feat(genre): measured classifier + mix-doctor widening for the missing families (PR-G7) by slittycode · Pull Request #216 · slittycode/ableton-sonic-analyzer

slittycode · 2026-07-13T10:31:16Z

What

PR-G7, the final piece of the genre-generalization program: the classifier and mix-doctor learn the missing electronic families, with every measured number sourced from real reference audio (93 GiantSteps clips run through the full Phase 1 pipeline — method, caveats, and provenance in audits/genre-target-sourcing-2026-07-13.md, committed here).

Changes

Nine new _GENRE_SIGNATURES — measured: gabber, hardstyle, electro, downtempo (GiantSteps hardcore n=17, hard-dance n=11, electro-house n=15, chill-out n=15); knowledge-based and flagged in-line: trap, footwork, two-step, speed-garage, halftime-dnb (2015 Beatport has no reference audio for these; same provenance as the original table). two-step's low sidechain window (0.1–0.35) is the audit's uk-garage fix.
Three measured retunes — breaks, dubstep, psytrance: all three controls' old windows NEVER fired on real audio (real breaks read acid-techno 8/15; measured dubstep bass decay is 0.11–0.40 s vs the assumed 0.6–1.2 s wobble; measured psy centroid 860–1910 vs the 2200–5500 window).
Octave-aware bpm axis in analyze_genre_detail: when PR-G3's bpmOctaveEvidence contests the shipped bpm, each signature's bpm window scores against both hypotheses (alternate discounted 10%). Forced by measurement: every hardcore clip ships a halved bpm (84–128 for true 156–190) — a gabber signature can never fire on the shipped value alone. Never touches bpm itself (invariant feat: value-first backport — genre profiles, MixDoctor, broken-pipe fixes #1).
Family plumbing on all four cross-boundary sites (tripwire Preserve signal monitor playback during analysis and harden run flow #3): _GENRE_FAMILY_MAP gains garage (fixing the audit's unmapped uk-garage/bassline/garage) plus hardcore, trap, electro, downtempo; the TS genreFamily union + validFamilies allowlist mirror it; JSON_SCHEMA.md now pins the closed set.
Four measured mix-doctor profiles (electro, downtempo, hardstyle, gabber) with p25/p75/p50 band targets and family fallbacks in mapLegacyToEnhanced. Knowledge-based genres get no profiles — targets are not invented; trap falls back to its hiphop lineage.

Measured validation — real-clip classifier replay (old table → new)

Beatport label	Old top read	New top read
breaks (15)	acid-techno 8	breaks 7
chill-out (15)	hiphop 5 / acid 4	downtempo 12
dubstep (10)	hiphop 4 / trance 4	dubstep 4
electro-house (15)	acid-techno 8	electro 8
hard-dance (11)	hiphop 3	hardstyle 6 (+gabber 1)
hardcore (17)	hiphop 8	gabber 10
psy-trance (10)	acid 5 / hiphop 4	psytrance 4

Correct-family top-1: ~0/93 → 51/93. Honest limits: LOFI previews under-read the top octave (bright-side windows widened, flagged for re-sourcing), and measured pumpingStrength runs 0.15–0.42 everywhere vs the 0.35–0.75 many ORIGINAL entries assume — full-table recalibration is a named follow-up, not smuggled into this PR.

Validation

Backend: full unittest discover — 1366 tests, OK (3 skipped; exit codes captured directly) (6 new classifier tests incl. gabber-fires-via-octave-evidence and no-op-when-evidence-supports-shipped).
Synthetic eval: 114/0 active on this branch base (pre-G6 38-clip corpus — G7 touches no eval surface; CI validates the merged ref against the 40-clip main) (genre isn't gated there; regression assurance).
Golden: passes without re-baseline.
CI-mirror fundamentals gate: 15/15, allPassed.
UI npm run verify: 51 smoke passed.

Full measured percentiles (93 clips)

