2026-09-26 m3 spectra aasist (audit)

The library

Rendered from docs/specs/2026-09-26_m3-spectra-aasist-audit.md; the markdown file is the record.

Dimension	Grade	Notes
Plan adherence	Acceptable	Amended implementation and artifacts match; execution order deviated slightly.
Scope discipline	Excellent	M3 commits stay within the declared file list.
Test coverage	Acceptable	Spec records passing tests; independent pytest verification blocked by sandbox restrictions.
Review compliance	Excellent	Prior findings addressed, including exact model-output row validation.
Freeze integrity	Acceptable	Skipped: no P1/P2/P3 freeze hashes present.
Regression check	Acceptable	Ruff passes; artifacts verified; full test rerun unavailable. Direction script unchanged.
Documentation	Excellent	Report, STATUS and disclosure cover the shipped behavior and limitations.
Overall	Acceptable

Plan adherence — downgrade to Acceptable. The full scoring run started at 05:22, before In-the-Wild finished at 05:28 and its numbers were handed off at 05:29. This differs from the requested stress-readout-before-full-run sequence. The stress results arrived before publication, and the final metadata contains the completed readout.
Test coverage / Regression check — downgrade to Acceptable. uv run pytest failed during cache initialization; direct virtualenv pytest failed during collection because no writable temporary directory was available. These are environment failures, not demonstrated regressions. Recorded test results could not be independently confirmed. Ruff passed; independent checks confirmed export integrity, all crop draws, direction-baseline membership and AUC, and the reported holdout metrics.
Plan adherence — no additional downgrade; pre-existing limitation. The supplied fold file separates bona fide groups, but LJ Speech is grouped by chapter rather than speaker. Its single speaker consequently appears in inner and holdout rows. This is explicitly documented in the existing fold builder and was not introduced by M3.