swap is kept only if every tracked transmitter's share of *total outgoing synaptic weight* stays within {{bal_tol}} of the real graph's. The node count of each label is preserved exactly, the weighted share to within the tolerance. This is the ladder's rank-3 rung. +* **`sign_permute`**; the plain permutation of the node label array. It is a **joint target-set-and-sign null**, not a rung: see below. +* **`weight_permute`**; synapse counts are permuted across edges. Topology and transmitters are kept; the pairing between wiring and strength is destroyed. +* **`rewire_degree_preserving`**; double-edge swaps, roughly 10 × |E| accepted swaps. Every node keeps its in-degree, out-degree and transmitter; the wiring pattern is destroyed. +* **`erdos_renyi`**; the same node and edge counts, endpoints drawn uniformly, weights drawn from the empirical weight list. Only the size of the graph and its census survive. -**Why `sign_permute` is not a transmitter-identity null.** The v0.6 supplement stated that permuting labels preserves the transmitter histogram "and therefore the global excitation/inhibition balance, preserved *exactly*". That is wrong, and the error was load-bearing: out-strength is heterogeneous across nodes, so the *weighted* balance moves. Measured (T23), the cholinergic share of total outgoing synaptic weight on the `named` cut is {{bal_real_ach_share}} on the real graph against {{bal_plain_ach_mean}} ± {{bal_plain_ach_sd}} across {{bal_n}} permutations, range {{bal_plain_ach_min}}–{{bal_plain_ach_max}}, putting the real graph at the {{bal_plain_ach_percentile}}th percentile of its own null. Because a gain patch acts *through* transmitter identity, the mode moves the drug's target set and the sign matrix at once, and the extra variance makes it conservative. The weight-matched null holds every tracked share to within {{bal_matched_max_deviation}}. Substituting it does not change imidacloprid's permutation probability but it shrinks the equivalence gap from {{dep_named_imidacloprid_gap_sign_permute}} Hz to {{dep_named_imidacloprid_gap_sign_permute_weight_matched}} Hz against a margin of {{dep_named_imidacloprid_delta}} Hz. +**Why `sign_permute` is not a transmitter-identity null.** The v0.6 supplement stated that permuting labels preserves the transmitter histogram "and therefore the global excitation/inhibition balance, preserved *exactly*". That is wrong, and the error was load-bearing: out-strength is heterogeneous across nodes, so the *weighted* balance moves. Measured (T23), the cholinergic share of total outgoing synaptic weight on the `named` cut is {{bal_real_ach_share}} on the real graph against {{bal_plain_ach_mean}} ± {{bal_plain_ach_sd}} across {{bal_n}} permutations, range {{bal_plain_ach_min}}–{{bal_plain_ach_max}}, putting the real graph at the {{bal_plain_ach_percentile}}th percentile of its own null. Because a gain patch acts *through* transmitter identity, the mode moves the drug's target set and the sign matrix at once, and its changed variance can affect rejection rates; conservativeness is not established in general. The weight-matched null holds every tracked share to within {{bal_matched_max_deviation}}. Substituting it changes the null ensemble and its probability, while shrinking the median gap from {{dep_named_imidacloprid_gap_sign_permute}} rate-model units to {{dep_named_imidacloprid_gap_sign_permute_weight_matched}} rate-model units against a margin of {{dep_named_imidacloprid_delta}} rate-model units. Every mode preserves the node set, the node order and the seed block, and never creates self-loops or duplicate edges. The effect is always `treated − vehicle` with **both arms on the same graph**. The reported probability is the empirical two-sided permutation *p*, `(k+1)/(n+1)`, with resolution `1/(n+1)`; on a landscape the same `n` shuffled graphs are reused by every cell (a paired design), and shuffle *i* depends only on `(seed, i)`, so each cell's numbers are identical to running its profile alone. **Why `z` is secondary.** A null distribution built on degraded graphs is not normal, and for the Erdős–Rényi mode its variance can collapse, producing numerically enormous *z* values that mean only "decisively different". -### S2.1 The three-way verdict and the equivalence margin +### S2.1 The three-way diagnostic and median-gap tolerance -Failing to reject a null is not evidence of equivalence, and none of the software's return values says it is. Each mode gets exactly one of `distinguishable`, `equivalent_within_tolerance` or `indeterminate`, against a margin δ prespecified as {{dep_named_imidacloprid_delta_frac}} of the **vehicle readout** of the real graph. The vehicle readout is the right scale because it is in the units of the readout, does not depend on the drug, on the null ensemble or on the number of shuffles, and is the same for every mode of a cell, so an equivalence claim made against it is comparable across the ladder. It is one tenth of the 50 % relative change the selectivity module treats as a circuit response. A cell whose baseline is undefined or zero can form no margin; its non-rejected modes are then `indeterminate`, which is the correct answer rather than a missing one. +Failing to reject a null is not evidence of equivalence. The implementation first assesses the empirical permutation probability. For a non-rejected mode, it compares the absolute real-effect to null-median gap against a configured tolerance δ = {{dep_named_imidacloprid_delta_frac}} times the real graph's vehicle readout. Passing this point-gap check returns the legacy API string `equivalent_within_tolerance`; throughout this paper this is a **descriptive within-margin diagnostic**, not inferential equivalence. No confidence interval for the median gap, TOST procedure, distributional closeness criterion or error-controlled equivalence decision is computed. A noisy or broad null can pass the criterion. A cell with an undefined or zero vehicle value cannot form a positive margin and remains `indeterminate`. The vehicle-based tolerance is a modelling convention, not a measured biological relevance bound or choice registered before inspecting results. -**The confirmatory profile in full.** For imidacloprid at 1 µM on `named`, at {{dep_n}} permutations against a margin of {{dep_named_imidacloprid_delta}} Hz, the per-mode probabilities and verdicts are: weight permutation *p* = {{dep_named_imidacloprid_p_weight_permute}}, {{dep_named_imidacloprid_verdict_weight_permute}}; degree-preserving rewiring *p* = {{dep_named_imidacloprid_p_rewire_degree_preserving}}, {{dep_named_imidacloprid_verdict_rewire_degree_preserving}}; the weight-matched transmitter null *p* = {{dep_named_imidacloprid_p_sign_permute_weight_matched}}, {{dep_named_imidacloprid_verdict_sign_permute_weight_matched}}; the joint label-permutation null *p* = {{dep_named_imidacloprid_p_sign_permute}}, {{dep_named_imidacloprid_verdict_sign_permute}}; and the Erdős–Rényi control *p* = {{dep_named_imidacloprid_p_erdos_renyi}}, {{dep_named_imidacloprid_verdict_erdos_renyi}}. The resulting necessary level is `{{dep_named_imidacloprid_level}}` with verdict {{dep_named_imidacloprid_level_verdict}}. +**The reference profile in full.** For imidacloprid at 1 µM on `named`, at {{dep_n}} permutations against a margin of {{dep_named_imidacloprid_delta}} rate-model units, the per-mode probabilities and verdicts are: weight permutation *p* = {{dep_named_imidacloprid_p_weight_permute}}, {{dep_named_imidacloprid_verdict_weight_permute}}; degree-preserving rewiring *p* = {{dep_named_imidacloprid_p_rewire_degree_preserving}}, {{dep_named_imidacloprid_verdict_rewire_degree_preserving}}; the weight-matched transmitter null *p* = {{dep_named_imidacloprid_p_sign_permute_weight_matched}}, {{dep_named_imidacloprid_verdict_sign_permute_weight_matched}}; the joint label-permutation null *p* = {{dep_named_imidacloprid_p_sign_permute}}, {{dep_named_imidacloprid_verdict_sign_permute}}; and the Erdős–Rényi control *p* = {{dep_named_imidacloprid_p_erdos_renyi}}, {{dep_named_imidacloprid_verdict_erdos_renyi}}. The resulting necessary level is `{{dep_named_imidacloprid_level}}` with verdict {{dep_named_imidacloprid_level_verdict}}. -The **necessary information level** is the weakest rung the test could not distinguish from the real cut, reported with its verdict. Ranks 2 and 3 are not strictly nested — `weight_permute` keeps transmitter identity and loses the weight-topology pairing, the transmitter nulls the reverse — so a richer model can be distinguishable where a weaker one is not. Such cells are **non-monotone** and are flagged: {{dep_land_non_monotone}} of {{dep_land_cells}} on the `named` cut ({{dep_land_non_monotone_cells}}) and {{dep_land_taste_non_monotone}} of {{dep_land_taste_cells}} on `taste_motor`. Their level must be read as "the cheapest graph model this test cannot tell apart from the real one", never as "everything above it is indistinguishable too". +The **necessary information level** is the weakest rung the test could not distinguish from the real cut, reported with its verdict. Ranks 2 and 3 are not strictly nested; `weight_permute` keeps transmitter identity and loses the weight-topology pairing, the transmitter nulls the reverse; so a richer model can be distinguishable where a weaker one is not. Such cells are **non-monotone** and are flagged: {{dep_land_non_monotone}} of {{dep_land_cells}} on the `named` cut ({{dep_land_non_monotone_cells}}) and {{dep_land_taste_non_monotone}} of {{dep_land_taste_cells}} on `taste_motor`. Their level must be read as "the cheapest graph model this test cannot tell apart from the real one", never as "everything above it is indistinguishable too". -### S2.2 Effect floor, multiplicity, and confirmatory versus exploratory +### S2.2 Effect floor and multiplicity -**Effect floor.** An absolute floor of 1e-6 Hz against a ~6.6 Hz baseline admits effects at numerical-noise level, so the effective floor is `max(absolute, 1 % × |vehicle readout|)`. Landscapes report class counts under both. On the `named` cut the relative floor moves {{dep_land_below_relative_floor}} cells into `no-effect` ({{dep_land_below_relative_floor_cells}}), changing the counts from {{dep_land_topology_dependent_absfloor}} / {{dep_land_composition_dominated_absfloor}} / {{dep_land_no_effect_absfloor}} to {{dep_land_topology_dependent}} / {{dep_land_composition_dominated}} / {{dep_land_no_effect}}. +**Effect floor.** An absolute floor of 1e-6 rate-model units against a ~6.6 rate-model units baseline admits effects at numerical-noise level, so the effective floor is `max(absolute, 1 % × |vehicle readout|)`. Landscapes report class counts under both. On the `named` cut the relative floor moves {{dep_land_below_relative_floor}} cells into `no-effect` ({{dep_land_below_relative_floor_cells}}), changing the counts from {{dep_land_topology_dependent_absfloor}} / {{dep_land_composition_dominated_absfloor}} / {{dep_land_no_effect_absfloor}} to {{dep_land_topology_dependent}} / {{dep_land_composition_dominated}} / {{dep_land_no_effect}}. -**Multiplicity.** A landscape is a screen. Benjamini-Hochberg is applied across the structural tests of the run ({{dep_land_n_structural_tests}} on `named`, {{dep_land_taste_n_structural_tests}} on `taste_motor`), the class is decided on the adjusted probabilities, and both counts are reported. Because an empirical permutation probability cannot fall below `1/(n+1)`, the result also reports how many tests must sit at that floor before anything can be rejected, so that a corrected count of zero is never mistaken for a finding (`fdr_can_reject` = {{dep_land_fdr_can_reject}} here). +**Multiplicity.** A landscape is a screen. Benjamini-Hochberg is applied across the component structural tests of the run ({{dep_land_n_structural_tests}} on `named`, {{dep_land_taste_n_structural_tests}} on `taste_motor`), the class is decided from those adjusted probabilities, and both adjusted-test and unadjusted counts are reported. A topology-dependent cell is an OR decision over its two structural modes, so test-level adjustment does not by itself establish FDR control for the resulting cell labels or counts. The shared shuffle draws make the tests dependent; reuse alone does not establish the positive-regression condition required for the usual Benjamini-Hochberg guarantee. The adjusted values are therefore a multiplicity sensitivity analysis. Because an empirical permutation probability cannot fall below `1/(n+1)`, the result also reports how many tests must sit at that floor before anything can be rejected, so that an adjusted-test count of zero is never mistaken for a finding (`fdr_can_reject` = {{dep_land_fdr_can_reject}} here). -**Confirmatory versus exploratory.** The imidacloprid and fipronil profiles at *n* ≥ 1000 are prespecified planned comparisons and carry no correction; everything else is exploratory and says so in its warnings. Fipronil at 1 µM is the clean illustration: confirmatory *p* = {{dep_named_fipronil_min_structural_p}}, but inside the exploratory landscape its adjusted probability is {{dep_land_fipronil_q}} and the cell does not survive correction ({{dep_land_fipronil_class}} corrected against {{dep_land_fipronil_class_raw}} raw). +**Reference profiles versus exploratory screens.** The imidacloprid and fipronil profiles at *n* ≥ 1000 receive higher computational effort and are reported with nominal probabilities. They were refined during iterative development and are not pre-registered confirmations. Fipronil at 1 µM has reference-profile *p* = {{dep_named_fipronil_min_structural_p}}; inside the library screen its adjusted probability is {{dep_land_fipronil_q}}, with {{dep_land_fipronil_class}} corrected versus {{dep_land_fipronil_class_raw}} raw labels. These differ in both permutation effort and adjustment, so the change cannot be attributed to adjustment alone. Labels based on non-rejection do not establish equivalence, including after adjustment. ### S2.3 Does the instrument work? Ground-truth recovery and power Without this, a negative from the ladder is indistinguishable from an under-powered test. A synthetic cut of {{val_recovery_nodes}} nodes and about {{val_recovery_edges}} edges is built with log-normal weights and a transmitter composition close to the real cut's, and a directed cycle of cholinergic nodes passing through the driven seeds is planted in it at a known strength. The cycle is recurrent cholinergic gain, so a patch that collapses `g_ach` removes an amplification that exists only while the cycle is intact: the drug *contrast*, not merely the rate, depends on the wiring. Strength 0 plants nothing and is the negative control. Nothing here touches the connectome or the compound library. -{{val_recovery_statement}} Over the power grid ({{val_power_replicates}} independently generated graphs per cell, planted strengths {{val_power_strengths}}, permutation counts {{val_power_ns}}) the detection rate by strength is {{val_power_by_strength}} and by permutation count {{val_power_by_n}}; the empirical false-positive rate on the unplanted control is {{val_power_control_hits}} of {{val_power_control_runs}} ({{val_power_fpr}}). Effect size dominates: raising the permutation count buys far less than raising the planted strength. Full grid: T22, Fig. F16. +{{val_recovery_statement}} Over the power grid ({{val_power_replicates}} independently generated graphs per cell, planted strengths {{val_power_strengths}}, permutation counts {{val_power_ns}}) the detection rate by strength is {{val_power_by_strength}} and by permutation count {{val_power_by_n}}; the empirical false-positive rate on the unplanted control is {{val_power_control_hits}} of {{val_power_control_runs}} ({{val_power_fpr}}). The grid is small, and changes with strength and permutation count are descriptive. Six detections from six networks would still have a two-sided Wilson 95% interval of approximately 0.61–1.00; zero false positives in eighteen independent trials would allow an upper bound near 0.18. Dependence between reused networks/settings further limits pooled interpretation. The unplanted case also omits cycle edges and associated transmitter relabelling, so it is not a perfectly matched zero-strength intervention. Full grid: T22, Fig. F16. **Permutation counts in this document.** {{dep_n}} per mode for the `named` profiles (resolution {{dep_p_floor}}), {{dep_n_taste}} for the `taste_motor` path arm, {{dep_n_landscape}} for each of the {{dep_land_cells}} `named` landscape cells and {{dep_n_landscape_taste}} for each of the {{dep_land_taste_cells}} `taste_motor` cells. The sweep (T11) shows the verdict at α = 0.05 settling from *n* ≈ {{dep_verdict_stable_from_n}}–{{dep_verdict_stable_to_n}}, and a mid-range *p* reaching ±0.02 of its final value between *n* = {{dep_p_within_0p02_from_n}} and {{dep_p_within_0p02_worst_n}}. -### S2.4 Where does the verdict settle with the size of the cut? +### S2.4 How do labels vary across the examined cut ladder? -The verdict reversed between the two committed cuts, and two cuts are not a trend. `flylab.analysis.scale.dependence_vs_scale` runs the *same* dependence profile (`dependence_profile`, unchanged) across a ladder of nested cuts of {{scale_ladder_sizes}} cells grown from the taste-motor seed set by one fixed rule, so a verdict can be compared across size without the construction changing underneath it. The two committed cuts are built by a different recipe, so the clean scale axis is the `scale_*` rungs alone; a sequence that interleaves the committed cuts mixes a size effect with a construction effect, and `verdict_stability(recipe="scale_ladder")` reports each separately. A verdict counts as *settled* only if it is the same on the two largest cuts run with no change above that point — one rung of agreement is n = 1 and is not stability. +The label reversed between the two committed cuts, and two cuts are not a trend. `flylab.analysis.scale.dependence_vs_scale` repeats the same profile across nested cuts of {{scale_ladder_sizes}} cells grown from the taste-motor seeds by one fixed rule. Holding that recipe fixed removes one construction difference, but size, edge count, density, composition, paths, readout support and permutation effort still co-vary. The two committed cuts use another recipe, so interleaving them with `scale_*` adds a construction confound. The legacy *settled* field is only a matching-label indicator: it is true when the two largest cuts run have the same label with no later change; it is not statistical stability or convergence. -Cost is the constraint, not correctness. From the per-shuffle and per-edge constants measured on this machine (T26), a profile at *n* = {{scale_frontier_n_paper}} on the {{scale_top_rung}} rung is {{scale_top_profile_hours}} hours on one core and the landscape over the specification family is {{scale_top_landscape_days}} days per specification; a whole-CNS profile at the weight-5 floor is {{scale_whole_cns_profile_hours}} hours and the dense LIF matrix stops fitting in memory well before it. The permutation count must therefore fall with the cut, and `permutation_budget` refuses to go below the resolution floor `1/α − 1`, so every rung records the *n* it was run at. +Computational cost restricts permutation resolution and therefore the strength of the statistical conclusions. From the per-shuffle and per-edge constants measured on this machine (T26), a profile at *n* = {{scale_frontier_n_paper}} on the {{scale_top_rung}} rung is {{scale_top_profile_hours}} hours on one core and the landscape over the specification family is {{scale_top_landscape_days}} days per specification; a whole-CNS profile at the weight-5 floor is {{scale_whole_cns_profile_hours}} hours and the dense LIF matrix stops fitting in memory well before it. The permutation count must therefore fall with the cut, and `permutation_budget` refuses to go below the resolution floor `1/α − 1`, so every rung records the *n* it was run at. -The ladder's smallest rung, `scale_1k`, is already the decisive contrast with the two cuts of T19: {{scale_1k_nodes}} nodes, {{scale_1k_edges}} edges, mean degree {{scale_1k_mean_degree}}, in-star {{scale_1k_in_star}} — *fewer* nodes than `named` and a densely recurrent cut nothing like it. +The ladder's smallest rung, `scale_1k`, is a descriptive contrast with the two cuts of T19: {{scale_1k_nodes}} nodes, {{scale_1k_edges}} edges, mean degree {{scale_1k_mean_degree}}, in-star {{scale_1k_in_star}}; fewer nodes than `named`, but a much denser non-in-star extract. -**Result** (T27; the study record is read from `scale_study.json` beside the paper artifacts, or from `data/derived/` when it is promoted there). Over the {{scale_study_n_cuts}} cuts run ({{scale_study_cuts}}), the composition-dominated verdict survives on {{scale_study_n_composition_dominated}} — {{scale_study_composition_dominated_cuts}} — so {{scale_study_density_not_size}}. Per-rung sequences: {{scale_study_imidacloprid_sequence}}. Fipronil is topology-dependent on every rung. In summary: {{scale_study_statement}} On the clean `scale_*` ladder the verdicts settle: {{scale_study_settled_clean_ladder}}. At {{scale_study_every_null_cuts}} the necessary information level is the real network itself, i.e. the prediction is distinguishable from *every* rung of the ladder. Imidacloprid's relative effect stays within {{scale_study_imidacloprid_relative_range}} of the vehicle readout across the ladder, so the verdict changes while the quantity being measured does not. +**Result** (T27; the study record is read from `scale_study.json` beside the paper artifacts, or from `data/derived/` when it is promoted there). Across the {{scale_study_n_cuts}} cuts examined ({{scale_study_cuts}}), the composition-dominated label occurs on {{scale_study_n_composition_dominated}} cut, {{scale_study_composition_dominated_cuts}}. Because `scale_1k` has fewer nodes but a different label, node count alone does not order these observations. The imidacloprid sequence is {{scale_study_imidacloprid_sequence}}; fipronil is labelled topology-dependent on every examined rung. The legacy matching-label indicator on the `scale_*` subset is {{scale_study_settled_clean_ladder}}; this records final-label agreement, not convergence. At {{scale_study_every_null_cuts}} all tested null modes are rejected at the rung's stated effort. Imidacloprid's relative effect stays within {{scale_study_imidacloprid_relative_range}} of the vehicle readout, but similar aggregate magnitude does not make the cell populations, drive support or other extract properties identical. -**Two caveats, both reported rather than absorbed.** The permutation budget falls with the edge count ({{scale_study_n_range}} draws, resolution {{scale_study_resolution_range}}): {{scale_study_coarse_note}}, so those rungs confirm a verdict already settled lower down rather than establishing it. And `named` and `taste_motor` are hops-limited neighbourhoods while the `scale_*` rungs are nested ranked-BFS cuts, so a difference *between* the two families confounds scale with construction; differences *within* the `scale_*` rungs do not. +**Two caveats, both reported rather than absorbed.** The permutation budget falls with edge count ({{scale_study_n_range}} draws, resolution {{scale_study_resolution_range}}): {{scale_study_coarse_note}}, so the top-rung labels are resolution-limited observations, not independent confirmations. And `named` and `taste_motor` are hops-limited neighbourhoods while `scale_*` uses nested ranked-BFS cuts. Comparisons between those families confound recipe with structure; comparisons within `scale_*` hold the recipe fixed but do not isolate size, density, recurrence or the other properties that change together. ## S3. Runtimes and their calibration -**Rate runtime.** `flylab/circuit/rate.py` holds a dense signed weight matrix `W[j,i] = sign(nt_i) × synapses(i→j)`, with the signs of Table T21. **That matrix is then row-normalised**: `W_signed = W / max(Σ_i |W[j,i]|, 1)`, one denominator per postsynaptic cell. A gain patch afterwards scales the columns by the presynaptic transmitter's gain, so `W_eff = W_signed · diag(g)` and **each cell's recurrent input is a composition-weighted average of its presynaptic gains, normalised by its own total absolute input**. The update is a clipped leaky iteration `r ← clip((1 − α)·r + α·g_nav·(d + W_eff·r), 0, r_max)` with α = 0.3, 80 steps, `r_max` = 300 Hz and *d* the per-node external drive in Hz. It is deterministic, and one evaluation of the {{named_nodes}}-node graph costs about 50 ms with single-threaded BLAS — which is what makes {{dep_n}}-permutation profiles and an {{dep_land_cells}}-cell landscape affordable at all. +**Rate runtime.** `flylab/circuit/rate.py` holds a dense signed weight matrix `W[j,i] = sign(nt_i) × synapses(i→j)`, with the signs of Table T21. **That matrix is then row-normalised**: `W_signed = W / max(Σ_i |W[j,i]|, 1)`, one denominator per postsynaptic cell. A gain patch afterwards scales the columns by the presynaptic transmitter's gain, so `W_eff = W_signed · diag(g)` and **each cell's recurrent input is a composition-weighted average of its presynaptic gains, normalised by its own total absolute input**. The update is a clipped leaky iteration `r ← clip((1 − α)·r + α·g_nav·(d + W_eff·r), 0, r_max)` with α = 0.3, 80 steps, `r_max` = 300 rate-model units and *d* the per-node external drive in rate-model units. It is deterministic, and one evaluation of the {{named_nodes}}-node graph costs about 50 ms with single-threaded BLAS; which is what makes {{dep_n}}-permutation profiles and an {{dep_land_cells}}-cell landscape affordable at all. The normalisation is an **asserted modelling choice**, not a measurement, and it is consequential for RQ2: a readout in which each cell averages its presynaptic gains builds part of "composition-dominated" into itself. Three modes are implemented and swept in §3.3 (T25): @@ -113,37 +113,37 @@ The normalisation is an **asserted modelling choice**, not a measurement, and it | `none` | 1 | {{norm_note_none}} | | `degree` | `max(in-degree, 1)` | {{norm_note_degree}} | -The composition-versus-full rank correlation at 1 µM is {{abl_rho_composition_row_abs}} under `row_abs`, {{abl_rho_composition_none}} under `none` and {{abl_rho_composition_degree}} under `degree` ({{abl_rho_composition_degree_1e8}} at 10 nM): the composition verdict survives removing the normalisation but not replacing it with a degree-corrected one. Levels are not comparable across modes — the unnormalised operator is supercritical on this cut, so the `r_max` clip rather than the wiring bounds the rates — only the orderings are. +The composition-versus-full rank correlation at 1 µM is {{abl_rho_composition_row_abs}} under `row_abs`, {{abl_rho_composition_none}} under `none` and {{abl_rho_composition_degree}} under `degree` ({{abl_rho_composition_degree_1e8}} at 10 nM): the composition verdict survives removing the normalisation but not replacing it with a degree-corrected one. Levels are not comparable across modes; the unnormalised operator is supercritical on this cut, so the `r_max` clip rather than the wiring bounds the rates; only the orderings are. -**LIF runtime.** `flylab/circuit/lif.py` is a current-based leaky integrate-and-fire network with the parameters of Shiu *et al.* [7]: τ_m = 20 ms, V_rest = −52 mV, V_th = −45 mV, V_reset = −52 mV, refractory 2.2 ms, exponential synaptic current with τ_s = 5 ms, forward Euler at dt = 0.1 ms. A presynaptic spike on edge *i*→*j* injects `w_scale × synapses × sign(nt_i) × gain(nt_i)` mV into *j*. +**LIF runtime.** `flylab/circuit/lif.py` is a current-based leaky integrate-and-fire network with the parameters of Shiu *et al.* [6]: τ_m = 20 ms, V_rest = −52 mV, V_th = −45 mV, V_reset = −52 mV, refractory 2.2 ms, exponential synaptic current with τ_s = 5 ms, forward Euler at dt = 0.1 ms. A presynaptic spike on edge *i*→*j* injects `w_scale × synapses × sign(nt_i) × gain(nt_i)` mV into *j*. -Two choices are modelling decisions, not literature values. *The synaptic scale is a one-time calibration scan*: `w_scale` = 0.05 mV per synapse and a drive weight of 0.5 mV per event were chosen once, by scanning, so that vehicle MN9 cells fire in a plausible band under the standard {{veto_sugar_hz}} Hz seed drive. No recording was used and no error was minimised against data, so this does not consume the one fitted parameter the project allows — but it is a tuned constant. *A background Poisson drive is required, and it is a choice*: excising a 1-hop neighbourhood removes every input from outside it, so every non-seed cell receives an independent Poisson drive at 0.65 × the seed rate, standing in for the amputated inputs. This is the most consequential free choice in the LIF arm; it sets the operating point and it is why the two engines report sugar-driven MN9 at {{veto_rate_vehicle_sugar}} Hz (rate) against {{veto_lif_vehicle_sugar}} Hz (LIF). +Two choices are modelling decisions, not literature values. *The synaptic scale is a one-time calibration scan*: `w_scale` = 0.05 mV per synapse and a drive weight of 0.5 mV per event were chosen once, by scanning, so that vehicle MN9 cells fire in a plausible band under the standard {{veto_sugar_hz}} Hz input drive seed drive. No recording was used and no error was minimised against data, so this does not consume the one fitted parameter the project allows; but it is a tuned constant. *A background Poisson drive is required, and it is a choice*: excising a 1-hop neighbourhood removes every input from outside it, so every non-seed cell receives an independent Poisson drive at 0.65 × the seed rate, standing in for the amputated inputs. This is the most consequential free choice in the LIF arm; it sets the operating point and it is why the two engines report sugar-driven MN9 at {{veto_rate_vehicle_sugar}} rate-model units (rate) against {{veto_lif_vehicle_sugar}} LIF spikes/s (LIF). ## S4. Supporting capabilities These are implemented, tested and used, but are supporting capabilities rather than contributions of this paper. -**Exposure.** `flylab/pharm/exposure.py` maps a dose and route onto *C(t)* and hence engagement(t) in one compartment. Its constants are class placeholders chosen so that the three routes differ in the direction the literature describes, and are fitted to no *Drosophila* ADME dataset because that dataset does not exist: of the seven quantities such a model needs, `data/literature/fly_pharmacokinetics.yaml` records only metabolite identity as well established [52], [54], [55], [58]. Adult haemolymph volume, adult body mass, cuticular penetration rate and any *Drosophila* elimination half-life are `null`; what is recorded is 25 nL recoverable per adult [57], 174 ± 81 nL in third-instar larvae [56], a water-partitioning measurement [53], an ethanol entry that could not supply a rate [65], and an imidacloprid half-life of 4.5–5 h in *honeybee* [58] — the only verified insect half-life for a library compound. Two structural warnings travel with every profile: haemolymph is the lowest-radioactivity compartment for imidacloprid, so a one-compartment haemolymph model under-estimates CNS exposure [59]; and a substantial metabolic route runs through the gut microbiome, so a single-enzyme clearance term is wrong for this compound [54]. +**Exposure.** `flylab/pharm/exposure.py` maps a dose and route onto *C(t)* and hence engagement(t) in one compartment. Its constants are class placeholders chosen so that the three routes differ in the direction the literature describes, and are fitted to no *Drosophila* ADME dataset because that dataset does not exist: of the seven quantities such a model needs, `data/literature/fly_pharmacokinetics.yaml` records only metabolite identity as well established [51], [53], [54], [57]. Adult haemolymph volume, adult body mass, cuticular penetration rate and any *Drosophila* elimination half-life are `null`; what is recorded is 25 nL recoverable per adult [56], 174 ± 81 nL in third-instar larvae [55], a water-partitioning measurement [52], an ethanol entry that could not supply a rate [64], and an imidacloprid half-life of 4.5–5 h in *honeybee* [57]; the only verified insect half-life for a library compound. Two structural warnings travel with every profile: haemolymph is the lowest-radioactivity compartment for imidacloprid, so a one-compartment haemolymph model under-estimates CNS exposure [58]; and a substantial metabolic route runs through the gut microbiome, so a single-enzyme clearance term is wrong for this compound [53]. -**Genotype.** `flylab/pharm/genotype.py` applies published target-site resistance alleles as multiplicative potency shifts from `data/literature/resistance_alleles.yaml` [13]–[25], enforcing one rule: **a shift applies per compound, never per receptor.** Two published findings make the naive implementation wrong — *Rdl* A301S shifts GABA potency {{geno_gaba_rdl_A301S_nlug_fold}}-fold but fipronil only {{geno_fipronil_rdl_A301S_nlug_fold}}-fold on the expressed planthopper receptor [15], and *para* M918T shifts deltamethrin {{geno_deltamethrin_para_M918T_superkdr_fold}}-fold while leaving DDT at {{geno_ddt_para_M918T_superkdr_fold}}-fold [22]. An allele/compound pair with no sourced number is left unchanged and listed under `unshifted`, never extrapolated. One row is deliberately ugly: *Rdl* A301G in *D. simulans* carries a {{geno_fipronil_rdl_A302G_dsim_fold}}-fold shift derived from a whole-animal resistance ratio, which drives fipronil engagement to {{geno_fipronil_rdl_A302G_dsim_occ}} and is almost certainly an overstatement of the target-site change [19]. The notebook says so. Panel: T9, F9. +**Genotype.** `flylab/pharm/genotype.py` applies published target-site resistance alleles as multiplicative potency shifts from `data/literature/resistance_alleles.yaml` [12]–[24], enforcing one rule: **a shift applies per compound, never per receptor.** Two published findings make the naive implementation wrong; *Rdl* A301S shifts GABA potency {{geno_gaba_rdl_A301S_nlug_fold}}-fold but fipronil only {{geno_fipronil_rdl_A301S_nlug_fold}}-fold on the expressed planthopper receptor [14], and *para* M918T shifts deltamethrin {{geno_deltamethrin_para_M918T_superkdr_fold}}-fold while leaving DDT at {{geno_ddt_para_M918T_superkdr_fold}}-fold [21]. An allele/compound pair with no sourced number is left unchanged and listed under `unshifted`, never extrapolated. One row is deliberately ugly: *Rdl* A301G in *D. simulans* carries a {{geno_fipronil_rdl_A302G_dsim_fold}}-fold shift derived from a whole-animal resistance ratio, which drives fipronil engagement to {{geno_fipronil_rdl_A302G_dsim_occ}} and is almost certainly an overstatement of the target-site change [18]. The notebook says so. Panel: T9, F9. -**Mixtures.** `flylab/pharm/mixtures.py` implements Bliss independence [45] and Loewe additivity [48], [49] as explicit null models, following the response-surface framing of Greco *et al.* [46], and selects between them per pair: Loewe where the components share a receptor key, Bliss where their targets are disjoint. The three-case comparison against the honeybee dataset [51] is a qualitative concordance check and not validation. Table T5. +**Mixtures.** `flylab/pharm/mixtures.py` implements Bliss independence [44] and Loewe additivity [47], [48] as explicit null models, following the response-surface framing of Greco *et al.* [45], and selects between them per pair: Loewe where the components share a receptor key, Bliss where their targets are disjoint. The three-case comparison against the honeybee dataset [50] is a qualitative concordance check and not validation. Table T5. -**Expression weighting.** `flylab/pharm/expression.py` relaxes the assumption that every cell expresses the receptor, using class-level statements from single-cell atlases [39]–[44] mapped by hand onto MaleCNS superclasses, and interpolates each node's gain between 1.0 and the full scalar gain. It is a **sensitivity layer and never the primary model**: none of the cited atlases uses MaleCNS superclass labels, so every mapping row is tagged `cross_atlas_inference`; coverage is {{expr_coverage_overall}} of cells; and receptor expression in identified adult motor neurons is a confirmed gap, so MN9 carries the `unknown` default. Weighting shifts MN9 by {{expr_mn9_delta_hz}} Hz at 1 µM imidacloprid, which we report as the size of the uncertainty rather than as a correction. +**Expression weighting.