Sixteen seams where the page said one thing in prose and another in a list, a cell or a link

[wshallwshall]

The rest of the coherence review, after the five defects that made an implementer build the wrong artifact landed in 5e2ef34 (#33). Thirty edits. No argument on the page is reversed.

What was wrong, by class:

• Duplication. Four passages told twice, three near-verbatim across the Part 1 boundary — including the page's only first-person narration, and a cost argument whose two copies had already drifted apart in wording.
• Reference. "This repository" named three different trees. Two of the paths it points at do not exist here, and an agent can act on that by searching its own tree for them.
• Navigation. The pre-flight reading list named two sections and summarised three items, the third delivered by neither. Part 2's manifest accounted for three of the six forcing functions it delivers. The paste command said "begin at step 1" where the agent's slice supplies two different step ones.
• Unproved claims. Three absolutes the page's own checklist requires be proved or weakened, two competing claims of page-level primacy, and one passage saying no body assesses you against ASVS while another describes the organizations that do.
• Vocabulary. "Verified" carried two senses producing different numbers from one instruction. A requirement ID was called a control ID — in the section arguing a requirement is not a mechanism. "Cell", the page's most-used term at 125 occurrences, was never defined.
• The three residues from Five rules disagreed with the page that states them, and one of them could never fire #33 are closed: the rule field's domain, the line field's advisory read, and review row 5 checking all three parts of an absence claim.

How this was produced. Six batched drafters, an adversarial checker on each, then a reconciler doing what no batch could — cross-batch collision detection and ratchet arithmetic. Six edits were amended before landing: two edits to the same paragraph were merged, one coined a fifth name for the decision procedure, one removed a job name at one end of the page while another reintroduced the same ambiguity 1,400 lines earlier, and one wrote a fresh unproven completeness claim into the page whose failure table ranks that defect first. None of those was visible from inside a single batch.

What was deliberately not done. verified_at is not renamed — the name is published and entered the checklist two commits ago, so the correction is made at the definition site instead. The Part 1 / Part 2 label mismatch stays, because dropping that vocabulary costs a second file: HOUSE-STYLE.md pins that section as the only marker that Part 1 ends. Four further items are recorded and left open rather than half-closed.

Measured. Long sentences 219 → 217 against a baseline of 219, inside the slack band, so the baseline does not move. Fat cells 30 → 30, at baseline with zero slack. Words 61,422 → 61,484. Zero doubled blank lines, zero non-ASCII bytes, every internal anchor still resolving. 146 tests green, Word copy rebuilt with pandoc 3.10.

🤖 Generated with Claude Code