Apply corrections, typeset the report, check the ledgers in CI

[lvelho]

Four parts specified; three done, one blocked. Four commits.

⚠ Part B could not be done

The specification reads "Insert the text the maintainer supplies" followed by the literal line [PASTE THE SECTION TEXT HERE]. The text was never pasted. Nothing was inserted and nothing was guessed at. The insertion point is untouched and ready.

Its checkable claim was checked anyway, and it fails as worded. Part B was to assert that no specification text survives anywhere in bio-3d-vision at 5f8e39f. Three fragments do:

location	what survives
exp007/preregistration.md:38	quotes it directly — the specification says "six arms, one stimulus"
exp009/preregistration.md:19	quotes a clause, then argues it is wrong
exp007/findings.md:46	"which neither exp005 nor this specification stated"

The weaker claim is true and is the one to make: no specification was ever committed.

The pattern sharpens the argument rather than weakening it. Every surviving fragment is text an artifact was arguing with. Specification text reaches the record only where a preregistration or findings file had to disagree with it to justify what it did instead — that is, only where the specification was executed. A rejected specification produces no artifact, so nothing is ever in a position to quote it. The survival mechanism is parasitic on the work happening, which is exactly why it cannot preserve a rejection. gap-004 carried the same overstatement and is corrected in place.

Part A — corrections

All 13 proposed replacements applied across 20 passages (five claims recur). Count matches the tally: 5 CORRECTED + 8 UNDER-QUALIFIED.

Running the fixed check after applying found one the record had not named: claim 5's figure recurs a third time in §6 — "here are the five open questions", sixty lines below the cited sentence. The check has now earned its place twice, both times on the same failure: a figure corrected where it was noticed and left where it was not.

Judgement calls: "five working days" applied to the four elapsed-time sentences, not the two register descriptors the entry says are right; claims 31 and 41 offered alternatives and the first was taken in both.

The six UNSUPPORTED claims are unchanged, and the draft now says so in the document rather than only in the ledger.

gap-003 closed — but the closure is the re-read, not the pin: claim 2 is a verbatim quotation cited to active-stereo CLAUDE.md:141, and the line was read again at 3f7a263 and still says what the claim says.

spec-defects.md gains two entries: the specified one (entry 4), and entry 5 — this specification's own unfilled placeholder, under the file's rule that a defect not recorded is paid for twice. The file now separates them as a second family: defects of assembly, where each clause reads correctly alone and the error is a relation between two, so every check in the family must be mechanical.

Part C — typeset · 22 pages, zero overfull or underfull boxes

Nine files, one per section plus appendix. Converted by a re-runnable script, not by hand.

The transit is lossless and checked rather than asserted: 11,628 words in, 11,628 out, zero differing blocks, similarity 1.0000.

• No bibliography — the draft cites no published work (checked: no author name, no "et al.", no year-in-parentheses). Citations I judge would be wanted are named in refs.bib as comments — four @software entries at the SHAs now in the ledger — and deliberately not written in.
• No abstract — writing one would mean composing prose that has never been through a claims pass, in the passage the template names as a report's most dangerous.
• What read worse in LaTeX: one monospace path. \texttt{report/claims-verified.md} cannot hyphenate, 17.6pt over; Markdown does not justify, so it reads fine there. \allowbreak was tried and did not take — recorded, because the obvious fix failing is the useful half. \sloppy scoped to that block does.
• CI needs no new packages.

Part D — od-004 closed, with no new dependency

A ledgers job parses both ledgers with ruby -ryaml. Ruby, not Python, is the whole reason this was open: YAML is not in Python's stdlib and PyYAML is not guaranteed on the runner, so the obvious check needed a pip install. Ruby ships Psych in its stdlib and is preinstalled. The smallest dependency that works is none.

It was watched failing before it was committed — and the command was extracted from the workflow file and run, not hand-typed, so the YAML quoting was part of what got tested.

Falsifiers

1. PASSES. Clean git clone → make -C report → 22 pages, exit 0, no manual install. CI compiles it.
2. PASSES. No figure stated two ways. Worth noting: "thirteen" now carries three referents that are all genuinely 13 (experiments, foreclosures, and the 13 applied corrections) — a collision, not an inconsistency. One false positive: the check reads LaTeX comments, where "twenty-two pages" appears.
3. FAILS — reported above, in detail, without the section text.
4. PASSES. 13 of 13 applied; count matches.

🤖 Generated with Claude Code

[lvelho]

Part B is now in — inserted verbatim as §6.3, between the devil's-advocate subsection and "And now the one that matters". PR now six commits; CI green on both jobs.

