Skip to content
Draft
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
10 changes: 5 additions & 5 deletions CITATION.cff
Original file line number Diff line number Diff line change
Expand Up @@ -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
Expand All @@ -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
Expand All @@ -60,7 +61,6 @@ keywords:
- sensitivity analysis
- specification robustness
- reproducible research
- 3Rs
identifiers:
- type: other
value: "Zenodo DOI to be minted on release"
Expand Down
22 changes: 11 additions & 11 deletions HANDOFF.md
Original file line number Diff line number Diff line change
Expand Up @@ -48,32 +48,32 @@ 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.

## Findings that constrain what you may claim

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.
Expand Down
4 changes: 2 additions & 2 deletions README.md
Original file line number Diff line number Diff line change
Expand Up @@ -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

Expand Down
Loading
Loading