** `flylab/pharm/expression.py` relaxes the assumption that every cell expresses the receptor, using class-level statements from single-cell atlases [38]–[43] mapped by hand onto MaleCNS superclasses, and interpolates each node's gain between 1.0 and the full scalar gain. It is a **sensitivity layer and never the primary model**: none of the cited atlases uses MaleCNS superclass labels, so every mapping row is tagged `cross_atlas_inference`; {{expr_coverage_overall}} is the fraction of cell–receptor slots with a non-unknown ordinal class mapping, not a fraction of cells with measured expression; and receptor expression in identified adult motor neurons is a confirmed gap, so MN9 carries the `unknown` default. Weighting shifts MN9 by {{expr_mn9_delta_hz}} rate-model units at 1 µM imidacloprid, which we report as the size of the uncertainty rather than as a correction. -**Model dose-response.** `flylab/assays/ensemble.py` builds the model's own concentration-response curve, fits a four-parameter Hill function by Nelder–Mead in log-concentration space and bootstraps it. For imidacloprid on the `named` cut the circuit midpoint is {{ic50_M}} with a bootstrap 95 % CI of {{ic50_lo_M}} to {{ic50_hi_M}} (r² = {{ic50_r2}}). The interval does not bracket the point estimate, and that is reported rather than hidden: the bootstrap resamples replicates within each concentration, the lower plateau of the ladder is short, and the midpoint is therefore poorly identified. This is in any case the **model's** circuit midpoint, not a receptor or an animal IC50. The one-at-a-time tornado at 1 µM finds exactly zero span for {{tornado_zero_span_params}} — the engagement is saturated — which the global analysis (§3.5) confirms and quantifies. +**Model dose-response.** `flylab/assays/ensemble.py` builds the model's own concentration-response curve, fits a four-parameter Hill function by Nelder–Mead in log-concentration space and bootstraps it. For imidacloprid on the `named` cut the circuit midpoint is {{ic50_M}} with a bootstrap 95 % CI of {{ic50_lo_M}} to {{ic50_hi_M}} (r² = {{ic50_r2}}). The interval does not bracket the point estimate, and that is reported rather than hidden: the bootstrap resamples replicates within each concentration, the lower plateau of the ladder is short, and the midpoint is therefore poorly identified. This is in any case the **model's** circuit midpoint, not a receptor or an animal IC50. The one-at-a-time tornado at 1 µM finds exactly zero span for {{tornado_zero_span_params}}; the engagement is saturated; which the global analysis (§3.5) confirms and quantifies. ## S5. Prospective predictions The following {{pred_n_hypotheses}} hypotheses are written before any experiment and labelled `software_prediction`; `live_result` is `null` in every row and stays null until a real table is imported by hand. They are **prospective, not pre-registered**: a mutable repository is not a registry, and the wording will change only if a tagged release is externally archived with a timestamp. Full table: T3. * **H1** (receptor engagement). At 1 µM, imidacloprid insect nicotinic engagement ≥ 0.9 while vertebrate α4β2 stays ≤ 0.2. Predicted gap {{pred_H1_effect}}. Checkable against published expressed-receptor work rather than against flies. -* **H2** (PER, `protocols/per.md`). The same dose lowers cholinergic drive in the MN9/DNp01 neighbourhood: predicted change in network mean rate {{pred_H2_effect}} Hz. **Specification-dependence caveat.** The *sign* of this prediction is a modelling choice, not a model output: §3.4 shows the same compound at the same engagement takes the network from {{stab_default_vehicle_hz}} Hz to {{stab_default_treated_hz}} Hz under the shipped biphasic rule and from {{stab_monotone_vehicle_hz}} Hz to {{stab_monotone_treated_hz}} Hz under a monotone rule of the same admissible family, and only {{stab_C1_retained}} of {{stab_family_size}} specifications retain the suppression direction. An experimenter should treat H2 as a signed interval over the family rather than as a point prediction, and a live result in either direction falsifies a specification rather than the integration. The control rate this needs could not be sourced; the nearest substitute is 40 % responders at 1 h food deprivation rising to 68 % at 2 h in 463 wild-type flies [61], which also shows that deprivation time is a larger nuisance variable than most effects FlyLab would try to detect. -* **H3** (climbing, `protocols/climbing.md`). Fipronil at 1 µM disinhibits the neighbourhood: predicted change {{pred_H3_effect}} Hz, opposite in sign to H2. **The same caveat applies transitively**: H3 is defined relative to H2, so its contrast inherits H2's specification dependence even though the fipronil arm itself is retained by {{stab_C2_retained}} of {{stab_family_size}} specifications. Negative geotaxis is the best-served endpoint [60], [64], [66]. -* **H4** (negative control). Diazepam engages vertebrate GABA-A and not insect RDL in this library, so it moves no fly readout: predicted effect {{pred_H4_effect}}. A non-zero live effect would falsify either the library row [17] or the assumption that the compound reaches the CNS unmetabolised. -* **H5** (within-animal control). Bitter input vetoes sugar-driven MN9 on the map-extracted path: predicted change {{pred_H5_effect}} Hz relative to sugar alone, matching the published direction [7], [63]. -* **H6**. Fipronil relieves the bitter veto on the map path: predicted rise in the veto ratio {{pred_H6_effect}}. The dependence analysis shows this readout is not distinguishable from its shuffled controls, so a live confirmation would support the pharmacology and **not** the wiring. -* **H7** (selectivity negative control). Picrotoxin perturbs the fly neighbourhood while its receptor selectivity index is ≈ 0: predicted change {{pred_H7_effect}} Hz. Any insect-selectivity claim the tool makes must fail for picrotoxin, and picrotoxin's lack of insect-over-vertebrate selectivity is independently attested [31]. +* **H2** (PER, `protocols/per.md`). The same dose lowers cholinergic drive in the MN9/DNp01 neighbourhood: predicted change in network mean rate {{pred_H2_effect}} rate-model units. **Specification-dependence caveat.** The *sign* of this prediction is a modelling choice, not a model output: §3.4 shows the same compound at the same engagement takes the network from {{stab_default_vehicle_hz}} rate-model units to {{stab_default_treated_hz}} rate-model units under the shipped biphasic rule and from {{stab_monotone_vehicle_hz}} rate-model units to {{stab_monotone_treated_hz}} rate-model units under a monotone rule of the same admissible family, and only {{stab_C1_retained}} of {{stab_family_size}} specifications retain the suppression direction. An experimenter should treat H2 as a signed interval over the family rather than as a point prediction, and a live result in either direction falsifies a specification rather than the integration. The control rate this needs could not be sourced; the nearest substitute is 40 % responders at 1 h food deprivation rising to 68 % at 2 h in 463 wild-type flies [60], which also shows that deprivation time is a larger nuisance variable than most effects FlyLab would try to detect. +* **H3** (climbing, `protocols/climbing.md`). Fipronil at 1 µM disinhibits the neighbourhood: predicted change {{pred_H3_effect}} rate-model units, opposite in sign to H2. **The same caveat applies transitively**: H3 is defined relative to H2, so its contrast inherits H2's specification dependence even though the fipronil arm itself is retained by {{stab_C2_retained}} of {{stab_family_size}} specifications. Negative geotaxis is the best-served endpoint [59], [63], [65]. +* **H4** (negative control). Diazepam engages vertebrate GABA-A and not insect RDL in this library, so it moves no fly readout: predicted effect {{pred_H4_effect}}. A non-zero live effect could reveal an omitted target or model failure rather than falsifying the absence of evidence in the library row [16] or the assumption that the compound reaches the CNS unmetabolised. +* **H5** (within-animal control). Bitter input vetoes sugar-driven MN9 on the map-extracted path: predicted change {{pred_H5_effect}} rate-model units relative to sugar alone, matching the published direction [6], [62]. +* **H6**. Fipronil relieves the bitter veto on the map path: predicted rise in the veto ratio {{pred_H6_effect}}. The dependence analysis shows this readout is not distinguishable from its shuffled controls, so a matching live response alone would not discriminate the tested wiring alternatives. This non-rejection has no endpoint-specific power calibration. +* **H7** (selectivity negative control). Picrotoxin perturbs the fly neighbourhood while its receptor selectivity index is ≈ 0: predicted change {{pred_H7_effect}} rate-model units. Any insect-selectivity claim the tool makes must fail for picrotoxin, and picrotoxin's lack of insect-over-vertebrate selectivity is independently attested [30]. Rows where the model's own output disagrees with the prospective direction: {{pred_direction_mismatches}}. @@ -151,7 +151,7 @@ Rows where the model's own output disagrees with the prospective direction: {{pr ## S6. Evidence typing: the full rules -The permitted transformation is a function of two facts — the **parameter type** the source measured and the **evidence distance** between that source and this compound at this receptor in this species. The distance grades are {{ev_distance_labels}}. The full (parameter type × distance) → model table is generated as T20; its shape is: +The permitted transformation is a function of two facts; the **parameter type** the source measured and the **evidence distance** between that source and this compound at this receptor in this species. The distance grades are {{ev_distance_labels}}. A Kd/Ki earns physical `binding_occupancy` only at E0 (the exact compound, receptor and species); at E1–E3 it is explicitly a binding-derived engagement proxy, and at E4 it is not modelled. The full (parameter type × distance) → model table is generated as T20; its shape is: | Parameter type | E0 | E1–E2 | E3 | E4 | |---|---|---|---|---| @@ -159,11 +159,11 @@ The permitted transformation is a function of two facts — the **parameter type | `EC50`, `IC50`, `Kb` | `functional_engagement` | `functional_engagement_proxy` | `functional_engagement_proxy` | `not_modelled` | | `relative_potency`, `class_order`, `unknown` | `not_modelled` | `not_modelled` | `not_modelled` | `not_modelled` | -A proxy computes the same Hill expression as its non-proxy sibling but reports itself as a transfer and carries a provenance warning naming what was transferred — the species, the preparation or the chemical class. Asking for a binding occupancy from an EC50, or for any numeric engagement from a placeholder row, raises `EvidenceTypeError`; the library validator rejects a placeholder that carries a value and a non-placeholder that does not. The library currently reports {{ev_library_problems}} validation problems. +A proxy computes the same Hill expression as its non-proxy sibling but reports itself as a transfer and carries a provenance warning naming what was transferred; the species, the preparation or the chemical class. Asking for a binding occupancy from an EC50, or for any numeric engagement from a placeholder or unsupported row, raises/refuses via `EvidenceTypeError`; the library validator rejects a placeholder that carries a value and a non-placeholder that does not. The library currently reports {{ev_library_problems}} validation problems. -**The soundness invariant.** *No numeric engagement reaching a circuit readout may originate from a row whose (parameter type, evidence distance) pair does not admit one.* `tests/test_pharm_evidence.py` checks it exhaustively rather than at a call site: over every row of the shipped library and every entry point that can produce an engagement, and with a companion check that dropping the inadmissible rows changes no gain vector. Mutating the rule table makes the test fail with a message naming the row and the pair. This, rather than the enum, is what the manuscript claims as a computing contribution. +**The soundness invariant.** *No numeric engagement reaching a circuit readout may originate from a row whose (parameter type, evidence distance) pair does not admit one.* `tests/test_pharm_evidence.py` performs a finite regression audit over every row of the shipped library through the enumerated comparison, dose-curve, uncertainty, subgraph, taste-assay and dashboard paths. Companion checks show that dropping inadmissible rows changes no gain vector, and deliberate invalid-transformation requests raise with the offending evidence pair. These tests cover the shipped rows and enumerated paths; they do not prove coverage of an unenumerated future entry point, nor do they mutate the rule table end to end. This enforced refusal plus finite audit, rather than the enum alone, is what the manuscript claims as a computing contribution; legacy untyped rows are normalised as class-extrapolated EC50s for backwards compatibility and are not evidence of a measured target. -**Census.** {{ev_rows}} rows over {{ev_compounds}} compounds and {{ev_receptor_keys}} receptor keys (Table T10, Figure F11). Parameter types: {{ev_param_EC50}} EC50, {{ev_param_IC50}} IC50, {{ev_param_Kd}} Kd, {{ev_param_unknown}} unknown. Evidence distances: {{ev_dist_E0}} E0, {{ev_dist_E1}} E1, {{ev_dist_E2}} E2, {{ev_dist_E3}} E3, {{ev_dist_E4}} E4. Engagement models: {{ev_model_binding_occupancy}} `binding_occupancy`, {{ev_model_binding_engagement_proxy}} `binding_engagement_proxy`, {{ev_model_functional_engagement}} `functional_engagement`, {{ev_model_functional_engagement_proxy}} `functional_engagement_proxy`, {{ev_model_not_modelled}} `not_modelled`. Rows entitled to a binding-occupancy model: {{ev_binding_row_names}} [77]. Rows demoted to a binding proxy by distance: {{ev_binding_proxy_row_names}}. +**Census.** {{ev_rows}} rows over {{ev_compounds}} compounds and {{ev_receptor_keys}} receptor keys (Table T10, Figure F11). Parameter types: {{ev_param_EC50}} EC50, {{ev_param_IC50}} IC50, {{ev_param_Kd}} Kd, {{ev_param_unknown}} unknown. Evidence distances: {{ev_dist_E0}} E0, {{ev_dist_E1}} E1, {{ev_dist_E2}} E2, {{ev_dist_E3}} E3, {{ev_dist_E4}} E4. Engagement models: {{ev_model_binding_occupancy}} `binding_occupancy`, {{ev_model_binding_engagement_proxy}} `binding_engagement_proxy`, {{ev_model_functional_engagement}} `functional_engagement`, {{ev_model_functional_engagement_proxy}} `functional_engagement_proxy`, {{ev_model_not_modelled}} `not_modelled`. Rows entitled to a binding-occupancy model: {{ev_binding_row_names}} [76]. Rows demoted to a binding proxy by distance: {{ev_binding_proxy_row_names}}. ## S7. Reproduction details @@ -184,7 +184,7 @@ A proxy computes the same Hill expression as its non-proxy sibling but reports i | dose-response bootstrap | {{ic50_n_boot}} resamples | 40 | | prediction replicates | {{pred_n_rep}} | 2 | -The middle column is read from `results.json`, so it describes the run that rendered this document rather than an aspiration. `--fast` changes statistical effort only, never the model: point estimates that do not depend on replicate or permutation count are identical, and a `--fast` *p* is coarser. Two tests guard the committed record: it must not come from a `--fast` run, and — new in this revision — **every statistical parameter it records must equal the shipped default**, because the v0.6 record was generated at a shuffle count the code no longer used and nothing detected it. +The middle column is read from `results.json`, so it describes the run that rendered this document rather than an aspiration. `--fast` changes statistical effort only, never the model: point estimates that do not depend on replicate or permutation count are identical, and a `--fast` *p* is coarser. Two tests guard the committed record: it must not come from a `--fast` run, and; new in this revision; **every statistical parameter it records must equal the shipped default**, because the v0.6 record was generated at a shuffle count the code no longer used and nothing detected it. Two of the analyses are started in worker processes when the run begins and joined at their own step, because they dominate the wall clock and depend on nothing the other steps produce. `--jobs` controls that and the permutation workers; results are identical at any value. @@ -196,13 +196,13 @@ Every assay returns a notebook (schema 0.3) whose `warnings` list ships with the These are summarised in §4 and given here at length. -**(i) The transmitter sign table is mostly asserted.** Acetylcholine at +1.0 is the normalisation every other entry is scaled against, and GABA at −1.0 is the convention that inhibition is as strong per synapse as excitation. The remaining five — {{sign_asserted_transmitters}} — are asserted magnitudes with no measurement behind them (T21). Glutamate in particular is signed inhibitory at 40 % of the stated strength, which is a hedge against the cells whose glutamatergic output is excitatory rather than a measurement of anything; any such cell is mis-signed in this model. None of these five is a member of the specification family, and the transmitter nulls do not bound them, because they permute *assignments* and leave the table untouched. Separately, the ablation ladder's level B inherits a whole-CNS excitation index that scores glutamate as partly excitatory while level D signs it inhibitory; the two levels of one ladder therefore disagree about a transmitter sign, and reconciling them moves the composition-versus-full rank correlation by at most {{abl_glutamate_max_delta_rho}}. +**(i) The transmitter sign table is mostly asserted.** Acetylcholine at +1.0 is the normalisation every other entry is scaled against, and GABA at −1.0 is the convention that inhibition is as strong per synapse as excitation. The remaining five; {{sign_asserted_transmitters}}; are asserted magnitudes with no measurement behind them (T21). Glutamate in particular is signed inhibitory at 40 % of the stated strength, which is a hedge against the cells whose glutamatergic output is excitatory rather than a measurement of anything; any such cell is mis-signed in this model. None of these five is a member of the specification family, and the transmitter nulls do not bound them, because they permute *assignments* and leave the table untouched. Separately, the ablation ladder's level B inherits a whole-CNS excitation index that scores glutamate as partly excitatory while level D signs it inhibitory; the two levels of one ladder therefore disagree about a transmitter sign, and reconciling them moves the composition-versus-full rank correlation by at most {{abl_glutamate_max_delta_rho}}. **(ii) The row normalisation is load-bearing** (§S3, §3.3). -**(iii) Receptor effects are applied across transmitter-defined synapses.** A gain on `g_ach` scales every cholinergic synapse, whether or not that synapse expresses the receptor the compound binds. This is the assumption the expression layer relaxes, and it is reported as a bound rather than an improvement: coverage is {{expr_coverage_overall}} of cells, every mapping row is a cross-atlas inference, and receptor expression in identified adult motor neurons — MN9's class — is a confirmed gap in the atlas literature, so the paper's principal readout carries the default weight. Weighting shifts MN9 by {{expr_mn9_delta_hz}} Hz at 1 µM imidacloprid. +**(iii) Receptor effects are applied across transmitter-defined synapses.** A gain on `g_ach` scales every cholinergic synapse, whether or not that synapse expresses the receptor the compound binds. This is the assumption the expression layer relaxes, and it is reported as a bound rather than an improvement: {{expr_coverage_overall}} is the fraction of cell–receptor slots with a non-unknown ordinal class mapping, not a fraction of cells with measured expression; every mapping row is a cross-atlas inference; and receptor expression in identified adult motor neurons; MN9's class; is a confirmed gap in the atlas literature, so the paper's principal readout carries the default weight. Weighting shifts MN9 by {{expr_mn9_delta_hz}} rate-model units at 1 µM imidacloprid. -**(iv) Hops-limited cuts.** A 1-hop neighbourhood with a weight floor is a fraction of MN9's real input; missing inhibition can flip the sign of any readout, and the disinhibition readout is bounded by the GABA share that survived the cut. Growing the cut is deliberately constrained: past about 1 MB of JSON the weight floor is raised rather than hops added, because a 2-hop cut of a motor neuron in a 25-million-edge CNS is most of the animal. §3.2 is the quantitative demonstration of how much the choice of cut matters — the dependence verdict itself reverses between the two committed cuts. +**(iv) Hops-limited cuts.** A 1-hop neighbourhood with a weight floor is a fraction of MN9's real input; missing inhibition can flip the sign of any readout, and the disinhibition readout is bounded by the GABA share that survived the cut. Growing the cut is deliberately constrained: past about 1 MB of JSON the weight floor is raised rather than hops added, because a 2-hop cut of a motor neuron in a 25-million-edge CNS is most of the animal. §3.2 is the quantitative demonstration of how much the choice of cut matters; the dependence verdict itself reverses between the two committed cuts. **(v) The two cuts differ in what they can express.** The `named` cut carries {{named_cells_acetylcholine}} cholinergic, {{named_cells_gaba}} GABAergic and {{named_cells_glutamate}} glutamatergic cells, and contains {{named_cells_octopamine}} octopaminergic cells and {{taste_motor_cells_octopamine}} on `taste_motor`, so a compound whose only insect mechanism acts through `g_oct` cannot move either readout for substrate reasons rather than pharmacological ones; {{named_share_unclear_cells}} of `named` cells carry no consensus transmitter prediction and are presynaptically inert. Full structural census: T19. @@ -214,9 +214,9 @@ These are the domain-side results the method is demonstrated on. They are suppor **The evidence census in full.** Of {{ev_rows}} rows over {{ev_compounds}} compounds and {{ev_receptor_keys}} receptor keys, {{ev_param_EC50}} carry an EC50, {{ev_param_IC50}} an IC50 and {{ev_param_Kd}} a binding constant. By evidence distance the split is {{ev_dist_E0}} / {{ev_dist_E1}} / {{ev_dist_E2}} / {{ev_dist_E3}} / {{ev_dist_E4}} rows at E0–E4. At the 1 µM working concentration the two headline compounds engage their insect targets at {{imi_insect_occ}} and {{fip_insect_occ}} against {{imi_vert_occ}} and {{fip_vert_occ}} at the vertebrate counterparts ({{imi_ec50_ratio}}-fold and {{fip_ec50_ratio}}-fold; {{window_log10_imidacloprid}} log₁₀ units of window for the first). Placeholder rows contribute nothing: the negative-control row reports "{{dia_insect_rdl_occ}}". -**Sensory-veto readout.** Driving the {{veto_n_sweet_grn}} sweet GRN body IDs at {{veto_sugar_hz}} Hz and reading MN9 reproduces the published direction of a bitter veto [7], [63] on both runtimes ({{veto_rate_vehicle_ratio}} rate, {{veto_lif_vehicle_ratio}} LIF in vehicle; {{veto_rate_fipronil_ratio}} and {{veto_lif_fipronil_ratio}} under fipronil). The two runtimes agree on **direction only** — sugar-driven MN9 is {{veto_rate_vehicle_sugar}} Hz on the rate engine against {{veto_lif_vehicle_sugar}} Hz on the LIF engine, a difference following from the LIF background drive (§S3) — and we do not present that agreement as cross-model validation. +**Sensory-veto readout.** Driving the {{veto_n_sweet_grn}} sweet GRN body IDs at {{veto_sugar_hz}} Hz input drive and reading MN9 reproduces the published direction of a bitter veto [6], [62] on both runtimes ({{veto_rate_vehicle_ratio}} rate, {{veto_lif_vehicle_ratio}} LIF in vehicle; {{veto_rate_fipronil_ratio}} and {{veto_lif_fipronil_ratio}} under fipronil). The `LB3b/c` sweet and `LB1a-d` bitter labels are working hypotheses from published gustatory nomenclature, not MaleCNS modality annotations. The two runtimes agree on **direction only**; sugar-driven MN9 is {{veto_rate_vehicle_sugar}} rate-model units on the rate engine against {{veto_lif_vehicle_sugar}} LIF spikes/s on the LIF engine, a difference following from the LIF background drive (§S3); and we do not present that agreement as cross-model validation. -**A properly powered negative.** At {{dep_n_taste}} permutations the veto ratio on the extracted sensory path is not distinguishable from any shuffled control (max |z| = {{dep_taste_fipronil_max_abs_z}}), equivalent within tolerance for {{dep_taste_fipronil_equivalent_modes}} and indeterminate for {{dep_taste_fipronil_indeterminate_modes}}. Because §3.1 characterises the instrument's power, this is reportable as a negative rather than as an absence of evidence. The comparator compound has no veto ratio at this dose at all — the readout is *undefined*, not zero, and is reported as undefined. +**Non-rejection with unquantified endpoint-specific power.** At {{dep_n_taste}} permutations the sensory-veto ratio is not distinguished from the tested controls (max |z| = {{dep_taste_fipronil_max_abs_z}}). The descriptive median-gap criterion passes for {{dep_taste_fipronil_equivalent_modes}} and is indeterminate for {{dep_taste_fipronil_indeterminate_modes}}. The synthetic mean-rate study in §3.1 does not establish power for this different graph and ratio endpoint, so the result cannot rule out a wiring effect. The comparator's ratio is undefined at this dose because its denominator is silenced; it is not zero. **Selectivity landscape.** The split is by mechanism rather than potency: {{landscape_amplify_compounds}} are amplified by the circuit and {{landscape_buffer_compounds}} buffered, subject to the threshold caveat of §3.4. Compounds never reaching the circuit threshold on either cut ({{landscape_unscored_compounds}}) are reported **unscored** rather than buffered, and one of them is unscored for a substrate reason rather than a pharmacological one: chlordimeform's only insect mechanism acts through octopamine, and neither cut contains an octopaminergic cell. diff --git a/papers/figures/F1_architecture.png b/papers/figures/F1_architecture.png index 6f5386d..c72c45d 100644 Binary files a/papers/figures/F1_architecture.png and b/papers/figures/F1_architecture.png differ diff --git a/papers/figures/F1_architecture.svg b/papers/figures/F1_architecture.svg index 5abcf7e..ca041e2 100644 --- a/papers/figures/F1_architecture.svg +++ b/papers/figures/F1_architecture.svg @@ -6,11 +6,11 @@ - 2026-09-20T13:54:55.572622 + 2026-09-21T17:08:57.397634 image/svg+xml - Matplotlib v3.11.2, https://matplotlib.org/ + Matplotlib v3.10.8, https://matplotlib.org/ @@ -21,3731 +21,3678 @@ - - + - + - + - + - + - + - + - + - - + + - - + + - - + + - - + + - - + + - - + + - - + + - - + + - - - - - - - - - - - - - - + + + + + + + + + + + + + - - - - - - - - - - - - - - - - - - - - - - - - + + + + + + + + + + + + + + + + + + + + + + + + - + - - - - - - - - - - - - - + + + + + + + + + + + + - - - - - - - - - - - - - - - - - - - - + + + + + + + + + + + + + + + + + + + + - - + + - - - - - - - - - - - - - - - - - - + + + + + + + + + + + + + + + + + - - + + - - - - - - - - - - - - - - - - - + + + + + + + + + + + + + + + + - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + - + - - - - - - - + + + + + + - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + - - - - - - - - - - - - - - - - - + + + + + + + + + + + + + + + + - + - - - - - - + + + + + - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + - + - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + - - + - - - - - - - - - - - - - - - - - - - - + + + + + + + + + + + + + + + + + + + + - + - - - - - - + + + + + - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + - - + + - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + - + - - - - - - - - - - + + + + + + + + + - - - - - - - - - - - - - - - - - - - - - - - + + + + + + + + + + + + + + + + + + + + + + + - + - - - - + + + - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + - + - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - + + + + + + + + + + + + + + + + + + + + + + + + + + + + + - - - - - - - - - - - - - - - - - - - - - - - - - - - + + + + + + + + + + + + + + + + + + + + + + + + + + + - - - - - - - - - - - - - - - + + + + + + + + + + + + + + + - + - - + - - - - - - - - - - - - - - - - - - - - - - - - - - - - + + + + + + + + + + + + + + + + + + + + + + + + + + + + - + - - + - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + - - - - - - - - - - - - - - - - - - - - - + + + + + + + + + + + + + + + + + + + + + + + + - - - - - - - - - - - - - - - - - - - - - - - - - - - - - + + + + + + + + + + + + + + + + + + + + + + + + + + + + + - - - - - - - - - - - - - - - - - - - - - - + + + + + + + + + + + + + + + + + + + + + + - - - - - - - - - - - - - - - - - - - - - - - - - - - - - + + + + + + + + + + + + + + + + + + + + + + + + + + + + + - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + - - - + + - - - - - - - - - - - - - - - - - - - - - - - - - - + + + + + + + + + + + + + + + + + + + + + + + + + + - + - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + - - - - - - - - - - - + + + + + + + + + + + - - - - - - - - + + + + + + + + - - - - - - - - - + + + + + + + + + - - - - - - - - - - + + + + + + + + + + - + - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - + + + + + + + + + + + + + + + + + + + + + + + + + + + + - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + - + diff --git a/papers/figures/F2_dual_scorecard.png b/papers/figures/F2_dual_scorecard.png index 2d682a3..6bd14fd 100644 Binary files a/papers/figures/F2_dual_scorecard.png and b/papers/figures/F2_dual_scorecard.png differ diff --git a/papers/figures/F3_occupancy_curves.png b/papers/figures/F3_occupancy_curves.png index d107d27..46e5a8f 100644 Binary files a/papers/figures/F3_occupancy_curves.png and b/papers/figures/F3_occupancy_curves.png differ diff --git a/papers/figures/F5_bitter_veto.png b/papers/figures/F5_bitter_veto.png index d9626fc..29fac29 100644 Binary files a/papers/figures/F5_bitter_veto.png and b/papers/figures/F5_bitter_veto.png differ diff --git a/papers/figures/F9_genotype_panel.png b/papers/figures/F9_genotype_panel.png index b83b413..8bd68cf 100644 Binary files a/papers/figures/F9_genotype_panel.png and b/papers/figures/F9_genotype_panel.png differ diff --git a/papers/figures/captions.md b/papers/figures/captions.md index 3eca8c6..a284c4e 100644 --- a/papers/figures/captions.md +++ b/papers/figures/captions.md @@ -4,32 +4,32 @@ Generated by `scripts/reproduce_paper.py`. Do not edit by hand. **F10_rank_validation.** Left: the model's compound ordering against each published ordering that the teaching library can cover, with Spearman's rho per entry and the number of compounds it orders. 8 entry-assay combinations are evaluable and 16 are skipped because the library cannot cover them; 4 of the evaluable ones order exactly two compounds, where rho can only be +-1, so no bare mean is quoted: over the 4 entries with more than two compounds the mean is 0.756, and as a concordance count the result is 8 concordant against 0 discordant orderings (two-sided binomial p = 0.008). Right: the two inversions the validation module records rather than fixes. Nitenpyram sits near the top of the library's insect nAChR potency ordering and is the least potent neonicotinoid in *Drosophila* whole-animal bioassays, and clothianidin/imidacloprid are ordered the opposite way from the bioassays. Both are receptor-EC50 versus whole-animal-potency differences: the whole-animal number contains uptake, penetration and metabolism, which FlyLab does not model. -**F11_evidence_types.** What the compound library rests on, row by row. Left: the parameter each cited source actually measured. Middle: the ordered evidence distance between that source and this compound at this receptor in this species. Right: the transformation the type system then permits, which is a function of the two. Green marks the 10 on-target (E0) rows and the 1 of them that may use a physical binding-occupancy model; amber marks the proxy models, where the same Hill expression is computed but reported as a transfer with a warning naming the gap; red marks rows that carry no modellable evidence and return N/A rather than a small number. The type system, not a convention, enforces this: asking for an engagement from an untyped row raises rather than returning a value, and a soundness invariant is checked exhaustively over every row and every entry point. +**F11_evidence_types.** What the compound library rests on, row by row. Left: the parameter each cited source actually measured. Middle: the ordered evidence distance between that source and this compound at this receptor in this species. Right: the transformation the evidence dispatch rule permits, which is a function of the two. Green marks the 10 on-target (E0) rows and the 1 of them that may use a physical binding-occupancy model; amber marks the proxy models, where the same Hill expression is computed but reported as a transfer with a warning naming the gap; red marks rows that carry no modellable evidence and return N/A rather than a small number. Runtime validation enforces this: asking for an engagement from an untyped row raises rather than returning a value, and a soundness invariant is audited over shipped rows and enumerated pipeline paths. -**F12_dependence_landscape.** Every compound of the library at four concentrations, on both committed cuts, classified by the weakest graph model the permutation test could **not** distinguish from the real cut -- never by a model that reproduces the effect. Classes are decided on Benjamini-Hochberg adjusted probabilities across each landscape's structural tests, with a prespecified relative effect floor of 1 % of the vehicle readout. Left, the 1-hop `named` cut: 11 of 84 cells need the wiring pattern, 42 are not distinguishable from any structure-preserving degradation, and 31 do not move the readout at all. Right, the denser `taste_motor` cut: 54 topology-dependent, 0 composition-dominated, 30 no-effect. The composition-dominated majority does not survive the change of substrate, and the `named` cut is an in-star whose degree-preserving rewire is close to the identity (T19). +**F12_dependence_landscape.** Every compound of the library at four concentrations, on both committed cuts, classified from structural component-test decisions and the configured effect floor. Classes are decided on Benjamini-Hochberg adjusted probabilities across each landscape's structural tests, with a prespecified relative effect floor of 1 % of the vehicle readout. Left, the 1-hop `named` cut: 11 of 84 cells are classified topology-dependent, 42 are not distinguishable from any structure-preserving degradation, and 31 fall below the configured effect floor. Right, the denser `taste_motor` cut: 54 topology-dependent, 0 composition-dominated, 30 no-effect. The composition-dominated majority does not survive the change of substrate, and the `named` cut is an in-star whose degree-preserving rewire is close to the identity (T19). BH adjusts structural component tests; cell/conclusion-level FDR control is not established, and the dependence assumptions remain unverified. Adjustment cannot strengthen non-rejection. -**F13_ablation.