Claim Audit - 2026-05-27

Claim Audit - 2026-05-27

This audit records the corrections made after reviewing the computational
choreography docs against MotionMixApp and Comp-Core source.

Main Correction

The old docs over-presented research vocabulary as deployed architecture. The
rewritten docs now separate:

• verified runtime code;
• local artifacts;
• research terms;
• future training plans;
• philosophical framing.

Diffusion Correction

User memory was correct: the current MotionMixApp generation path is better
described as a conditioning encoder plus one-step flow path.

Verified source:

• `MotionMixApp/Services/DiffusionService.swift`
• `MotionMixApp/MLModels/ConditioningEncoder.mlpackage`
• `MotionMixApp/MLModels/FlowGenerator1Step.mlpackage`
• `MotionMixApp/MASTER-TASKS.md`
• `MotionMixApp/CLAUDE.md`

Correct wording:

DiffusionService contains a conditioned one-step flow/token generation path with
phone-hub and rule-based fallbacks.

Incorrect wording:

The deployed app uses a full multi-step diffusion sampler conditioned on the full
128D body state.

Important caveat: the current CoreML encoder input is 104D, so the app truncates
the larger dynamics vector before `ConditioningEncoder`.

Second-pass proof added after user correction:

DiffusionService.swift lines 11-24: historical name, current implementation is
ConditioningEncoder + FlowGenerator1Step, not in-app multi-step DDPM/DDIM.

DiffusionService.swift lines 258-321: runtime packs [1,104] dynamics, receives
[1,768] embedding, samples [1,384,81] Gaussian noise, runs FlowGenerator1Step,
then argmaxes logits into a TokenGrid.

CoreML model signatures:
ConditioningEncoder: dynamics [1,104] -> embedding [1,768]
FlowGenerator1Step: noise [1,384,81] + conditioning [1,768] -> logits [1,384,81]

Therefore "diffusion" is a legacy service name in the app. The verified runtime
claim is conditioned one-step flow/token generation with fallbacks.

SAN Correction

Verified source:

• `MotionMixApp/Services/SANService.swift`
• `MotionMixApp/Services/SANTrajectoryLogger.swift`
• `Comp-Core/core/audio-media/cc-echelon/crates/cc-brain/src/san/`

Correct wording:

SAN is Somatic Adaptive Network, a Rust/Swift adaptive output pipeline with
bundled local weights, a manifest loader, and trajectory logging.

Current local manifest audit:

san_manifest.json entries: 76
total parameters: 164,248
san_weights.bin: 656,992 bytes

Incorrect without a matching training report:

SAN is 135K parameters.
SAN trained on 5,408 pairs.
SAN validation loss is 0.028.
SAN replaced all heuristics.

Additional caution: `SANService.mixFactor` defaults to `0.0`, and Rust
`SANConfig::default()` also starts at pure heuristic blend. Do not write that
SAN drives heard audio unless the specific consumer path and nonzero mix factor
are verified.

The inspected Rust source currently has a doc/code vocabulary mismatch: top
comments still mention raw 104D in places, while `SANConfig::default()` uses
`input_dim = 128` and `flatten_latent(..., 128)` is used by the 128D FFI path.
Docs should describe a family of related dynamics vectors, not one invariant
shape everywhere.

LIM-RPS Correction

`LIM-RPS` is a real historical/research term in the repo, especially in
`motion-bridge/src/lim_rps.rs`. It is not the canonical name for the entire
current runtime.

Use source names:

• `EchelonCore`
• `LatentUpdater`
• `SimpleLatentUpdater`
• `LearnedLatentUpdater`
• `DellLatentUpdater`
• `SANPipeline`
• `DiffusionService`
• `ClaimBridge`

DELL Correction

`DellLatentUpdater` exists, but source comments say the current implementation
creates synthetic embeddings from available sensor data. Do not call it a fully
trained full-body production model unless a specific artifact and call path prove
that.

N'Ko / MAOE Correction

Verified ASR anchor:

trajectory-biased Transformer CTC

MAOE status:

post-ASR correction/admissibility/routing layer

Do not say MAOE is the trained acoustic ASR anchor unless a future experiment
proves it.

Use the N'Ko handoff language when describing the ASR result: N'Ko Trajectory
CTC, UnifiedCTCHead, Whisper large-v3 encoder features, 6-layer Transformer with
trajectory bias injection, 20.57
an admissibility/correction/control-plane route around ASR hypotheses.

Camera / Mocopi Correction

The docs now state the live-system requirement:

camera-only gestures must work;
mocopi improves confidence but is not mandatory;
K11 remains the final Rekordbox command safety gate.

Files Rewritten In This Pass

• `00-COMPCORE-SOURCE-OF-TRUTH.md`
• `01-system-overview/what-is-this.md`
• `01-system-overview/how-it-all-fits.md`
• `01-system-overview/full-stack-map.md`
• `02-body-as-input/overview.md`
• `02-body-as-input/sensor-fusion.md`
• `02-body-as-input/imu-and-motion.md`
• `02-body-as-input/mocopi-bones.md`
• `02-body-as-input/femto-bolt-depth.md`
• `03-latent-representation/overview.md`
• `03-latent-representation/dell-architecture.md`
• `03-latent-representation/echelon-bridge.md`
• `04-generative-output/overview.md`
• `04-generative-output/motion-lexicon.md`
• `07-nko-synthesis/overview.md`
• `07-nko-synthesis/parallel-architectures.md`
• `07-nko-synthesis/convergence-vision.md`
• `07-nko-synthesis/embodied-to-digital.md`
• `07-nko-synthesis/maoe-routing.md`
• `08-philosophy/body-as-computer.md`
• `08-philosophy/latent-aesthetics.md`
• `08-philosophy/cultural-sovereignty.md`
• `08-philosophy/griot-protocol.md`
• `09-reference/critical-files.md`
• `09-reference/glossary.md`

Remaining Scan Hits

Some legacy phrases still appear intentionally in negative examples such as
"Do not use" or "Incorrect wording." Those are guardrails, not active claims.