93 rows, 0 errors

=== breaks (n=15)
  bpm              p10=  125.10 p25=  128.40 p50=  130.00 p75=  137.05 p90=  140.00 n=15
  subBass          p10=  -14.16 p25=  -11.10 p50=  -10.10 p75=   -8.35 p90=   -7.56 n=15
  crestFactor      p10=    9.86 p25=   10.80 p50=   11.70 p75=   12.95 p90=   14.16 n=15
  onsetRate        p10=    4.27 p25=    5.03 p50=    5.49 p75=    6.02 p90=    6.49 n=15
  spectralCentroid p10= 1303.92 p25= 1411.75 p50= 1572.30 p75= 2069.60 p90= 2226.04 n=15
  pumpingStrength  p10=    0.22 p25=    0.27 p50=    0.29 p75=    0.32 p90=    0.34 n=15
  bassDecayMs      p10=   82.40 p25=  110.50 p50=  206.00 p75=  262.50 p90=  470.40 n=15
  rt60             p10=    1.92 p25=    2.17 p50=    2.59 p75=    2.97 p90=    3.00 n=14
  kickThd          p10=    0.38 p25=    0.45 p50=    0.47 p75=    0.65 p90=    0.82 n=15
  lufsIntegrated   p10=  -11.70 p25=  -10.80 p50=   -9.70 p75=   -8.50 p90=   -7.54 n=15
  plr              p10=    9.06 p25=   10.25 p50=   10.70 p75=   11.60 p90=   12.74 n=15
  lowBass          p10=  -14.14 p25=  -13.40 p50=  -11.50 p75=  -11.05 p90=  -10.14 n=15
  lowMids          p10=  -22.56 p25=  -21.45 p50=  -19.50 p75=  -17.20 p90=  -15.94 n=15
  mids             p10=  -19.84 p25=  -18.75 p50=  -17.60 p75=  -16.30 p90=  -13.32 n=15
  upperMids        p10=  -24.96 p25=  -23.15 p50=  -21.30 p75=  -18.85 p90=  -17.54 n=15
  highs            p10=  -26.88 p25=  -24.80 p50=  -21.90 p75=  -20.35 p90=  -19.56 n=15
  brilliance       p10=  -35.46 p25=  -32.05 p50=  -27.10 p75=  -25.20 p90=  -23.24 n=15
  classifier reads today: [('acid-techno', 8), ('trance', 3), ('rock', 1), ('progressive-house', 1)]

=== chill-out (n=15)
  bpm              p10=   83.34 p25=   92.45 p50=   99.80 p75=  117.20 p90=  126.10 n=15
  subBass          p10=  -20.20 p25=  -17.85 p50=  -12.80 p75=   -9.75 p90=   -7.76 n=15
  crestFactor      p10=   11.00 p25=   12.50 p50=   14.60 p75=   15.65 p90=   15.96 n=15
  onsetRate        p10=    2.79 p25=    3.40 p50=    4.39 p75=    5.05 p90=    5.60 n=15
  spectralCentroid p10=  615.28 p25=  752.80 p50=  885.00 p75= 1366.70 p90= 1589.62 n=15
  pumpingStrength  p10=    0.27 p25=    0.29 p50=    0.31 p75=    0.34 p90=    0.37 n=15
  bassDecayMs      p10=   81.60 p25=  110.00 p50=  194.00 p75=  332.50 p90=  398.80 n=15
  rt60             p10=    1.64 p25=    2.29 p50=    2.63 p75=    2.95 p90=    3.00 n=14
  kickThd          p10=    0.50 p25=    0.54 p50=    0.65 p75=    0.84 p90=    0.87 n=15
  lufsIntegrated   p10=  -14.76 p25=  -13.10 p50=  -12.00 p75=  -11.25 p90=   -9.84 n=15
  plr              p10=   11.14 p25=   11.85 p50=   12.30 p75=   12.90 p90=   14.30 n=15
  lowBass          p10=  -20.00 p25=  -17.15 p50=  -16.30 p75=  -13.35 p90=  -10.32 n=15
  lowMids          p10=  -24.06 p25=  -22.05 p50=  -20.20 p75=  -17.45 p90=  -13.64 n=15
  mids             p10=  -21.48 p25=  -20.70 p50=  -19.50 p75=  -18.75 p90=  -15.22 n=15
  upperMids        p10=  -32.20 p25=  -30.70 p50=  -28.70 p75=  -25.40 p90=  -23.52 n=15
  highs            p10=  -36.34 p25=  -35.50 p50=  -34.00 p75=  -29.20 p90=  -26.06 n=15
  brilliance       p10=  -46.44 p25=  -42.85 p50=  -38.60 p75=  -32.65 p90=  -29.56 n=15
  classifier reads today: [('hiphop', 5), ('acid-techno', 4), ('rock', 3), ('organic-house', 1)]