** Each level predicts the same 21 compounds with one layer of information removed, and is scored by how well it reproduces the full model's *ordering* of the library (levels have different units, so values are not comparable; the correlation is signed, so a perfect inversion scores -1 rather than 1). Crosses mark the historical depression-only level-C rule, which applied a multiplier of at most 1.0 to every transmitter alike and so could not express disinhibition: its poor correlation was structural, not a finding about connectomes, and the conclusion drawn from it is withdrawn. With a direction-aware rule level C rises to 0.961 at 1 uM from 0.242. The composition level's agreement with the full model must be read against the matched reference distribution of T24, not on its own. +**F13_ablation.** Each level predicts the same 21 compounds with one layer of information removed, and is scored by how well it reproduces the full model's *ordering* of the library (levels have different units, so values are not comparable; the correlation is signed, so a perfect inversion scores -1 rather than 1). Crosses mark the historical depression-only level-C rule, which applied a multiplier of at most 1.0 to every transmitter alike and so could not express disinhibition: its poor correlation was structural, not a finding about connectomes, and the conclusion drawn from it is withdrawn. With a direction-aware rule level C rises to 0.961 at 1 uM from 0.242. The composition level's agreement with the full model must be read against the matched reference distribution of T24, not on its own. Rate-engine quantities labelled Hz in legacy fields or axes are rate-model units, not calibrated physiological firing rates. -**F14_conclusion_stability.** Each row is a conclusion this paper could state; each column is one prespecified, admissible way of turning receptor engagement into synaptic gain. Green retains the conclusion, red reverses it, amber means the readout does not exist under that specification. The topology rows are decided on 100 permutations per specification and on Benjamini-Hochberg adjusted probabilities across the 100 structural tests of the run; until v0.6.1 the specification never reached the engine those rows were computed on, so their uniformity was guaranteed by construction rather than measured. Retained by every specification: C2_rdl_disinhibition, C3_imidacloprid_topology_not_distinguishable, C4_fipronil_topology_exceeds, C7_map_bitter_veto. The nicotinic suppression conclusion is reversed by every monotone rule, which is how a result that depends on a modelling choice looks when it is tested rather than asserted; and for a conclusion shaped as a failure to reject, the fraction retained is only as strong as the equivalence behind it (T14's `n_equivalent_within_tolerance` against `n_indeterminate`). +**F14_conclusion_stability.** Each row is a conclusion this paper could state; each column is one fixed, admissible way of turning receptor engagement into synaptic gain. Green retains the conclusion, red reverses it, amber means the readout does not exist under that specification. The topology rows are decided on 100 permutations per specification and on Benjamini-Hochberg adjusted probabilities across the 100 structural tests of the run; until v0.6.1 the specification never reached the engine those rows were computed on, so their uniformity was guaranteed by construction rather than measured. Retained by every specification: C2_rdl_disinhibition, C3_imidacloprid_topology_not_distinguishable, C4_fipronil_topology_exceeds, C7_map_bitter_veto. The nicotinic suppression conclusion is reversed by every monotone rule, which is how a result that depends on a modelling choice looks when it is tested rather than asserted; and for a conclusion shaped as a failure to reject, the fraction retained is a non-rejection, not evidence of equivalence (T14's `n_equivalent_within_tolerance` against `n_indeterminate`). The legacy within-tolerance label is a point-gap diagnostic, not a formal equivalence test. BH adjusts structural component tests; cell/conclusion-level FDR control is not established, and the dependence assumptions remain unverified. Adjustment cannot strengthen non-rejection. -**F15_uncertainty_voi.** Left: the share of the variance of the neighbourhood mean rate attributable to each assumption, first-order and total-order, from a Saltelli cross-sample with Jansen estimators, with bootstrap first-order confidence intervals. A factor is **resolved** only when its interval excludes zero; the starred rows are not, and a factor's index exceeding the noise floor does not resolve it. The dashed line is that floor, defined as the largest magnitude among the *negative* first-order estimates: a true first-order index cannot be negative, so that magnitude is a measured lower bound on the estimator's error, and on this run it is several times the declared null factor's own draw (0.046 set by gain_coef, against lif_seed at -0.007). Right: the same shares scaled back into the readout's variance and mapped onto the experiment that would resolve each assumption; a negative estimate is clipped to zero for the decision value, because a value of information cannot be negative. Green marks a factor that needs no experiment at all, only a re-analysis of data already held. +**F15_uncertainty_voi.** Left: the share of the variance of the neighbourhood mean rate attributable to each assumption, first-order and total-order, from a Saltelli cross-sample with Jansen estimators, with bootstrap first-order confidence intervals. A factor is **resolved** only when its interval excludes zero; the starred rows are not, and a factor's index exceeding the negative-estimate diagnostic does not resolve it. The dashed line is that floor, defined as the largest magnitude among the *negative* first-order estimates: a true first-order index cannot be negative, so that magnitude is a measured lower bound on the estimator's error, and on this run it is several times the declared null factor's own draw (0.046 set by gain_coef, against lif_seed at -0.007). Right: the same shares scaled back into the readout's variance and mapped onto the experiment that would resolve each assumption; a negative estimate is clipped to zero for the decision value, because a value of information cannot be negative. Green marks a factor that needs no experiment at all, only a re-analysis of data already held. The largest negative estimate is an observed error diagnostic, not a calibrated uncertainty bound. VOI concerns output variance under assumed model input ranges, not the benefit of an experiment on a fly. Rate-engine quantities labelled Hz in legacy fields or axes are rate-model units, not calibrated physiological firing rates. -**F16_instrument_validation.** The dependence ladder pointed at a graph whose answer is known in advance. A recurrent cholinergic cycle of known strength is planted in a synthetic cut of the same size, density and transmitter composition as the `named` cut; a gain patch that collapses `g_ach` then removes an amplification that exists only while the cycle is intact, so the drug contrast depends on the wiring by construction. Left: the share of independently generated graphs the ladder classifies topology-dependent, against planted strength and permutation count. Right: the recovery experiment itself, with red marking the strengths the ladder recovered. The unplanted control is the empirical false-positive rate (0 of 18). Without this panel, a negative from the ladder would be indistinguishable from an under-powered test. +**F16_instrument_validation.** Synthetic recovery checks with an imposed cholinergic cycle and a gain patch. The synthetic network is not size- or density-matched to the named connectome cut. The zero-strength case omits cycle edges and associated transmitter relabelling. Detection frequencies in this small grid have broad uncertainty; networks are reused across permutation counts. These checks do not establish general calibration or power for the real taste-motor ratio endpoint. Counts and settings are shown in the artifact. Rate-engine quantities labelled Hz in legacy fields or axes are rate-model units, not calibrated physiological firing rates. -**F1_architecture.** FlyLab dataflow. A compound and a free concentration enter the occupancy engine, which reads the sourced library (optionally shifted by a genotype, combined with a mixture partner, or driven by an exposure profile). Occupancy splits into two scorecards. The insect side passes through the mechanism table to a gain patch on transmitter classes of a named MaleCNS cut; the vertebrate side is scored at the same dose and never touches a circuit. The circuit runtime (rate or LIF) produces named-cell readouts, which the analysis and validation layers turn into empirical permutation probabilities with their three-way verdicts, selectivity indices, rank correlations and prospective predictions. Everything lands in one notebook JSON with a provenance block. +**F1_architecture.** FlyLab dataflow. A compound and a free concentration enter the engagement engine, which reads the sourced library (optionally shifted by a genotype, combined with a mixture partner, or driven by an exposure profile). Engagement splits into two scorecards. The insect side passes through the mechanism table to a gain patch on transmitter classes of a named MaleCNS cut; the vertebrate side is scored at the same dose and never touches a circuit. The circuit runtime (rate or LIF) produces named-cell readouts, which the analysis and validation layers turn into empirical permutation probabilities with their three-way verdicts, selectivity indices, rank correlations and prospective predictions. Everything lands in one notebook JSON with a provenance block. -**F2_dual_scorecard.** Dual scorecard for imidacloprid and fipronil at 1 µM. Bars are Hill occupancy of the insect target (blue) and its vertebrate counterpart (orange) for each of the five receptor pairs declared in `library.yaml:selectivity_pairs`. Hatched bars rest on a class placeholder: the compound has no sourced EC50 at that pair and the row is inert. Imidacloprid reaches 0.991 at insect nAChR against 0.091 at vertebrate a4b2; fipronil reaches 0.985 at insect RDL against 0.476 at native heteromeric vertebrate GABA-A. Teaching-tier EC50s: these are literature-order values, not measured constants. +**F2_dual_scorecard.** Dual scorecard for imidacloprid and fipronil at 1 µM. Bars are Hill engagement of the insect target (blue) and its vertebrate counterpart (orange) for each of the five receptor pairs declared in `library.yaml:selectivity_pairs`. Hatched entries indicate unsupported parameters and are interpreted as N/A. Imidacloprid reaches 0.991 at insect nAChR against 0.091 at vertebrate a4b2; fipronil reaches 0.985 at insect RDL against 0.476 at native heteromeric vertebrate GABA-A. Teaching-tier EC50s: these are literature-order values, not measured constants. -**F3_occupancy_curves.** Hill occupancy against free concentration for the four-compound panel, one insect target (solid, blue) and its vertebrate counterpart (dashed, orange) per compound. The shaded band is the selectivity window between the two EC50s, annotated in log10 units; the dotted vertical line is the 1 µM working concentration used elsewhere in the paper. Ivermectin's window is the GluCl/GlyR pair, deltamethrin's the Nav pair. Curves are the teaching library, not measured concentration-response data. +**F3_occupancy_curves.** Hill engagement against free concentration for the four-compound panel, one insect target (solid, blue) and its vertebrate counterpart (dashed, orange) per compound. The shaded band is the selectivity window between the two EC50s, annotated in log10 units; the dotted vertical line is the 1 µM working concentration used elsewhere in the paper. Ivermectin's window is the GluCl/GlyR pair, deltamethrin's the Nav pair. Curves are the teaching library, not measured concentration-response data. **F4_taste_motor_graph.** Left: the committed MaleCNS v1.0 taste-motor cut (1841 cells, 19066 edges; 1 hop around MN9, DNp01 and the labellar GRN types, plus induced partner-partner edges of at least 10 synapses). Node and edge colour is the predicted consensus transmitter; layout is the deterministic superclass-column layout from `flylab.analysis.layout`. Seeds are marked: sweet GRNs (LB3b/c, up triangles), bitter GRNs (LB1a-d, down triangles), DNp01 (squares) and MN9 (stars). Right: the five strongest sweet and five strongest bitter paths of at most 3 synapses from a labellar GRN to MN9, ranked by the absolute product of signed synapse weights; there are 2595 such sweet and 2372 such bitter paths in total. The sweet/bitter assignment of LB types is a working hypothesis, not a MaleCNS annotation. -**F5_bitter_veto.** MN9 firing on the map-extracted labellar-GRN path, driven through the real MaleCNS sweet (34) and bitter (38) GRN body IDs at 150 Hz, never by driving MN9 directly. Left: leaky rate engine. Right: Shiu-style LIF engine with a background Poisson drive. In vehicle, adding bitter drive takes MN9 to zero on both engines (veto ratio 0.00). Fipronil at 1 µM blocks RDL, raises the sugar-only rate and partially relieves the veto (ratio 0.41 rate, 0.16 LIF). Imidacloprid silences MN9 altogether, so its ratio is undefined rather than zero. Absolute rates differ between engines by construction and are not calibrated to any recording. +**F5_bitter_veto.** MN9 firing on the map-extracted labellar-GRN path, driven through the real MaleCNS sweet (34) and bitter (38) GRN body IDs at a nominal drive of 150 rate-model input units or LIF input spikes/s, never by driving MN9 directly. Left: leaky rate engine. Right: Shiu-style LIF engine with a background Poisson drive. In vehicle, adding bitter drive takes MN9 to zero on both engines (veto ratio 0.00). Fipronil at 1 µM blocks RDL, raises the sugar-only rate and partially relieves the veto (ratio 0.41 rate, 0.16 LIF). Imidacloprid silences MN9 altogether, so its ratio is undefined rather than zero. Absolute rates differ between engines by construction and are not calibrated to any recording. Left-axis values are rate-model units; right-axis values are simulated LIF spikes/s, not measured fly firing rates. -**F6_dependence_profile.** Histograms are the drug effect (treated minus vehicle on the network mean rate) measured on 1000 degraded copies of the 1-hop MN9/DNp01 `named` cut; the red line is the same contrast on the real cut. Columns run in increasing order of destruction: transmitter labels permuted, synapse weights permuted, degree-preserving double-edge rewiring, Erdos-Renyi. Titles carry the empirical two-sided permutation p and the three-way verdict against the prespecified equivalence margin (delta = 0.332 Hz, 5 % of the vehicle readout): `distinguishable` means the test rejected, `equivalent within tolerance` means the gap from the null median is below the margin, and `indeterminate` means neither -- a failure to reject is not evidence that the degraded graph gives the same effect. Plain label permutation is shown because it is informative, but it moves the weighted excitation/inhibition balance as well as transmitter identity (T23), so it is a joint null and not a rung of the ladder. Imidacloprid is composition-dominated (necessary level: degree_sequence, equivalent_within_tolerance); fipronil is topology-dependent (necessary level: wiring_with_weighted_transmitter_balance). +**F6_dependence_profile.** Histograms are the drug effect (treated minus vehicle on the network mean rate) measured on 1000 degraded copies of the 1-hop MN9/DNp01 `named` cut; the red line is the same contrast on the real cut. Columns run in increasing order of destruction: transmitter labels permuted, synapse weights permuted, degree-preserving double-edge rewiring, Erdos-Renyi. Titles carry the empirical two-sided permutation p and the three-way verdict against the configured median-gap tolerance (delta = 0.332 Hz, 5 % of the vehicle readout): `distinguishable` means the test rejected, `equivalent within tolerance` means the gap from the null median is below the margin, and `indeterminate` means neither -- a failure to reject is not evidence that the degraded graph gives the same effect. Plain label permutation is shown because it is informative, but it moves the weighted excitation/inhibition balance as well as transmitter identity (T23), so it is a joint null and not a rung of the ladder. Imidacloprid is composition-dominated (necessary level: degree_sequence, equivalent_within_tolerance); fipronil is topology-dependent (necessary level: wiring_with_weighted_transmitter_balance). The legacy within-tolerance label is a point-gap diagnostic, not a formal equivalence test. Rate-engine quantities labelled Hz in legacy fields or axes are rate-model units, not calibrated physiological firing rates. -**F7_selectivity_landscape.** Each point is one compound on one committed graph. The x axis is the receptor selectivity index from the library alone; the y axis is the circuit index, the log10 width of the window between the concentration that moves the simulated network mean rate by 50% and the concentration at which the most potent vertebrate target reaches 20% occupancy. Points below the dashed identity line are buffered by the circuit, points above are amplified. The split is by mechanism, not by potency: chlorpyrifos_oxon, ddt, deltamethrin, permethrin (Nav modulators and the AChE inhibitor) amplify, mean gap +0.29 log10 units, while the nAChR-acting set (acetamiprid, acetylcholine, clothianidin, imidacloprid, nicotine, nitenpyram) buffers, mean gap -0.87. Compounds with no circuit index (chlordimeform, dieldrin, fipronil, gaba, ivermectin, picrotoxin, spinosad, sulfoxaflor, thiamethoxam) never reach the 50% threshold on this cut and are omitted. +**F7_selectivity_landscape.** Each point is one compound on one committed graph. The x axis is the receptor selectivity index from the library alone; the y axis is the circuit index, the log10 width of the window between the concentration that moves the simulated network mean rate by 50% and the concentration at which the most potent vertebrate target reaches 20% engagement. Points below the dashed identity line are buffered by the circuit, points above are amplified. The split is by mechanism, not by potency: chlorpyrifos_oxon, ddt, deltamethrin, permethrin (Nav modulators and the AChE inhibitor) amplify, mean gap +0.29 log10 units, while the nAChR-acting set (acetamiprid, acetylcholine, clothianidin, imidacloprid, nicotine, nitenpyram) buffers, mean gap -0.87. Compounds with no circuit index (chlordimeform, dieldrin, fipronil, gaba, ivermectin, picrotoxin, spinosad, sulfoxaflor, thiamethoxam) never reach the 50% threshold on this cut and are omitted. -**F8_dose_response_sensitivity.** Left: the model's own dose-response for imidacloprid on the 1-hop MN9/DNp01 cut. Points are 4 replicates per concentration with log-normal EC50 jitter (sd 0.3 log10 units) and drive jitter; the curve is a 4-parameter Hill fit, and the band is a 200-resample bootstrap CI on the fitted IC50 (34.7 nM). This is the *model's* circuit IC50, not a receptor IC50 and not an animal IC50. Right: one-at-a-time sensitivity of the same readout at 1 µM, each parameter halved and doubled. The gain-rule coefficient dominates; the EC50 and Hill coefficient have zero span because occupancy is already saturated at this dose. +**F8_dose_response_sensitivity.** Left: the model's own dose-response for imidacloprid on the 1-hop MN9/DNp01 cut. Points are 4 replicates per concentration with log-normal EC50 jitter (sd 0.3 log10 units) and drive jitter; the curve is a 4-parameter Hill fit, and the band is a 200-resample bootstrap CI on the fitted IC50 (34.7 nM). This is the *model's* circuit IC50, not a receptor IC50 and not an animal IC50. Right: one-at-a-time sensitivity of the same readout at 1 µM, each parameter halved and doubled. The gain-rule coefficient dominates; the EC50 and Hill coefficient have zero span because engagement is already saturated at this dose. Rate-engine quantities labelled Hz in legacy fields or axes are rate-model units, not calibrated physiological firing rates. -**F9_genotype_panel.** Target-receptor occupancy at 1 uM for wild type and each published resistance allele that has a sourced fold-shift for that compound. Grey bars are alleles with no sourced number for the compound: they are left identical to wild type and reported, never extrapolated. Left and centre: fipronil and imidacloprid. Right: the specificity check that motivates the per-compound rule. *para* M918T (super-kdr) shifts deltamethrin 100-fold but leaves DDT at 1-fold, and *Rdl* A301S shifts GABA 33.9-fold while leaving fipronil at 1.1-fold. A toggle that shifted a whole receptor uniformly would be wrong in both cases. +**F9_genotype_panel.** Target-receptor engagement at 1 uM for wild type and each published resistance allele that has a sourced fold-shift for that compound. Grey bars are alleles with no sourced number for the compound: they are left identical to wild type and reported, never extrapolated. Left and centre: fipronil and imidacloprid. Right: the specificity check that motivates the per-compound rule. *para* M918T (super-kdr) shifts deltamethrin 100-fold but leaves DDT at 1-fold, and *Rdl* A301S shifts GABA 33.9-fold while leaving fipronil at 1.1-fold. A toggle that shifted a whole receptor uniformly would be wrong in both cases. diff --git a/papers/results.json b/papers/results.json index ede5eaf..3213c5b 100644 --- a/papers/results.json +++ b/papers/results.json @@ -1,6 +1,6 @@ { "schema": "flylab-paper-results/1", - "generated_utc": "2026-09-20T15:48:02+00:00", + "generated_utc": "2026-09-22T10:00:06+00:00", "fast": false, "seed": 0, "jobs": 4, @@ -1658,15 +1658,15 @@ "step": "dependence" }, "dep_named_imidacloprid_confirmatory": { - "value": true, - "text": "yes", + "value": false, + "text": "False", "unit": null, "note": null, "step": "dependence" }, "dep_named_imidacloprid_design": { - "value": "confirmatory (prespecified)", - "text": "confirmatory (prespecified)", + "value": "exploratory high-effort reference profile", + "text": "exploratory high-effort reference profile", "unit": null, "note": null, "step": "dependence" @@ -2063,15 +2063,15 @@ "step": "dependence" }, "dep_named_fipronil_confirmatory": { - "value": true, - "text": "yes", + "value": false, + "text": "False", "unit": null, "note": null, "step": "dependence" }, "dep_named_fipronil_design": { - "value": "confirmatory (prespecified)", - "text": "confirmatory (prespecified)", + "value": "exploratory high-effort reference profile", + "text": "exploratory high-effort reference profile", "unit": null, "note": null, "step": "dependence" @@ -2467,7 +2467,7 @@ }, "dep_taste_fipronil_confirmatory": { "value": false, - "text": "no", + "text": "False", "unit": null, "note": null, "step": "dependence" @@ -3167,8 +3167,8 @@ "step": "dependence" }, "dep_land_statement": { - "value": "exploratory screen, FDR-controlled: 11 of 84 cells are topology-dependent after Benjamini-Hochberg at fdr_alpha = 0.05 across 168 structural tests (14 before correction).", - "text": "exploratory screen, FDR-controlled: 11 of 84 cells are topology-dependent after Benjamini-Hochberg at fdr_alpha = 0.05 across 168 structural tests (14 before correction).", + "value": "exploratory screen, classified from BH-adjusted component tests; no cell-level FDR guarantee: 11 of 84 cells are topology-dependent after Benjamini-Hochberg at fdr_alpha = 0.05 across 168 structural tests (14 before correction).", + "text": "exploratory screen, classified from BH-adjusted component tests; no cell-level FDR guarantee: 11 of 84 cells are topology-dependent after Benjamini-Hochberg at fdr_alpha = 0.05 across 168 structural tests (14 before correction).", "unit": null, "note": null, "step": "dependence" @@ -3607,8 +3607,8 @@ "step": "dependence" }, "dep_land_taste_statement": { - "value": "exploratory screen, FDR-controlled: 54 of 84 cells are topology-dependent after Benjamini-Hochberg at fdr_alpha = 0.05 across 168 structural tests (54 before correction).", - "text": "exploratory screen, FDR-controlled: 54 of 84 cells are topology-dependent after Benjamini-Hochberg at fdr_alpha = 0.05 across 168 structural tests (54 before correction).", + "value": "exploratory screen, classified from BH-adjusted component tests; no cell-level FDR guarantee: 54 of 84 cells are topology-dependent after Benjamini-Hochberg at fdr_alpha = 0.05 across 168 structural tests (54 before correction).", + "text": "exploratory screen, classified from BH-adjusted component tests; no cell-level FDR guarantee: 54 of 84 cells are topology-dependent after Benjamini-Hochberg at fdr_alpha = 0.05 across 168 structural tests (54 before correction).", "unit": null, "note": null, "step": "dependence" @@ -7448,15 +7448,15 @@ "step": "scale" }, "paper_words_body": { - "value": 6971, - "text": "6971", + "value": 5665, + "text": "5665", "unit": "words", "note": null, "step": "paper" }, "paper_words": { - "value": 10424, - "text": "10\u202f424", + "value": 9185, + "text": "9185", "unit": null, "note": null, "step": "paper" @@ -7469,8 +7469,8 @@ "step": "paper" }, "supplement_words": { - "value": 6676, - "text": "6676", + "value": 6970, + "text": "6970", "unit": "words", "note": null, "step": "paper" @@ -7524,10 +7524,10 @@ { "name": "F2_dual_scorecard.png", "path": "figures/F2_dual_scorecard.png", - "bytes": 293669, + "bytes": 244157, "step": "scorecard", "figure": "F2_dual_scorecard", - "caption": "Dual scorecard for imidacloprid and fipronil at 1 \u00b5M. Bars are Hill occupancy of the insect target (blue) and its vertebrate counterpart (orange) for each of the five receptor pairs declared in `library.yaml:selectivity_pairs`. Hatched bars rest on a class placeholder: the compound has no sourced EC50 at that pair and the row is inert. Imidacloprid reaches 0.991 at insect nAChR against 0.091 at vertebrate a4b2; fipronil reaches 0.985 at insect RDL against 0.476 at native heteromeric vertebrate GABA-A. Teaching-tier EC50s: these are literature-order values, not measured constants." + "caption": "Dual scorecard for imidacloprid and fipronil at 1 \u00b5M. Bars are Hill engagement of the insect target (blue) and its vertebrate counterpart (orange) for each of the five receptor pairs declared in `library.yaml:selectivity_pairs`. Hatched entries indicate unsupported parameters and are interpreted as N/A. Imidacloprid reaches 0.991 at insect nAChR against 0.091 at vertebrate a4b2; fipronil reaches 0.985 at insect RDL against 0.476 at native heteromeric vertebrate GABA-A. Teaching-tier EC50s: these are literature-order values, not measured constants." }, { "name": "F11_evidence_types.png", @@ -7535,15 +7535,15 @@ "bytes": 180063, "step": "evidence", "figure": "F11_evidence_types", - "caption": "What the compound library rests on, row by row. Left: the parameter each cited source actually measured. Middle: the ordered evidence distance between that source and this compound at this receptor in this species. Right: the transformation the type system then permits, which is a function of the two. Green marks the 10 on-target (E0) rows and the 1 of them that may use a physical binding-occupancy model; amber marks the proxy models, where the same Hill expression is computed but reported as a transfer with a warning naming the gap; red marks rows that carry no modellable evidence and return N/A rather than a small number. The type system, not a convention, enforces this: asking for an engagement from an untyped row raises rather than returning a value, and a soundness invariant is checked exhaustively over every row and every entry point." + "caption": "What the compound library rests on, row by row. Left: the parameter each cited source actually measured. Middle: the ordered evidence distance between that source and this compound at this receptor in this species. Right: the transformation the evidence dispatch rule permits, which is a function of the two. Green marks the 10 on-target (E0) rows and the 1 of them that may use a physical binding-occupancy model; amber marks the proxy models, where the same Hill expression is computed but reported as a transfer with a warning naming the gap; red marks rows that carry no modellable evidence and return N/A rather than a small number. Runtime validation enforces this: asking for an engagement from an untyped row raises rather than returning a value, and a soundness invariant is audited over shipped rows and enumerated pipeline paths." }, { "name": "F3_occupancy_curves.png", "path": "figures/F3_occupancy_curves.png", - "bytes": 399837, + "bytes": 375928, "step": "curves", "figure": "F3_occupancy_curves", - "caption": "Hill occupancy against free concentration for the four-compound panel, one insect target (solid, blue) and its vertebrate counterpart (dashed, orange) per compound. The shaded band is the selectivity window between the two EC50s, annotated in log10 units; the dotted vertical line is the 1 \u00b5M working concentration used elsewhere in the paper. Ivermectin's window is the GluCl/GlyR pair, deltamethrin's the Nav pair. Curves are the teaching library, not measured concentration-response data." + "caption": "Hill engagement against free concentration for the four-compound panel, one insect target (solid, blue) and its vertebrate counterpart (dashed, orange) per compound. The shaded band is the selectivity window between the two EC50s, annotated in log10 units; the dotted vertical line is the 1 \u00b5M working concentration used elsewhere in the paper. Ivermectin's window is the GluCl/GlyR pair, deltamethrin's the Nav pair. Curves are the teaching library, not measured concentration-response data." }, { "name": "F4_taste_motor_graph.png", @@ -7556,10 +7556,10 @@ { "name": "F5_bitter_veto.png", "path": "figures/F5_bitter_veto.png", - "bytes": 145962, + "bytes": 149905, "step": "veto", "figure": "F5_bitter_veto", - "caption": "MN9 firing on the map-extracted labellar-GRN path, driven through the real MaleCNS sweet (34) and bitter (38) GRN body IDs at 150 Hz, never by driving MN9 directly. Left: leaky rate engine. Right: Shiu-style LIF engine with a background Poisson drive. In vehicle, adding bitter drive takes MN9 to zero on both engines (veto ratio 0.00). Fipronil at 1 \u00b5M blocks RDL, raises the sugar-only rate and partially relieves the veto (ratio 0.41 rate, 0.16 LIF). Imidacloprid silences MN9 altogether, so its ratio is undefined rather than zero. Absolute rates differ between engines by construction and are not calibrated to any recording." + "caption": "MN9 firing on the map-extracted labellar-GRN path, driven through the real MaleCNS sweet (34) and bitter (38) GRN body IDs at a nominal drive of 150 rate-model input units or LIF input spikes/s, never by driving MN9 directly. Left: leaky rate engine. Right: Shiu-style LIF engine with a background Poisson drive. In vehicle, adding bitter drive takes MN9 to zero on both engines (veto ratio 0.00). Fipronil at 1 \u00b5M blocks RDL, raises the sugar-only rate and partially relieves the veto (ratio 0.41 rate, 0.16 LIF). Imidacloprid silences MN9 altogether, so its ratio is undefined rather than zero. Absolute rates differ between engines by construction and are not calibrated to any recording. Left-axis values are rate-model units; right-axis values are simulated LIF spikes/s, not measured fly firing rates." }, { "name": "F6_dependence_profile.png", @@ -7567,7 +7567,7 @@ "bytes": 334549, "step": "dependence", "figure": "F6_dependence_profile", - "caption": "Histograms are the drug effect (treated minus vehicle on the network mean rate) measured on 1000 degraded copies of the 1-hop MN9/DNp01 `named` cut; the red line is the same contrast on the real cut. Columns run in increasing order of destruction: transmitter labels permuted, synapse weights permuted, degree-preserving double-edge rewiring, Erdos-Renyi. Titles carry the empirical two-sided permutation p and the three-way verdict against the prespecified equivalence margin (delta = 0.332 Hz, 5 % of the vehicle readout): `distinguishable` means the test rejected, `equivalent within tolerance` means the gap from the null median is below the margin, and `indeterminate` means neither -- a failure to reject is not evidence that the degraded graph gives the same effect. Plain label permutation is shown because it is informative, but it moves the weighted excitation/inhibition balance as well as transmitter identity (T23), so it is a joint null and not a rung of the ladder. Imidacloprid is composition-dominated (necessary level: degree_sequence, equivalent_within_tolerance); fipronil is topology-dependent (necessary level: wiring_with_weighted_transmitter_balance)." + "caption": "Histograms are the drug effect (treated minus vehicle on the network mean rate) measured on 1000 degraded copies of the 1-hop MN9/DNp01 `named` cut; the red line is the same contrast on the real cut. Columns run in increasing order of destruction: transmitter labels permuted, synapse weights permuted, degree-preserving double-edge rewiring, Erdos-Renyi. Titles carry the empirical two-sided permutation p and the three-way verdict against the configured median-gap tolerance (delta = 0.332 Hz, 5 % of the vehicle readout): `distinguishable` means the test rejected, `equivalent within tolerance` means the gap from the null median is below the margin, and `indeterminate` means neither -- a failure to reject is not evidence that the degraded graph gives the same effect. Plain label permutation is shown because it is informative, but it moves the weighted excitation/inhibition balance as well as transmitter identity (T23), so it is a joint null and not a rung of the ladder. Imidacloprid is composition-dominated (necessary level: degree_sequence, equivalent_within_tolerance); fipronil is topology-dependent (necessary level: wiring_with_weighted_transmitter_balance). The legacy within-tolerance label is a point-gap diagnostic, not a formal equivalence test. Rate-engine quantities labelled Hz in legacy fields or axes are rate-model units, not calibrated physiological firing rates." }, { "name": "F12_dependence_landscape.png", @@ -7575,7 +7575,7 @@ "bytes": 207481, "step": "dependence", "figure": "F12_dependence_landscape", - "caption": "Every compound of the library at four concentrations, on both committed cuts, classified by the weakest graph model the permutation test could **not** distinguish from the real cut -- never by a model that reproduces the effect. Classes are decided on Benjamini-Hochberg adjusted probabilities across each landscape's structural tests, with a prespecified relative effect floor of 1 % of the vehicle readout. Left, the 1-hop `named` cut: 11 of 84 cells need the wiring pattern, 42 are not distinguishable from any structure-preserving degradation, and 31 do not move the readout at all. Right, the denser `taste_motor` cut: 54 topology-dependent, 0 composition-dominated, 30 no-effect. The composition-dominated majority does not survive the change of substrate, and the `named` cut is an in-star whose degree-preserving rewire is close to the identity (T19)." + "caption": "Every compound of the library at four concentrations, on both committed cuts, classified from structural component-test decisions and the configured effect floor. Classes are decided on Benjamini-Hochberg adjusted probabilities across each landscape's structural tests, with a prespecified relative effect floor of 1 % of the vehicle readout. Left, the 1-hop `named` cut: 11 of 84 cells are classified topology-dependent, 42 are not distinguishable from any structure-preserving degradation, and 31 fall below the configured effect floor. Right, the denser `taste_motor` cut: 54 topology-dependent, 0 composition-dominated, 30 no-effect. The composition-dominated majority does not survive the change of substrate, and the `named` cut is an in-star whose degree-preserving rewire is close to the identity (T19). BH adjusts structural component tests; cell/conclusion-level FDR control is not established, and the dependence assumptions remain unverified. Adjustment cannot strengthen non-rejection." }, { "name": "F16_instrument_validation.png", @@ -7583,7 +7583,7 @@ "bytes": 