LUME Chain 4 — Music V6 Retrain Pipeline (mode + emotion conditioning)

LUME Chain 4 — Music V6 Retrain Pipeline (mode + emotion conditioning)

Status: RELEASED (validated plan + capture protocol + retrain scripts; live training NOT executed)
Subject: Replace V5 (2-track) ConditioningEncoder + FlowGenerator1Step with V6 trained on 30-track diverse capture data, conditioned on (LUME mode, emotion) so each Sky Garden / Turquoise Alcove / Radiant Underground / Iridescent Beauty / Aurora Veil produces its own music character. V6 also consumes full 128D dynamics (closes the V5 truncation shim at DiffusionService.swift line ~258).
Started: 2026-05-08
Released: 2026-05-08
Chain owner: Mohamed
Execution model: META:OMEGA + META:HYDRA collapsed (single pass), 8-lens reviewed
Prerequisites:
- Chain 1 (Echelon Layer 4 + 128D Temporal Closure) — released, see `RELEASE-CHAIN-1.md`. Live verification of Chain 1 is REQUIRED before V6 .mlpackages get deployed (the dispatch fixes are upstream of any model swap)
- HD1 mounted on Mac4 with `/Volumes/HD1/training-phrases/v6/` writable
- iPhone 16 Pro Max + iPhone 16 Plus available for capture sessions
- Mac5 (Tailscale `[ip]`) accessible for training jobs

---

What this release contains

A VALIDATED PLAN + CAPTURE PROTOCOL + READY-TO-RUN RETRAIN SCRIPTS. It does NOT contain trained V6 weights or deployed V6 models. Live execution is gated on:

1. 30 capture sessions × 15 min run by a human dancer (~7.5 hours over 1 week)
2. Mac5 GPU running ConditioningEncoder + FlowGenerator retrain (~30 min + 4-8 hours overnight)
3. CoreML export to .mlpackage (15 min)
4. iOS smoke test + deploy + on-device verification (1 hour)

When all 4 are green and V6-ACCEPTANCE.md hard gates pass, Chain 4 transitions from RELEASED to FULLY SHIPPED.

This chain is training-data + ML — the novelty is in DATASET DIVERSITY + CONDITIONING SCHEMA, not architecture. FlowGenerator1Step architecture is unchanged from V5; only its inputs are.

---

Five LUME modes targeted (V6 capture)

Mode encoding: 8-bit one-hot (5 active + 3 reserved for V7+).

Five emotion states

still / open / inward / energetic / release. 5-bit one-hot.

Conditioning schema (CRITICAL correction from review)

V6 conditioning vector = `dynamics_128 ⊕ mode_8 ⊕ emotion_5 = 141D`.

Mode + emotion are session metadata, NOT body state. They concat OUTSIDE the 128D body contract. The 128D layout from `MotionMixApp/CLAUDE.md` stays untouched. ConditioningEncoder shape: `141 → 256 → 512 → 768`.

This corrects the original brief's framing of "fitting modes inside the 128D vector."

---

Pipeline artifacts

All under `Desktop/omega-output/music-v6-retrain-20260509/`:

• `03-review.md` — 8-lens review, 0 unresolved CRITICAL findings
• `pipeline/CAPTURE-PROTOCOL.md` — per-session protocol, 30-track manifest, mode×emotion coverage matrix
• `pipeline/V6-TRAINING-FORMAT.md` — NPZ schema, conditioning layout, train/val split rule
• `pipeline/V6-ACCEPTANCE.md` — 9 hard gates spanning training/inference/iOS/subjective
• `pipeline/scripts/build_v6_pairs.py` — JSONL+NPZ → V6 pairs
• `pipeline/scripts/train_san_v6.py` — SAN V6 (135K params, MLX, cross-track val)
• `pipeline/scripts/retrain_conditioning_encoder.py` — V6 encoder 141→768 (PyTorch)
• `pipeline/scripts/retrain_flow_generator.py` — V6 flow joint train (PyTorch)
• `pipeline/scripts/export_v6_to_coreml.py` — coremltools 9.0+ export with V5 fallback
• `pipeline/scripts/v6_smoke_test.py` — pre-deploy verification

All scripts compile cleanly (`python3 -m py_compile`).

---

What's NOT included (forward-deferred)

• Actual 30 capture sessions (human-only execution)
• Actual Mac5 training run (overnight GPU)
• Actual .mlpackage exports
• Actual iOS deploy
• Subjective per-mode listening test results

These require physical execution after Chain 1 is live-verified.

---

Why Chain 1 is the deploy gate

Chain 1 fixes dispatch/slot bugs that prevent trained models from being USED in the live runtime path. V6 produces new .mlpackages but those .mlpackages route through the same DiffusionService → SAN → AudioEngine pipeline. If Chain 1 is not live, V6 weights sit on disk unused.

Training itself can run in parallel with Chain 1 verification (data pipeline does not need iOS runtime). Deploy gate: Chain 1 verified working with V5 weights → swap to V6 → re-verify on device.

---

Reviewed under 8 lenses

• Feasibility: 30 sessions × 15 min = 7.5 human-hours feasible over 1 week
• Coherence: 128D contract preserved; mode/emotion live OUTSIDE dynamics vector
• Risk: Rollback flag (`useV6Models`) + V5 fallback bundle preserved
• Dependency: V6 deploy gates on Chain 1 verification
• Novelty: Mode-conditional generation is genuinely new behavior, not just preset switching
• Executability: Capture is human-only; everything else automatable
• Entanglement: "V6 weights" disambiguated — diffusion-side V6 vs SAN-side V6 are separate
• Paradox fuel: V5's 0.028 val loss likely overfit; V6 cross-track val loss expected ≥0.040 honestly

Verdict: GO. 0 unresolved CRITICAL.

---

Closing note

Chain 4 produces a VALIDATED PLAN + EXECUTABLE PIPELINE, not a trained model. The bottleneck is human dance time (capture sessions). Once captures land + Chain 1 verified, Mac5 train + iOS deploy is ~10 hours start to finish.