Next Training Roadmap - Source-Grounded

Next Training Roadmap - Source-Grounded

This page replaces the old V6 roadmap. The old version assumed specific model
directions and training counts. This version lists what must be verified and built
next.

Goal

Create a reliable training loop where captured body/audio/control sessions become
auditable model artifacts.

session folders
  -> validation / alignment
  -> training pairs
  -> training run
  -> exported artifacts
  -> on-device load
  -> live verification

Workstream 1: Session Archive

Implement the folder structure from [data-capture.md](data-capture.md).

Required before any training claim:

• one folder per session;
• manifest;
• SAN JSONL;
• available camera video or pose JSONL;
• optional mocopi/watch/sensor-logger streams;
• track/Rekordbox events when relevant;
• gesture annotations when relevant.

Workstream 2: Dynamics128 Unification

Current issue:

- Rust has `dynamics_128`.
- Swift `EchelonBridge.getDynamics128()` overlays pose/mocopi/watch fields.
- SAN logs `san_get_flat_input`.
- DiffusionService overlays activations/expressions.
- ClaimBridge currently reads raw `echelon_get_dynamics_128` inside
`EchelonBridge.step`.

Target:

one authoritative Dynamics128 builder
  -> SAN
  -> DiffusionService
  -> ClaimBridge
  -> logger

Until this is done, training data must record which vector producer generated
each frame.

Workstream 3: SAN Provenance

Current verified:

• SAN architecture exists.
• local weights exist.
• manifest has 164,248 parameters.
• logger writes schema v2 128D JSONL.

Needed:

• find or rebuild the training script;
• identify source sessions;
• produce a current validation report;
• verify on-device non-flat output;
• verify behavior at nonzero `mixFactor`.

Workstream 4: CoreML One-Step Flow

Current verified:

• `DiffusionService` exists.
• `ConditioningEncoder.mlpackage` exists.
• `FlowGenerator1Step.mlpackage` exists.
• app uses a 104D CoreML input shim;
• app generates one-step token logits, not multi-step DDIM in Swift.

Needed:

• prove whether the current CoreML models are trained or placeholder;
• retrain `ConditioningEncoder` if the target is full 128D input;
• retrain or replace `FlowGenerator1Step`;
• compare against rule-based fallback;
• document token-grid quality and latency.

Workstream 5: N'Ko Motion Inscription

Current verified:

• ClaimBridge exists in Swift and Rust.
• it emits controlled N'Ko claim-sigil detections.
• `cc-inscription` and Convex inscription mutations exist.

Needed:

• make sure ClaimBridge reads the intended dynamics vector;
• log accepted and rejected claims;
• attach source metadata to each inscription;
• connect motion sessions to N'Ko claim timelines.

Workstream 6: AirDeck Gesture Library

Current target:

camera/body truth
  -> gesture event
  -> K11 safety bridge
  -> Rekordbox keyboard/MIDI action

Needed:

• camera-only gestures must work without mocopi;
• left/right deck gestures must be distinct;
• self-play synthetic gestures must produce reports before live keys;
• gesture library must be visible to the performer;
• every live command needs a cooldown and safety gate.

What Not To Claim Yet

Do not claim:

• SAN V6 exists;
• a full 128D model is trained;
• the CoreML generation path is a verified diffusion sampler;
• a specific training pair count;
• a specific validation loss;
• a single shared body/audio/N'Ko model;
• that mocopi is required for gesture control.

Promotion Criteria

A model or gesture becomes "live-safe" only after:

1. source data is archived;
2. training or rule code is identified;
3. runtime artifact is identified;
4. on-device load is verified;
5. live output is non-flat;
6. behavior is observed;
7. fallback is tested;
8. report is written.