WS3a: asa-dsp ↔ Essentia loudness parity harness (checkpoint)

[slittycode]

What this is

The WS3a parity checkpoint for wiring packages/loudness-spectro-wasm (asa-dsp, openmeters-derived BS.1770) into the product. Before asa-dsp can back any ASA loudness path (WS3b backend, WS3c browser readout), it must agree with Essentia — ASA's authoritative Phase 1 loudness source, since Phase 1 measurements are ground truth (PURPOSE.md invariant #1).

scripts/essentia_parity.py runs a deterministic synthetic corpus (tones, 20 Hz→20 kHz sweep, white/pink noise, decorrelated stereo) through both paths at each file's native rate (no resampling — that alone breaks ±0.1 LU EBU conformance):

• asa-dsp via the native measure-cli binary (source-identical to the WASM core, so WS3a needs no wasm toolchain — per the plan).
• Essentia via AudioLoader → LoudnessEBUR128, mirroring apps/backend/analyze_core.analyze_loudness, plus a dBTP true peak from TruePeakDetector.

Primary gate: integrated LUFS within ±0.1 LU on every file.

Result — borderline (see docs/essentia-parity-report.md)

Realistic signals (tones, noise, stereo)	7/8 within ±0.02 LU (most ±0.002) — essentially perfect BS.1770 parity
Sole breach	log_sweep_20_20k at −0.102 LU — 0.002 over the gate
Worst delta	0.102 LU

The one miss is a full-range log sweep — a pathological, non-program stress signal sweeping into the near-Nyquist region, not music. The verdict in the committed report is reported exactly as the harness computed it (FAIL by that single borderline case); the substantive finding is that the cores agree on every realistic signal.

Decision pending (the hard checkpoint)

Per the plan, WS3a is a STOP point. The question for the owner: treat the sweep breach as (a) a pathological-edge artifact and clear WS3b/WS3c to proceed behind a default-off flag (loudness default never flips to asa-dsp until real-program parity is also proven), or (b) a strict hard stop. Awaiting direction before stacking WS3b/WS3c.

Notes

• Independent of fix(loudness): Phase 1 v2 — truePeak→dBTP, bpmConfidence→0–1, coherent PLR #128 (the Phase 1 v2 loudness migration); forked from main.
• Harness + report only — no product code touched. asa-dsp remains unwired.
• The corpus is generated, not committed (broadband noise stresses K-weighting most); the report is reproducible from the seed.

https://claude.ai/code/session_01Wq1vE1D7o3W9xZKSHXz43Q

---

Generated by Claude Code

[slittycode]

Verdict: COMMENT

Summary

Harness + docs only — packages/loudness-spectro-wasm/scripts/essentia_parity.py, the committed checkpoint report, and a README addition. No product code touched; asa-dsp remains unwired. Phase 1 contract unchanged, no phase boundary violations. One genuine code oddity in run() and one secondary-metric gap worth flagging before WS3c. CI was still in progress at review time.

Findings

Worth considering — run() calls build_report() and discards the result (scripts/essentia_parity.py:289–290)

def run(...):
    ...
    _text, passed = build_report(rows, tolerance)  # result thrown away
    return rows, passed

Then in main() (~line 310):

rows, passed = run(corpus, args.tolerance)
...
text, passed = build_report(rows, args.tolerance)  # called again, shadows passed

build_report() runs twice and passed from run() is immediately overwritten. Not a bug — the function is pure so both calls agree — but it's confusing and the _text discard is a hint that run() shouldn't own this call at all. Simplest fix: drop the build_report() call from run() and have it return only rows.

Worth considering — white_noise true-peak gap (+0.59 dBTP)

From the committed report secondary table: Essentia −6.61 dBTP vs asa-dsp −6.02 dBTP on white_noise. The caveats section correctly flags true peak as "a sanity signal, not a gate," but if WS3c lands a browser true-peak readout the 0.59 dBTP gap on broadband content needs a gate or at minimum an explicit "known divergence, do not expose" note before that work proceeds. The gap is benign for WS3a but should be addressed in the WS3b/WS3c scoping.

Test results

CI was in progress (Backend, Frontend, Loudness WASM all in_progress) at review time. All changed files are under packages/loudness-spectro-wasm/, which is standalone and unwired. No unit tests for the harness — consistent with the established pattern for every evaluate_*.py script under apps/backend/scripts/; the harness is itself the measurement gate.

Phase boundary check

Clean. No Phase 1 output types modified. No apps/backend/ or apps/ui/ files touched. The reported FAIL verdict is an honest, measurement-cited finding, not an override of Phase 1 data.

---

Generated by Claude Code