203094, "step": "dependence", "figure": "F16_instrument_validation", - "caption": "The dependence ladder pointed at a graph whose answer is known in advance. A recurrent cholinergic cycle of known strength is planted in a synthetic cut of the same size, density and transmitter composition as the `named` cut; a gain patch that collapses `g_ach` then removes an amplification that exists only while the cycle is intact, so the drug contrast depends on the wiring by construction. Left: the share of independently generated graphs the ladder classifies topology-dependent, against planted strength and permutation count. Right: the recovery experiment itself, with red marking the strengths the ladder recovered. The unplanted control is the empirical false-positive rate (0 of 18). Without this panel, a negative from the ladder would be indistinguishable from an under-powered test." + "caption": "Synthetic recovery checks with an imposed cholinergic cycle and a gain patch. The synthetic network is not size- or density-matched to the named connectome cut. The zero-strength case omits cycle edges and associated transmitter relabelling. Detection frequencies in this small grid have broad uncertainty; networks are reused across permutation counts. These checks do not establish general calibration or power for the real taste-motor ratio endpoint. Counts and settings are shown in the artifact. Rate-engine quantities labelled Hz in legacy fields or axes are rate-model units, not calibrated physiological firing rates." }, { "name": "F13_ablation.png", @@ -7591,7 +7591,7 @@ "bytes": 185319, "step": "ablation", "figure": "F13_ablation", - "caption": "Each level predicts the same 21 compounds with one layer of information removed, and is scored by how well it reproduces the full model's *ordering* of the library (levels have different units, so values are not comparable; the correlation is signed, so a perfect inversion scores -1 rather than 1). Crosses mark the historical depression-only level-C rule, which applied a multiplier of at most 1.0 to every transmitter alike and so could not express disinhibition: its poor correlation was structural, not a finding about connectomes, and the conclusion drawn from it is withdrawn. With a direction-aware rule level C rises to 0.961 at 1 uM from 0.242. The composition level's agreement with the full model must be read against the matched reference distribution of T24, not on its own." + "caption": "Each level predicts the same 21 compounds with one layer of information removed, and is scored by how well it reproduces the full model's *ordering* of the library (levels have different units, so values are not comparable; the correlation is signed, so a perfect inversion scores -1 rather than 1). Crosses mark the historical depression-only level-C rule, which applied a multiplier of at most 1.0 to every transmitter alike and so could not express disinhibition: its poor correlation was structural, not a finding about connectomes, and the conclusion drawn from it is withdrawn. With a direction-aware rule level C rises to 0.961 at 1 uM from 0.242. The composition level's agreement with the full model must be read against the matched reference distribution of T24, not on its own. Rate-engine quantities labelled Hz in legacy fields or axes are rate-model units, not calibrated physiological firing rates." }, { "name": "F7_selectivity_landscape.png", @@ -7599,7 +7599,7 @@ "bytes": 248793, "step": "landscape", "figure": "F7_selectivity_landscape", - "caption": "Each point is one compound on one committed graph. The x axis is the receptor selectivity index from the library alone; the y axis is the circuit index, the log10 width of the window between the concentration that moves the simulated network mean rate by 50% and the concentration at which the most potent vertebrate target reaches 20% occupancy. Points below the dashed identity line are buffered by the circuit, points above are amplified. The split is by mechanism, not by potency: chlorpyrifos_oxon, ddt, deltamethrin, permethrin (Nav modulators and the AChE inhibitor) amplify, mean gap +0.29 log10 units, while the nAChR-acting set (acetamiprid, acetylcholine, clothianidin, imidacloprid, nicotine, nitenpyram) buffers, mean gap -0.87. Compounds with no circuit index (chlordimeform, dieldrin, fipronil, gaba, ivermectin, picrotoxin, spinosad, sulfoxaflor, thiamethoxam) never reach the 50% threshold on this cut and are omitted." + "caption": "Each point is one compound on one committed graph. The x axis is the receptor selectivity index from the library alone; the y axis is the circuit index, the log10 width of the window between the concentration that moves the simulated network mean rate by 50% and the concentration at which the most potent vertebrate target reaches 20% engagement. Points below the dashed identity line are buffered by the circuit, points above are amplified. The split is by mechanism, not by potency: chlorpyrifos_oxon, ddt, deltamethrin, permethrin (Nav modulators and the AChE inhibitor) amplify, mean gap +0.29 log10 units, while the nAChR-acting set (acetamiprid, acetylcholine, clothianidin, imidacloprid, nicotine, nitenpyram) buffers, mean gap -0.87. Compounds with no circuit index (chlordimeform, dieldrin, fipronil, gaba, ivermectin, picrotoxin, spinosad, sulfoxaflor, thiamethoxam) never reach the 50% threshold on this cut and are omitted." }, { "name": "F8_dose_response_sensitivity.png", @@ -7607,15 +7607,15 @@ "bytes": 182736, "step": "dose", "figure": "F8_dose_response_sensitivity", - "caption": "Left: the model's own dose-response for imidacloprid on the 1-hop MN9/DNp01 cut. Points are 4 replicates per concentration with log-normal EC50 jitter (sd 0.3 log10 units) and drive jitter; the curve is a 4-parameter Hill fit, and the band is a 200-resample bootstrap CI on the fitted IC50 (34.7 nM). This is the *model's* circuit IC50, not a receptor IC50 and not an animal IC50. Right: one-at-a-time sensitivity of the same readout at 1 \u00b5M, each parameter halved and doubled. The gain-rule coefficient dominates; the EC50 and Hill coefficient have zero span because occupancy is already saturated at this dose." + "caption": "Left: the model's own dose-response for imidacloprid on the 1-hop MN9/DNp01 cut. Points are 4 replicates per concentration with log-normal EC50 jitter (sd 0.3 log10 units) and drive jitter; the curve is a 4-parameter Hill fit, and the band is a 200-resample bootstrap CI on the fitted IC50 (34.7 nM). This is the *model's* circuit IC50, not a receptor IC50 and not an animal IC50. Right: one-at-a-time sensitivity of the same readout at 1 \u00b5M, each parameter halved and doubled. The gain-rule coefficient dominates; the EC50 and Hill coefficient have zero span because engagement is already saturated at this dose. Rate-engine quantities labelled Hz in legacy fields or axes are rate-model units, not calibrated physiological firing rates." }, { "name": "F9_genotype_panel.png", "path": "figures/F9_genotype_panel.png", - "bytes": 179353, + "bytes": 173103, "step": "genotype", "figure": "F9_genotype_panel", - "caption": "Target-receptor occupancy at 1 uM for wild type and each published resistance allele that has a sourced fold-shift for that compound. Grey bars are alleles with no sourced number for the compound: they are left identical to wild type and reported, never extrapolated. Left and centre: fipronil and imidacloprid. Right: the specificity check that motivates the per-compound rule. *para* M918T (super-kdr) shifts deltamethrin 100-fold but leaves DDT at 1-fold, and *Rdl* A301S shifts GABA 33.9-fold while leaving fipronil at 1.1-fold. A toggle that shifted a whole receptor uniformly would be wrong in both cases." + "caption": "Target-receptor engagement at 1 uM for wild type and each published resistance allele that has a sourced fold-shift for that compound. Grey bars are alleles with no sourced number for the compound: they are left identical to wild type and reported, never extrapolated. Left and centre: fipronil and imidacloprid. Right: the specificity check that motivates the per-compound rule. *para* M918T (super-kdr) shifts deltamethrin 100-fold but leaves DDT at 1-fold, and *Rdl* A301S shifts GABA 33.9-fold while leaving fipronil at 1.1-fold. A toggle that shifted a whole receptor uniformly would be wrong in both cases." }, { "name": "F10_rank_validation.png", @@ -7631,7 +7631,7 @@ "bytes": 263855, "step": "stability", "figure": "F14_conclusion_stability", - "caption": "Each row is a conclusion this paper could state; each column is one prespecified, admissible way of turning receptor engagement into synaptic gain. Green retains the conclusion, red reverses it, amber means the readout does not exist under that specification. The topology rows are decided on 100 permutations per specification and on Benjamini-Hochberg adjusted probabilities across the 100 structural tests of the run; until v0.6.1 the specification never reached the engine those rows were computed on, so their uniformity was guaranteed by construction rather than measured. Retained by every specification: C2_rdl_disinhibition, C3_imidacloprid_topology_not_distinguishable, C4_fipronil_topology_exceeds, C7_map_bitter_veto. The nicotinic suppression conclusion is reversed by every monotone rule, which is how a result that depends on a modelling choice looks when it is tested rather than asserted; and for a conclusion shaped as a failure to reject, the fraction retained is only as strong as the equivalence behind it (T14's `n_equivalent_within_tolerance` against `n_indeterminate`)." + "caption": "Each row is a conclusion this paper could state; each column is one fixed, admissible way of turning receptor engagement into synaptic gain. Green retains the conclusion, red reverses it, amber means the readout does not exist under that specification. The topology rows are decided on 100 permutations per specification and on Benjamini-Hochberg adjusted probabilities across the 100 structural tests of the run; until v0.6.1 the specification never reached the engine those rows were computed on, so their uniformity was guaranteed by construction rather than measured. Retained by every specification: C2_rdl_disinhibition, C3_imidacloprid_topology_not_distinguishable, C4_fipronil_topology_exceeds, C7_map_bitter_veto. The nicotinic suppression conclusion is reversed by every monotone rule, which is how a result that depends on a modelling choice looks when it is tested rather than asserted; and for a conclusion shaped as a failure to reject, the fraction retained is a non-rejection, not evidence of equivalence (T14's `n_equivalent_within_tolerance` against `n_indeterminate`). The legacy within-tolerance label is a point-gap diagnostic, not a formal equivalence test. BH adjusts structural component tests; cell/conclusion-level FDR control is not established, and the dependence assumptions remain unverified. Adjustment cannot strengthen non-rejection." }, { "name": "F15_uncertainty_voi.png", @@ -7639,20 +7639,20 @@ "bytes": 192844, "step": "uncertainty", "figure": "F15_uncertainty_voi", - "caption": "Left: the share of the variance of the neighbourhood mean rate attributable to each assumption, first-order and total-order, from a Saltelli cross-sample with Jansen estimators, with bootstrap first-order confidence intervals. A factor is **resolved** only when its interval excludes zero; the starred rows are not, and a factor's index exceeding the noise floor does not resolve it. The dashed line is that floor, defined as the largest magnitude among the *negative* first-order estimates: a true first-order index cannot be negative, so that magnitude is a measured lower bound on the estimator's error, and on this run it is several times the declared null factor's own draw (0.046 set by gain_coef, against lif_seed at -0.007). Right: the same shares scaled back into the readout's variance and mapped onto the experiment that would resolve each assumption; a negative estimate is clipped to zero for the decision value, because a value of information cannot be negative. Green marks a factor that needs no experiment at all, only a re-analysis of data already held." + "caption": "Left: the share of the variance of the neighbourhood mean rate attributable to each assumption, first-order and total-order, from a Saltelli cross-sample with Jansen estimators, with bootstrap first-order confidence intervals. A factor is **resolved** only when its interval excludes zero; the starred rows are not, and a factor's index exceeding the negative-estimate diagnostic does not resolve it. The dashed line is that floor, defined as the largest magnitude among the *negative* first-order estimates: a true first-order index cannot be negative, so that magnitude is a measured lower bound on the estimator's error, and on this run it is several times the declared null factor's own draw (0.046 set by gain_coef, against lif_seed at -0.007). Right: the same shares scaled back into the readout's variance and mapped onto the experiment that would resolve each assumption; a negative estimate is clipped to zero for the decision value, because a value of information cannot be negative. Green marks a factor that needs no experiment at all, only a re-analysis of data already held. The largest negative estimate is an observed error diagnostic, not a calibrated uncertainty bound. VOI concerns output variance under assumed model input ranges, not the benefit of an experiment on a fly. Rate-engine quantities labelled Hz in legacy fields or axes are rate-model units, not calibrated physiological firing rates." }, { "name": "F1_architecture.png", "path": "figures/F1_architecture.png", - "bytes": 288705, + "bytes": 349802, "step": "architecture", "figure": "F1_architecture", - "caption": "FlyLab dataflow. A compound and a free concentration enter the occupancy engine, which reads the sourced library (optionally shifted by a genotype, combined with a mixture partner, or driven by an exposure profile). Occupancy splits into two scorecards. The insect side passes through the mechanism table to a gain patch on transmitter classes of a named MaleCNS cut; the vertebrate side is scored at the same dose and never touches a circuit. The circuit runtime (rate or LIF) produces named-cell readouts, which the analysis and validation layers turn into empirical permutation probabilities with their three-way verdicts, selectivity indices, rank correlations and prospective predictions. Everything lands in one notebook JSON with a provenance block." + "caption": "FlyLab dataflow. A compound and a free concentration enter the engagement engine, which reads the sourced library (optionally shifted by a genotype, combined with a mixture partner, or driven by an exposure profile). Engagement splits into two scorecards. The insect side passes through the mechanism table to a gain patch on transmitter classes of a named MaleCNS cut; the vertebrate side is scored at the same dose and never touches a circuit. The circuit runtime (rate or LIF) produces named-cell readouts, which the analysis and validation layers turn into empirical permutation probabilities with their three-way verdicts, selectivity indices, rank correlations and prospective predictions. Everything lands in one notebook JSON with a provenance block." }, { "name": "F1_architecture.svg", "path": "figures/F1_architecture.svg", - "bytes": 128097, + "bytes": 126262, "step": "architecture", "figure": "F1_architecture", "caption": null @@ -7697,7 +7697,7 @@ { "name": "T20_transformation_rule.md", "path": "tables/T20_transformation_rule.md", - "bytes": 3132, + "bytes": 3239, "rows": 40, "step": "evidence" }, @@ -7781,7 +7781,7 @@ { "name": "T6_dependence_profile.md", "path": "tables/T6_dependence_profile.md", - "bytes": 2682, + "bytes": 3091, "rows": 15, "step": "dependence" }, @@ -7823,7 +7823,7 @@ { "name": "T22_instrument_validation.md", "path": "tables/T22_instrument_validation.md", - "bytes": 1483, + "bytes": 1495, "rows": 16, "step": "dependence" }, @@ -7837,7 +7837,7 @@ { "name": "T12_dependence_landscape.md", "path": "tables/T12_dependence_landscape.md", - "bytes": 9982, + "bytes": 10169, "rows": 84, "step": "dependence" }, @@ -7851,7 +7851,7 @@ { "name": "T12b_dependence_landscape_taste_motor.md", "path": "tables/T12b_dependence_landscape_taste_motor.md", - "bytes": 10227, + "bytes": 10414, "rows": 84, "step": "dependence" }, @@ -7893,7 +7893,7 @@ { "name": "T13_ablation.md", "path": "tables/T13_ablation.md", - "bytes": 7570, + "bytes": 7695, "rows": 84, "step": "ablation" }, @@ -7907,7 +7907,7 @@ { "name": "T13b_ablation_information.md", "path": "tables/T13b_ablation_information.md", - "bytes": 2318, + "bytes": 2443, "rows": 20, "step": "ablation" }, @@ -8005,7 +8005,7 @@ { "name": "T14_conclusion_stability.md", "path": "tables/T14_conclusion_stability.md", - "bytes": 1759, + "bytes": 2050, "rows": 7, "step": "stability" }, @@ -8033,7 +8033,7 @@ { "name": "T16_uncertainty_budget.md", "path": "tables/T16_uncertainty_budget.md", - "bytes": 1765, + "bytes": 2112, "rows": 10, "step": "uncertainty" }, @@ -8047,7 +8047,7 @@ { "name": "T17_value_of_information.md", "path": "tables/T17_value_of_information.md", - "bytes": 1971, + "bytes": 2096, "rows": 9, "step": "uncertainty" }, @@ -8089,42 +8089,62 @@ { "name": "T27_scale_study.md", "path": "tables/T27_scale_study.md", - "bytes": 2783, + "bytes": 2969, "rows": 14, "step": "scale" } ], "notes": [ "[mechanisms] 5 of the 7 transmitter sign magnitudes the rate engine iterates (glutamate -0.4, histamine -0.5, dopamine +0.2, serotonin +0.2, octopamine +0.2) are asserted rather than measured, are exempt from the specification family and are not bounded by the label-permuting null. They are now published in T21 rather than left in the source.", - "[graph] the `named` cut is an in-star: 85% of its 1360 edges terminate on its 4 seed cells, only 184 of 1126 nodes receive any input and only 25% of edges leave a node that itself receives input (mean degree 1.21). A degree-preserving rewire of such a graph is close to the identity, so a negative topology verdict on this cut is weak evidence. The `taste_motor` cut (mean degree 10.36) is the substrate to generalise from.", + "[graph] the `named` cut is an in-star: 85% of its 1360 edges terminate on its 4 seed cells, only 184 of 1126 nodes receive any input and only 25% of edges leave a node that itself receives input (mean degree 1.21). A degree-preserving rewire of such a graph is close to the identity, so a negative topology verdict on this cut is weak evidence. The `taste_motor` cut (mean degree 10.36) is a denser comparison, not evidence of whole-network generality.", "[graph] the `named` cut contains no octopaminergic cell at all, so a compound whose only insect mechanism acts through g_oct (chlordimeform) cannot move this readout for substrate reasons rather than pharmacological ones.", - "[veto] imidacloprid at 1 uM silences MN9 on the map path on both engines (sugar-driven MN9 = 0 Hz), so its bitter-veto ratio is undefined, not 0. g_ach hits the 0.05 floor, which removes the cholinergic GRN->MN9 drive itself. This is a property of the agonist patch rule, and it is why the taste-map arm of the null panel is undefined for imidacloprid.", - "[dependence] the transmitter null matters to the equivalence claim and not to the probability: imidacloprid's gap from the null median is 2.65 Hz under the plain label permutation and 0.24 Hz under the weight-matched null, against a prespecified margin of 0.332 Hz. The plain permutation is a joint target-set-and-sign null and is reported as one.", - "[dependence] the fipronil bitter-veto ratio on the taste-motor path is a negative at n=300 shuffles (max |z| = 0.220, p resolution 0.0033): the veto ratio is model behaviour, not evidence about the wiring, and the verdicts say which modes are equivalent within tolerance (none) and which are merely indeterminate (erdos_renyi, rewire_degree_preserving, sign_permute, sign_permute_weight_matched, weight_permute).", - "[dependence] plain transmitter-label permutation does NOT preserve the weighted excitation/inhibition balance: the cholinergic share of total outgoing synaptic weight is 0.619 on the real `named` cut against 0.544 +- 0.027 over 1000 permutations, putting the real graph at the 100th percentile of its own null. The weight-matched null holds every tracked share within 0.002.", - "[dependence] instrument validation: the ladder recovered 2 of 3 planted topology-dependent effects at n = 200 and returned no false positive on the unplanted control. Detection over the planted grid by permutation count is n=50: 0.61, n=200: 0.61, n=1000: 0.67 and by planted strength strength 0.25: 0.00-0.17; strength 0.5: 0.67-0.83; strength 1: 1.00-1.00; the empirical false-positive rate on the unplanted control is 0 of 18. Effect size dominates the permutation count: full power at planted strength 1, detection at or below one in four at 0.25.", - "[dependence] multiplicity changes the headline count on the `named` cut: 14 cells are topology-dependent on the raw permutation p and 11 survive Benjamini-Hochberg across the 168 structural tests of the landscape. The corrected count is the one the paper quotes.", - "[dependence] the central RQ2 result REVERSES on a cut that has topology: repeated on `taste_motor` (1841 nodes, 19066 edges, mean degree 10.4) the landscape returns 0 composition-dominated cells of 84, against 42 of 84 on the in-star `named` cut. The composition-dominated majority is a property of the substrate, not of receptor perturbation on connectomes.", - "[ablation] the ablation ladder's ranking is concentration-dependent: with the historical depression-only level-C rule the weakest level is 10 nM: C_topology_only; 100 nM: A_receptor_only; 1 \u00b5M: A_receptor_only; 10 \u00b5M: A_receptor_only. Any sentence of the form 'the topology-only level is the worst of the four' is true of one column and false of the others.", - "[ablation] the level-C conclusion is withdrawn: with a direction-aware generic rule the topology-only baseline's rank correlation with the full model at 1 uM rises from 0.242 (depression-only floor, which cannot express disinhibition and whose every entry is at most zero) to 0.961. 'The connectome without the pharmacology is not a cheap substitute' was a statement about a sign-broken baseline.", - "[ablation] the composition-versus-full correlation carries no compound-level information: shuffling the compound-to-gain assignment leaves it identically 0.988, and pharmacology-free pseudo-compounds whose gain vectors have the shape the mechanism rules produce already reach a median of 0.847. Both levels are functions of the same gain vector, so the number is a statement about the map from gains to readouts.", - "[ablation] the composition verdict is not robust to the engine's row normalisation: the B-versus-D rank correlation at 1 uM is 0.988 under the shipped `row_abs` mode and 0.666 under a degree-corrected one. The normalisation is an undocumented modelling choice that makes each cell's recurrent input a composition-weighted average of its presynaptic gains.", + "[veto] imidacloprid at 1 uM silences MN9 on the map path on both engines (sugar-driven MN9 = 0 rate-model units), so its bitter-veto ratio is undefined, not 0. g_ach hits the 0.05 floor, which removes the cholinergic GRN->MN9 drive itself. This is a property of the agonist patch rule, and it is why the taste-map arm of the null panel is undefined for imidacloprid.", + "[dependence] the transmitter null changes both the descriptive median-gap check and the ensemble probabilities: imidacloprid's gap from the null median is 2.65 rate-model units under the plain label permutation and 0.24 rate-model units under the weight-matched null, against a prespecified margin of 0.332 rate-model units. The plain permutation is a joint target-set-and-sign null and is reported as one. These diagnostics do not establish statistical equivalence, cell-level FDR control or endpoint-specific power.", + "[dependence] the fipronil bitter-veto ratio on the taste-motor path is a negative at n=300 shuffles (max |z| = 0.220, p resolution 0.0033): the veto ratio is model behaviour, not evidence about the wiring, and the verdicts say which modes are equivalent within tolerance (none) and which are merely indeterminate (erdos_renyi, rewire_degree_preserving, sign_permute, sign_permute_weight_matched, weight_permute). These diagnostics do not establish statistical equivalence, cell-level FDR control or endpoint-specific power.", + "[dependence] plain transmitter-label permutation does NOT preserve the weighted excitation/inhibition balance: the cholinergic share of total outgoing synaptic weight is 0.619 on the real `named` cut against 0.544 +- 0.027 over 1000 permutations, putting the real graph at the 100th percentile of its own null. The weight-matched null holds every tracked share within 0.002. These diagnostics do not establish statistical equivalence, cell-level FDR control or endpoint-specific power.", + "[dependence] instrument validation: the ladder recovered 2 of 3 planted topology-dependent effects at n = 200 and returned no false positive on the unplanted control. Detection over the planted grid by permutation count is n=50: 0.61, n=200: 0.61, n=1000: 0.67 and by planted strength strength 0.25: 0.00-0.17; strength 0.5: 0.67-0.83; strength 1: 1.00-1.00; the empirical false-positive rate on the unplanted control is 0 of 18. All sampled networks were detected at planted strength 1, detection at or below one in four at 0.25. These diagnostics do not establish statistical equivalence, cell-level FDR control or endpoint-specific power.", + "[dependence] multiplicity changes the headline count on the `named` cut: 14 cells are topology-dependent on the raw permutation p and 11 survive Benjamini-Hochberg across the 168 structural tests of the landscape. The corrected count is the one the paper quotes. These diagnostics do not establish statistical equivalence, cell-level FDR control or endpoint-specific power.", + "[dependence] the central RQ2 result REVERSES on a cut that has topology: repeated on `taste_motor` (1841 nodes, 19066 edges, mean degree 10.4) the landscape returns 0 composition-dominated cells of 84, against 42 of 84 on the in-star `named` cut. The composition-dominated majority is a property of the substrate, not of receptor perturbation on connectomes. These diagnostics do not establish statistical equivalence, cell-level FDR control or endpoint-specific power.", + "[ablation] the ablation ladder's ranking is concentration-dependent: with the historical depression-only level-C rule the weakest level is 10 nM: C_topology_only; 100 nM: A_receptor_only; 1 \u00b5M: A_receptor_only; 10 \u00b5M: A_receptor_only. Any sentence of the form 'the topology-only level is the worst of the four' is true of one column and false of the others. Jointly permuting paired scores preserves correlation by construction; the observed correlation exceeds the matched reference 95th percentile.", + "[ablation] the level-C conclusion is withdrawn: with a direction-aware generic rule the topology-only baseline's rank correlation with the full model at 1 uM rises from 0.242 (depression-only floor, which cannot express disinhibition and whose every entry is at most zero) to 0.961. 'The connectome without the pharmacology is not a cheap substitute' was a statement about a sign-broken baseline. Jointly permuting paired scores preserves correlation by construction; the observed correlation exceeds the matched reference 95th percentile.", + "[ablation] the composition-versus-full correlation must be interpreted against shared assumptions and its reference distribution: shuffling the compound-to-gain assignment leaves it identically 0.988, and pharmacology-free pseudo-compounds whose gain vectors have the shape the mechanism rules produce already reach a median of 0.847. Both levels are functions of the same gain vector, so the number is a statement about the map from gains to readouts. Jointly permuting paired scores preserves correlation by construction; the observed correlation exceeds the matched reference 95th percentile.", + "[ablation] the composition verdict is not robust to the engine's row normalisation: the B-versus-D rank correlation at 1 uM is 0.988 under the shipped `row_abs` mode and 0.666 under a degree-corrected one. The normalisation is an undocumented modelling choice that makes each cell's recurrent input a composition-weighted average of its presynaptic gains. Jointly permuting paired scores preserves correlation by construction; the observed correlation exceeds the matched reference 95th percentile.", "[landscape] no circuit selectivity index exists for chlordimeform, dieldrin, fipronil, gaba, ivermectin, picrotoxin, spinosad, sulfoxaflor, thiamethoxam: on this neighbourhood their largest relative change in the mean rate never reaches the 50% threshold (RDL/GluCl blockers are bounded by the GABA/glutamate share of the cut and by the 0.05 gain floor).", - "[dose] at 1 uM imidacloprid the model is exactly insensitive to a 2x change in the insect_nAChR EC50 and in the Hill coefficient (span 0 Hz): occupancy is saturated, so the gain-rule coefficient is the only parameter that still moves the readout. A dose-response experiment, not a single high dose, is what constrains the EC50.", + "[dose] at 1 uM imidacloprid the model is exactly insensitive to a 2x change in the insect_nAChR EC50 and in the Hill coefficient (span 0 rate-model units): occupancy is saturated, so the gain-rule coefficient is the only parameter that still moves the readout. A dose-response experiment, not a single high dose, is what constrains the EC50.", "[validation] the literature-concordance mean is not quotable bare: 4 of the 8 evaluable comparisons order exactly two compounds, where Spearman's rho can only be +-1 and the two-sided p is 1.0 whatever the model does. Over the 4 comparisons with more than two compounds the mean is 0.756; as a concordance count it is 8 concordant against 0 discordant orderings (two-sided binomial p = 0.008).", - "[stability] the C3 retention rate is no longer vacuous and is no longer uniform: at n_shuffles = 100 (resolution 0.0099, against the previous 6 at which p <= 0.05 was arithmetically unattainable) the conclusion is retained by 25 of 25 specifications, but only 17 of those reach equivalence within the prespecified margin; 8 are merely indeterminate.", - "[stability] the suppression conclusion is specification-dependent: 10 of 25 prespecified specifications reverse it. Under linear@1.00x the same compound at the same engagement EXCITES the network (10.1 Hz treated vs 6.631 Hz vehicle) instead of suppressing it. Suppression is a property of the biphasic desensitisation term, not of the pharmacology-connectome integration.", - "[stability] the supplement's claim that the monotone rules 'buffer at least as much, never less' is false: 3 of the 6 reversing specifications with a defined index buffer LESS than the shipped rule (linear@1.50x, saturating@1.25x, saturating@1.50x), and the aggregate of -0.893 is carried by the extreme of a range running -0.199 to -1.734.", - "[stability] T8 and T15 have different scopes and must be read that way: T8 scores the whole library on both cuts, T15 only the 10 compounds of the two mechanism classes the conclusion is about. acetylcholine carries a circuit index in T8 and is outside T15's compound set entirely.", - "[stability] the amplify/buffer split is NOT stable across the 3x3 threshold grid: nav_ache fails at circuit_frac=0.50,vert_limit=0.10. The paper quotes the range over the grid, never the point estimate.", - "[uncertainty] the estimator's noise floor is 0.046, set by the most negative first-order estimate (gain_coef), not by the declared null factor lif_seed at -0.007. Only gain_transform, weight_threshold have a first-order confidence interval that excludes zero; every other factor is unresolved at this sample size, and a negative point estimate is estimator error rather than a negative contribution.", + "[stability] the C3 retention rate uses an attainable rejection threshold but is a non-rejection criterion: at n_shuffles = 100 (resolution 0.0099, against the previous 6 at which p <= 0.05 was arithmetically unattainable) the conclusion is retained by 25 of 25 specifications, but only 17 of those pass the descriptive median-gap tolerance; 8 are merely indeterminate. Adjusted non-rejection is not strengthened evidence for absence of topology dependence.", + "[stability] the suppression conclusion is specification-dependent: 10 of 25 prespecified specifications reverse it. Under linear@1.00x the same compound at the same engagement EXCITES the network (10.1 rate-model units treated vs 6.631 rate-model units vehicle) instead of suppressing it. Suppression is a property of the biphasic desensitisation term, not of the pharmacology-connectome integration. Adjusted non-rejection is not strengthened evidence for absence of topology dependence.", + "[stability] the supplement's claim that the monotone rules 'buffer at least as much, never less' is false: 3 of the 6 reversing specifications with a defined index buffer LESS than the shipped rule (linear@1.50x, saturating@1.25x, saturating@1.50x), and the aggregate of -0.893 is carried by the extreme of a range running -0.199 to -1.734. Adjusted non-rejection is not strengthened evidence for absence of topology dependence.", + "[stability] T8 and T15 have different scopes and must be read that way: T8 scores the whole library on both cuts, T15 only the 10 compounds of the two mechanism classes the conclusion is about. acetylcholine carries a circuit index in T8 and is outside T15's compound set entirely. Adjusted non-rejection is not strengthened evidence for absence of topology dependence.", + "[stability] the amplify/buffer split is NOT stable across the 3x3 threshold grid: nav_ache fails at circuit_frac=0.50,vert_limit=0.10. The paper quotes the range over the grid, never the point estimate. Adjusted non-rejection is not strengthened evidence for absence of topology dependence.", + "[uncertainty] the largest observed negative-estimate magnitude is 0.046, set by the most negative first-order estimate (gain_coef), not by the declared null factor lif_seed at -0.007. Only gain_transform, weight_threshold have a first-order confidence interval that excludes zero; every other factor is unresolved at this sample size, and a negative point estimate is estimator error rather than a negative contribution.", "[claims] the provenance chain behind a circuit readout has 9 links and exactly 1 of them (malecns_edges) is a measurement of the system being simulated; the rest are literature reuse, asserted modelling choices and computation.", - "[scale] scale study: 7 cuts (scale_1k, named, taste_motor, scale_5k, scale_10k, scale_25k, scale_50k); verdicts settled on the clean ladder: yes. imidacloprid: the verdict is 'topology-dependent' on the largest 5 of 7 cuts tested, first reached at 'taste_motor'. fipronil: the verdict is 'topology-dependent' on the largest 7 of 7 cuts tested, first reached at 'scale_1k'." + "[scale] scale study: 7 cuts (scale_1k, named, taste_motor, scale_5k, scale_10k, scale_25k, scale_50k); labels match on the largest tested clean-ladder rungs: yes. imidacloprid: the verdict is 'topology-dependent' on the largest 5 of 7 cuts tested, first reached at 'taste_motor'. fipronil: the verdict is 'topology-dependent' on the largest 7 of 7 cuts tested, first reached at 'scale_1k'. This is a cached descriptive comparison, not a new scale-study rerun, convergence proof or isolated recurrence effect." ], "partial": false, "honesty": [ "Every number in this file is model_derived: it comes from the teaching EC50 library, the gain patch rules and one or two hops-limited MaleCNS cuts. Nothing here was measured in a living fly.", "live_lab is null in every notebook this pipeline produced.", "--fast reduces replicate and shuffle counts only; z-scores and CIs from a fast run are noisier than the