A/B studio — legacy vs $50K analog chain
Render a track through the new studio chain (PolyBLEP oscillators +
Moog ladder + SSL bus comp + plate reverb + tape sat + 10-stage
master bus), compare against the currently-shipping audio via 3×
Gemini consensus. Only promotes the studio version if it beats
legacy by prefer_legacy_delta (default 0.3) on overall score.
The studio chain summary
The chain models a $50,000 hardware studio:
- Oscillators — PolyBLEP saw / square / triangle, supersaw (JP-8000 style), wavetable v2 with morphing, hard-sync, analog drift
- Filters — Moog ladder (LPF24, via Pedalboard’s LadderFilter), state-variable, MS-20, Steiner-Parker, vactrol LPG (Buchla)
- Modulation — LFO with 7 shapes, DUSG (Serge slope), envelope follower, S&H, slew limiter, sidechain keyer, ModMatrix
- Effects — BBD chorus, plate reverb (EMT-140), spring reverb, tape sat v2 with wow/flutter, transformer (Neve) sat, tube (12AX7) sat, phaser, flanger
- Master bus — 10 stages, 5 presets (see below)
Eligibility
The A/B runner only acts on tracks with sound_class ∈ {synth-pure, synth-ambient, mechanical-ambient}. Organic content (purrs, bird
calls) doesn’t benefit from analog modeling — it benefits from real
recordings (see elevenlabs-sfx and freesound pipelines).
Parameters
- track_id — catalog track id to A/B
- judge_runs (3, must be odd) — Gemini consensus runs per candidate
- prefer_legacy_delta (0.3) — studio must beat legacy by this delta on the overall score to promote
- studio_preset — override the auto-selected mastering preset
Decision matrix
| Outcome | Action |
|---|---|
studio_avg ≥ legacy_avg + 0.3 AND studio_ships ≥ 2 of 3 | Promote studio |
studio_avg ≥ legacy_avg + 0.1 AND studio_ships ≥ 2 of 3 | Keep legacy (tiebreaker) |
studio_ships < 2 of 3 | Keep legacy (didn’t ship at all) |
| Both fail consensus | Flag for human review |
Cost + runtime
- 3 judges × 2 candidates × ~$0.04 ≈ $0.30 / track
- 5 minutes wall-clock typical (Gemini latency + sometimes retries)
When to reach for it
- A catalog synth track is shipping at quality < 9/10
- The DSP rendering hasn’t changed but the studio chain has
- You suspect the legacy version is “fine but flat” — the studio chain adds analog warmth, harmonic richness, depth from reverb