`.abcd/development/release-gate/manifest.json` describes itself as "the

.abcd/development/release-gate/manifest.json describes itself as "the reproducibility anchor (iss-122): two honest runs of the same tier mean the same thing because the doc list, the directions, the checker count, and the prompt are fixed here rather than chosen per run".

Three full-tier runs on 2026-08-23, same manifest, same 28 checkers:

Runs 2 and 3 are a controlled measurement. The brief is byte-identical between those two commits — git diff a2c77e5 6e4d537 -- .abcd/development/brief/ is empty — so the detector's entire subject was held constant. It returned 21 more findings, a 17% swing, on the same bytes. Per-file the swings are larger than the total: 05-prompt-quality.md 0 to 5, 02-verification-matrix.md 1 to 7, 06-capture.md 4 to 9, and in the other direction 11-history.md 3 to 0 and 18-ideate.md 4 to 1.

What IS stable is roughly which chapters are implicated: 28 files appear in both runs 2 and 3, with 4 unique to run 2 and 10 to run 3. So the detector broadly agrees on WHICH chapters have drifted and disagrees on HOW MANY discrepancies each holds and WHAT CLASS each is — and even the file set is only mostly stable, not fixed.

The manifest is not wrong to exist and it is not lying about what it pins. It pins the inputs, and it does: scope, depth, doc list, prompt. What it cannot pin is the output, because the checkers are LLM agents. The defect is the sentence claiming the consequence — "two honest runs of the same tier mean the same thing" — which is a statement about outputs that the mechanism cannot deliver.

The consequence is practical, not theoretical:

1. A receipt's failing count is not a measurement. Comparing 37 (v0.6.1) to 147 cannot support "drift got worse", and a later run showing 90 would not evidence improvement. The 0.6.2 receipt records 147 and says so in its own provenance, because that number will otherwise sit in the ledger beside v0.6.1's 37 and invite exactly this comparison.
2. The class distribution cannot drive scope decisions. An intent that scoped itself to "the 58 false-claims" would have scoped itself to 48 a few hours later.
3. The stable half is still useful, and is the half to build on: which chapters drift is reproducible, and that is what locates a generation seam.

Fix is a record correction, not a mechanism change: the manifest's comment should claim what it delivers — comparable SCOPE, not comparable findings — and the runbook should say a receipt's finding count is an observation from one run rather than a metric. Refines iss-2608231346137587, whose corpus this governs; that issue's own class counts carry the same caveat.

This is the enforcement-claims-are-facts class again (iss-2608230847432286): a record stating an assurance the machinery underneath it does not provide. It was found only because a release recovery happened to run the same gate twice in one afternoon, which is not something the process would otherwise ever do.

Grounds

• pursued: no record states that two runs of one tier yield comparable findings or that a receipt's count measures drift; a surviving sentence saying runs 'mean the same thing', or a receipt comparison presented as a trend, would show it wrong