Review-and-fix of #201 (Codex's PR-A1 implementation of the accuracy program). The structure was right, but neither committed gate had ever been run: asa verify's backend leg was red on main (3 failing default-manifest checks) and the committed synthetic manifest failed 52 of 140 checks when executed. CI never caught it because the workflow only runs unittest discover.

What was broken (all verified by running the gates)

• drum_stem_known_counts: snare detector counted the 16 kicks (kick click/pitch-sweep energy lands in the 120–2000 Hz band, and Codex placed snares on kick beats so coincident onsets merge); hat detector found 4–8 of 32 hats (below its floor). Counts read 15 and 8 vs truths 8 and 32.
• four_on_floor_clear_128: BPM measured exactly right (128.0, Percival 128.4, agreement) yet failed quality:tempo — 4-bar clips cap the extractor's confidence below the 0.7 bar.
• Synthetic manifest: mono_bass_transcription missing --transcribe (noteF1 = 0), chord fixtures spelled sharps against the Viterbi's flat vocab (F# major scored 0.25), key comparison lacked enharmonic folding (Db minor ≠ C# Minor), BPM checked on transient-free chord clips (read 61–98 vs 96–140 truths), spec feat(ui): re-analyze CTA for legacy Session Musician runs #29 silently dropped by a [:28] slice, ground-truth-for-later (hitTimes, swingPercent) embedded inside expected.

Fixes

Generator — engineered band-disjoint one-shots (kick = 50 Hz sine with a load-bearing 20 ms fade-in and 0.20 s length; snare/hat = double-filtered noise bursts); grid clips are a bare accented-kick pulse (off-beat hats bin ambiguously in analyze_time_signature's ±half-beat counting and flipped 4/4 reads to 3/4 or 6/8); swing clips keep swung hats but check BPM only; chord labels computed from pitch classes, flat-spelled; per-kind expected blocks with everything else under truth; --transcribe on the bass fixture; all 29 specs render; order-independent determinism.

Harness — enharmonic pitch-class folding for key + chord label comparison; knownGaps: per-track check names that still run and report scores but don't gate (checksInformational in the summary), keeping measured baseline weaknesses visible as improvement targets.

Trust layer — bpmAgreement=True with confidence ≥ 0.5 is now authoritative: a BPM measured exactly right and cross-confirmed by an independent estimator was being hedged as "cross-check is not strong enough", which was factually wrong. Covered by new tests; golden unaffected.

CI — the default-manifest fundamentals gate (render + evaluate_fundamentals.py --fail-on-skip) now runs in the backend job, closing the hole that let #201 merge with a red gate.

Results

• Default manifest (the asa verify / CI path): 15/15 checks green.
• Synthetic manifest: 29 clips, 54/54 active checks green, 0 failed, 8 informational knownGaps (odd-meter detection + downstream downbeats, tempo octave at 174 BPM, 7/8 tempo wobble) — quantified in audits/accuracy-baseline-2026-07-03.md, the accuracy program's "before" column.
• Notable baseline finding: the beat grid is strong everywhere (F1 0.95–0.98 even in 3/4, 6/8, 7/8); meter → downbeats is the weak layer — exactly the case for running the pre-registered beat_this gate next.
• Backend suite: 1283 tests OK. Both gate commands exit 0 and reproduce deterministically.

🤖 Generated with Claude Code

https://claude.ai/code/session_01T4wfz87k6kzJqkKLKZE7YL

Generated by Claude Code