Training And Learning - Source-Grounded Overview

Training And Learning - Source-Grounded Overview

The system currently has two separate things that must not be confused:

1. data capture - MotionMixApp logs 128D SAN input and output frames;
2. model training - an offline process that must produce weight artifacts and
validation evidence.

The code proves capture and runtime loading exist. It does not, by itself, prove
the current models are well trained.

Verified Capture Path

Source: `MotionMixApp/Services/SANTrajectoryLogger.swift`.

SANService.step(core:)
  -> san_get_flat_input(...)
  -> san_get_output(...)
  -> SANTrajectoryLogger.logFrameDirect(...)
  -> Documents/san-training/<session-id>.jsonl
  -> optional HTTP NDJSON POST to http://[ip]:9471/san-frame

Frame schema:

schema_version = 2
dims = 128
input = [128 floats]
outputs = audio, camera, phrase, pattern, gesture, phase, regime, latency

Verified Runtime Artifacts

Local SAN artifacts:

MotionMixApp/MotionMixApp/Resources/san_weights.bin
MotionMixApp/MotionMixApp/Resources/san_manifest.json

Current inspected manifest:

entries: 76
parameters: 164,248
weights size: 656,992 bytes

CoreML generation artifacts:

MotionMixApp/MotionMixApp/MLModels/ConditioningEncoder.mlpackage
MotionMixApp/MotionMixApp/MLModels/FlowGenerator1Step.mlpackage
MotionMixApp/MotionMixApp/Models/MotionToMusic.mlpackage

But `MASTER-TASKS.md` describes the CoreML generation models as shell/zero-init
placeholders at that checkpoint, so quality must be re-verified before claiming
trained generation.

What Must Be Proven Before Saying "The Model Learned"

For SAN:

• exact training data path;
• frame count;
• train/validation split;
• training script path;
• checkpoint path;
• exported `san_weights.bin` and `san_manifest.json`;
• parameter count;
• validation metrics;
• on-device load count;
• live non-flat output;
• `mixFactor` greater than zero in the behavior being evaluated.

For CoreML flow generation:

• exact training data path;
• proof that `ConditioningEncoder` is not a placeholder;
• proof that `FlowGenerator1Step` is not a placeholder;
• output token-grid sanity checks;
• comparison against rule-based fallback;
• latency on device.

For N'Ko motion inscription:

• ClaimBridge detections with source dynamics;
• Convex inscription records;
• confidence thresholds and rejected claims;
• mapping from claim sigil to rendered N'Ko line.

Current Learning Lanes

SAN capture lane
  iPhone runtime -> 128D JSONL -> future SAN training

CoreML generation lane
  dynamics -> ConditioningEncoder -> one-step flow -> token grid

N'Ko inscription lane
  dynamics -> ClaimBridge -> controlled N'Ko claim record

AirDeck gesture lane
  camera/body truth -> gesture detector -> K11 safety bridge -> Rekordbox

These lanes can share data later, but they are not currently one trained unified
model.

Docs In This Section