The flagged stop condition did not fire — but it would have

The specification warned that the section asserts no specification text survives anywhere in bio-3d-vision at 5f8e39f, and said to stop if that failed. It would have failed — three fragments survive, quoted inside artifacts that were being committed anyway.

The supplied text does not say that. It says the specification "is the only artifact in the loop that is never committed" — the true and weaker claim, which re-checks clean: zero tracked files are specifications, and no filename matches spec, task, brief, instruction, relay or prompt. The distinction is load-bearing and the section gets it right. Recorded as A3 so nobody re-derives the stronger version later.

Its five checkable claims were verified before insertion

All CONFIRMED, recorded as A1–A5 in claims-verified.md. New prose making factual claims doesn't get an exemption for being new. mn-004's tally is unchanged — these are numbered separately.

A5 disagrees with my own ledger, and both are right

The section says "five of the six unsupported claims"; gap-004 said four. Different selectors:

• Four — unrecordable by construction, the workflow guarantees no artifact (17 and 20 live in specifications; 12 is a rate over conversational turns; 25 is a cleared session's search).
• Five — "describes an event the workflow guarantees will leave no trace", the section's own selector, which also admits 45: the afternoons a foreclosure cost were events, and nothing logs duration.

Recorded as two counts with their selectors rather than resolved by picking one — which is exactly what mn-001/mn-002 required, for the same reason.

Claim 38 is the odd one out under either selector: not an event, but a proportion of a document that still exists. So it is unsupported and measurable — the only one of the six a successor could close by going to look.

Unchanged

22 pages, zero overfull/underfull boxes. Transit re-verified lossless: 12,184 words in, 12,184 out, zero differing blocks. The section lands as §6.3 and does not disturb §6's opening claim that "the last one is the serious one" — And now the one that matters is still last.

[lvelho]

Revised Part B is in — replaced in draft.md and regenerated into 06-where-it-failed.tex, so both stay committed and the translation stays diffable. 23 pages, zero overfull/underfull boxes, CI green on both jobs.

Falsifiers

1. PASSES. Clean git clone → make -C report → 23 pages, exit 0, no manual install. CI compiles it.
2. PASSES. No figure stated two ways. "Thirteen" carries three referents (experiments, foreclosures, applied corrections), "eighteen" two (claims, plan steps), "three" two (failures that arrived late, rejections) — all genuinely those values, all contextually distinguished.
3. The rewritten claim holds. No specification was ever committed re-checks clean: zero tracked files are specifications, no filename matches spec/task/brief/instruction/relay/prompt. A6 and A7 do not hold — below.
4. PASSES. 13 of 13 applied earlier; mn-004's tally unchanged, and the new section's verdicts are numbered A1–A7 separately.

Two claims in the new text did not check out

A6 — "Three passages quote specification text" → CORRECTED to four, and one of the three named isn't one of them.

passage	reports spec content?
exp007/preregistration.md:38	yes — quotes "six arms, one stimulus"
exp009/preregistration.md:19	yes — quotes, then argues it's wrong
exp007/findings.md:123	yes — and it's the best example, omitted. "The specification said 18 steps for A, A′, D and E", under a heading reading "A deviation from the specification, and why"
exp007/findings.md:32	yes
exp007/findings.md:46	no — records what the specification did not state. An absence, not a quotation.

A7 — "Every one is text that an artifact was arguing with" → UNDER-QUALIFIED. exp007/findings.md:32 agrees: "The specification asked for this to be visible rather than inferred, and it is."

The conclusion is unaffected, which is why this is cheap. The load-bearing sentence is the next one — the survival mechanism is parasitic on the work happening — and it holds for all four. Agreement or disagreement, the fragment exists because the specification was executed. "Arguing with" is too narrow a mechanism for a conclusion that only needs "executed". Both replacements proposed; prose unedited.

A6 is my error propagating

My earlier search returned four hits and my report said three. findings.md:123 was in the output and never opened, and your text was written from that summary.

spec-defects.md §6 records both instances — Chat's grep for REPOSITORY STATE ASSUMED, which could only ever have found a committed specification and never a quoted one, and mine, better-aimed and still read from the expectation rather than the output. Two stations, same error, three iterations apart.

And the od-003 check had the same defect

It was case-sensitive — saw "three passages", missed "Three passages". Measured: 252 hits against 301, missing 49 (16%), and the missed class was every sentence-initial figure. Fixed with -i plus tolower in the grouping (without the second, "Three" and "three" count as two figures). It had been run three times and reported as passing. Found by hand, not by the check — which is §6's defect inside the tool built to catch §6's defect.

[lvelho]

Option 2 done — converter committed, tooling back with it. od-007 closed. All three CI jobs green on de575e4; 23 pages, zero boxes; clean checkout still builds with TeX alone.

What landed

tools/md2tex.py generates report/sections/*.tex from report/draft.md. It lived outside the repo, which made main.tex's claim that the .tex is regenerable true and unverifiable at the same time.

fc-002 and fc-003 are annotated, not reversed, per CLAUDE.md. What came back is what fc-002's scope said would: ruff, ruff format, mypy, pytest, as a code job. What did not: no [build-system], no package, no dependencies, and src//experiments/ stay closed.

fc-003's order was followed. Its scope says restore the pinned-clone guard first, and it is the first test for that reason. It fired immediately — on the docstrings of the two files that restored it, which named the directory in prose. I rewrote the prose and left the guard crude. Loosening a check to admit its first true positive is how a check stops being one.

The second guard is what earns the reopening. test_typeset_sections_match_the_draft regenerates into a temp dir and compares bytes, so main.tex's standing claim can now fail. Both guards were watched failing before commit, and the sync test writes only to a temp dir — a failing test can never repair what it checks.

Two findings worth your attention

fc-002's scope named the wrong trigger. It said the entry stops holding "the moment one generated figure is wanted". No figure was wanted; a converter arrived. fc-003 said "the moment any .py file lands" and covered it exactly. fc-002 described the use it was picturing; fc-003 described the condition. A scope clause that names an instance is silent on the case that actually occurs — and looks wrong when it isn't.

The gates failed on their first CI run, and it was od-004's failure again. pip install ".[dev]" asks pip to build the directory as a package; there's no build backend by design, so setuptools rejected pyproject.toml. It passed locally because the install step never ran there — the tools were already present. Same shape as the Psych 3/Psych 4 split three commits earlier: local success meant nothing.

Fixed with PEP 735 [dependency-groups] and pip install --group dev — not by adding a version and a backend, which would manufacture the package the file exists to avoid. Then I re-ran all four gates against the versions CI actually installs (ruff 0.16.6, mypy 2.3.1, pytest 9.1.1 — all newer than local). Checking locally again would have made the same mistake a third time.

Costs paid

The script was scratch-quality (semicolons, % formatting) and had to be rewritten to pass the gates it arrived under — output verified byte-identical before and after, and again after ruff format.

Five statements elsewhere became false and are corrected in the same commit: CLAUDE.md's Commands section and definition-of-done clause 2, .gitignore's note that the guard was gone, README's "no Python", and report/figures/README.md's restore instructions. A decision that invalidates five statements is one whose cost is mostly in the prose.

[lvelho]

A6 and A7 closed. All three CI jobs green on d8c7052; 23 pages, zero boxes, transit lossless at 12,335 words.

claim	was	now
A6	"Three passages … quote specification text", listing one that records an absence	Four passages, list rebuilt. exp007/findings.md:46 dropped; :123 and :32 added. All four re-read at 5f8e39f and each characterisation checked against its text before the sentence was written.
A7	"Every one is text that an artifact was arguing with"	"…had to reckon with — mostly to depart from it, once to confirm it."
(no verdict)	"reaches the record only where … had to disagree with it"	"had to account for it — usually to justify departing from it, once to note that it was met."

The third row is the finding

A7's verdict named one sentence; the defect was in two — sixty words apart, in the same subsection. The paragraph after A7 restated the same overreach in different words.

The mechanical repeated-figure check could not have caught it. It repeats an idea, not a number, and every word in the restatement differs. The only thing that finds this class is applying a correction and then re-reading the paragraph around it for the same claim in other words. Recorded in claims-verified.md as a practice: when a characterisation is corrected, read forward for its restatements — the same move as grepping a corrected figure for its other occurrences, but nothing automates it.

Your conclusion is untouched. The load-bearing condition was always that the specification had to be executed and produce an artifact, and that holds for all four passages; "arguing with" was a narrower mechanism than the conclusion needed.

One slip of mine, also fixed

The draft status block and main.tex still read "its five checkable claims … all confirmed, recorded as A1–A5". That was written for the first version of the section and never updated when the revision took it to seven claims with two non-confirmed. Both now say A1–A7 — five confirmed, one corrected, one under-qualified, both since closed.

gap-004 already carried the corrected count of four, so ledger and prose now agree. Before this change they didn't, and the ledger was right — which is the third time in this PR that the record caught the prose rather than the other way round.

Where the report stands

Every CORRECTED and UNDER-QUALIFIED claim in the document is now closed — 13 from the main pass, 2 from the addendum. The six UNSUPPORTED claims remain, unchanged and flagged in the prose itself, because no replacement can be written from a record that does not contain them. That is gap-004 and od-006, and it is the subject of §6.3.