2026-09-26 m4 fusion v2 m5 scorer (audit)

The library

Rendered from docs/specs/2026-09-26_m4-fusion-v2-m5-scorer-audit.md; the markdown file is the record.

Dimension	Grade	Notes
Plan adherence	Fail	A5 forbids copying/logging the PARITY artifact when acceptance fails; both remain.
Scope discipline	Excellent	Capability added without switching the default or changing the detector contract.
Test coverage	Fail	A5’s full-set M5-logit tolerance is exceeded on five rows.
Review compliance	Acceptable	Plan-review fixes implemented; prior audit blockers remain unresolved below.
Freeze integrity	Acceptable	Skipped: no P1/P2/P3 freeze hashes found.
Regression check	Acceptable	Final-code suite reports 510 passed; six focused tests independently passed. Draft TSV hash matches.
Documentation	Excellent	Scoped docs describe the implementation, unchanged default, and outstanding deviation. M5 is already disclosed.
Overall	Fail	A5 remains unmet and unwaived.

Test coverage — downgrade to Fail. Independently recomputing M5 live-versus-export differences across 1,671 cached rows gives maximum 0.0350065, with five rows exceeding 0.02. Probability parity and zero flips do not satisfy the separate logit gate. Meet the existing gate, or obtain Nathan’s explicit acceptance of a documented amendment; the proposed replacement thresholds are not yet approved.
Plan adherence — downgrade to Fail. A5 explicitly requires no PARITY copy or log row upon failure. The artifact and log row remain. Relabelling them “validation evidence” does not resolve that requirement. Pending an authorized amendment, retain evidence under outputs/, remove the submission-directory copy, and withdraw its log entry with an audit trail.
Regression check / Test coverage — no additional downgrade. Independent verification passed six focused tests covering constants identity, both candidate TSV polarities, v1 arithmetic, strict E boundaries, and error thresholds. The full-suite result of 510 passed is recorded evidence, not independently rerun here. Repository lint still reports three issues solely in the unrelated src/hearsay/analyzer.py; these are not attributed to this change.