=== dubstep (n=10)
  bpm              p10=   69.77 p25=   69.93 p50=  124.95 p75=  139.97 p90=  140.00 n=10
  subBass          p10=   -9.91 p25=   -9.05 p50=   -7.50 p75=   -6.47 p90=   -4.57 n=10
  crestFactor      p10=    8.07 p25=    8.93 p50=   10.50 p75=   11.10 p90=   11.45 n=10
  onsetRate        p10=    3.37 p25=    4.40 p50=    4.84 p75=    5.24 p90=    5.35 n=10
  spectralCentroid p10=  912.39 p25= 1151.65 p50= 1350.30 p75= 2286.70 p90= 2552.87 n=10
  pumpingStrength  p10=    0.26 p25=    0.29 p50=    0.31 p75=    0.35 p90=    0.35 n=10
  bassDecayMs      p10=  107.90 p25=  191.25 p50=  262.50 p75=  281.50 p90=  401.40 n=10
  rt60             p10=    2.23 p25=    2.32 p50=    3.00 p75=    3.00 p90=    3.00 n=9
  kickThd          p10=    0.35 p25=    0.38 p50=    0.53 p75=    0.60 p90=    0.64 n=10
  lufsIntegrated   p10=  -10.63 p25=   -9.43 p50=   -8.55 p75=   -7.92 p90=   -5.25 n=10
  plr              p10=    6.79 p25=    9.72 p50=    9.80 p75=   10.07 p90=   11.17 n=10
  lowBass          p10=  -14.21 p25=  -13.25 p50=  -10.80 p75=  -10.07 p90=   -8.89 n=10
  lowMids          p10=  -22.79 p25=  -20.60 p50=  -19.45 p75=  -18.07 p90=  -17.59 n=10
  mids             p10=  -21.63 p25=  -17.62 p50=  -15.90 p75=  -14.80 p90=  -12.51 n=10
  upperMids        p10=  -25.05 p25=  -23.27 p50=  -20.65 p75=  -17.62 p90=  -14.80 n=10
  highs            p10=  -26.74 p25=  -24.68 p50=  -22.40 p75=  -20.50 p90=  -19.52 n=10
  brilliance       p10=  -33.90 p25=  -31.18 p50=  -27.90 p75=  -24.40 p90=  -22.79 n=10
  classifier reads today: [('hiphop', 4), ('trance', 4), ('acid-techno', 1), ('psytrance', 1)]

=== electro-house (n=15)
  bpm              p10=  124.88 p25=  126.95 p50=  128.10 p75=  128.35 p90=  129.52 n=15
  subBass          p10=   -9.28 p25=   -8.60 p50=   -7.90 p75=   -7.20 p90=   -6.70 n=15
  crestFactor      p10=    8.96 p25=    9.35 p50=   10.00 p75=   10.55 p90=   10.90 n=15
  onsetRate        p10=    2.91 p25=    3.25 p50=    3.81 p75=    4.42 p90=    5.09 n=15
  spectralCentroid p10= 1295.18 p25= 1399.55 p50= 1741.90 p75= 1868.65 p90= 2008.90 n=15
  pumpingStrength  p10=    0.22 p25=    0.24 p50=    0.27 p75=    0.32 p90=    0.33 n=15
  bassDecayMs      p10=  109.80 p25=  135.50 p50=  175.00 p75=  213.50 p90=  231.00 n=15
  rt60             p10=    1.64 p25=    1.86 p50=    2.38 p75=    2.63 p90=    2.73 n=10
  kickThd          p10=    0.47 p25=    0.56 p50=    0.62 p75=    0.68 p90=    0.74 n=15
  lufsIntegrated   p10=   -9.84 p25=   -8.80 p50=   -7.30 p75=   -6.70 p90=   -5.28 n=15
  plr              p10=    7.34 p25=    7.95 p50=    9.00 p75=   10.00 p90=   10.42 n=15
  lowBass          p10=  -14.06 p25=  -11.70 p50=  -11.00 p75=   -9.25 p90=   -7.92 n=15
  lowMids          p10=  -19.62 p25=  -18.70 p50=  -16.90 p75=  -16.40 p90=  -15.60 n=15
  mids             p10=  -16.32 p25=  -15.45 p50=  -15.00 p75=  -13.30 p90=  -11.88 n=15
  upperMids        p10=  -21.04 p25=  -20.10 p50=  -18.90 p75=  -17.20 p90=  -16.22 n=15
  highs            p10=  -25.10 p25=  -23.50 p50=  -21.00 p75=  -19.60 p90=  -18.88 n=15
  brilliance       p10=  -30.84 p25=  -28.35 p50=  -26.00 p75=  -25.25 p90=  -24.54 n=15
  classifier reads today: [('acid-techno', 8), ('progressive-house', 3), ('hiphop', 1), ('industrial-techno', 1)]

