First claims-verification pass over draft.md §§1-8.5

[lvelho]

Two stale items fixed in their own commit (9b58d1a), then the pass (c1f...).

42 claims — 22 confirmed, 5 corrected, 8 under-qualified, 6 unsupported, 1 partial (mn-004, the only measured entry this repository will ever have).

Resolved against .reference/ clones: bio-3d-vision @ 5f8e39f and math-ai-method @ 81a638e, both pinned and verified by rev-parse; active-stereo and bioeye at default branches — unpinned, gap-003.

The two stale items

• od-005 — §8.6 is written, and so are 8.7 and 8.8. The entry tracked one named artifact and three arrived.
• od-002 — resolved at source, and both ledger entries were right. The instance's first pass tallies 22 claims and then says "Of the 18 substantive claims, 7 stand as written." Claims 19–22 are bibliographic citations under a ## Citations heading: 22 − 4 = 18, and striking the three confirmed citations from the CONFIRMED list of ten leaves exactly seven. The defect was the draft's sentence, which named neither the pass nor the subset.
• The spec predicted mn-002's 21-of-22 arithmetic gap would survive. It does not — the template's prose enumerates four verdict classes of five; the missing PARTIAL is claim 19, Bajcsy 1988, whose page range does not resolve. mn-003.

The recurring defect is not the instance's

The instance's pass reported scope-and-metric. This document is second-order — it restates figures already condensed once — and its characteristic error is transcription:

the record says	the draft says	claim
"twice in seven iterations"	"shaped every specification for a week"	24
ledger marked inherited on day one, 42 entries	"applied to numbers on day two"	22
commit: "close four cold-start gaps"	"four remain open"	41
exp012 is the criterion re-analysis	"four of thirteen were about the stimulus"	29

The number survives; the noun attached to it does not.

Falsifier 2 fired — and neither offered branch is the reason

8 : 5, against the instance's 8 : 3. But under-qualified is not depressed (8 in both) — corrected is elevated, which is the second-order structure above. Testable prediction, recorded before anyone tests it: a third-order document should show more corrections still.

Falsifier 3 fired hardest, and the six are not scattered

Every unsupported claim describes an event the workflow guarantees leaves no artifact. The sharpest is claim 17, "three specifications were rejected on false assumptions" — which §4.2 calls "the clearest case in the project of a rule paying for itself". docs/workflow.md states that a specification is the only artifact of the loop that leaves no trace, and that none was ever committed. One rejection survives (fc-012), only because it reshaped a foreclosure taken anyway, and it was not a stale-state rejection at all.

The method records what work produced and has no channel for work it correctly prevented — so its strongest evidence is the kind it systematically destroys. gap-004, od-006. It cannot be backfilled.

od-003 fixed and validated

The shipped grep returned one line, 52 1023:7: uniq -c -f2 counts whitespace fields after sort -t: ordered colon-separated ones. It also matches only numerals while this document spells its figures out. The replacement was validated against a known answer before being trusted and earned its place on the first run — it found a third location of the eighteen-claims figure (draft.md:391) that the two sanctioned amendments left inconsistent, and found that "thirteen" carries three referents and "eighteen" two.

What needs the maintainer

No prose was corrected — every non-confirmed claim carries a proposed replacement sentence. One item is urgent: draft.md:391 still reads "Of eighteen claims in the first draft, seven stood as written" while :870 and :1219 were amended, so the document now states the figure two ways. That is the repeated-claim defect, live, in the document that ships the check.

🤖 Generated with Claude Code