2026-09-26 post draft heads (audit)

The library

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

Dimension	Grade	Notes
Plan adherence	Acceptable	Outputs and diagnostics match; calibration deviation explicitly reconciled with the unchanged M1b recipe.
Scope discipline	Excellent	Changes confined to the experiment, repair tooling, tests, and associated documentation.
Test coverage	Acceptable	Recorded suite: 775 passed, 1 skipped. Independent full rerun blocked by read-only environment.
Review compliance	Excellent	Prior findings addressed, including both architecture calibration statements.
Freeze integrity	Acceptable	Skipped: no P1/P2/P3 hashes supplied.
Regression check	Acceptable	No regression identified; diagnostics reproduce, experiment logged, shipped TSV hash unchanged.
Documentation	Excellent	Report, disclosure, and calibration corrections describe the completed change.
Overall	Acceptable

Plan adherence — Acceptable downgrade. Platt calibration uses pooled inner-OOF labels rather than being fold-local as originally specified. The post-run correction explains this inherited M1b behavior. The saved positive slope preserves ranking metrics; outer-holdout, stress, and test labels remain excluded from fitting.
Test coverage / Regression check — Acceptable verification limitation. The recorded full-suite result could not be independently repeated: uv cache access and pytest temporary/cache writes were blocked by the read-only sandbox, with four collection errors. Using Python 3.12.13 directly, the documentation tests passed 52/52, and ruff check . --no-cache passed. Independent checks reproduced the reported diagnostics, export counts and path sets, finite scores, and identical shipped/final TSV hashes. These environment errors are not evidence of a change-induced regression.