paper's." - ] + ], + "review_metadata": { + "mode": "editorial replay; not a completed numerical rerun", + "numerical_source_git_sha": "f95fd88a2165e66d5203db6c6fa8c6a020c5a8e0", + "review_source_git_sha": "24431be9597646988c2862a18ea55f3d87d0db05", + "changed_design_fields": [ + "dep_land_design", + "dep_land_statement", + "dep_land_taste_design", + "dep_land_taste_statement", + "dep_named_fipronil_confirmatory", + "dep_named_fipronil_design", + "dep_named_imidacloprid_confirmatory", + "dep_named_imidacloprid_design", + "dep_taste_fipronil_confirmatory", + "dep_taste_fipronil_design" + ], + "numerical_value_payloads_preserved": true, + "table_csv_bytes_preserved": true, + "full_rerun_status": "interrupted during dependence analysis; no completed rerun claimed" + } } diff --git a/papers/reviews/current_reference_audit.md b/papers/reviews/current_reference_audit.md new file mode 100644 index 0000000..1b57d51 --- /dev/null +++ b/papers/reviews/current_reference_audit.md @@ -0,0 +1,17 @@ +# Current reference audit + +Audit date: 2026-09-20. Scope: the two manuscript templates only. + +## Corrections applied + +- `[2]` is now Okabe, Goto, Hirayama and Nishimura, “Virtual rat web: A versatile simulation tool for pharmacology education in a variety of settings,” *European Journal of Pharmacology* 997 (2025):177618, DOI [`10.1016/j.ejphar.2025.177618`](https://doi.org/10.1016/j.ejphar.2025.177618), PMID [`40222443`](https://pubmed.ncbi.nlm.nih.gov/40222443/). +- `[3]` is now Thiyagarajan, “PharmVR-Immersive Virtual Reality based Animal Simulator Software,” *Preprints.org* (2025), DOI [`10.20944/preprints202507.0959.v1`](https://doi.org/10.20944/preprints202507.0959.v1). The source PDF ([primary record](https://www.preprints.org/frontend/manuscript/b156435455d18d42e6573a5aa04425a8/download_pub)) labels it a concept paper and “NOT PEER-REVIEWED.” +- The former `[4]`, “Virtual simulation in pharmacology and physiology education: an evaluation,” was removed: no matching article, author list, DOI, PMID, or JEEHP record could be verified. +- All citations and bibliography entries formerly numbered `[5]`–`[77]` were shifted down by one. The resulting list is `[1]`–`[76]`. +- New `[9]` is Eckstein et al., “Neurotransmitter classification from electron microscopy images at synaptic sites in *Drosophila melanogaster*,” *Cell* 187 (2024):2574–2594.e23, DOI [`10.1016/j.cell.2024.03.016`](https://doi.org/10.1016/j.cell.2024.03.016), PMID [`38729112`](https://pubmed.ncbi.nlm.nih.gov/38729112/), PMCID [`PMC11106717`](https://pmc.ncbi.nlm.nih.gov/articles/PMC11106717/). This is the primary transmitter-classification source; `FlyConnectome` was not a verifiable publication with the old title. +- New `[10]` identifies Patrick Mineault’s “Deconstructing viral fly sims” as a *NeuroAI archive* newsletter essay dated 14 September 2026, with the exact URL [`neuroai.science/p/are-flies-playing-beat-saber`](https://www.neuroai.science/p/are-flies-playing-beat-saber). It is explicitly not described as a technical note or peer-reviewed evidence. +- New `[76]` is valid: Tomizawa, Latli and Casida, “Novel neonicotinoid-agarose affinity column for *Drosophila* and *Musca* nicotinic acetylcholine receptors,” *Journal of Neurochemistry* 67 (1996):1669–1676, DOI [`10.1046/j.1471-4159.1996.67041669.x`](https://doi.org/10.1046/j.1471-4159.1996.67041669.x), PMID [`8858952`](https://pubmed.ncbi.nlm.nih.gov/8858952/). The PubMed abstract reports *Drosophila* head-membrane [3H]imidacloprid binding Kd values of 1–2 nM. + +## Wording limits retained + +The supplement now calls `{{expr_coverage_overall}}` a fraction of cell–receptor slots with a non-unknown ordinal class mapping, not measured expression coverage. It also states that `LB3b/c` “sweet” and `LB1a-d` “bitter” are working hypotheses from gustatory nomenclature, not MaleCNS modality annotations. The evidence section reserves physical `binding_occupancy` for exact compound/receptor/species E0 Kd/Ki evidence and treats transferred constants as proxies. diff --git a/papers/reviews/current_revision_decisions.md b/papers/reviews/current_revision_decisions.md new file mode 100644 index 0000000..9125ab6 --- /dev/null +++ b/papers/reviews/current_revision_decisions.md @@ -0,0 +1,62 @@ +# Review decisions for the current manuscript + +Base: main `38cad67`. This revision first compared the newer Fable/Opus changes with the earlier `round2_referee_report.md` and the Grok review supplied by the author. Earlier counts and defects were not assumed to describe the current manuscript. Review is internal and AI-assisted, not journal peer review. + +## Review rounds + +1. Independent statistical review of the updated manuscript, analysis code and saved results, plus a separate reference audit. Reports: `current_round1_statistical.md` and `current_reference_audit.md`. +2. Independent rereview of the revised templates against implementation details. Report: `current_round2_statistical.md`. This found residual component-test versus cell-level multiplicity claims, asymmetric interpretation of non-rejection, stale confirmatory labels, a contradictory C4 warning, and overstated invariant-test coverage. +3. Integrating review of manuscript, supplement, generator, project claim guidance and submission checklist. Accepted changes were checked against their implementation; remaining limitations are recorded below. This round also repaired prose and checked that the earlier reviews were applied to the current evidence rather than their superseded snapshots. + +The statistical and implementation reviews used GPT-5.6 Sol with high reasoning; the focused reference audit used GPT-5.6 Luna with high reasoning. The integrating reviewer retained responsibility for accepting changes. Some earlier agent turns stopped at a usage limit and were resumed; an interrupted review was not counted as completed evidence. + +## Earlier referee report + +| Issue | Current disposition | +|---|---| +| Row normalisation and sparse-cut generalisation | Newer upstream work already documents normalisation, compares the denser taste-motor cut, and supplies scale/normalisation analyses. Retained. Further narrowed the claim: recurrence is not isolated from density, composition and extraction recipe. | +| Biased transmitter-label null | Upstream added a transmitter-outweight-matched null and documented imbalance of plain label permutation. Retained the distinction; no claim that plain label permutation isolates sign alone. | +| Vacuous specification analysis | Upstream fixed gain propagation and increased permutation effort. Retained. Further clarified that adjusted non-rejection does not strengthen C3, whereas C4 is a thresholded rejection predicate. | +| Non-rejection described as reproduction | Removed remaining affirmative equivalence/reproduction interpretations. The legacy within-tolerance identifier is a descriptive point-gap check, not a confidence-bounded equivalence test. | +| Direction-blind ablation and absolute rank correlation | Upstream corrected both. Retained the direction-aware comparator and signed rank correlation. | +| Composition/full correlation and reference | Retained the matched reference. The observed correlation exceeds its reference 95th percentile; joint permutation of both vectors is a mathematical invariance, not proof of no compound information. | +| Uncertainty, interaction residual and drive ranking | Retained upstream corrections. The largest negative estimate is a diagnostic, not a calibrated error bound or general noise floor. | +| Effect floor, nonmonotone ladder and library/substrate composition | Retained explicit reporting conventions, nonmonotonicity caveat, transmitter census and restricted interpretation. No claim that a null effect establishes biological inactivity. | +| Evidence provenance and pharmacological interpretation | Retained the newer evidence-distance rules and native Drosophila binding row. Narrowed invariant claims to tested rows and enumerated paths. | + +## Grok review + +| Comment | Decision | +|---|---| +| M1: false chloride-blocker class | Already corrected upstream. Keep mechanisms distinct; no chemical-class discovery claim. | +| M2: submitted header and missing metadata | Working draft, not submitted. Author identities, affiliations, ORCIDs and corresponding contact require human confirmation. No identity or DOI fabricated. | +| M3: exploratory landscape census | Earlier 20/84 count is obsolete. Retain current counts from generated results, described as exploratory classes based on BH-adjusted component tests. This is neither demonstrated cell-level FDR control nor validated arbitrary-dependence control. | +| M4: type-system inflation and explicit rules | Describe validated tagged records, dispatch/refusal rules and finite regression coverage. Add admissible/rejected examples in main Methods. This is evidence gating and labelling, not a new binding model. | +| M5: vertebrate comparison | Table I identifies a vertebrate receptor panel; no second nervous-system simulation or organism safety claim. | +| M6: incomparable Hz | Distinguish rate-model units, LIF spikes/s and input spike rates in prose, captions and plotted labels. Machine-readable legacy field names need not imply physical calibration. | +| Minor: LB3/LB1 modality | Explicit literature-based working assignment, not MaleCNS ground truth. | +| Minor: expression coverage | Abstract includes coverage and unresolved MN9. Coverage is of cell/receptor mapping, not the fraction of experimentally measured cells. | +| Minor: VOI | Caption and text identify variance under assumed model input ranges, not value of an experiment on a fly. | +| Minor: 3Rs | Removed unsupported index/keyword claim. No demonstrated replacement of animal experiments. | +| Minor: stale commit header | Removed prose commit pin; actual numerical provenance remains in generated records. An archival tagged release remains pending. | +| Minor: references | Corrected Virtual Rat citation, explicitly identified PharmVR as a preprint concept, replaced an unverifiable technical note, and labelled newsletter commentary as informal. See source audit. | +| Minor: abstract ablation result | Included the high composition-only/full-model agreement under shared assumptions. | + +## New findings accepted + +- Extract sensitivity is observed; recurrence is a candidate associated feature, not an isolated cause or validated predictor. +- Matching labels on the two largest scale rungs at 20 permutations do not establish convergence. Their probability estimates sit at the resolution floor. +- Six synthetic networks per setting give imprecise detection rates. The zero-strength control also changes edges and transmitter labels. This grid does not establish general calibration or power for the taste-motor ratio endpoint. +- BH is applied to structural component tests. A cell is classified by an OR over those tests, so component-test adjustment does not automatically control cell-level FDR. Shared random draws do not establish the usual BH dependence condition. +- High permutation effort and compound identity do not make a retrospectively selected profile confirmatory. +- Missing permutation probability cannot yield a within-tolerance verdict. The helper now returns indeterminate rather than accepting the point gap or failing while formatting a missing value. +- The recovery network has 250 nodes and is not size- or density-matched to the named cut; the main text and captions now say so. +- The standard pipeline imports the scale-study cache. Large source cuts and cache source/input hashes are missing, so replay is distinguished from independent recomputation. + +## Remaining limits and author actions + +The manuscript remains a computational-methods working draft. No animal outcome, clinical safety inference, independent circuit validation or formal statistical equivalence has been established. Formal equivalence, controlled recurrence interventions, stronger null calibration and a fully source-pinned scale archive would require additional research, not prose edits. The existing scale-study JSON remains a historical computed record; its legacy labels are explained rather than silently rewriting numerical evidence. + +Before submission, confirm author metadata and AI-use disclosure, create a genuine archival deposit, confirm venue formatting and prepare submission materials. A DOI alone would not constitute preregistration; that requires a frozen protocol before data collection. These are not fabricated to make the checklist appear complete. + +Validation details are recorded in `current_verification.md`. The full numerical rerun was interrupted; the delivered update preserves the committed default-effort estimates and explicitly records editorial replay. It does not claim a completed independent numerical reproduction. diff --git a/papers/reviews/current_round1_statistical.md b/papers/reviews/current_round1_statistical.md new file mode 100644 index 0000000..13fc6b5 --- /dev/null +++ b/papers/reviews/current_round1_statistical.md @@ -0,0 +1,61 @@ +# Current manuscript: Round 1 independent statistical review + +**Snapshot reviewed:** repository `b82b8ed`; manuscript/results provenance `f95fd88`. This is a fresh review of the current source and artifacts. Earlier findings that the revision has fixed are not repeated. + +**Recommendation: major revision.** The planted-effect calibration, relative effect floor, FDR layer, weight-matched transmitter null, and explicit scale study are substantive improvements. The current headline still overreaches in three places: the code's within-margin rule is not a valid equivalence test; the scale study does not identify recurrence as the cause of the verdict; and the synthetic power surface is too small to support the general design rules claimed for it. + +## Blocking + +### B1. `equivalent_within_tolerance` is a point-estimate heuristic, not statistical evidence of equivalence + +**Repository evidence.** `flylab/analysis/dependence.py:697-759` assigns `equivalent_within_tolerance` when two conditions hold: an ordinary difference-style permutation test does not reject, and the point gap `|real_effect - median(null)|` is below δ. No uncertainty interval is formed for the null median or for the gap. With this rule, even a very small shuffle sample can be called equivalent whenever its sample median happens to lie near the real effect. In landscapes, `_decision_p` uses the BH-adjusted probability (`dependence.py:790-804`), so multiplicity adjustment can change a raw rejection into non-rejection and thereby help trigger an “equivalent” label. FDR correction of a difference test cannot supply evidence for equivalence. + +The rule also makes difference and equivalence mutually exclusive, although they answer different questions and can both hold. T6 supplies a live example: fipronil versus degree-preserving rewiring has `p = 0.00699`, gap `0.243 Hz`, and δ `0.332 Hz`; it is within the stated practical margin but the code labels it `distinguishable`. The same occurs for weight permutation (`p = 0.00599`, gap `0.310 < 0.332`). If δ defines practical sameness, a small difference-test p should not automatically overrule it. + +**Implication.** The abstract's “prespecified equivalence margin,” §2.3's “same effect to within the margin,” 281 landscape equivalence verdicts, and the 17/25 C3 equivalence count state more than the procedure establishes. The 5% margin is also a modelling convention justified by another model threshold, not an empirically grounded smallest important difference; calling it “prespecified” should mean fixed before examining these results, not merely encoded before the latest run. + +**Required fix.** Choose the estimand first: equivalence of the real effect to the null ensemble's median, or a distributional statement about null realizations. For median equivalence, require a confidence interval for the null median (bootstrap or order-statistic interval) to lie wholly inside `[real_effect - δ, real_effect + δ]`; an equivalent two-one-sided construction is also acceptable. For a distributional claim, prespecify the required mass inside the margin and estimate it with uncertainty. Report difference evidence and equivalence evidence on separate axes, and correct the equivalence family directly if many equivalence claims are made. Until then rename the current output `within_margin_point_estimate` and do not call it equivalence. + +### B2. The data rule out a simple node-count explanation; they do not establish “recurrence, not size” + +**Repository evidence.** The clean `scale_*` ladder changes node set, edge count, mean degree, transmitter composition, seed share, vehicle rate, and recurrence together. Every clean rung is topology-dependent, so there is no within-recipe verdict variation from which to identify a structural predictor. The only composition-dominated cut is `named`, built by a different extraction recipe. `flylab/analysis/scale.py:602-695` correctly describes the structure comparison as observational and says that a minority class of one cannot distinguish candidate statistics; lines 615-619 explicitly require a controlled cut with the same node set and altered edge structure. `papers/scale_study.json` contains no such intervention. + +The manuscript's key size counterexample is especially non-diagnostic. At `scale_1k`, imidacloprid is called topology-dependent because **weight permutation** rejects (`p = 0.00670`), while degree-preserving rewiring, the mode that destroys who-connects-to-whom while retaining degrees, does not (`p = 0.493`) and is labelled within tolerance. Its necessary level is paradoxically `size_and_composition` because the ladder is non-monotone. Thus `scale_1k` does not show that destroying recurrence changes the answer; it shows sensitivity to the weight-topology pairing. + +**Implication.** The repeated headline “recurrence, not size, predicts which way” in the abstract, §3.2, Discussion, and Conclusion is causal/selective language unsupported by the study. The evidence shows that node count alone is insufficient and that the verdict differs across structurally and procedurally different extracts. It does not select recurrence over density, composition, degree distribution, weight assignment, seed coverage, or extraction recipe. + +**Required fix.** Replace the headline with: “the verdict changes across extracts and does not follow node count monotonically.” Describe recurrence as a candidate associated feature. To retain the stronger claim, run the controlled experiment already specified by `scale.py`: same node set, transmitter labels, degree/weight distributions, readout, and permutation budget, with recurrence selectively broken or varied. A factorial synthetic study varying size and recurrence independently would also support the methodological claim. + +### B3. The top-of-ladder verdict has not statistically “settled” + +**Repository evidence.** `scale_25k` and `scale_50k` each use only 20 permutations. For imidacloprid, both structural probabilities equal the smallest possible value, `1/21 = 0.047619`, and every mode has `stabilised = false` in `papers/scale_study.json`. Nevertheless `verdict_stability` calls the sequence settled whenever the last two categorical labels agree (`scale.py:475-580`). Its `under_powered_cuts` check only detects a resolution *coarser* than alpha; a floor barely below 0.05 passes. Agreement between two thresholded results at the only rejectable value is not evidence that a trend has stabilized. + +**Required fix.** Call the upper rungs exploratory and resolution-limited, and replace “settled” with “the observed class was unchanged across the tested clean rungs.” Do not use the two `n = 20` rungs to establish stability. A settled claim needs a prespecified precision or convergence criterion met on at least two upper rungs, ideally with a common adequate permutation budget or sequential Monte Carlo bounds around the p-values. + +## Major + +### M1. The synthetic experiment is a useful positive control, but the claimed power and false-positive performance are too precise + +**Repository evidence.** T22 uses six independently generated graphs per strength-by-*n* cell. Consequently every detection estimate moves in steps of 0.167. Illustrative 95% Wilson intervals are broad: `4/6 = 0.667` gives `[0.300, 0.903]`, `5/6 = 0.833` gives `[0.436, 0.970]`, and `6/6` gives `[0.610, 1.000]`. The control has 0 detections in 18 runs, whose 95% Wilson upper bound is 0.176; it does not estimate the false-positive rate as zero with useful precision. Reusing the same six graph seeds across permutation counts is a sensible paired design, but those cells are not independent replications. + +The control is also not identical apart from loop strength: `synthetic_cut` adds ten cycle edges and relabels the ring nodes cholinergic only when strength is positive (`dependence.py:1577-1597`). That is a valid planted positive control, but it varies degree, composition, weight allocation, and recurrence together. It tests one motif and one gain mechanism, not general instrument validity. + +**Required fix.** Report binomial counts and confidence intervals in T22/F16. Replace “validated,” “full power,” “effect size dominates,” and “permutations past a few hundred are wasted” with “positive-control calibration” and the narrower observations from this grid. Increase graph replicates substantially before claiming a power surface or false-positive rate. Use a matched control containing the same nodes, labels, degrees, and cycle-edge weights but with the loop selectively broken; then vary graph size and loop strength independently. + +### M2. BH validity under the observed dependence has not been demonstrated + +**Repository evidence.** The supplement asserts that sharing one shuffle stream makes the structural tests positively dependent and therefore licenses Benjamini-Hochberg. A common random stream does not by itself establish positive regression dependence: compound effects have different signs and nonlinear gain rules, so test statistics can be positively or negatively associated. No dependence diagnostic or proof is supplied. The same concern applies to the 100 structural tests in the specification run. + +**Required fix.** Either establish the dependence condition required by BH for these statistics, use Benjamini-Yekutieli for arbitrary dependence, or calibrate family error/FDR empirically with joint null simulations. State the exact family once: landscapes currently adjust two structural modes per cell, while other reported ladder modes remain descriptive. + +### M3. The scale result is re-renderable but not independently recomputable from the repository + +**Repository evidence.** Only `malecns_scale_1k.json` is committed. The 5k-50k paths recorded in `papers/scale_study.json` point to `/root/.flylab/...` and are absent in this checkout. The study JSON has no code SHA, cut hashes, library hash, generation timestamp, or per-cut seed/provenance block. `scripts/reproduce_paper.py:4622-4635, 4717-4728` explicitly reads this cached result rather than rerunning the study. Current manuscript/results provenance is `f95fd88`, while the reviewed source is `b82b8ed`. + +**Implication.** One command can re-render T27 from a committed cache; it cannot regenerate the scale analysis from its source graphs. “Every figure, table and number” and RQ4 need to distinguish artifact replay from analysis reproduction. + +**Required fix.** Add a schema/provenance block to `scale_study.json` with code, library, map, cut hashes, analysis parameters, and seeds; publish the large cuts as durable artifacts or provide a verified workflow that recreates them. Have the paper pipeline validate the cache against those hashes and the checked-out release. Rephrase the current claim as “all ordinary analyses rerun; the scale table is deterministically re-rendered from a committed, provenance-checked study record” until the external cuts are available. + +## Statistical language that can remain + +The following current claims are supported if kept scoped: the verdict differs between `named` and `taste_motor`; node count alone does not order the observed verdicts; the scale ladder stays topology-dependent over the tested rungs; the instrument detects the strong planted loop in this synthetic design and is insensitive to its weak version; and a negative dependence result must be reported with the extract and runtime on which it was measured. diff --git a/papers/reviews/current_round2_statistical.md b/papers/reviews/current_round2_statistical.md new file mode 100644 index 0000000..35283fb --- /dev/null +++ b/papers/reviews/current_round2_statistical.md @@ -0,0 +1,43 @@ +# Current manuscript: round-2 statistical review + +Scope: current `papers/IJRC_FlyLab_draft.md.in` and `papers/SUPPLEMENT.md.in`, checked against the current analysis code and `papers/results.json`. The revised templates now appropriately qualify the median-gap diagnostic, the small synthetic grid, recurrence, and cached scale study. The remaining actionable issues are below. + +## 1. Major: the multiplicity target does not match the reported cell-level claim + +**Evidence.** `_apply_fdr` runs Benjamini–Hochberg over the two structural **mode tests** per cell (`flylab/analysis/dependence.py:2300-2329`), stores the minimum adjusted mode probability as the cell's `q_value` (`:2337-2347`), and then reclassifies the cell. A topology-dependent cell is an OR decision: at least one structural mode rejects. The specification analysis does the same explicitly (`flylab/analysis/robustness.py:1166-1174`). Nevertheless, the generated statements call **11/84** and **54/84 cells** “FDR-controlled” (`papers/results.json:3169-3174`, `:3609-3614`; originating at `flylab/analysis/dependence.py:2550-2554`). + +**Why it matters.** BH control for 168 component hypotheses does not automatically become FDR control for 84 compound-level OR decisions. For example, a true-positive cell may contribute two true rejected modes while a null cell contributes one false mode; the test-level false-discovery proportion can then be smaller than the cell-level proportion. In addition, reused shuffle draws establish dependence, not the positive-dependence condition needed for ordinary BH; the supplement acknowledges this only for the specification run (`papers/SUPPLEMENT.md.in:48`), not for the landscapes. + +**Fix.** Either (a) describe these as “classes based on BH-adjusted component tests” and remove “FDR-controlled” at the cell/count level, explicitly treating BH as a sensitivity analysis under unverified dependence, or (b) form one valid cell-level global-null probability (for example, a within-cell multiplicity-adjusted minimum or joint permutation maximum statistic), then adjust those 84 cell probabilities. Do not present the current count as cell-level FDR-controlled without such a change. + +## 2. Major: “C3 survives correction” reverses the meaning of multiplicity adjustment + +**Evidence.** The manuscript says both topology conclusions “still hold” after BH (`papers/IJRC_FlyLab_draft.md.in:167`), and the supplement says both “survive the correction” (`papers/SUPPLEMENT.md.in:48`). C3 is retained precisely when neither structural test rejects (`flylab/analysis/robustness.py:1027-1045`, `:1330-1331`). Its record is 25/25 both adjusted and unadjusted, with only 17 median-gap passes and 8 indeterminate cases (`papers/results.json`, keys `stab_C3_retained`, `stab_C3_retained_uncorrected`, `stab_C3_equivalent`, `stab_C3_indeterminate`). + +**Why it matters.** Increasing component probabilities through multiplicity adjustment can only make a failure-to-reject predicate easier to retain. Thus “survives correction” is meaningful for positive C4 rejections, but it cannot strengthen C3 or serve as robustness evidence for absence of topology dependence. The following caveat about non-rejection is correct but does not repair that asymmetric logic. + +**Fix.** Report the two directions separately: “C4 remains rejected in 25/25 specifications after adjustment; C3 has no structural rejection in 25/25, unchanged from the nominal analysis, with 17 descriptive median-gap passes and 8 indeterminate.” Avoid “both survive correction.” + +## 3. Major: C4's audit warning describes a statistic the implementation no longer uses + +**Evidence.** The run warning says “C4 is an effect-size contrast with no threshold” (`flylab/analysis/robustness.py:1257-1262`). In fact, C4 is defined as `class == topology-dependent`, requiring at least one structural permutation test to reject (`:1047-1060`), and its final stability value is overwritten from BH-adjusted rejection decisions (`:1311-1331`). The result record correctly calls it a “permutation classification (rejection), FDR-corrected,” so the warning contradicts both code and reported evidence. + +**Why it matters.** This is exactly the sort of self-description mismatch the paper presents its auditability machinery as preventing. It can also mislead readers into treating 25/25 as a threshold-free effect-size comparison. + +**Fix.** Replace the warning and stale C4 doc/name language with the actual predicate: a thresholded OR over the two structural permutation tests after the stated adjustment. Add a test that the emitted warning/readout agrees with `CONCLUSIONS[C4]` and the post-BH decision path. This wording fix need not change the numeric result. + +## 4. Major: retrospective reference profiles remain labelled “confirmatory (prespecified)” in artifacts and API output + +**Evidence.** The supplement correctly says the imidacloprid and fipronil profiles “were refined during iterative development and are not independent preregistered confirmations” (`papers/SUPPLEMENT.md.in:82`). The implementation nevertheless declares those two compounds confirmatory (`flylab/analysis/dependence.py:79-80`, `:365-369`) and automatically sets the flag solely from compound identity and `n >= PAPER_N` (`:1280-1283`), producing `design = "confirmatory (prespecified)"` (`:1465-1467`). Those contradictory labels are committed in `papers/results.json:1660-1672` and `:2065-2077`. + +**Why it matters.** Computational effort and nominal single-profile testing do not make a retrospectively selected profile confirmatory. The metadata can be consumed independently of the narrowed prose and overstate the inferential status. + +**Fix.** Rename these outputs to “high-effort reference profile” (or equivalent), make prospective/confirmatory status an explicit caller-supplied field backed by a dated protocol rather than inferred from compound and sample size, and regenerate the affected record. Nominal probabilities can still be reported as such. + +## 5. Moderate: the soundness-invariant novelty claim overstates the mutation and exhaustiveness evidence + +**Evidence.** The manuscript claims an invariant checked “over every row and every entry point” plus “a mutation check” (`papers/IJRC_FlyLab_draft.md.in:77`); the supplement says mutating the rule table makes the test fail with a row-specific message (`papers/SUPPLEMENT.md.in:164`). The main test does cover all shipped rows through a hand-maintained set of five route groups (`tests/test_pharm_evidence.py:428-499`). However, `test_the_invariant_would_catch_a_violation` neither mutates the rule table nor reruns the pipeline invariant: it makes two direct invalid calls to `check_transformation` and asserts they raise (`:502-516`). A maintained enumeration is also not proof that all present and future entry points are covered. + +**Why it matters.** The invariant is presented as a primary computing contribution. The tests support a finite regression audit of the shipped library and enumerated paths, but not the stronger mutation-testing and whole-pipeline exhaustiveness claims. + +**Fix.