A full-tier `iss35-brief-surface-crosscheck` at content commit `7a4ee00`

A full-tier iss35-brief-surface-crosscheck at content commit 7a4ee00 returned 125 discrepancies: 58 false-claim, 38 undocumented-surface, 15 stale-count, 7 criterion-violation, 7 fictional-layout. Exactly one points at a file the release bundle carries (commands/memory.md, fixed in this cut); the other 124 are .abcd/development/brief/, which ships in no release (adr-28), so no installer is misled. Contributors reading the record are.

The class counts above are one run's observation, not a measurement. A second full-tier run hours later, on a tree differing only by seven shipped-doc fixes, returned 126 findings with false-claim at 48 and undocumented-surface at 50 — the same chapters, redistributed. See iss-2608231409595789. What is reproducible is WHICH chapters drift (29 of 33 files common to both runs); the counts and classes are not. Scope the intent on the stable half.

The corpus is preserved at .abcd/.work.local/scratch/brief-drift-2026-08-23/crosscheck-findings.json with the run's content commit, manifest hash and tier. It earns promotion to a dated research note when the intent below is filed. Do not fix the 124 by hand before mining it — the whole value of this run is that it is evidence.

What the corpus already says

124 of 125 findings resolve to a blaming commit via git blame on their file:line. The dates are the finding: 66 of them were last written in 2026-07 and 58 in 2026-08 — evenly spread, in every class. So this is not one bad merge that outran its record; it is a steady rate. Roughly half the false claims were written in July and survived four releases — v0.4.2, v0.5.1, v0.6.0, v0.6.1 — each of which recorded a PROMOTE receipt against this same detector, at 26, 28 and 37 findings respectively.

That is the real defect, and it is the proxy-gate class again (iss-2608230847432286): the gate ran, reported, went green, and the thing it measures got steadily worse. receipt_gate never coupled the verdict to the finding count, and surface_coverage — the deterministic half — checks that a registered sub-verb has a ROW in the chapter's table, never that the prose beside the row is true. It blocked a commit today for a missing history drain row while 124 false prose claims sat green beside it.

Diagnosis

The brief mixes two kinds of content with two different failure modes in the same paragraphs:

• Shape — flags, sub-verbs, exit codes, schema fields, counts, file layouts. Derivable from the command tree, therefore checkable, therefore should never be hand-written. false-claim and stale-count are shape by definition, and most undocumented-surface is the binary growing something the chapter's shape section never learned.
• Intent — why a surface exists, what it refuses to do, which trade was made. Not derivable, and cannot drift against code, because it is not a claim about code shape.

The comparison that makes it concrete, from the same repo on the same day: docs/reference/cli/commands.md is generated and drift-tested, and the same reviewers found ZERO discrepancies across its 775 lines. The brief is hand-written prose and returned 124. The brief holds a second, hand-maintained copy of a fact the command tree already owns — one-canonical-primitive violated at two copies rather than three.

Direction, not yet decided

Generate or deterministically check the shape; let the prose carry only the why. A chapter becomes a generated shape block plus hand-written rationale, and the shape then cannot drift.

Explicitly rejected: a "you changed a surface, so touch its brief chapter" gate. It forces an edit without forcing correctness, which is a phantom gate — the exact class this issue is an instance of.

Routing

This warrants an intent, and the maintainer has said so. The four-piece split below was hand-run and confirmed by the maintainer on 2026-08-23, with the intent to be filed after the 0.6.2 cut — so the next session files from this routing rather than re-running the interview:

• capability -> intent: the brief's surface chapters become a generated reflection of the shipped surface;
• trust rule -> ADR + brief invariant: shape claims are derived, never hand-authored;
• stance -> principle: none new, one-canonical-primitive already says it;
• plumbing -> brief: the generation mechanism and where the seam sits.

Verdict: SPLIT, confirmed. Open questions the corpus should answer first: did the surface change and the brief change land in the same commit (the hypothesis that locates the seam); which chapters drift fastest normalised by surface churn; and how many releases each false claim survived.

Grounds

• pursued: a shape claim nobody writes by hand cannot drift; a false-claim or stale-count finding about a flag or sub-verb a generated appendix covers, in the next full-tier crosscheck, would show it wrong