2026-09-26 post draft gate sweep (audit)

The library

Rendered from docs/specs/2026-09-26_post-draft-gate-sweep-audit.md; the markdown file is the record.

Dimension	Grade	Notes
Plan adherence	Excellent	Implements the governing manifest and amendments; final report records KEEP.
Scope discipline	Excellent	No detector-contract, scoring-path, or shipped-artifact changes.
Test coverage	Acceptable	Independently ran 92 tests successfully; 1 skipped, 15 filesystem-dependent cases deselected. Ruff passes.
Review compliance	Acceptable	Previous fixes are present; model/export identity remains an explicitly accepted limitation.
Freeze integrity	Acceptable	No P1/P2/P3 hashes present; original manifest text remains intact, with appended amendments.
Regression check	Acceptable	Data self-checks pass. Both TSV hashes and fusion_v2 constants match recorded values; no fusion_v3 exists. Full-suite rerun unavailable.
Documentation	Acceptable	Final outcome and WavLM disclosure updated; minor historical wording remains.
Overall	Acceptable

Test coverage / Regression check — downgrade to Acceptable. The read-only sandbox blocks uv cache writes and filesystem test fixtures. Independent verification produced 92 passed, 1 skipped, 15 deselected, including the available data self-checks. The recorded 107 passed, 1 skipped targeted run and earlier 722 passed full-suite run were not independently reproduced in full.
Review compliance — downgrade to Acceptable. The supplied model directory is not cryptographically tied to the evaluated export; precheck_write checks for meta.json, whose hash is recorded only when writing. This is the explicitly accepted C6 residual risk. No candidate passes, so it does not affect this KEEP decision.
Documentation — downgrade to Acceptable, cosmetic only. The script’s opening describes H_noise as “withdrawn,” and the sweep report’s historical “Decision rule applied” paragraph still says “unless H_noise passes at ~20:30.” The opening and final Run 2 section clearly establish that H_noise failed. Update or mark those remaining sentences as historical.
Regression check / Documentation — no additional downgrade. The conditional KEEP log row is absent, but this audit occurred before the specified 19:30 ruling window. Append it after Nathan’s ruling or the applicable cutoff. The existing final TSV remains byte-identical to the shipped file.