=== hard-dance (n=11)
  bpm              p10=   74.90 p25=   91.90 p50=  144.50 p75=  148.00 p90=  148.00 n=11
  subBass          p10=  -10.90 p25=  -10.20 p50=   -9.40 p75=   -7.30 p90=   -6.40 n=11
  crestFactor      p10=    8.60 p25=    9.75 p50=   11.10 p75=   11.70 p90=   11.90 n=11
  onsetRate        p10=    2.52 p25=    3.20 p50=    3.67 p75=    4.46 p90=    4.85 n=11
  spectralCentroid p10= 1660.00 p25= 1898.45 p50= 2250.90 p75= 2394.40 p90= 2603.10 n=11
  pumpingStrength  p10=    0.18 p25=    0.23 p50=    0.27 p75=    0.31 p90=    0.33 n=11
  bassDecayMs      p10=  121.00 p25=  155.00 p50=  198.00 p75=  282.00 p90=  298.00 n=11
  rt60             p10=    1.76 p25=    1.84 p50=    2.27 p75=    2.75 p90=    2.90 n=5
  kickThd          p10=    0.56 p25=    0.58 p50=    0.69 p75=    0.73 p90=    0.76 n=11
  lufsIntegrated   p10=   -9.20 p25=   -8.90 p50=   -8.10 p75=   -6.50 p90=   -5.60 n=11
  plr              p10=    6.80 p25=    8.10 p50=    9.50 p75=    9.70 p90=   10.20 n=11
  lowBass          p10=  -14.10 p25=  -13.40 p50=  -12.50 p75=  -11.60 p90=  -11.10 n=11
  lowMids          p10=  -20.50 p25=  -18.60 p50=  -18.00 p75=  -16.10 p90=  -14.00 n=11
  mids             p10=  -16.50 p25=  -16.05 p50=  -14.50 p75=  -12.75 p90=  -10.60 n=11
  upperMids        p10=  -19.20 p25=  -19.05 p50=  -18.30 p75=  -16.80 p90=  -14.40 n=11
  highs            p10=  -22.60 p25=  -21.75 p50=  -20.60 p75=  -19.05 p90=  -18.00 n=11
  brilliance       p10=  -29.00 p25=  -28.00 p50=  -25.20 p75=  -24.60 p90=  -21.90 n=11
  classifier reads today: [('hiphop', 3), ('techno', 2), ('acid-techno', 2), ('pop', 1)]

=== hardcore-hard-techno (n=17)
  bpm              p10=   84.54 p25=   85.10 p50=   92.40 p75=  104.00 p90=  120.94 n=17
  subBass          p10=  -11.38 p25=   -9.00 p50=   -8.30 p75=   -7.30 p90=   -5.82 n=17
  crestFactor      p10=    7.72 p25=    8.30 p50=    9.20 p75=    9.90 p90=   11.76 n=17
  onsetRate        p10=    2.94 p25=    3.27 p50=    4.03 p75=    4.39 p90=    5.03 n=17
  spectralCentroid p10= 1907.02 p25= 2096.10 p50= 2408.10 p75= 2490.10 p90= 2874.44 n=17
  pumpingStrength  p10=    0.15 p25=    0.20 p50=    0.24 p75=    0.30 p90=    0.33 n=17
  bassDecayMs      p10=   85.00 p25=  111.00 p50=  189.00 p75=  227.00 p90=  354.60 n=17
  rt60             p10=    1.12 p25=    1.32 p50=    1.76 p75=    1.79 p90=    2.52 n=5
  kickThd          p10=    0.52 p25=    0.55 p50=    0.68 p75=    0.70 p90=    0.74 n=17
  lufsIntegrated   p10=   -8.26 p25=   -6.70 p50=   -6.10 p75=   -5.00 p90=   -4.42 n=17
  plr              p10=    6.14 p25=    6.60 p50=    7.20 p75=    8.50 p90=    9.56 n=17
  lowBass          p10=  -12.32 p25=  -11.50 p50=   -9.00 p75=   -7.80 p90=   -6.62 n=17
  lowMids          p10=  -20.66 p25=  -17.80 p50=  -16.00 p75=  -14.40 p90=  -12.66 n=17
  mids             p10=  -16.40 p25=  -14.80 p50=  -13.70 p75=  -11.80 p90=  -10.48 n=17
  upperMids        p10=  -19.10 p25=  -17.20 p50=  -16.70 p75=  -15.20 p90=  -13.96 n=17
  highs            p10=  -20.80 p25=  -20.10 p50=  -16.90 p75=  -16.50 p90=  -15.00 n=17
  brilliance       p10=  -27.64 p25=  -25.60 p50=  -23.40 p75=  -20.60 p90=  -20.32 n=17
  classifier reads today: [('hiphop', 8), ('acid-techno', 4), ('rock', 2), ('techno', 1)]