** Narrow the prose to “all shipped rows across the enumerated comparison, curve, uncertainty, assay, and dashboard paths, plus deliberate invalid-transformation tests.” Alternatively, add a real mutation test that alters the transformation rule/row, runs the end-to-end audit, and demonstrates the claimed diagnostic; add an architectural enforcement point if exhaustive path coverage is intended. diff --git a/papers/reviews/current_verification.md b/papers/reviews/current_verification.md new file mode 100644 index 0000000..cc9c292 --- /dev/null +++ b/papers/reviews/current_verification.md @@ -0,0 +1,25 @@ +# Verification of the internal-review revision + +Date: 2026-09-21. Base manuscript: main `38cad67`. Reviewed source: `3bd92cc2bb759774884887c7c0fd6741fe260ef9`. This note describes completed checks, not journal peer review or submission readiness. + +## Completed + +- Full default test suite: **1221 passed, 2 skipped, 12 deselected**, with eight dependency/fork deprecation warnings, in 364.84 seconds. The repository excludes slow tests by default. Command: `OPENBLAS_NUM_THREADS=1 OMP_NUM_THREADS=1 MKL_NUM_THREADS=1 python -m pytest -q`. +- Focused paper/replay checks: **21 passed, 1 deselected**. These cover rendering contracts, both cached-scale rendering branches, numerical/provenance preservation and repeatable editorial replay. +- Commit-attribution policy passed in an isolated checkout of this branch. The working repository retains pre-rewrite local recovery refs, so its all-refs scan includes superseded history; the proposed branch does not. +- Manuscript and supplement rendered from their templates. Main body: **6872 words**. All quoted keys resolve. Bibliography: 76 consecutive entries, with no missing citation numbers. No em dashes in the rendered documents. +- `git diff --check` passed. Updated receptor-panel and rate/LIF plots were visually inspected for labels and clipping. +- The architecture, receptor scorecard, curves and genotype plots were regenerated at ordinary effort. Their value payloads exactly match the saved numerical record. The rate/LIF plot was also regenerated before the full attempt was interrupted. +- The stored record is not `--fast`. Its declared effort remains 1000 permutations for the named profiles and named landscape, 300 for the taste-motor landscape, 100 per stability specification, and Sobol base size 1024. These are inherited recorded settings, not a claim that this review independently reran them all. + +## Numerical preservation and reproduction limit + +The full default-effort rerun was interrupted during the dependence step. It reached the reference-profile, transmitter-balance and synthetic-recovery outputs, but did not complete the complete landscape and paper pipeline. No completed full numerical reproduction is claimed. + +The delivered `results.json` preserves the numerical values from the previously committed default-effort record and its original recorded source provenance (`f95fd88a2165e66d5203db6c6fa8c6a020c5a8e0`). That historical source identifier is retained as recorded, not replaced with a new hash to imply a rerun. Comparing all value payloads with main shows changes only to design-status/interpretation fields and document word counts. All table CSV files are byte-identical to main. + +`scripts/paper_review_metadata.py` performs an explicit editorial replay. It corrects captions, notes and retrospective/confirmatory labels while checking that other value payloads and table CSV bytes are unchanged. The same caption/note corrections are applied by future ordinary pipeline runs. `results.json.review_metadata` records the distinction and the reviewed source SHA. A paper-only render does not itself establish analysis reproducibility. + +The scale table remains a replay of `scale_study.json`. The uncommitted large source cuts and missing cache input/source hashes remain limitations. Formal equivalence, cell-level FDR control, endpoint-specific power and an isolated causal recurrence effect have not been established by these edits. + +The submission checklist still requires verified author metadata, author review of sources and AI disclosure, a genuine archival deposit, venue formatting and submission materials. No author identity, DOI, animal result or completed rerun was fabricated. diff --git a/papers/reviews/round4_editorial_decisions.md b/papers/reviews/round4_editorial_decisions.md new file mode 100644 index 0000000..0556f7a --- /dev/null +++ b/papers/reviews/round4_editorial_decisions.md @@ -0,0 +1,30 @@ +# Round 4 editorial decisions + +Scope: readability revision of `papers/IJRC_FlyLab_draft.md.in`. No generated manuscript, numerical record, figure, table, code, reference entry or supplement source was edited. No numerical analysis or document render was run in this pass. + +## Accepted changes + +- Shortened the title to state the transferable method without a subtitle. +- Added the supplied author name, affiliation, email and ORCID to the working source. +- Rewrote the abstract to lead with the extract-conditional inversion, runtime-normalisation sensitivity and two instruments that previously flattered their conclusions. It remains a claim-plus-limit abstract and omits expression coverage, which is retained in the main limitations. +- Reduced the article from four research questions to three. Reproducibility remains §3.7 as engineering support, not a research question, because the commands, cached-scale boundary, backend parity limit and interrupted-verification boundary are necessary for readers. +- Reordered §3.2 so the cross-extract inversion precedes the detailed single-extract profiles. The two-extract landscape remains in the main text because it is the direct evidence for the paper's central methodological result; downstream domain readouts remain in the supplement. +- Reordered §3.3 so the matched reference and runtime-normalisation sensitivity qualify the raw composition/full correlation before it is interpreted. +- Reordered §3.4 so corrected specification results precede the historical failure, while retaining the C3/C4 adjustment asymmetry. +- Removed repeated caveat language where the same limitation is already stated in Methods, Results and §S9. Unique substantive limits were retained. +- Added a five-step transferable audit protocol for other literature-parameterised network domains, explicitly labelled as proposed steps rather than implementations in those domains. + +## Scientific boundaries retained + +- Dependence labels remain conditional on extract, readout, runtime, null family and statistical effort; recurrence is not isolated causally. +- A point-gap check after non-rejection is not statistical equivalence. +- The synthetic grid is limited positive-control evidence, not general calibration or endpoint-specific power. +- Composition/full agreement is reported with its matched reference and normalisation dependence. +- Nicotinic suppression remains gain-rule dependent, and threshold classifications remain unstable. +- Expression mapping coverage remains `{{expr_coverage_overall}}`, with MN9 unresolved. +- There is no independent circuit-level or live-animal validation, no animal dose prediction and no cross-species safety claim. +- The scale table remains a replayed saved study, and the latest full verification attempt was interrupted during dependence analysis. + +## Editorial constraints + +All existing result placeholders and the reference list were preserved. The IJRC Research Article framing remains. No em dash was introduced. The supplement was not changed because no unique paragraph needed relocation after the main-text compression. diff --git a/papers/submission/FlyLab_manuscript.docx b/papers/submission/FlyLab_manuscript.docx new file mode 100644 index 0000000..f19d836 Binary files /dev/null and b/papers/submission/FlyLab_manuscript.docx differ diff --git a/papers/submission/FlyLab_manuscript.json b/papers/submission/FlyLab_manuscript.json new file mode 100644 index 0000000..5c8a861 --- /dev/null +++ b/papers/submission/FlyLab_manuscript.json @@ -0,0 +1,19 @@ +{ + "operation": "Word export only; numerical analyses not rerun", + "inputs_sha256": { + "papers/IJRC_FlyLab_draft.md": "daf31354c7b8bff83c7319aab05d7ec915731aad03e105fdc2d1a44338bea917", + "papers/figures/captions.md": "0e1ffc3302939aee0efd000f040c301464980c63f55ebc0497762659abd4d94b", + "scripts/export_paper_word.py": "06ff06bafd0bc81d6e11adfff71bd95de309bea426a635a143d3046e5733e448", + "papers/figures/F6_dependence_profile.png": "838219d044a94200adbf6b04270c510a4ec3b3e625bb23b262c2fd4e8b942918", + "papers/figures/F13_ablation.png": "f526c9d315fc0be13ef8216ce5ab824c8001ad3486fef8a6454d302cf4ed5560", + "papers/figures/F14_conclusion_stability.png": "f4a8474416dc694c703696343717436ae3b2ad39abcdb23142fbb39b0510c2f2", + "papers/figures/F11_evidence_types.png": "71c27c1084780bd0d502d9584386ba006aaab28a74bfdbfc690d4dbb8b2b7287" + }, + "output_sha256": "4c6d5da13b9acad3254b0af2d1ddf162d7f83094441c416b57a6322dd217e8ed", + "main_figure_archive_ids": [ + "F6", + "F13", + "F14", + "F11" + ] +} diff --git a/papers/submission/README.md b/papers/submission/README.md new file mode 100644 index 0000000..8387cf4 --- /dev/null +++ b/papers/submission/README.md @@ -0,0 +1,26 @@ +# Word manuscript preparation + +`FlyLab_manuscript.docx` is the editable, author-identified preparation copy of the reviewed manuscript. It is not a submitted paper and no submission has been made. + +Author details supplied by the author on 22 September 2026: + +- Michél Nguyen +- University of the People +- Corresponding email: michel_ng@icloud.com +- ORCID: https://orcid.org/0000-0001-6834-4422 + +## Rebuild + +Run `python scripts/export_paper_word.py` with Python packages `python-docx` and the `pandoc` executable available. It reads the rendered Markdown, existing figure images and captions. It does not run a simulation or change numerical results. Render and visually inspect the exported document before delivery. + +The Word manuscript embeds four main result figures, numbered 1 to 4 in reading order. They correspond to archive identifiers F6, F13, F14 and F11. Other figure and table references use supplementary SF and ST identifiers and point to the existing `papers/figures/`, `papers/tables/` and `papers/SUPPLEMENT.md` artifacts. Those supporting files must accompany a later submission; the Word manuscript alone is not the full evidence package. + +## Venue checks before submission + +Checked against https://ijrcom.org/index.php/ijrc/about/submissions on 22 September 2026. A4, 20 mm margins, single spacing and no page numbering are used. The Research Articles section suggests 6000 to 8000 words excluding references and figures; a separate general section gives 5000 to 10000. The concise revision falls below the article-specific typical range and within the general range. Confirm its length with the editor; text was not added merely to reach 6000 words. The automated 5000 to 8000 guard is an editorial check, not certification of venue compliance. + +The journal states double-blind review. Prepare an anonymised manuscript and separate author information as required by the submission portal. The current file intentionally retains the author information requested by the author. + +The reference instructions mix IEEE style with Word endnote and superscript instructions. This preparation copy retains the manuscript's numbered IEEE-style citations and reference list. Confirm the editor's current template and citation requirements before converting references or declaring final format compliance. Postal address and telephone information, where required by the portal, have not been invented. + +An archival release and DOI, complete numerical rerun, final author verification of references and declarations, cover letter, and final supplementary-file packaging remain outstanding. The inherited numerical record and the interrupted-rerun disclosure are preserved; Word export is not numerical reproduction. diff --git a/papers/tables/T12_dependence_landscape.md b/papers/tables/T12_dependence_landscape.md index dc7938c..a8ee9c0 100644 --- a/papers/tables/T12_dependence_landscape.md +++ b/papers/tables/T12_dependence_landscape.md @@ -1,4 +1,4 @@ -**T12_dependence_landscape.** Connectome-dependence landscape on the `named` cut: every compound of the library at 4 concentrations, 1000 shuffles per mode, the same shuffled graphs reused by every cell (a paired design). `class` is decided on the Benjamini-Hochberg adjusted probabilities across the structural tests of the whole landscape and `class_raw` on the uncorrected ones; `class_absolute_floor` drops the prespecified relative effect floor (1 % of the vehicle readout) so the cost of that convention is visible. `necessary_level` is the weakest graph model on the information ladder that this test could **not** distinguish from the real cut -- never a claim that the model reproduces the effect -- and `non_monotone` marks the cells where a richer model on the ladder *was* distinguishable, so the level must be read as 'the cheapest graph model this test cannot tell apart from the real one' and not as 'everything above it is indistinguishable too'. +**T12_dependence_landscape.** Connectome-dependence landscape on the `named` cut: every compound of the library at 4 concentrations, 1000 shuffles per mode, the same shuffled graphs reused by every cell (a paired design). `class` is decided on the Benjamini-Hochberg adjusted probabilities across the structural tests of the whole landscape and `class_raw` on the uncorrected ones; `class_absolute_floor` drops the prespecified relative effect floor (1 % of the vehicle readout) so the cost of that convention is visible. `necessary_level` is the weakest graph model on the information ladder that this test could **not** distinguish from the real cut -- never a claim that the model reproduces the effect -- and `non_monotone` marks the cells where a richer model on the ladder *was* distinguishable, so the level must be read as 'the cheapest graph model this test cannot tell apart from the real one' and not as 'everything above it is indistinguishable too'. BH adjusts structural component tests; cell/conclusion-level FDR control is not established, and the dependence assumptions remain unverified. Adjustment cannot strengthen non-rejection. | compound | conc_M | real_effect | class | class_raw | q_value | necessary_level | non_monotone | |---|---|---|---|---|---|---|---| diff --git a/papers/tables/T12b_dependence_landscape_taste_motor.md b/papers/tables/T12b_dependence_landscape_taste_motor.md index 8a9b24e..a00aee4 100644 --- a/papers/tables/T12b_dependence_landscape_taste_motor.md +++ b/papers/tables/T12b_dependence_landscape_taste_motor.md @@ -1,4 +1,4 @@ -**T12b_dependence_landscape_taste_motor.** The same landscape repeated on the denser `taste_motor` cut (1841 nodes, 19066 edges, mean degree 10.4) at 300 shuffles per mode. The `named` cut is an in-star (T19) and a degree-preserving rewire of it is close to the identity; this table is what the same instrument says about a cut that has wiring to destroy. The two do not agree, and that disagreement is the paper's main RQ2 result. +**T12b_dependence_landscape_taste_motor.** The same landscape repeated on the denser `taste_motor` cut (1841 nodes, 19066 edges, mean degree 10.4) at 300 shuffles per mode. The `named` cut is an in-star (T19) and a degree-preserving rewire of it is close to the identity; this table is what the same instrument says about a cut that has wiring to destroy. The two do not agree, and that disagreement is the paper's main RQ2 result. BH adjusts structural component tests; cell/conclusion-level FDR control is not established, and the dependence assumptions remain unverified. Adjustment cannot strengthen non-rejection. | compound | conc_M | real_effect | class | class_raw | q_value | necessary_level | non_monotone | |---|---|---|---|---|---|---|---| diff --git a/papers/tables/T13_ablation.md b/papers/tables/T13_ablation.md index edcf888..c7b7d0e 100644 --- a/papers/tables/T13_ablation.md +++ b/papers/tables/T13_ablation.md @@ -1,4 +1,4 @@ -**T13_ablation.** Ablation ladder. Each level is a prediction for the same compound-concentration cell from a model that has been denied one layer of information: receptor engagement only (dimensionless), mechanism gains on the graph's transmitter composition only (excitation index), the real cut with one generic multiplier instead of mechanism-specific gains (Hz), and the full model (Hz). The generic multiplier is direction-aware: generic in magnitude, with one bit of sign taken from the mechanism table, which is the minimum a connectome-without-pharmacology model needs to order a library containing disinhibitors. `C_topology_only_floor` is the historical depression-only rule, kept as a floor: every one of its entries is at most zero, so it cannot express disinhibition and is not a competitive baseline. Levels have different units, so they are compared by ordering, never by value. +**T13_ablation.** Ablation ladder. Each level is a prediction for the same compound-concentration cell from a model that has been denied one layer of information: receptor engagement only (dimensionless), mechanism gains on the graph's transmitter composition only (excitation index), the real cut with one generic multiplier instead of mechanism-specific gains (Hz), and the full model (Hz). The generic multiplier is direction-aware: generic in magnitude, with one bit of sign taken from the mechanism table, which is the minimum a connectome-without-pharmacology model needs to order a library containing disinhibitors. `C_topology_only_floor` is the historical depression-only rule, kept as a floor: every one of its entries is at most zero, so it cannot express disinhibition and is not a competitive baseline. Levels have different units, so they are compared by ordering, never by value. Rate-engine quantities labelled Hz in legacy fields or axes are rate-model units, not calibrated physiological firing rates. | compound | class | conc_M | A_receptor_only | B_composition_only | C_topology_only | D_full_flylab | |---|---|---|---|---|---|---| diff --git a/papers/tables/T13b_ablation_information.md b/papers/tables/T13b_ablation_information.md index cbe502e..962df33 100644 --- a/papers/tables/T13b_ablation_information.md +++ b/papers/tables/T13b_ablation_information.md @@ -1,4 +1,4 @@ -**T13b_ablation_information.** How much of the full model's ordering of the library each ablated level recovers, per concentration. `reproduces_full_ordering` is applied to the **signed** rank correlation: an ordering that is a perfect inversion of the full model's reproduces nothing, and the previous absolute-value test would have credited it. +**T13b_ablation_information.** How much of the full model's ordering of the library each ablated level recovers, per concentration. `reproduces_full_ordering` is applied to the **signed** rank correlation: an ordering that is a perfect inversion of the full model's reproduces nothing, and the previous absolute-value test would have credited it. Rate-engine quantities labelled Hz in legacy fields or axes are rate-model units, not calibrated physiological firing rates. | conc_M | level | unit | spearman_rho_vs_full | pearson_r_vs_full | residual_rms_standardised | residual_rms_hz | reproduces_full_ordering | information_added_vs_previous | |---|---|---|---|---|---|---|---|---| diff --git a/papers/tables/T14_conclusion_stability.md b/papers/tables/T14_conclusion_stability.md index 83ba0ab..e3ad740 100644 --- a/papers/tables/T14_conclusion_stability.md +++ b/papers/tables/T14_conclusion_stability.md @@ -1,4 +1,4 @@ -**T14_conclusion_stability.** Conclusion stability matrix: every principal conclusion re-derived under each of 25 prespecified engagement-to-gain specifications (biphasic, monotone-linear and saturating shapes at several coefficient scales). `n_undecidable` counts specifications under which the readout does not exist (for example a ratio whose denominator is silenced); the headline fraction counts those against the conclusion. The two topology rows are decided on 100 permutations per specification (resolution 0.0099) and on Benjamini-Hochberg adjusted probabilities across the 100 structural tests of the run, with the uncorrected count beside them. Until v0.6.1 the specification context rebound the gain function on the assay modules but not on the engine the permutation path resolves, so every specification fed the *default* gains to its nulls and the matrix's topology rows were empty; they are now computed per specification. For a conclusion shaped as a failure to reject, `n_equivalent_within_tolerance` is the part of the retention that is evidence of equivalence and `n_indeterminate` the part that is only a non-rejection. +**T14_conclusion_stability.** Conclusion stability matrix: every principal conclusion re-derived under each of 25 fixed engagement-to-gain specifications (biphasic, monotone-linear and saturating shapes at several coefficient scales). `n_undecidable` counts specifications under which the readout does not exist (for example a ratio whose denominator is silenced); the headline fraction counts those against the conclusion. The two topology rows are decided on 100 permutations per specification (resolution 0.0099) and on Benjamini-Hochberg adjusted probabilities across the 100 structural tests of the run, with the uncorrected count beside them. Until v0.6.1 the specification context rebound the gain function on the assay modules but not on the engine the permutation path resolves, so every specification fed the *default* gains to its nulls and the matrix's topology rows were empty; they are now computed per specification. For a conclusion shaped as a failure to reject, `n_equivalent_within_tolerance` is the part passing a descriptive median-gap check, not statistical equivalence and `n_indeterminate` the part that is only a non-rejection. The legacy within-tolerance label is a point-gap diagnostic, not a formal equivalence test. BH adjusts structural component tests; cell/conclusion-level FDR control is not established, and the dependence assumptions remain unverified. Adjustment cannot strengthen non-rejection. | conclusion | n_specs | n_retained | n_lost | n_undecidable | n_retained_uncorrected | n_equivalent_within_tolerance | n_indeterminate | |---|---|---|---|---|---|---|---| diff --git a/papers/tables/T16_uncertainty_budget.md b/papers/tables/T16_uncertainty_budget.md index c12812a..aec598d 100644 --- a/papers/tables/T16_uncertainty_budget.md +++ b/papers/tables/T16_uncertainty_budget.md @@ -1,4 +1,4 @@ -**T16_uncertainty_budget.** Global uncertainty budget for the neighbourhood mean rate (imidacloprid at 1 µM, 11264 model evaluations, Jansen estimators on a Saltelli cross-sample). S1 is the variance share resolving that factor alone would remove; ST includes its interactions. Estimates are **unclipped**, so a factor whose true index is zero can come out negative; the largest such magnitude is the estimator's measured noise floor and is reported with the factor that set it. A factor counts as resolved only when its bootstrap interval excludes zero, which is a stronger test than exceeding that floor. The shares do not sum to one; the remainder is the interaction row, and it is reported both on the raw estimates and with the negative ones clipped, because leaving them in counts estimator noise as interaction. A null factor with no effect on the deterministic rate engine is included on purpose as a control. +**T16_uncertainty_budget.** Global uncertainty budget for the neighbourhood mean rate (imidacloprid at 1 µM, 11264 model evaluations, Jansen estimators on a Saltelli cross-sample). S1 is the variance share resolving that factor alone would remove; ST includes its interactions. Estimates are **unclipped**, so a factor whose true index is zero can come out negative; the largest such magnitude is the estimator's measured negative-estimate diagnostic and is reported with the factor that set it. A factor counts as resolved only when its bootstrap interval excludes zero, which is a stronger test than exceeding that floor. The shares do not sum to one; the remainder is the interaction row, and it is reported both on the raw estimates and with the negative ones clipped, because leaving them in counts estimator noise as interaction. A null factor with no effect on the deterministic rate engine is included on purpose as a control. The largest negative estimate is an observed error diagnostic, not a calibrated uncertainty bound. VOI concerns output variance under assumed model input ranges, not the benefit of an experiment on a fly. Rate-engine quantities labelled Hz in legacy fields or axes are rate-model units, not calibrated physiological firing rates. | source | kind | share_of_variance_S1 | share_with_interactions_ST | variance_removed_hz2 | ci_low | ci_high | |---|---|---|---|---|---|---| diff --git a/papers/tables/T17_value_of_information.md b/papers/tables/T17_value_of_information.md index 6ac91f5..8677a02 100644 --- a/papers/tables/T17_value_of_information.md +++ b/papers/tables/T17_value_of_information.md @@ -1,4 +1,4 @@ -**T17_value_of_information.** Value of information. VOI_j = S_j x Var(Y) is the model variance that would disappear if assumption j were resolved exactly while everything else stayed as uncertain as it is; the upper bound uses the total-order index and is what resolving j *last* would buy. A value of information cannot be negative, so a negative first-order estimate is clipped to zero for the decision value and kept unclipped in `voi_fraction_raw`. `state` says whether the sample could resolve the factor at all: an unresolved factor carries no ranking claim, which is a statement about the sample size rather than about the factor. This is variance of a model output under assumed input ranges, not an expected gain in accuracy about a living fly. +**T17_value_of_information.** Value of information. VOI_j = S_j x Var(Y) is the model variance that would disappear if assumption j were resolved exactly while everything else stayed as uncertain as it is; the upper bound uses the total-order index and is what resolving j *last* would buy. A value of information cannot be negative, so a negative first-order estimate is clipped to zero for the decision value and kept unclipped in `voi_fraction_raw`. `state` says whether the sample could resolve the factor at all: an unresolved factor carries no ranking claim, which is a statement about the sample size rather than about the factor. This is variance of a model output under assumed input ranges, not an expected gain in accuracy about a living fly. Rate-engine quantities labelled Hz in legacy fields or axes are rate-model units, not calibrated physiological firing rates. | rank | factor | state | voi_fraction_of_var | voi_var_hz2 | experiment | cost | |---|---|---|---|---|---|---| diff --git a/papers/tables/T20_transformation_rule.md b/papers/tables/T20_transformation_rule.md index 2dc6235..f20b68e 100644 --- a/papers/tables/T20_transformation_rule.md +++ b/papers/tables/T20_transformation_rule.md @@ -1,4 +1,4 @@ -**T20_transformation_rule.** The transformation rule of the evidence type system, materialised. The permitted transformation is a function of two facts: the parameter the source measured and the ordered distance E0-E4 between that source and this compound at this receptor in this species. A binding constant yields physical fractional occupancy only at E0; transferred, it becomes a labelled proxy carrying a warning that names the gap. Anything at E4 is `not_modelled` and returns N/A. `tests/test_pharm_evidence.py` checks the soundness invariant exhaustively over every library row and every entry point: no number reaching a circuit readout originates from a row whose (param_type, distance) pair does not admit one. +**T20_transformation_rule.** The validated evidence dispatch rule, materialised. The permitted transformation is a function of two facts: the parameter the source measured and the ordered distance E0-E4 between that source and this compound at this receptor in this species. A binding constant yields physical fractional occupancy only at E0; transferred, it becomes a labelled proxy carrying a warning that names the gap. Anything at E4 is `not_modelled` and returns N/A. The finite tests in `tests/test_pharm_evidence.py` enumerate every row in the shipped library and each listed engagement-to-circuit path: no number reaching a circuit readout originates from a row whose (param_type, distance) pair does not admit one. This is an audit of that finite implementation surface, not a proof over arbitrary future rows or paths. | param_type | evidence_distance_label | engagement_model | model_strength | |---|---|---|---| diff --git a/papers/tables/T22_instrument_validation.md b/papers/tables/T22_instrument_validation.md index bcd66e4..6b26805 100644 --- a/papers/tables/T22_instrument_validation.md +++ b/papers/tables/T22_instrument_validation.md @@ -1,4 +1,4 @@ -**T22_instrument_validation.** Ground-truth recovery and power for the dependence ladder. A recurrent cholinergic cycle of known strength is planted in a synthetic cut of the same size, density and transmitter composition as the `named` cut, and a gain patch that collapses `g_ach` removes an amplification that exists only while the cycle is intact -- so the drug *contrast*, not merely the rate, depends on the wiring. `loop_strength = 0` plants nothing and is the negative control, whose detection rate is the empirical false-positive rate. Nothing here touches the connectome or the compound library: the experiment is about the instrument. +**T22_instrument_validation.** Synthetic recovery checks with an imposed cholinergic cycle and a gain patch. The synthetic network is not size- or density-matched to the named connectome cut. The zero-strength case omits cycle edges and associated transmitter relabelling. Detection frequencies in this small grid have broad uncertainty; networks are reused across permutation counts. These checks do not establish general calibration or power for the real taste-motor ratio endpoint. Counts and settings are shown in the artifact. Rate-engine quantities labelled Hz in legacy fields or axes are rate-model units, not calibrated physiological firing rates. | experiment | loop_strength | n | replicates | detection_rate | class | |---|---|---|---|---|---| diff --git a/papers/tables/T27_scale_study.md b/papers/tables/T27_scale_study.md index e48876e..b49806e 100644 --- a/papers/tables/T27_scale_study.md +++ b/papers/tables/T27_scale_study.md @@ -1,4 +1,4 @@ -**T27_scale_study.** The same dependence profile across the cuts the scaling study could afford, smallest first: the cut's structure, the permutation count it was run at (which falls with the edge count, so the resolution coarsens up the ladder), the per-mode empirical probabilities, the necessary information level and the class. `named` and `taste_motor` are built by a different recipe from the `scale_*` rungs, so the clean scale axis is the `scale_*` rows alone. Stability inside the range tested is not stability. The largest rung here is far short of the whole CNS, and the permutation count falls with scale, so 'settled' means 'did not move over the cuts this study could afford'. +**T27_scale_study.** Cached dependence-profile records across the cuts the separate scaling study could afford, smallest first: the cut's structure, the permutation count it was run at (which falls with the edge count, so the resolution coarsens up the ladder), the per-mode empirical probabilities, the necessary information level and the class. `named` and `taste_motor` are built by a different recipe from the `scale_*` rungs, so the clean scale axis is the `scale_*` rows alone. This reproduction step reads and summarises those cached rows; it does not recompute the profiles. The legacy `settled` field records matching labels on the largest observed cuts, not statistical stability or convergence. This table replays a saved study; large-cut profiles are not recomputed. Matching labels at the largest rungs do not establish convergence or isolate recurrence causally. | cut | n_nodes | n_edges | mean_degree | compound | n | p_degree | p_weight | p_sign_wm | necessary_information_level | class | |---|---|---|---|---|---|---|---|---|---|---| diff --git a/papers/tables/T6_dependence_profile.md b/papers/tables/T6_dependence_profile.md index 17cdc61..db4ce88 100644 --- a/papers/tables/T6_dependence_profile.md +++ b/papers/tables/T6_dependence_profile.md @@ -1,4 +1,4 @@ -**T6_dependence_profile.** Connectome-dependence profile. `real_effect` is treated minus vehicle on the real MaleCNS cut; the null is the same contrast on `n` degraded copies of that cut, with the seed block and node order held fixed. `p_two_sided` is the empirical permutation probability (k+1)/(n+1) and is the statistic to read first; `p_resolution` is its floor, and `z` is a standardised distance from a usually non-normal null. `verdict` is the only claim the design supports: `distinguishable` when the test rejects, `equivalent_within_tolerance` when the gap from the null median is below the prespecified margin `delta`, and `indeterminate` otherwise -- a failure to reject is never evidence of equivalence. `sign_permute` is a **joint** target-set-and-sign null and is off the ladder (`on_ladder = False`); the rank-3 rung is the weight-matched transmitter null, which holds each transmitter's share of total outgoing weight fixed (T23). +**T6_dependence_profile.** Connectome-dependence profile. `real_effect` is treated minus vehicle on the real MaleCNS cut; the null is the same contrast on `n` degraded copies of that cut, with the seed block and node order held fixed. `p_two_sided` is the empirical permutation probability (k+1)/(n+1) and is the statistic to read first; `p_resolution` is its floor, and `z` is a standardised distance from a usually non-normal null. `verdict` records `distinguishable` when the test rejects. The legacy API label `equivalent_within_tolerance` means only that the point estimate of the absolute real-effect to null-median gap is below the configured margin `delta`; it is a descriptive point-gap diagnostic, not a formal equivalence test. `indeterminate` means neither condition was met. No interval for the gap, TOST, distributional closeness criterion, or error-controlled equivalence decision is computed. `sign_permute` is a **joint** target-set-and-sign null and is off the ladder (`on_ladder = False`); the rank-3 rung is the weight-matched transmitter null, which holds each transmitter's share of total outgoing weight fixed (T23). The legacy within-tolerance label is a point-gap diagnostic, not a formal equivalence test. Rate-engine quantities labelled Hz in legacy fields or axes are rate-model units, not calibrated physiological firing rates. | compound | assay | mode | on_ladder | real_effect | p_two_sided | verdict | abs_gap_from_null_median | delta | |---|---|---|---|---|---|---|---|---| diff --git a/scripts/export_paper_word.py b/scripts/export_paper_word.py new file mode 100644 index 0000000..f14c23b --- /dev/null +++ b/scripts/export_paper_word.py @@ -0,0 +1,190 @@ +#!/usr/bin/env python3 +"""Export the reviewed Markdown manuscript to editable Word, without rerunning science. + +Requires pandoc and python-docx. Run with the document-authoring Python runtime. +The four embedded result figures are numbered in reading order; other F/T IDs +remain supplementary artifact identifiers. Numerical source files are read only. +""" +from __future__ import annotations + +import re +import hashlib +import json +import subprocess +import tempfile +from pathlib import Path + +from docx import Document +from docx.enum.table import WD_CELL_VERTICAL_ALIGNMENT +from docx.oxml import OxmlElement +from docx.oxml.ns import qn +from docx.shared import Mm, Pt, RGBColor + +ROOT = Path(__file__).resolve().parents[1] +PAPERS = ROOT / "papers" +OUTPUT = PAPERS / "submission" / "FlyLab_manuscript.docx" +# Section -> archived figure ID. These are deliberately a small main-text set. +FIGURES = {"3.2": "F6", "3.3": "F13", "3.4": "F14", "3.6": "F11"} + + +def manuscript_markdown() -> str: + source = (PAPERS / "IJRC_FlyLab_draft.md").read_text() + source = source.split("### Figures and tables")[0].rstrip().removesuffix("---").rstrip() + # Submission status and build explanations belong in the repository, not title matter. + source = re.sub(r"^\*\*Computational-methods Research Article\..*?\n", "", source, flags=re.M) + source = re.sub(r"^> Simulation-result numeric spans.*?\n", "> No number in this paper was measured in a living animal.\n", source, flags=re.M) + source = source.replace("https://github.com/pinkysworld/FlyLab\nSoftware version", "https://github.com/pinkysworld/FlyLab. Software version") + captions = dict(re.findall(r"^\*\*(F\d+_[^.]+)\.\*\* (.+)$", (PAPERS / "figures/captions.md").read_text(), re.M)) + sections = re.split(r"(?=^### 3\.)", source, flags=re.M) + main_ids = {identifier: str(i + 1) for i, identifier in enumerate(FIGURES.values())} + for i, section in enumerate(sections): + match = re.match(r"### (3\.\d+)\b", section) + if not match or match[1] not in FIGURES: + continue + identifier = FIGURES[match[1]] + name = next(key for key in captions if key.split("_")[0] == identifier) + figure = f"\n\n![{name}]({PAPERS / 'figures' / (name + '.png')})\n\n**Figure {main_ids[identifier]}.** {captions[name]}\n\n" + # Last result section shares a chunk with Discussion; insert before it. + boundary = re.search(r"^## ", section, re.M) + if boundary: + sections[i] = section[:boundary.start()].rstrip() + figure + section[boundary.start():] + else: + sections[i] = section.rstrip() + figure + source = "".join(sections) + # Only references in prose, never paths or alt text, get remapped. + lines = [] + for line in source.splitlines(): + if not line.startswith("!["): + line = re.sub(r"\bF(\d+)\b", lambda m: main_ids.get(m[0], "S" + m[0]), line) + line = re.sub(r"\bT(\d+[a-z]?)\b", r"ST\1", line) + lines.append(line) + source = "\n".join(lines) + source = source.replace("| Graph model |", "**Table II. Graph degradations and retained information.**\n\n| Graph model |", 1) + # DOI strings become usable links without changing citation text or numbering. + source = re.sub(r"DOI (10\.[^\s]+?)(?=\.?(?:\s|$))", lambda m: "DOI [" + m[1].rstrip(".") + "](https://doi.org/" + m[1].rstrip(".") + ")", source) + note = ("Supplementary figures and tables retain the archive identifiers SF and ST. " + "Their captions, data and the supplementary methods are available with the " + "[supporting artifacts](https://github.com/pinkysworld/FlyLab/tree/codex/flylab-reviewed-manuscript/papers). " + "Figures 1–4 correspond to archive figures F6, F13, F14 and F11, respectively.\n\n") + source = source.replace("## References", note + "## References", 1) + return source + + +def style_document(doc: Document) -> None: + styles = {style.name: style for style in doc.styles} + for section in doc.sections: + section.page_width, section.page_height = Mm(210), Mm(297) + section.top_margin = section.bottom_margin = Mm(20) + section.left_margin = section.right_margin = Mm(20) + for style in doc.styles: + if style.type in (1, 2): + style.font.name = "Times New Roman" + style.font.color.rgb = RGBColor(0, 0, 0) + normal = styles["Normal"] + normal.font.size = Pt(11) + normal.paragraph_format.line_spacing = 1 + normal.paragraph_format.space_after = Pt(6) + for name, size in (("Title", 18), ("Heading 1", 13), ("Heading 2", 12), ("Heading 3", 11)): + style = styles[name] + style.font.size = Pt(size) + style.font.bold = True + style.paragraph_format.space_before = Pt(12) + style.paragraph_format.space_after = Pt(6) + style.paragraph_format.keep_with_next = True + for name in ("Source Code", "Verbatim Char"): + if name in doc.styles: + doc.styles[name].font.name = "DejaVu Sans Mono" + doc.styles[name].font.size = Pt(9) + paragraphs = doc.paragraphs + for i, paragraph in enumerate(paragraphs): + if i == 0: + paragraph.style = doc.styles["Title"] + paragraph.paragraph_format.widow_control = True + if paragraph._p.xpath(".//w:drawing"): + paragraph.paragraph_format.keep_with_next = True + paragraph.alignment = 1 + if re.match(r"Figure \d+\.", paragraph.text): + paragraph.paragraph_format.space_after = Pt(10) + for run in paragraph.runs: + run.font.size = Pt(9.5) + if re.match(r"Table [IVX]+\.", paragraph.text): + paragraph.paragraph_format.keep_with_next = True + if re.match(r"\[\d+\]", paragraph.text): + paragraph.paragraph_format.space_after = Pt(4) + for run in paragraph.runs: + run.font.size = Pt(10) + for shape in doc.inline_shapes: + ratio = min(Mm(170) / shape.width, Mm(100) / shape.height) + shape.width = int(shape.width * ratio) + shape.height = int(shape.height * ratio) + for table in doc.tables: + table.autofit = False + n = len(table.columns) + widths = [Mm(78), Mm(92)] if n == 2 else [Mm(50)] + [Mm(120 / (n - 1))] * (n - 1) + for column, width in zip(table.columns, widths): + column.width = width + for r, row in enumerate(table.rows): + if r == 0: + marker = OxmlElement("w:tblHeader") + row._tr.get_or_add_trPr().append(marker) + for c, cell in enumerate(row.cells): + cell.width = widths[c] + cell.vertical_alignment = WD_CELL_VERTICAL_ALIGNMENT.CENTER + props = cell._tc.get_or_add_tcPr() + borders = OxmlElement("w:tcBorders") + for edge in ("top", "left", "bottom", "right"): + el = OxmlElement("w:" + edge) + for key, value in (("val", "single"), ("sz", "4"), ("color", "D9D9D9")): + el.set(qn("w:" + key), value) + borders.append(el) + props.append(borders) + margins = OxmlElement("w:tcMar") + for edge in ("top", "left", "bottom", "right"): + el = OxmlElement("w:" + edge) + el.set(qn("w:w"), "80") + el.set(qn("w:type"), "dxa") + margins.append(el) + props.append(margins) + if r == 0: + shade = OxmlElement("w:shd") + shade.set(qn("w:fill"), "E7E6E6") + props.append(shade) + for paragraph in cell.paragraphs: + paragraph.paragraph_format.space_after = Pt(3) + paragraph.paragraph_format.space_before = Pt(3) + paragraph.paragraph_format.keep_with_next = False + for run in paragraph.runs: + run.font.size = Pt(9 if n > 3 else 10) + if r == 0: + run.font.bold = True + doc.core_properties.author = "Michél Nguyen" + doc.core_properties.last_modified_by = "Michél Nguyen" + doc.core_properties.title = paragraphs[0].text + doc.core_properties.comments = "" + + +def main() -> None: + OUTPUT.parent.mkdir(parents=True, exist_ok=True) + with tempfile.TemporaryDirectory(prefix="flylab-word-") as temporary: + md = Path(temporary) / "manuscript.md" + md.write_text(manuscript_markdown()) + subprocess.run(["pandoc", str(md), "--from=markdown-implicit_figures+tex_math_dollars", "--to=docx", "--output", str(OUTPUT)], check=True) + doc = Document(OUTPUT) + style_document(doc) + doc.save(OUTPUT) + assert len(doc.inline_shapes) == 4 + assert not any("[[MISSING:" in p.text for p in doc.paragraphs) + inputs = [PAPERS / "IJRC_FlyLab_draft.md", PAPERS / "figures/captions.md", Path(__file__)] + inputs.extend(next((PAPERS / "figures").glob(identifier + "_*.png")) for identifier in FIGURES.values()) + record = { + "operation": "Word export only; numerical analyses not rerun", + "inputs_sha256": {str(path.relative_to(ROOT)): hashlib.sha256(path.read_bytes()).hexdigest() for path in inputs}, + "output_sha256": hashlib.sha256(OUTPUT.read_bytes()).hexdigest(), + "main_figure_archive_ids": list(FIGURES.values()), + } + OUTPUT.with_suffix(".json").write_text(json.dumps(record, indent=2) + "\n") + print(OUTPUT) + + +if __name__ == "__main__": + main() diff --git a/scripts/manifest.py b/scripts/manifest.py index 11f6d19..41c55a3 100644 --- a/scripts/manifest.py +++ b/scripts/manifest.py @@ -92,8 +92,10 @@ "patterns": ( "papers/results.json", "papers/*.md", + "papers/*.md.in", "papers/figures/*", "papers/tables/*", + "papers/submission/*", ), "description": "the committed paper artifacts regenerated by scripts/reproduce_paper.py", }, diff --git a/scripts/paper_review_metadata.py b/scripts/paper_review_metadata.py new file mode 100644 index 0000000..54c444b --- /dev/null +++ b/scripts/paper_review_metadata.py @@ -0,0 +1,160 @@ +"""Refresh review wording on saved paper artifacts without rerunning analyses. + +The numerical-source provenance is retained. This is explicitly a metadata replay, +not numerical reproduction. All non-editorial value payloads are checked unchanged. +""" +from __future__ import annotations + +import copy +import json +from pathlib import Path +import re +import subprocess +import sys + +ROOT = Path(__file__).resolve().parents[1] +if str(ROOT) not in sys.path: + sys.path.insert(0, str(ROOT)) + + +def review_caption(name: str, caption: str) -> str: + """Correct interpretation of legacy captions without replacing data values.""" + text = caption.replace('—', ';') + text = text.replace('prespecified equivalence margin', 'configured median-gap tolerance') + text = text.replace('prespecified, admissible', 'fixed, admissible') + text = text.replace('prespecified engagement-to-gain', 'fixed engagement-to-gain') + text = text.replace('the part of the retention that is evidence of equivalence', + 'the part passing a descriptive median-gap check, not statistical equivalence') + text = text.replace('only as strong as the equivalence behind it', + 'a non-rejection, not evidence of equivalence') + text = text.replace('checked exhaustively over every row and every entry point', + 'audited over shipped rows and enumerated pipeline paths') + text = text.replace('The type system, not a convention, enforces this:', + 'Runtime validation enforces this:') + text = text.replace('the transformation the type system then permits', + 'the transformation the evidence dispatch rule permits') + if name.startswith(('F1_', 'F2_', 'F3_', 'F7_', 'F8_', 'F9_')): + text = text.replace('occupancy', 'engagement').replace('Occupancy', 'Engagement') + if name.startswith('F2_'): + text = text.replace('Hatched bars rest on a class placeholder: the compound has no sourced EC50 at that pair and the row is inert.', + 'Hatched entries indicate unsupported parameters and are interpreted as N/A.') + if name.startswith('F12_'): + text = text.replace('classified by the weakest graph model the permutation test could **not** distinguish from the real cut -- never by a model that reproduces the effect.', + 'classified from structural component-test decisions and the configured effect floor.') + text = re.sub(r'(\d+ of \d+) cells need the wiring pattern', r'\1 cells are classified topology-dependent', text) + text = text.replace('do not move the readout at all', 'fall below the configured effect floor') + if name.startswith(('F16_', 'T22_')): + text = ('Synthetic recovery checks with an imposed cholinergic cycle and a gain patch. ' + 'The synthetic network is not size- or density-matched to the named connectome cut. ' + 'The zero-strength case omits cycle edges and associated transmitter relabelling. ' + 'Detection frequencies in this small grid have broad uncertainty; networks are reused ' + 'across permutation counts. These checks do not establish general calibration or ' + 'power for the real taste-motor ratio endpoint. Counts and settings are shown in the artifact.') + if name.startswith(('F6_', 'T6_', 'F14_', 'T14_')): + note = 'The legacy within-tolerance label is a point-gap diagnostic, not a formal equivalence test.' + if note not in text: + text += ' ' + note + if name.startswith(('F12_', 'T12', 'F14_', 'T14_')): + note = ('BH adjusts structural component tests; cell/conclusion-level FDR control is not established, ' + 'and the dependence assumptions remain unverified. Adjustment cannot strengthen non-rejection.') + if note not in text: + text += ' ' + note + if name.startswith(('F15_', 'T16_')): + text = text.replace('noise floor', 'negative-estimate diagnostic') + note = ('The largest negative estimate is an observed error diagnostic, not a calibrated uncertainty bound. ' + 'VOI concerns output variance under assumed model input ranges, not the benefit of an experiment on a fly.') + if note not in text: + text += ' ' + note + if name.startswith('T27_'): + note = ('This table replays a saved study; large-cut profiles are not recomputed. Matching labels at ' + 'the largest rungs do not establish convergence or isolate recurrence causally.') + if note not in text: + text += ' ' + note + if name.startswith('F5_'): + text = text.replace('at 150 Hz', 'at a nominal drive of 150 rate-model input units or LIF input spikes/s') + note = 'Left-axis values are rate-model units; right-axis values are simulated LIF spikes/s, not measured fly firing rates.' + elif name.startswith(('F6_', 'F8_', 'F13_', 'F15_', 'F16_', 'T6_', 'T13', 'T16_', 'T17_', 'T22_')): + note = 'Rate-engine quantities labelled Hz in legacy fields or axes are rate-model units, not calibrated physiological firing rates.' + else: + note = '' + if note and note not in text: + text += ' ' + note + return text + + +def review_note(note: str) -> str: + text = note.replace(' Hz', ' rate-model units') + text = text.replace('is the substrate to generalise from', 'is a denser comparison, not evidence of whole-network generality') + text = text.replace('matters to the equivalence claim and not to the probability', 'changes both the descriptive median-gap check and the ensemble probabilities') + text = text.replace('changes the descriptive point-gap diagnostic, not the probability', 'changes both the descriptive point-gap diagnostic and the ensemble probabilities') + text = text.replace('Effect size dominates the permutation count: full power at planted strength', 'All sampled networks were detected at planted strength') + text = text.replace('carries no compound-level information', 'must be interpreted against shared assumptions and its reference distribution') + text = text.replace('reach equivalence within the prespecified margin', 'pass the descriptive median-gap tolerance') + text = text.replace('is no longer vacuous and is no longer uniform', 'uses an attainable rejection threshold but is a non-rejection criterion') + text = text.replace("the estimator's noise floor", 'the largest observed negative-estimate magnitude') + text = text.replace('verdicts settled on the clean ladder', 'labels match on the largest tested clean-ladder rungs') + qualifiers = { + '[dependence]': 'These diagnostics do not establish statistical equivalence, cell-level FDR control or endpoint-specific power.', + '[ablation]': 'Jointly permuting paired scores preserves correlation by construction; the observed correlation exceeds the matched reference 95th percentile.', + '[stability]': 'Adjusted non-rejection is not strengthened evidence for absence of topology dependence.', + '[scale]': 'This is a cached descriptive comparison, not a new scale-study rerun, convergence proof or isolated recurrence effect.', + } + for prefix, qualification in qualifiers.items(): + if text.startswith(prefix) and qualification not in text: + text += ' ' + qualification + return text + + +def refresh(outdir: Path = ROOT / 'papers') -> dict: + path = outdir / 'results.json' + data = json.loads(path.read_text()) + before = copy.deepcopy(data['values']) + allowed = set() + for key, row in data['values'].items(): + if key.startswith('dep_') and key.endswith('_confirmatory'): + row['value'] = False + row['text'] = 'False' + allowed.add(key) + elif key.startswith('dep_') and key.endswith('_design'): + row['value'] = row['text'] = ('exploratory high-effort reference profile' + if '_named_' in key else 'exploratory') + allowed.add(key) + elif key in ('dep_land_statement', 'dep_land_taste_statement'): + row['value'] = row['text'] = str(row['value']).replace( + 'FDR-controlled', 'classified from BH-adjusted component tests; no cell-level FDR guarantee') + allowed.add(key) + for row in data['figures']: + if row.get('caption'): + row['caption'] = review_caption(row.get('figure', row['name']), row['caption']) + file = outdir / row['path'] + if file.exists(): + row['bytes'] = file.stat().st_size + data['notes'] = [review_note(note) for note in data.get('notes', [])] + csv_before = {p: p.read_bytes() for p in (outdir / 'tables').glob('*.csv')} + for row in data['tables']: + file = outdir / row['path'] + if file.suffix == '.md' and file.exists(): + first, rest = file.read_text().split('\n', 1) + prefix, caption = first.split('** ', 1) + file.write_text(prefix + '** ' + review_caption(file.stem, caption) + '\n' + rest) + if file.exists(): + row['bytes'] = file.stat().st_size + assert all(p.read_bytes() == content for p, content in csv_before.items()) + for key in before: + if key not in allowed: + assert data['values'][key] == before[key], key + data['review_metadata'] = { + 'mode': 'editorial replay; not a completed numerical rerun', + 'numerical_source_git_sha': data.get('provenance', {}).get('git_sha'), + 'review_source_git_sha': subprocess.check_output(['git', 'rev-parse', 'HEAD'], cwd=ROOT, text=True).strip(), + 'changed_design_fields': sorted(allowed), + 'numerical_value_payloads_preserved': True, + 'table_csv_bytes_preserved': True, + 'full_rerun_status': 'interrupted during dependence analysis; no completed rerun claimed', + } + path.write_text(json.dumps(data, indent=2) + '\n') + return data['review_metadata'] + + +if __name__ == '__main__': + print(json.dumps(refresh(), indent=2)) diff --git a/scripts/reproduce_paper.py b/scripts/reproduce_paper.py index 93857f4..2fc0d0e 100644 --- a/scripts/reproduce_paper.py +++ b/scripts/reproduce_paper.py @@ -222,8 +222,10 @@ def say(self, msg: str) -> None: def note(self, msg: str) -> None: """A finding worth printing and recording (e.g. a contradicted claim).""" - self.notes.append(f"[{self._step}] {msg}") - self.say(f" ! {msg}") + from scripts.paper_review_metadata import review_note + note = review_note(f"[{self._step}] {msg}") + self.notes.append(note) + self.say(f" ! {note}") # -- recording -------------------------------------------------------- def put( @@ -331,13 +333,20 @@ def save_fig(ctx: Ctx, fig, name: str, caption: str, svg: bool = False) -> None: """Write ``name.png`` (300 dpi) and optionally ``name.svg``; record both.""" import matplotlib.pyplot as plt + from scripts.paper_review_metadata import review_caption + caption = review_caption(name, caption) ctx.ensure_dirs() + if name.startswith(("F1_", "F2_", "F3_", "F7_", "F8_", "F9_")): + from matplotlib.text import Text + for artist in fig.findobj(match=Text): + artist.set_text(artist.get_text().replace("occupancy", "engagement").replace("Occupancy", "Engagement")) png = ctx.figdir / f"{name}.png" fig.savefig(png) out = [png] if svg: s = ctx.figdir / f"{name}.svg" fig.savefig(s) + s.write_text("\n".join(line.rstrip() for line in s.read_text().splitlines()) + "\n") out.append(s) plt.close(fig) for p in out: @@ -363,6 +372,8 @@ def save_table( md_fields: Iterable[str] | None = None, ) -> None: """Write ``name.csv`` (all fields) and ``name.md`` (a readable subset).""" + from scripts.paper_review_metadata import review_caption + caption = review_caption(name, caption) ctx.ensure_dirs() fields = list(fields) csv_path = ctx.tabdir / f"{name}.csv" @@ -742,14 +753,14 @@ def step_scorecard(ctx: Ctx) -> None: ax.set_title(f"{title} at {_fmt_M(PAPER_CONC)}") ax.grid(axis="y", lw=0.5, zorder=0) ax.set_axisbelow(True) - axes[0].set_ylabel("receptor engagement (0-1); occupancy only for Kd/Ki rows)") + axes[0].set_ylabel("receptor engagement (0-1)") handles = [ plt.Rectangle((0, 0), 1, 1, color=INSECT, label="insect target"), plt.Rectangle((0, 0), 1, 1, color=VERTEBRATE, label="vertebrate counterpart"), - plt.Rectangle((0, 0), 1, 1, facecolor="white", edgecolor=TEXT2, hatch="///", label="class placeholder (no sourced EC50)"), + plt.Rectangle((0, 0), 1, 1, facecolor="white", edgecolor=TEXT2, hatch="///", label="not modelled (no sourced parameter)"), ] fig.legend(handles=handles, loc="lower center", ncol=3, bbox_to_anchor=(0.5, -0.12)) - fig.suptitle("Dual scorecard: same compound, same concentration, two organisms", y=1.02) + fig.suptitle("Insect and vertebrate receptor panels at the same concentration", y=1.02) save_fig( ctx, fig, @@ -1014,7 +1025,7 @@ def step_graph(ctx: Ctx) -> None: "node that itself receives input (mean degree %.2f). A degree-preserving " "rewire of such a graph is close to the identity, so a negative topology " "verdict on this cut is weak evidence. The `taste_motor` cut (mean degree " - "%.2f) is the substrate to generalise from." + "%.2f) provides a denser comparison, not a basis for whole-network generalisation." % ( ctx.text("named_share_onto_seeds"), ctx.get("named_edges") or 0, @@ -1264,17 +1275,22 @@ def step_veto(ctx: Ctx) -> None: r = nb["readouts"] data[engine][label] = r tag = f"{engine}_{label}" - ctx.put(f"veto_{tag}_sugar", _f(r["mn9_sugar_hz"]), unit="Hz") - ctx.put(f"veto_{tag}_sugar_bitter", _f(r["mn9_sugar_bitter_hz"]), unit="Hz") + output_unit = "rate-model units" if engine == "rate" else "LIF spikes/s" + ctx.put(f"veto_{tag}_sugar", _f(r["mn9_sugar_hz"]), unit=output_unit) + ctx.put(f"veto_{tag}_sugar_bitter", _f(r["mn9_sugar_bitter_hz"]), unit=output_unit) ctx.put(f"veto_{tag}_ratio", _f(r["bitter_veto_ratio"])) ctx.put("veto_n_sweet_grn", int(data["rate"]["vehicle"]["n_sweet_grn"])) ctx.put("veto_n_bitter_grn", int(data["rate"]["vehicle"]["n_bitter_grn"])) - ctx.put("veto_sugar_hz", float(data["rate"]["vehicle"]["sugar_hz"]), unit="Hz") + ctx.put( + "veto_sugar_hz", + float(data["rate"]["vehicle"]["sugar_hz"]), + unit="rate-model input units / LIF input spikes/s", + ) if data["rate"]["imidacloprid"]["bitter_veto_ratio"] is None: ctx.note( "imidacloprid at 1 uM silences MN9 on the map path on both engines " - "(sugar-driven MN9 = 0 Hz), so its bitter-veto ratio is undefined, not 0. " + "(sugar-driven MN9 output = 0), so its bitter-veto ratio is undefined, not 0. " "g_ach hits the 0.05 floor, which removes the cholinergic GRN->MN9 drive " "itself. This is a property of the agonist patch rule, and it is why the " "taste-map arm of the null panel is undefined for imidacloprid." @@ -1302,7 +1318,11 @@ def step_veto(ctx: Ctx) -> None: ) ax.set_xticks(x) ax.set_xticklabels(labels) - ax.set_ylabel("MN9 rate (Hz)") + ax.set_ylabel( + "MN9 output (rate-model units)" + if engine == "rate" + else "MN9 output (LIF spikes/s)" + ) ax.set_title(f"{engine} engine") ax.set_ylim(0, max(sugar + both) * 1.3 + 0.1) ax.grid(axis="y", lw=0.5) @@ -1318,15 +1338,17 @@ def step_veto(ctx: Ctx) -> None: "F5_bitter_veto", "MN9 firing on the map-extracted labellar-GRN path, driven through the real " f"MaleCNS sweet ({ctx.get('veto_n_sweet_grn')}) and bitter " - f"({ctx.get('veto_n_bitter_grn')}) GRN body IDs at {ctx.get('veto_sugar_hz'):.0f} Hz, " + f"({ctx.get('veto_n_bitter_grn')}) GRN body IDs with nominal drive " + f"{ctx.get('veto_sugar_hz'):.0f} rate-model input units (rate) or input spikes/s (LIF), " "never by driving MN9 directly. Left: leaky rate engine. Right: Shiu-style LIF " "engine with a background Poisson drive. In vehicle, adding bitter drive takes " "MN9 to zero on both engines (veto ratio 0.00). Fipronil at " f"{_fmt_M(PAPER_CONC)} blocks RDL, raises the sugar-only rate and partially " f"relieves the veto (ratio {ctx.get('veto_rate_fipronil_ratio'):.2f} rate, " f"{ctx.get('veto_lif_fipronil_ratio'):.2f} LIF). Imidacloprid silences MN9 " - "altogether, so its ratio is undefined rather than zero. Absolute rates differ " - "between engines by construction and are not calibrated to any recording.", + "altogether, so its ratio is undefined rather than zero. Rate-engine outputs " + "are rate-model units; LIF outputs are simulated spikes/s. Neither is calibrated " + "to a physiological recording.", ) @@ -1501,16 +1523,17 @@ def step_evidence(ctx: Ctx) -> None: "engagement_model_note", "param_type_note", ], - "The transformation rule of the evidence type system, materialised. The " + "The validated evidence dispatch rule, materialised. The " "permitted transformation is a function of two facts: the parameter the " "source measured and the ordered distance E0-E4 between that source and " "this compound at this receptor in this species. A binding constant yields " "physical fractional occupancy only at E0; transferred, it becomes a labelled " "proxy carrying a warning that names the gap. Anything at E4 is `not_modelled` " - "and returns N/A. `tests/test_pharm_evidence.py` checks the soundness " - "invariant exhaustively over every library row and every entry point: no " + "and returns N/A. The finite tests in `tests/test_pharm_evidence.py` enumerate " + "every row in the shipped library and each listed engagement-to-circuit path: no " "number reaching a circuit readout originates from a row whose " - "(param_type, distance) pair does not admit one.", + "(param_type, distance) pair does not admit one. This is an audit of that " + "finite implementation surface, not a proof over arbitrary future rows or paths.", md_fields=[ "param_type", "evidence_distance_label", @@ -1607,16 +1630,17 @@ def step_evidence(ctx: Ctx) -> None: "What the compound library rests on, row by row. Left: the parameter each " "cited source actually measured. Middle: the ordered evidence distance between " "that source and this compound at this receptor in this species. Right: the " - "transformation the type system then permits, which is a function of the two. " + "transformation the evidence dispatch rule permits, a function of the two. " f"Green marks the {ctx.get('ev_dist_E0')} on-target (E0) rows and the " f"{ctx.get('ev_n_binding_occupancy')} of them that may use a physical " "binding-occupancy model; amber marks the proxy models, where the same Hill " "expression is computed but reported as a transfer with a warning naming the " "gap; red marks rows that carry no modellable evidence and return N/A rather " - "than a small number. The type system, not a convention, enforces this: asking " - "for an engagement from an untyped row raises rather than returning a value, " - "and a soundness invariant is checked exhaustively over every row and every " - "entry point.", + "than a small number. Runtime validation enforces this: asking " + "for an engagement from an untyped row raises rather than returning a value. " + "Finite tests enumerate every row in the shipped library and each listed " + "engagement-to-circuit path; that audit does not establish correctness for " + "arbitrary future rows or paths.", ) @@ -1633,7 +1657,7 @@ def step_dependence(ctx: Ctx) -> None: difference rather than asserting it). *Non-rejection is no longer equivalence.* Every mode carries a three-way - verdict against a prespecified margin, and the landscape is + verdict plus a descriptive point-gap diagnostic against a configured margin, and the landscape is Benjamini-Hochberg corrected across its structural tests with both counts reported. @@ -1687,7 +1711,7 @@ def step_dependence(ctx: Ctx) -> None: def arm(compound: str, assay: str, readout: str, n_i: int, tag: str) -> dict[str, Any]: """One profile, plus the raw draws the histogram and the sweep need. - The profile is authoritative -- it carries the equivalence margin, the + The profile is authoritative -- it carries the configured point-gap margin, the three-way verdicts, the class and the ladder. ``null_distribution`` is called again only for the four ``nullmodels`` degradations, to keep their draws for F6 and their checkpoint prefixes for T11; shuffle *i* @@ -1745,13 +1769,17 @@ def arm(compound: str, assay: str, readout: str, n_i: int, tag: str) -> dict[str ctx.put( f"dep_{tag}_{compound}_effect", _f(prof["real_effect"]), - unit="Hz" if assay == "subgraph" else None, + unit="rate-model units" if assay == "subgraph" else None, ) ctx.put(f"dep_{tag}_{compound}_vehicle", _f(prof.get("real_vehicle"))) ctx.put(f"dep_{tag}_{compound}_class", cls["class"]) ctx.put(f"dep_{tag}_{compound}_level", lvl.get("level")) ctx.put(f"dep_{tag}_{compound}_level_verdict", lvl.get("verdict")) - ctx.put(f"dep_{tag}_{compound}_delta", _f(prof.get("delta")), unit="Hz") + ctx.put( + f"dep_{tag}_{compound}_delta", + _f(prof.get("delta")), + unit="rate-model units" if assay == "subgraph" else "ratio units", + ) ctx.put( f"dep_{tag}_{compound}_delta_frac", _f(prof.get("delta_frac")), @@ -1823,10 +1851,10 @@ def arm(compound: str, assay: str, readout: str, n_i: int, tag: str) -> dict[str ) if gap_plain is not None and gap_matched is not None: ctx.note( - "the transmitter null matters to the equivalence claim and not to the " - "probability: imidacloprid's gap from the null median is %.2f Hz under the " - "plain label permutation and %.2f Hz under the weight-matched null, against " - "a prespecified margin of %s Hz. The plain permutation is a joint " + "the transmitter null changes both the descriptive point-gap diagnostic and the " + "probability: imidacloprid's gap from the null median is %.2f rate-model units under the " + "plain label permutation and %.2f rate-model units under the weight-matched null, against " + "a configured margin of %s rate-model units. The plain permutation is a joint " "target-set-and-sign null and is reported as one." % (gap_plain, gap_matched, ctx.text("dep_named_imidacloprid_delta")) ) @@ -1835,8 +1863,8 @@ def arm(compound: str, assay: str, readout: str, n_i: int, tag: str) -> dict[str "the fipronil bitter-veto ratio on the taste-motor path is a negative at " "n=%d shuffles (max |z| = %s, p resolution %.4f): the veto ratio is model " "behaviour, not evidence about the wiring, and the verdicts say which " - "modes are equivalent within tolerance (%s) and which are merely " - "indeterminate (%s)." + "modes pass the descriptive within-margin point-gap check (%s) and which " + "are indeterminate (%s). The legacy label is not a formal equivalence result." % ( n_taste, ctx.text("dep_taste_fipronil_max_abs_z"), @@ -1879,11 +1907,14 @@ def arm(compound: str, assay: str, readout: str, n_i: int, tag: str) -> dict[str "cut, with the seed block and node order held fixed. `p_two_sided` is the " "empirical permutation probability (k+1)/(n+1) and is the statistic to read " "first; `p_resolution` is its floor, and `z` is a standardised distance from a " - "usually non-normal null. `verdict` is the only claim the design supports: " - "`distinguishable` when the test rejects, `equivalent_within_tolerance` when " - "the gap from the null median is below the prespecified margin `delta`, and " - "`indeterminate` otherwise -- a failure to reject is never evidence of " - "equivalence. `sign_permute` is a **joint** target-set-and-sign null and is " + "usually non-normal null. `verdict` records `distinguishable` when the test " + "rejects. The legacy API label `equivalent_within_tolerance` means only that " + "the point estimate of the absolute real-effect to null-median gap is below " + "the configured margin `delta`; it is a descriptive point-gap diagnostic, not " + "a formal equivalence test. `indeterminate` means neither condition was met. " + "No interval for the gap, TOST, distributional closeness criterion, or " + "error-controlled equivalence decision is computed. `sign_permute` is a " + "**joint** target-set-and-sign null and is " "off the ladder (`on_ladder = False`); the rank-3 rung is the weight-matched " "transmitter null, which holds each transmitter's share of total outgoing " "weight fixed (T23).", @@ -2090,11 +2121,11 @@ def arm(compound: str, assay: str, readout: str, n_i: int, tag: str) -> dict[str ), ) ctx.note( - "instrument validation: %s Detection over the planted grid by permutation " - "count is %s and by planted strength %s; the empirical false-positive rate on " - "the unplanted control is %d of %d. Effect size dominates the permutation " - "count: full power at planted strength %s, detection at or below one in four " - "at %s." + "synthetic recovery check: %s Detection over the planted grid by permutation " + "count is %s and by planted strength %s; detections on " + "the unplanted control are %d of %d. All sampled networks were detected at " + "planted strength %s, with detection at or below one in four " + "at %s. These small-sample frequencies do not establish general power or calibration." % ( rec["statement"], ctx.text("val_power_by_n"), @@ -2195,7 +2226,8 @@ def landscape(graph: str, n_i: int, tag: str) -> dict[str, Any]: f"dep_{tag}_below_relative_floor_cells", [f"{r['compound']}@{r['conc_M']:.0e}" for r in s["below_relative_effect_floor"]], text=", ".join( - f"{r['compound']} {_fmt_M(r['conc_M'])} ({r['real_effect']:.3g} Hz)" + f"{r['compound']} {_fmt_M(r['conc_M'])} " + f"({r['real_effect']:.3g} rate-model units)" for r in s["below_relative_effect_floor"] ) or "none", @@ -2233,8 +2265,8 @@ def landscape(graph: str, n_i: int, tag: str) -> dict[str, Any]: for v, k in verdicts.items(): ctx.put(f"dep_{tag}_verdicts_{v}", int(k)) deltas = land.get("delta_range") or [None, None] - ctx.put(f"dep_{tag}_delta_min", _f(deltas[0]), unit="Hz") - ctx.put(f"dep_{tag}_delta_max", _f(deltas[1]), unit="Hz") + ctx.put(f"dep_{tag}_delta_min", _f(deltas[0]), unit="rate-model units") + ctx.put(f"dep_{tag}_delta_max", _f(deltas[1]), unit="rate-model units") # the cells the paper names one by one for cell in land["cells"]: if abs(cell["conc_M"] - PAPER_CONC) > 1e-18: @@ -2260,10 +2292,12 @@ def landscape(graph: str, n_i: int, tag: str) -> dict[str, Any]: if land["summary"]["n_topology_dependent"] != land["summary"]["n_topology_dependent_raw"]: ctx.note( - "multiplicity changes the headline count on the `named` cut: %d cells are " - "topology-dependent on the raw permutation p and %d survive " - "Benjamini-Hochberg across the %d structural tests of the landscape. The " - "corrected count is the one the paper quotes." + "multiplicity changes the headline count on the `named` cut: the raw " + "test-level rule labels %d cells topology-dependent, while the rule using " + "Benjamini-Hochberg-adjusted probabilities labels %d. BH is applied across " + "the %d individual structural tests, not across cells; a cell is labelled " + "topology-dependent when at least one of its adjusted structural tests " + "rejects. This is not a cell-level FDR statement." % ( land["summary"]["n_topology_dependent_raw"], land["summary"]["n_topology_dependent"], @@ -2365,10 +2399,14 @@ def landscape_rows(res: dict[str, Any]) -> list[dict[str, Any]]: landscape_fields, f"Connectome-dependence landscape on the `named` cut: every compound of the " f"library at {len(land['concs_M'])} concentrations, {n_land} shuffles per mode, " - "the same shuffled graphs reused by every cell (a paired design). `class` is " - "decided on the Benjamini-Hochberg adjusted probabilities across the " - "structural tests of the whole landscape and `class_raw` on the uncorrected " - "ones; `class_absolute_floor` drops the prespecified relative effect floor (1 % " + "the same shuffled graphs reused by every cell (a paired design). Benjamini-" + "Hochberg adjusts the individual structural-test probabilities across the " + "whole landscape; it is not a cell-level FDR procedure. `class` uses those " + "test-level adjusted probabilities and labels a cell topology-dependent when " + "at least one adjusted structural test rejects; `class_raw` uses the nominal " + "probabilities. `q_value` is the smallest adjusted structural-test probability " + "for the cell, not a cell-level q-value. `class_absolute_floor` drops the " + "prespecified relative effect floor (1 % " "of the vehicle readout) so the cost of that convention is visible. " "`necessary_level` is the weakest graph model on the information ladder that " "this test could **not** distinguish from the real cut -- never a claim that " @@ -2386,7 +2424,9 @@ def landscape_rows(res: dict[str, Any]) -> list[dict[str, Any]]: f"The same landscape repeated on the denser `taste_motor` cut " f"({ctx.get('taste_nodes')} nodes, {ctx.get('taste_edges')} edges, mean degree " f"{ctx.get('taste_motor_mean_degree') or 0.0:.1f}) at {n_land_taste} shuffles per " - "mode. The `named` cut is an in-star (T19) and a degree-preserving rewire of it " + "mode. As in T12, BH adjusts individual structural tests across the landscape, " + "not cells; a cell is topology-dependent when at least one adjusted structural " + "test rejects. The `named` cut is an in-star (T19) and a degree-preserving rewire of it " "is close to the identity; this table is what the same instrument says about a " "cut that has wiring to destroy. The two do not agree, and that disagreement " "is the paper's main RQ2 result.", @@ -2420,11 +2460,12 @@ def landscape_rows(res: dict[str, Any]) -> list[dict[str, Any]]: if col == 0: ax.set_ylabel(f"{compound}\nshuffles") if row == 1: - ax.set_xlabel("effect on mean rate (Hz)") + ax.set_xlabel("effect on mean output (rate-model units)") axes[0, 0].legend(fontsize=7) fig.suptitle( f"Connectome-dependence profile, {n} permutations per mode " - f"(network mean rate, {_fmt_M(PAPER_CONC)}); p floor 1/(n+1) = {1.0 / (n + 1):.4f}", + f"(network mean rate-model output, {_fmt_M(PAPER_CONC)}); " + f"p floor 1/(n+1) = {1.0 / (n + 1):.4f}", y=1.02, ) fig.tight_layout() @@ -2433,17 +2474,18 @@ def landscape_rows(res: dict[str, Any]) -> list[dict[str, Any]]: fig, "F6_dependence_profile", "Histograms are the drug effect (treated minus vehicle on the network mean " - f"rate) measured on {n} degraded copies of the 1-hop MN9/DNp01 `named` cut; the " + f"rate-model output) measured on {n} degraded copies of the 1-hop MN9/DNp01 `named` cut; the " "red line is the same contrast on the real cut. Columns run in increasing order " "of destruction: transmitter labels permuted, synapse weights permuted, " "degree-preserving double-edge rewiring, Erdos-Renyi. Titles carry the " - "empirical two-sided permutation p and the three-way verdict against the " - f"prespecified equivalence margin (delta = " - f"{ctx.text('dep_named_imidacloprid_delta')} Hz, 5 % of the vehicle readout): " - "`distinguishable` means the test rejected, `equivalent within tolerance` means " - "the gap from the null median is below the margin, and `indeterminate` means " - "neither -- a failure to reject is not evidence that the degraded graph gives " - "the same effect. Plain label permutation is shown because it is informative, " + "empirical two-sided permutation p and the three-way legacy verdict against the " + f"configured point-gap margin (delta = " + f"{ctx.text('dep_named_imidacloprid_delta')} rate-model units, 5 % of the vehicle readout): " + "`distinguishable` means the test rejected; `equivalent within tolerance` means " + "only that the point gap from the null median is below the margin; and " + "`indeterminate` means neither. The within-margin label is a descriptive " + "point-gap diagnostic, not a formal equivalence result. Plain label permutation " + "is shown because it is informative, " "but it moves the weighted excitation/inhibition balance as well as transmitter " "identity (T23), so it is a joint null and not a rung of the ladder. " f"Imidacloprid is {imi['class']} (necessary level: " @@ -2519,8 +2561,9 @@ def landscape_rows(res: dict[str, Any]) -> list[dict[str, Any]]: "Every compound of the library at four concentrations, on both committed cuts, " "classified by the weakest graph model the permutation test could **not** " "distinguish from the real cut -- never by a model that reproduces the effect. " - "Classes are decided on Benjamini-Hochberg adjusted probabilities across each " - "landscape's structural tests, with a prespecified relative effect floor of 1 % " + "Benjamini-Hochberg adjusts the individual structural-test probabilities across " + "each landscape, not cells; the displayed class uses those adjusted tests and " + "a configured relative effect floor of 1 % " f"of the vehicle readout. Left, the 1-hop `named` cut: " f"{counts.get('topology-dependent', 0)} of {land['n_cells']} cells need the " f"wiring pattern, {counts.get('composition-dominated', 0)} are not " @@ -2560,7 +2603,7 @@ def landscape_rows(res: dict[str, Any]) -> list[dict[str, Any]]: ax2.set_xticks(range(len(labels))) ax2.set_xticklabels(labels) ax2.set_xlabel("planted loop strength") - ax2.set_ylabel("|drug effect| on the synthetic cut (Hz)") + ax2.set_ylabel("|drug effect| on the synthetic cut (rate-model units)") ax2.set_title( f"Ground-truth recovery at n = {rec['n']}\n" f"{rec['n_detected']} of {rec['n_positive']} planted effects recovered, " @@ -2577,16 +2620,19 @@ def landscape_rows(res: dict[str, Any]) -> list[dict[str, Any]]: "F16_instrument_validation", "The dependence ladder pointed at a graph whose answer is known in advance. A " "recurrent cholinergic cycle of known strength is planted in a synthetic cut of " - "the same size, density and transmitter composition as the `named` cut; a gain " + "approximately the size, density and transmitter composition of `named`; a gain " "patch that collapses `g_ach` then removes an amplification that exists only " "while the cycle is intact, so the drug contrast depends on the wiring by " "construction. Left: the share of independently generated graphs the ladder " "classifies topology-dependent, against planted strength and permutation count. " "Right: the recovery experiment itself, with red marking the strengths the " - "ladder recovered. The unplanted control is the empirical false-positive rate " + "ladder recovered. Detections on the unplanted control are " f"({ctx.get('val_power_control_hits')} of " - f"{ctx.get('val_power_control_runs')}). Without this panel, a negative from the " - "ladder would be indistinguishable from an under-powered test.", + f"{ctx.get('val_power_control_runs')}). Each grid cell has only " + f"{pw['replicates']} networks; uncertainty is broad and networks are reused " + "across permutation counts. The zero-strength control omits cycle edges and " + "transmitter relabelling. This is a limited recovery check, not general " + "calibration or endpoint-specific power for a real circuit.", ) def step_landscape(ctx: Ctx) -> None: @@ -2741,8 +2787,8 @@ def step_dose(ctx: Ctx) -> None: ctx.put("ic50_hi_M", float(ci["ic50"][1]), text=_fmt_M(ci["ic50"][1])) ctx.put("ic50_slope", float(fit["slope"])) ctx.put("ic50_r2", float(fit["r2"]), text=f"{fit['r2']:.4f}") - ctx.put("ic50_top_hz", float(fit["top"]), unit="Hz") - ctx.put("ic50_bottom_hz", float(fit["bottom"]), unit="Hz") + ctx.put("ic50_top_hz", float(fit["top"]), unit="rate-model units") + ctx.put("ic50_bottom_hz", float(fit["bottom"]), unit="rate-model units") ctx.put("ic50_in_range", bool(fit["in_range"])) ctx.put("ic50_readout", ic["readout"]) ctx.put( @@ -2754,14 +2800,14 @@ def step_dose(ctx: Ctx) -> None: tor = sensitivity("subgraph", "imidacloprid", PAPER_CONC, readout="mean_hz", seed=ctx.seed) tor = sorted(tor, key=lambda r: abs(float(r["span"])), reverse=True) for r in tor: - ctx.put(f"tornado_span_{r['param']}", float(r["span"]), unit="Hz") + ctx.put(f"tornado_span_{r['param']}", float(r["span"]), unit="rate-model units") ctx.put("tornado_top_param", tor[0]["param"]) zero_span = [r["param"] for r in tor if abs(float(r["span"])) < 1e-9] ctx.put("tornado_zero_span_params", zero_span, text=", ".join(zero_span) or "none") if "ec50" in zero_span: ctx.note( "at 1 uM imidacloprid the model is exactly insensitive to a 2x change in " - "the insect_nAChR EC50 and in the Hill coefficient (span 0 Hz): occupancy " + "the insect_nAChR EC50 and in the Hill coefficient (span 0 rate-model units): engagement " "is saturated, so the gain-rule coefficient is the only parameter that " "still moves the readout. A dose-response experiment, not a single high " "dose, is what constrains the EC50." @@ -2780,7 +2826,7 @@ def step_dose(ctx: Ctx) -> None: ax.axvline(fit["ic50"], color=CRIT, lw=1, ls="--") ax.set_xscale("log") ax.set_xlabel("imidacloprid concentration (M)") - ax.set_ylabel(f"network {ic['readout']} (Hz)") + ax.set_ylabel(f"network {ic['readout']} (rate-model units)") ax.set_title( f"Model IC50 = {_fmt_M(fit['ic50'])}\n" f"[{_fmt_M(ci['ic50'][0])}, {_fmt_M(ci['ic50'][1])}], slope {fit['slope']:.2f}, r2 {fit['r2']:.4f}" @@ -2800,7 +2846,9 @@ def step_dose(ctx: Ctx) -> None: ax2.axvline(base, color=TEXT1, lw=1) ax2.set_yticks(y) ax2.set_yticklabels(names) - ax2.set_xlabel(f"network mean rate (Hz); treated baseline at 1 uM = {base:.2f} Hz") + ax2.set_xlabel( + f"network mean output (rate-model units); treated baseline at 1 uM = {base:.2f}" + ) ax2.set_title("One-at-a-time sensitivity\n(each parameter halved and doubled)") ax2.grid(axis="x", lw=0.5) ax2.set_axisbelow(True) @@ -3391,11 +3439,11 @@ def step_expression(ctx: Ctx) -> None: nb = run_weighted_subgraph_assay("imidacloprid", PAPER_CONC) r = nb["readouts"] - ctx.put("expr_mn9_uniform_hz", float(r["uniform"]["mn9_hz"]), unit="Hz") - ctx.put("expr_mn9_weighted_hz", float(r["mn9_hz"]), unit="Hz") - ctx.put("expr_mn9_delta_hz", float(r["delta_vs_uniform"]["mn9_hz"]), unit="Hz") - ctx.put("expr_mean_uniform_hz", float(r["uniform"]["mean_hz"]), unit="Hz") - ctx.put("expr_mean_weighted_hz", float(r["mean_hz"]), unit="Hz") + ctx.put("expr_mn9_uniform_hz", float(r["uniform"]["mn9_hz"]), unit="rate-model units") + ctx.put("expr_mn9_weighted_hz", float(r["mn9_hz"]), unit="rate-model units") + ctx.put("expr_mn9_delta_hz", float(r["delta_vs_uniform"]["mn9_hz"]), unit="rate-model units") + ctx.put("expr_mean_uniform_hz", float(r["uniform"]["mean_hz"]), unit="rate-model units") + ctx.put("expr_mean_weighted_hz", float(r["mean_hz"]), unit="rate-model units") def step_ablation(ctx: Ctx) -> None: @@ -3460,7 +3508,7 @@ def step_ablation(ctx: Ctx) -> None: "spearman_rho_vs_full": rho, "pearson_r_vs_full": lvl.get("pearson_r_vs_full"), "residual_rms_standardised": lvl.get("residual_rms_standardised"), - "residual_rms_hz": lvl.get("residual_rms_hz"), + "residual_rms_rate_model_units": lvl.get("residual_rms_hz"), "reproduces_full_ordering": lvl.get("reproduces_full_ordering"), "information_added_vs_previous": lvl.get("information_added_vs_previous"), } @@ -3471,11 +3519,11 @@ def step_ablation(ctx: Ctx) -> None: { "conc_M": float(conc_key), "level": C_FLOOR_LEVEL, - "unit": floor.get("unit") or "Hz", + "unit": floor.get("unit") or "rate-model units", "spearman_rho_vs_full": floor.get("spearman_rho_vs_full"), "pearson_r_vs_full": floor.get("pearson_r_vs_full"), "residual_rms_standardised": floor.get("residual_rms_standardised"), - "residual_rms_hz": floor.get("residual_rms_hz"), + "residual_rms_rate_model_units": floor.get("residual_rms_hz"), "reproduces_full_ordering": floor.get("reproduces_full_ordering"), "information_added_vs_previous": None, } @@ -3492,7 +3540,11 @@ def step_ablation(ctx: Ctx) -> None: paper = tab["information_gain"][f"{1e-6:.3e}"] by_name = {lvl["level"]: lvl for lvl in paper["levels"]} ctx.put("abl_rho_topology_paper", _f(by_name["C_topology_only"].get("spearman_rho_vs_full"))) - ctx.put("abl_rms_topology_hz", _f(by_name["C_topology_only"].get("residual_rms_hz")), unit="Hz") + ctx.put( + "abl_rms_topology_hz", + _f(by_name["C_topology_only"].get("residual_rms_hz")), + unit="rate-model units", + ) ctx.put("abl_rho_composition_paper", _f(by_name["B_composition_only"].get("spearman_rho_vs_full"))) ctx.put("abl_rho_receptor_paper", _f(by_name["A_receptor_only"].get("spearman_rho_vs_full"))) paper_cmp = paper.get("generic_rule_comparison") or {} @@ -3748,8 +3800,8 @@ def step_ablation(ctx: Ctx) -> None: "compound-concentration cell from a model that has been denied one layer of " "information: receptor engagement only (dimensionless), mechanism gains on the " "graph's transmitter composition only (excitation index), the real cut with one " - "generic multiplier instead of mechanism-specific gains (Hz), and the full " - "model (Hz). The generic multiplier is direction-aware: generic in magnitude, " + "generic multiplier instead of mechanism-specific gains (rate-model units), " + "and the full model (rate-model units). The generic multiplier is direction-aware: generic in magnitude, " "with one bit of sign taken from the mechanism table, which is the minimum a " "connectome-without-pharmacology model needs to order a library containing " "disinhibitors. `C_topology_only_floor` is the historical depression-only rule, " @@ -3769,7 +3821,7 @@ def step_ablation(ctx: Ctx) -> None: "spearman_rho_vs_full", "pearson_r_vs_full", "residual_rms_standardised", - "residual_rms_hz", + "residual_rms_rate_model_units", "reproduces_full_ordering", "information_added_vs_previous", ], @@ -3890,8 +3942,9 @@ def step_stability(ctx: Ctx) -> None: "the C3 retention rate is no longer vacuous and is no longer uniform: at " "n_shuffles = %d (resolution %.4f, against the previous 6 at which p <= " "0.05 was arithmetically unattainable) the conclusion is retained by %d of " - "%d specifications, but only %d of those reach equivalence within the " - "prespecified margin; %d are merely indeterminate." + "%d specifications; %d of those pass the descriptive point-gap check " + "against the configured margin, while %d are non-rejections without " + "that point-gap result." % ( int(res["n_shuffles"]), float(res["shuffle_resolution"]), @@ -3925,8 +3978,8 @@ def step_stability(ctx: Ctx) -> None: with mechanism_spec(spec): nb = run_subgraph_assay("imidacloprid", PAPER_CONC) r = nb["readouts"] - ctx.put(f"stab_{label}_treated_hz", _f(r.get("mean_hz")), unit="Hz") - ctx.put(f"stab_{label}_vehicle_hz", _f((r.get("vehicle") or {}).get("mean_hz")), unit="Hz") + ctx.put(f"stab_{label}_treated_hz", _f(r.get("mean_hz")), unit="rate-model units") + ctx.put(f"stab_{label}_vehicle_hz", _f((r.get("vehicle") or {}).get("mean_hz")), unit="rate-model units") ctx.put(f"stab_{label}_g_ach", _f((nb.get("gains") or {}).get("g_ach"))) t, v = ctx.get("stab_monotone_treated_hz"), ctx.get("stab_monotone_vehicle_hz") if t and v: @@ -3934,7 +3987,7 @@ def step_stability(ctx: Ctx) -> None: ctx.note( "the suppression conclusion is specification-dependent: %d of %d " "prespecified specifications reverse it. Under %s the same compound at the " - "same engagement EXCITES the network (%s Hz treated vs %s Hz vehicle) " + "same engagement EXCITES the network (%s vs %s rate-model units, treated vs vehicle) " "instead of suppressing it. Suppression is a property of the biphasic " "desensitisation term, not of the pharmacology-connectome integration." % ( @@ -4040,7 +4093,7 @@ def step_stability(ctx: Ctx) -> None: "fraction_retained_decidable": r["fraction_retained_decidable"], "n_retained_uncorrected": r.get("n_retained_uncorrected"), "multiplicity": r.get("multiplicity"), - "n_equivalent_within_tolerance": (r.get("equivalence_breakdown") or {}).get( + "n_within_margin_point_gap": (r.get("equivalence_breakdown") or {}).get( "equivalent_within_tolerance" ), "n_indeterminate": (r.get("equivalence_breakdown") or {}).get("indeterminate"), @@ -4061,7 +4114,7 @@ def step_stability(ctx: Ctx) -> None: "fraction_retained_decidable", "n_retained_uncorrected", "multiplicity", - "n_equivalent_within_tolerance", + "n_within_margin_point_gap", "n_indeterminate", "failing_specs", "undecidable_specs", @@ -4080,8 +4133,9 @@ def step_stability(ctx: Ctx) -> None: "the permutation path resolves, so every specification fed the *default* gains " "to its nulls and the matrix's topology rows were empty; they are now computed " "per specification. For a conclusion shaped as a failure to reject, " - "`n_equivalent_within_tolerance` is the part of the retention that is evidence " - "of equivalence and `n_indeterminate` the part that is only a non-rejection.", + "`n_within_margin_point_gap` counts retained specifications whose point gap is " + "below the configured margin; it is a descriptive diagnostic, not a formal " + "equivalence result. `n_indeterminate` counts the remaining non-rejections.", md_fields=[ "conclusion", "n_specs", @@ -4089,7 +4143,7 @@ def step_stability(ctx: Ctx) -> None: "n_lost", "n_undecidable", "n_retained_uncorrected", - "n_equivalent_within_tolerance", + "n_within_margin_point_gap", "n_indeterminate", ], ) @@ -4252,9 +4306,9 @@ def step_stability(ctx: Ctx) -> None: + (", ".join(r["conclusion"] for r in res["rows"] if not r["fragile"]) or "none") + ". The nicotinic suppression conclusion is reversed by every monotone rule, " "which is how a result that depends on a modelling choice looks when it is " - "tested rather than asserted; and for a conclusion shaped as a failure to " - "reject, the fraction retained is only as strong as the equivalence behind it " - "(T14's `n_equivalent_within_tolerance` against `n_indeterminate`).", + "tested rather than asserted. For a conclusion shaped as a failure to reject, " + "T14 separates specifications passing the descriptive within-margin point-gap " + "check from indeterminate non-rejections; neither is a formal equivalence test.", ) @@ -4280,8 +4334,8 @@ def step_uncertainty(ctx: Ctx) -> None: ctx.put("unc_n_factors", int(res["n_factors"])) ctx.put("unc_evaluations", int(res["n_evaluations"])) ctx.put("unc_estimator", str(res["estimator"])) - ctx.put("unc_output_mean", _f(res["output_mean"]), unit="Hz") - ctx.put("unc_variance", _f(res["output_variance"]), unit="Hz^2") + ctx.put("unc_output_mean", _f(res["output_mean"]), unit="rate-model units") + ctx.put("unc_variance", _f(res["output_variance"]), unit="rate-model units squared") ctx.put("unc_interaction_share", _f(res["interaction_share"])) ctx.put("unc_interaction_share_clipped", _f(res["interaction_share_clipped"])) ctx.put("unc_sum_first_order", _f(res["sum_first_order"])) @@ -4330,7 +4384,7 @@ def step_uncertainty(ctx: Ctx) -> None: ctx.put("unc_resolution_statement", str(resolution["statement"])) if resolution["noise_floor_factor"] and resolution["noise_floor_factor"] != "lif_seed": ctx.note( - "the estimator's noise floor is %s, set by the most negative first-order " + "the largest negative-estimate magnitude is %s, from the first-order " "estimate (%s), not by the declared null factor lif_seed at %s. Only %s " "%s a first-order confidence interval that excludes zero; every other " "factor is unresolved at this sample size, and a negative point estimate " @@ -4368,19 +4422,19 @@ def step_uncertainty(ctx: Ctx) -> None: "kind": b.get("kind"), "share_of_variance_S1": b["share_of_variance"], "share_with_interactions_ST": b.get("share_with_interactions"), - "variance_removed_hz2": b["variance_removed"], + "variance_removed_rate_model_units_sq": b["variance_removed"], "ci_low": (b.get("ci") or [None, None])[0], "ci_high": (b.get("ci") or [None, None])[1], } for b in res["budget"] ], - ["source", "kind", "share_of_variance_S1", "share_with_interactions_ST", "variance_removed_hz2", "ci_low", "ci_high"], + ["source", "kind", "share_of_variance_S1", "share_with_interactions_ST", "variance_removed_rate_model_units_sq", "ci_low", "ci_high"], f"Global uncertainty budget for the neighbourhood mean rate " f"({res['compound']} at {_fmt_M(res['conc_M'])}, {res['n_evaluations']} model " "evaluations, Jansen estimators on a Saltelli cross-sample). S1 is the variance " "share resolving that factor alone would remove; ST includes its interactions. " "Estimates are **unclipped**, so a factor whose true index is zero can come out " - "negative; the largest such magnitude is the estimator's measured noise floor " + "negative; the largest such magnitude is an observed error diagnostic " "and is reported with the factor that set it. A factor counts as resolved only " "when its bootstrap interval excludes zero, which is a stronger test than " "exceeding that floor. The shares do not sum to one; the remainder is the " @@ -4391,10 +4445,10 @@ def step_uncertainty(ctx: Ctx) -> None: ) v = voi(result=res) - ctx.put("voi_var_total", _f(v["output_variance"]), unit="Hz^2") + ctx.put("voi_var_total", _f(v["output_variance"]), unit="rate-model units squared") for i, r in enumerate(v["rows"][:3], 1): ctx.put(f"voi_rank{i}_factor", r["factor"]) - ctx.put(f"voi_rank{i}_var", _f(r["voi_var"]), unit="Hz^2") + ctx.put(f"voi_rank{i}_var", _f(r["voi_var"]), unit="rate-model units squared") ctx.put(f"voi_rank{i}_fraction", _f(r["voi_fraction"])) ctx.put(f"voi_rank{i}_experiment", r["experiment"]) ctx.put(f"voi_rank{i}_cost", r["cost"]) @@ -4423,9 +4477,9 @@ def step_uncertainty(ctx: Ctx) -> None: "voi_fraction_of_var": r["voi_fraction"], "voi_fraction_raw": r.get("voi_fraction_raw"), "clipped_from_negative": r.get("clipped_from_negative"), - "voi_var_hz2": r["voi_var"], - "voi_upper_bound_var_hz2": r["voi_upper_bound_var"], - "sd_reduction_hz": r["sd_reduction"], + "voi_var_rate_model_units_sq": r["voi_var"], + "voi_upper_bound_var_rate_model_units_sq": r["voi_upper_bound_var"], + "sd_reduction_rate_model_units": r["sd_reduction"], "experiment": r["experiment"], "resolves": r["resolves"], "cost": r["cost"], @@ -4440,9 +4494,9 @@ def step_uncertainty(ctx: Ctx) -> None: "voi_fraction_of_var", "voi_fraction_raw", "clipped_from_negative", - "voi_var_hz2", - "voi_upper_bound_var_hz2", - "sd_reduction_hz", + "voi_var_rate_model_units_sq", + "voi_upper_bound_var_rate_model_units_sq", + "sd_reduction_rate_model_units", "experiment", "resolves", "cost", @@ -4458,7 +4512,7 @@ def step_uncertainty(ctx: Ctx) -> None: "which is a statement about the sample size rather than about the factor. This " "is variance of a model output under assumed input ranges, not an expected gain " "in accuracy about a living fly.", - md_fields=["rank", "factor", "state", "voi_fraction_of_var", "voi_var_hz2", "experiment", "cost"], + md_fields=["rank", "factor", "state", "voi_fraction_of_var", "voi_var_rate_model_units_sq", "experiment", "cost"], ) import numpy as np @@ -4481,7 +4535,7 @@ def step_uncertainty(ctx: Ctx) -> None: ax.text( floor, len(order) - 0.4, - f" estimator noise floor ({resolution['noise_floor_factor']})", + f" largest negative estimate ({resolution['noise_floor_factor']})", fontsize=7, color=CRIT, va="top", @@ -4507,8 +4561,8 @@ def step_uncertainty(ctx: Ctx) -> None: ax2.barh(y2, [r["voi_var"] for r in top], color=[GOOD if r.get("cost") == "none (compute only)" else SEQ[2] for r in top]) ax2.set_yticks(y2) ax2.set_yticklabels([r["factor"] for r in top], fontsize=8) - ax2.set_xlabel("model variance an exact answer would remove (Hz$^2$)") - ax2.set_title(f"Value of information (Var(Y) = {res['output_variance']:.3g} Hz$^2$)", fontsize=10) + ax2.set_xlabel("model variance an exact answer would remove (rate-model units squared)") + ax2.set_title(f"Value of information (Var(Y) = {res['output_variance']:.3g})", fontsize=10) ax2.grid(axis="x", lw=0.5) ax2.set_axisbelow(True) fig.tight_layout() @@ -4520,7 +4574,7 @@ def step_uncertainty(ctx: Ctx) -> None: "each assumption, first-order and total-order, from a Saltelli cross-sample " "with Jansen estimators, with bootstrap first-order confidence intervals. A " "factor is **resolved** only when its interval excludes zero; the starred rows " - "are not, and a factor's index exceeding the noise floor does not resolve it. " + "are not, and exceeding that diagnostic magnitude does not resolve a factor. " "The dashed line is that floor, defined as the largest magnitude among the " "*negative* first-order estimates: a true first-order index cannot be negative, " "so that magnitude is a measured lower bound on the estimator's error, and on " @@ -4620,14 +4674,14 @@ def step_claims(ctx: Ctx) -> None: def step_scale(ctx: Ctx) -> None: - """Where the dependence verdict might settle with the size of the cut. + """Summarise cached dependence labels across the available cut ladder. The verdict reversed between the two committed cuts, and nothing in the - method says where -- or whether -- it stabilises. A scaling study - (:func:`flylab.analysis.scale.dependence_vs_scale`) answers that on a + method says how labels would vary on other cuts. A separate scaling study + (:func:`flylab.analysis.scale.dependence_vs_scale`) records them on a ladder of nested cuts of 1k to 50k cells built by one fixed recipe. Only the 1k rung is small enough to commit; the rest are CI artifacts, so the - study is run outside this pipeline and its result is **read** here from + study is run outside this pipeline and its cached result is **read** here from ``/scale_study.json`` (or ``data/derived/scale_study.json``) when one exists. Without it, this step still records what is on disk, what the study would cost at each rung (T26) and an explicit "not yet measured" @@ -4735,13 +4789,37 @@ def step_scale(ctx: Ctx) -> None: ctx.put("scale_study_compounds", list(study.get("compounds") or []), text=", ".join(study.get("compounds") or [])) ctx.put("scale_study_settled", bool(stab.get("all_settled")), text="yes" if stab.get("all_settled") else "no") ctx.put("scale_study_settled_clean_ladder", bool(clean.get("all_settled")), text="yes" if clean.get("all_settled") else "no") - ctx.put("scale_study_statement", rep["statements"], text=" ".join(rep["statements"])) + label_statements: list[str] = [] + for compound, block in (stab.get("by_compound") or {}).items(): + cuts = list(block.get("cuts") or []) + sequence = list(block.get("sequence") or []) + if len(cuts) < 2: + statement = ( + f"{compound}: only one cut was observed, with label " + f"'{sequence[0] if sequence else 'unknown'}'." + ) + elif block.get("settled"): + statement = ( + f"{compound}: the largest {int(block.get('n_cuts_agreeing_at_top') or 0)} " + f"of {len(cuts)} observed cuts share the label '{block.get('final')}'." + ) + else: + statement = ( + f"{compound}: labels do not match across the largest observed cuts " + f"({sequence[-2:] if len(sequence) >= 2 else sequence})." + ) + label_statements.append(statement) + ctx.put( + "scale_study_statement", + label_statements, + text=" ".join(label_statements), + ) seqs = [] for comp, b in (clean.get("by_compound") or {}).items(): seqs.append(f"{comp}: " + " -> ".join(f"{c} ({v})" for c, v in zip(b["cuts"], b["sequence"]))) ctx.put("scale_study_sequences", seqs, text="; ".join(seqs) or "n/a") - # --- the paper-quotable shape of the result ------------------- + # --- the paper-quotable shape of the cached result ------------ # Rows, smallest cut first. The key methodological point is that the # inversion does NOT track node count: `scale_1k` has fewer nodes than # `named` and a mean degree twenty times higher, and it already comes @@ -4826,17 +4904,22 @@ def step_scale(ctx: Ctx) -> None: "T27_scale_study", rep["rows"], list(rep["columns"]), - "The same dependence profile across the cuts the scaling study could afford, " + "Cached dependence-profile records across the cuts the separate scaling " + "study could afford, " "smallest first: the cut's structure, the permutation count it was run at (which " "falls with the edge count, so the resolution coarsens up the ladder), the per-mode " "empirical probabilities, the necessary information level and the class. `named` " "and `taste_motor` are built by a different recipe from the `scale_*` rungs, so the " - "clean scale axis is the `scale_*` rows alone. " + str(stab.get("caveat") or ""), + "clean scale axis is the `scale_*` rows alone. This reproduction step reads " + "and summarises those cached rows; it does not recompute the profiles. The " + "legacy `settled` field records matching labels on the largest observed " + "cuts, not statistical stability or convergence.", md_fields=["cut", "n_nodes", "n_edges", "mean_degree", "compound", "n", "p_degree", "p_weight", "p_sign_wm", "necessary_information_level", "class"], ) ctx.note( - "scale study: %d cuts (%s); verdicts settled on the clean ladder: %s. %s" - % (len(cuts_run), ", ".join(cuts_run), "yes" if clean.get("all_settled") else "no", " ".join(rep["statements"])) + "cached scale study: %d cuts (%s); labels match across the largest observed " + "clean-ladder rungs under the legacy descriptive rule: %s. %s" + % (len(cuts_run), ", ".join(cuts_run), "yes" if clean.get("all_settled") else "no", " ".join(label_statements)) ) else: ctx.put("scale_study_status", "not yet run") @@ -4865,7 +4948,7 @@ def step_scale(ctx: Ctx) -> None: "scale_study_statement", None, text=( - "Where the verdict settles with the size of the cut has not been measured at " + "Dependence labels across the unobserved cut ladder have not been measured at " "this commit: the scale ladder's rungs on disk are " + (", ".join(on_disk) or "none") + " and the study over the full ladder had not been run." @@ -5046,7 +5129,7 @@ def sub(m: re.Match[str]) -> str: ("stability", "specification robustness + threshold grid (F14, T14, T14b, T15)", step_stability), ("uncertainty", "global uncertainty budget + VOI (F15, T16, T17)", step_uncertainty), ("claims", "claim provenance chain (T18)", step_claims), - ("scale", "where the dependence verdict settles with the cut (T26, T27 when the study exists)", step_scale), + ("scale", "summarise cached scale-study labels and the cost frontier (T26, T27 when available)", step_scale), ("architecture", "architecture schematic (F1)", step_architecture), ("paper", "render the draft and supplement from their templates", step_paper), ] @@ -5295,7 +5378,6 @@ def sub(m: re.Match[str]) -> str: "scale_study_n_range", "scale_study_resolution_range", "scale_study_settled_clean_ladder", - "scale_study_statement", "scale_top_landscape_days", "scale_top_profile_hours", "scale_top_rung", @@ -5319,8 +5401,6 @@ def sub(m: re.Match[str]) -> str: "stab_C3_retained_uncorrected", "stab_C3_undecidable", "stab_C4_claim", - "stab_C4_equivalent", - "stab_C4_indeterminate", "stab_C4_lost", "stab_C4_retained", "stab_C4_retained_uncorrected", diff --git a/tests/test_analysis_dependence.py b/tests/test_analysis_dependence.py index 6c0ac42..3f217d6 100644 --- a/tests/test_analysis_dependence.py +++ b/tests/test_analysis_dependence.py @@ -14,6 +14,7 @@ from flylab.analysis.dependence import ( CLASSES, CONFIRMATORY_COMPOUNDS, + REFERENCE_COMPOUNDS, DEFAULT_DELTA_FRAC, BALANCED_TRANSMITTERS, DEFAULT_MODES, @@ -457,7 +458,8 @@ def test_necessary_level_separates_equivalence_from_indeterminacy(): assert lvl["mode"] == "rewire_degree_preserving" assert lvl["verdict"] == "equivalent_within_tolerance" assert lvl["equivalent_within_tolerance"] is True - assert "equivalent to the real cut" in lvl["description"] + assert "legacy equivalent_within_tolerance identifier" in lvl["description"] + assert "formal equivalence test" in lvl["description"] # widen the gap past the margin: the same non-rejection is now indeterminate rows[1]["null_median"] = -5.0 @@ -506,7 +508,7 @@ def test_benjamini_hochberg_knows_when_resolution_forbids_rejection(): assert out3["can_reject"] is False -def test_landscape_reports_both_raw_and_fdr_controlled_counts(): +def test_landscape_reports_raw_and_bh_component_adjusted_counts(): land = dependence_landscape( compounds=["imidacloprid", "fipronil", "diazepam"], concs_M=(1e-7, 1e-6), @@ -540,8 +542,14 @@ def test_landscape_reports_both_raw_and_fdr_controlled_counts(): assert row.get("p_adjusted") is None assert cell["q_value"] is not None # the summary states what kind of run this was - assert land["summary"]["design"] in ("exploratory", "confirmatory") + assert land["summary"]["design"] == "exploratory" assert "Benjamini-Hochberg" in land["summary"]["statement"] + assert "component tests" in land["summary"]["statement"] + assert "not a guarantee of cell-level FDR control" in land["summary"]["statement"] + assert land["multiplicity_scope"] == "structural_component_tests" + assert land["cell_level_fdr_controlled"] is False + assert land["bh_dependence_assumptions_verified"] is False + assert land["fdr"]["cell_level_fdr_controlled"] is False json.dumps(land) @@ -567,7 +575,7 @@ def test_landscape_fdr_guards_an_unresolvable_grid(): assert "uninformative rather than zero" in land["summary"]["statement"] -def test_a_landscape_is_exploratory_and_a_prespecified_profile_is_not(): +def test_landscape_and_reference_profiles_default_to_exploratory(): # a real screen: 6 cells x 2 structural modes, so the adjusted threshold # 0.05/12 sits below what 60 shuffles can resolve land = dependence_landscape( @@ -582,12 +590,22 @@ def test_a_landscape_is_exploratory_and_a_prespecified_profile_is_not(): small = dependence_profile("fipronil", 1e-6, modes=("erdos_renyi",), n=25) assert small["confirmatory"] is False assert any("exploratory" in w for w in small["warnings"]) - # the prespecified confirmatory tests are named, and only they qualify - assert set(CONFIRMATORY_COMPOUNDS) == {"imidacloprid", "fipronil"} + # The historical constant is an API alias for retrospective reference + # profiles; compound identity and n never confer confirmatory status. + assert CONFIRMATORY_COMPOUNDS == REFERENCE_COMPOUNDS + reference = dependence_profile( + "fipronil", 1e-6, modes=("erdos_renyi",), n=PAPER_N + ) + assert reference["confirmatory"] is False + assert reference["high_effort_reference"] is True + assert reference["design"] == "exploratory high-effort reference profile" forced = dependence_profile( "fipronil", 1e-6, modes=("erdos_renyi",), n=25, confirmatory=True ) - assert forced["confirmatory"] is True and forced["design"] == "confirmatory (prespecified)" + assert forced["confirmatory"] is True + assert forced["confirmatory_basis"] == "caller_asserted_unverified" + assert forced["design"] == "confirmatory (caller asserted; protocol unverified)" + assert any("CALLER-ASSERTED AND UNVERIFIED" in w for w in forced["warnings"]) assert PAPER_N == 1000 @@ -769,3 +787,12 @@ def test_the_landscape_counts_its_non_monotone_ladders(): assert s["n_non_monotone_ladders"] <= land["n_cells"] for row in s["non_monotone_cells"]: assert row["richer_models_distinguishable"] + + +def test_missing_probability_cannot_establish_within_margin_verdict(): + """A point gap without a test must remain unassessed, including outside margin.""" + from flylab.analysis.dependence import mode_verdict + for median in (1.0, 10.0): + row = mode_verdict(real_effect=1.0, null_median=median, p=None, delta=0.1) + assert row["verdict"] == "indeterminate" + assert "unassessed" in row["verdict_reason"] diff --git a/tests/test_analysis_robustness.py b/tests/test_analysis_robustness.py index c08f426..123eaf6 100644 --- a/tests/test_analysis_robustness.py +++ b/tests/test_analysis_robustness.py @@ -394,6 +394,20 @@ def test_the_stated_readout_is_what_the_predicate_computed(tiny_stability): ) +def test_c4_audit_language_matches_its_thresholded_rejection_predicate(tiny_stability): + res = tiny_stability + c4 = CONCLUSIONS["C4_fipronil_topology_exceeds"] + assert "at least one" in c4["readout"] + assert "thresholded OR rejection" in c4["kind"] + warning_blob = " ".join(res["warnings"]) + assert "C4 is a thresholded OR predicate" in warning_blob + assert "effect-size contrast with no threshold" not in warning_blob + assert res["topology_multiplicity_scope"] == "structural_component_tests" + assert res["cell_level_fdr_controlled"] is False + assert res["bh_dependence_assumptions_verified"] is False + assert "does not verify the positive-dependence conditions" in warning_blob + + def test_the_specification_actually_reaches_the_permutation_engine(): """Regression guard for a silent defect: `mechanism_spec` rebinds `compute_gains` on the assay modules only, so the null-model engine never diff --git a/tests/test_paper_review_metadata.py b/tests/test_paper_review_metadata.py new file mode 100644 index 0000000..3f0d8fa --- /dev/null +++ b/tests/test_paper_review_metadata.py @@ -0,0 +1,37 @@ +"""Editorial replay must not modify scientific estimates or their source.""" +import copy +import json +from pathlib import Path +from scripts.paper_review_metadata import refresh, review_caption + + +def test_review_replay_preserves_estimates_provenance_and_table_data(tmp_path): + source = Path(__file__).resolve().parents[1] / 'papers' / 'results.json' + data = json.loads(source.read_text()) + before = copy.deepcopy(data) + (tmp_path / 'results.json').write_text(json.dumps(data)) + tables = tmp_path / 'tables' + tables.mkdir() + csv = tables / 'control.csv' + csv.write_bytes(b'compound,effect,p\nfipronil,0.9,0.001\n') + original_csv = csv.read_bytes() + metadata = refresh(tmp_path) + after = json.loads((tmp_path / 'results.json').read_text()) + assert after['provenance'] == before['provenance'] + allowed = set(metadata['changed_design_fields']) + for key, row in before['values'].items(): + if key not in allowed: + assert after['values'][key] == row + assert csv.read_bytes() == original_csv + once = copy.deepcopy(after) + refresh(tmp_path) + assert json.loads((tmp_path / 'results.json').read_text()) == once + + +def test_legacy_plot_labels_are_qualified_without_changing_estimates(): + caption = 'Effect = -6.173 Hz; configured median-gap tolerance = 0.332 Hz.' + revised = review_caption('F6_dependence_profile', caption) + assert caption in revised + assert 'not a formal equivalence test' in revised + assert 'not calibrated physiological firing rates' in revised + assert review_caption('F6_dependence_profile', revised) == revised diff --git a/tests/test_paper_word.py b/tests/test_paper_word.py new file mode 100644 index 0000000..bda0217 --- /dev/null +++ b/tests/test_paper_word.py @@ -0,0 +1,32 @@ +"""Check the committed Word copy without requiring the authoring dependencies.""" +import hashlib +import json +from pathlib import Path +from xml.etree import ElementTree as ET +from zipfile import ZipFile + +ROOT = Path(__file__).resolve().parents[1] +WORD = ROOT / "papers/submission/FlyLab_manuscript.docx" +W = "{http://schemas.openxmlformats.org/wordprocessingml/2006/main}" + + +def test_word_export_matches_inputs_and_payload(): + record = json.loads(WORD.with_suffix(".json").read_text()) + for path, expected in record["inputs_sha256"].items(): + assert hashlib.sha256((ROOT / path).read_bytes()).hexdigest() == expected, path + assert hashlib.sha256(WORD.read_bytes()).hexdigest() == record["output_sha256"] + assert record["main_figure_archive_ids"] == ["F6", "F13", "F14", "F11"] + + +def test_word_author_figures_and_references(): + with ZipFile(WORD) as archive: + document = ET.fromstring(archive.read("word/document.xml")) + text = "\n".join("".join(p.itertext()) for p in document.iter(W + "t")) + for item in ("Michél Nguyen", "University of the People", "michel_ng@icloud.com", "0000-0001-6834-4422"): + assert item in text + assert len(list(document.iter(W + "drawing"))) == 4 + for number in range(1, 77): + assert f"[{number}]" in text + assert "[[MISSING:" not in text + assert "Author name, affiliation" not in text + assert "—" not in text diff --git a/tests/test_reproduce.py b/tests/test_reproduce.py index 1176384..86b67f4 100644 --- a/tests/test_reproduce.py +++ b/tests/test_reproduce.py @@ -307,12 +307,17 @@ def test_full_fast_pipeline(tmp_path: Path) -> None: # -------------------------------------------------------------------------- @pytest.mark.skipif(not COMMITTED_RESULTS.exists(), reason="papers/results.json not built yet") def test_body_word_count_is_inside_the_venue_target() -> None: - """IJRC asks for 5000-7000 body words; the pipeline counts them.""" + """Editorial length guard, not a certification of journal format compliance. + + IJRC gives a general 5000-10000 range and a typical Research Articles + range of 6000-8000. Keep the concise draft between 5000 and 8000; final + article-specific length remains an editorial submission check. + """ values = json.loads(COMMITTED_RESULTS.read_text())["values"] row = values.get("paper_words_body") if row is None: pytest.skip("paper step has not run") - assert 5000 <= row["value"] <= 7000, f"body word count {row['value']} is outside 5000-7000" + assert 5000 <= row["value"] <= 8000, f"body word count {row['value']} is outside the editorial 5000-8000 guard" #: framings the peer review asked to have removed, and that must not creep back @@ -357,3 +362,24 @@ def test_manuscript_cites_the_adjacent_tools_it_is_positioned_against() -> None: text = TEMPLATE.read_text() for needle in ("FlyBrainLab", "10.7554/eLife.62362", "10.1038/s42003-024-06852-9"): assert needle in text, f"the manuscript no longer cites {needle}" + + +@pytest.mark.parametrize("cached", [False, True]) +def test_scale_renderer_records_descriptive_status_with_or_without_cache(tmp_path, cached): + """Both branches must emit the keyed record without rerunning large cuts.""" + from scripts.reproduce_paper import Ctx, step_scale + if cached: + (tmp_path / "scale_study.json").write_bytes( + (PAPERS / "scale_study.json").read_bytes() + ) + ctx = Ctx(outdir=tmp_path, quiet=True) + ctx._step = "scale" + step_scale(ctx) + row = ctx.values["scale_study_statement"] + assert row["step"] == "scale" + if cached: + assert row["value"] + assert "observed cuts share the label" in row["text"] + else: + assert row["value"] is None + assert "not been measured" in row["text"]