=== psy-trance (n=10)
  bpm              p10=   73.18 p25=   73.80 p50=  115.30 p75=  136.05 p90=  137.63 n=10
  subBass          p10=  -11.41 p25=  -10.70 p50=   -8.90 p75=   -7.25 p90=   -6.76 n=10
  crestFactor      p10=    9.85 p25=   10.33 p50=   11.45 p75=   12.85 p90=   13.41 n=10
  onsetRate        p10=    4.07 p25=    4.20 p50=    4.77 p75=    6.18 p90=    6.86 n=10
  spectralCentroid p10=  860.96 p25= 1391.08 p50= 1442.30 p75= 1719.92 p90= 1911.64 n=10
  pumpingStrength  p10=    0.28 p25=    0.30 p50=    0.31 p75=    0.34 p90=    0.35 n=10
  bassDecayMs      p10=   61.80 p25=   92.00 p50=  116.50 p75=  153.00 p90=  190.80 n=10
  rt60             p10=    1.90 p25=    2.60 p50=    2.83 p75=    3.00 p90=    3.00 n=6
  kickThd          p10=    0.38 p25=    0.44 p50=    0.59 p75=    0.70 p90=    0.78 n=10
  lufsIntegrated   p10=  -11.32 p25=  -10.60 p50=   -9.05 p75=   -8.25 p90=   -7.78 n=10
  plr              p10=    8.89 p25=    9.57 p50=   10.25 p75=   11.73 p90=   12.87 n=10
  lowBass          p10=  -14.13 p25=  -13.82 p50=  -12.25 p75=  -10.97 p90=  -10.47 n=10
  lowMids          p10=  -22.36 p25=  -22.05 p50=  -21.50 p75=  -18.75 p90=  -17.49 n=10
  mids             p10=  -21.19 p25=  -18.70 p50=  -17.00 p75=  -15.60 p90=  -15.01 n=10
  upperMids        p10=  -26.07 p25=  -24.20 p50=  -22.50 p75=  -20.82 p90=  -18.61 n=10
  highs            p10=  -27.97 p25=  -25.73 p50=  -22.85 p75=  -22.25 p90=  -20.48 n=10
  brilliance       p10=  -31.35 p25=  -29.88 p50=  -28.55 p75=  -26.52 p90=  -26.40 n=10
  classifier reads today: [('acid-techno', 5), ('hiphop', 4), ('trance', 1)]

🤖 Generated with Claude Code

…g families (accuracy PR-G7) Nine new signatures — measured from 93 GiantSteps reference clips (gabber, hardstyle, electro, downtempo) and knowledge-based flagged in-line (trap, footwork, two-step, speed-garage, halftime-dnb) — plus measured retunes of the three controls whose old windows never fired on real audio (breaks, dubstep, psytrance). The classifier's bpm axis becomes octave-aware via PR-G3's bpmOctaveEvidence (every measured hardcore clip ships a halved bpm; the alternate hypothesis is discounted 10% and never touches bpm itself). Family plumbing lands on all four cross-boundary sites: _GENRE_FAMILY_MAP (+garage/hardcore/ trap/electro/downtempo, fixing the unmapped garage ids), the TS genreFamily union + validFamilies, JSON_SCHEMA.md's now-closed set, and mapLegacyToEnhanced fallbacks. Four measured mix-doctor profiles; knowledge-based genres get none (targets are not invented). Real-clip replay: correct-family top-1 went from ~0/93 to 51/93. Provenance, method, and caveats (LOFI bright-side skew, preview length, rt60 saturation, table-wide pumping mismatch follow-up): audits/genre-target-sourcing-2026-07-13.md. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

slittycode merged commit 0ed792f into main Jul 13, 2026
5 checks passed

slittycode deleted the accuracy/genre-classifier-widening branch July 13, 2026 10:39

slittycode mentioned this pull request Jul 17, 2026

feat(ui): surface the recommendations.v1 contract (trust pill + table) + retry catch #185

Closed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

feat(genre): measured classifier + mix-doctor widening for the missing families (PR-G7) - #216

feat(genre): measured classifier + mix-doctor widening for the missing families (PR-G7)#216
slittycode merged 1 commit into
mainfrom
accuracy/genre-classifier-widening

slittycode commented Jul 13, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

Conversation

slittycode commented Jul 13, 2026

What

Changes

Measured validation — real-clip classifier replay (old table → new)

Validation

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant