Grand Diomande Research · Full HTML Reader

Full Stack Map

This map uses source-verified component names. It intentionally avoids old summary claims such as "SAN 135K", "5,408 pairs", "validation loss 0.028", or "LIM-RPS is the production brain" unless a source file proves them.

Embodied Trajectory Systems architecture technical paper candidate score 28 .md

Full Public Reader

Full Stack Map

This map uses source-verified component names. It intentionally avoids old
summary claims such as "SAN 135K", "5,408 pairs", "validation loss 0.028", or
"LIM-RPS is the production brain" unless a source file proves them.

Hardware And Hosts

text
SENSORS
  iPhones              camera, IMU, app-side sensors, SensorLogger HTTP push
  Apple Watches        HR, wrist motion, compass/barometer/location where enabled
  Sony mocopi          BLE to phone, UDP to Mac4
  Orbbec/Femto cameras RGB/depth/pose assist depending on host support

HOSTS
  Mac4                 Unity visuals, Sony/mocopi intake, local LUMM feed
  K11                  Rekordbox, AirDeck, pose viewer, command safety bridge
  Mac1 / MotionMix hub body-state endpoints, BodyTruth, dev coordination
  iPhone apps          MotionMixApp, camera nodes, sensor capture

MotionMixApp Runtime

text
MotionMixApp
  EchelonBridge.swift
    -> pending SensorFrameFFI queue
    -> Rust EchelonCore via FFI
    -> latentState / lexicon / uiState / sectionState
    -> SANService
    -> ClaimBridgeService
    -> audio bridge

  MocopiFeatureExtractor.swift
    -> 27-bone mocopi input
    -> 24D compressed body features

  SANTrajectoryLogger.swift
    -> schema-v2 JSONL
    -> Documents/san-training/<session-id>.jsonl
    -> optional NDJSON push to K11 endpoint

  DiffusionService.swift
    -> PhoneHubClient path if confident
    -> CoreML ConditioningEncoder + FlowGenerator1Step path
    -> rule-based fallback path

Rust / Comp-Core Runtime

text
Comp-Core/core/audio-media/cc-echelon
  crates/cc-brain
    EchelonCore
    LatentUpdater trait
    SimpleLatentUpdater
    LearnedLatentUpdater
    DellLatentUpdater
    SANPipeline
    ClaimBridge

  crates/echelon-audio
    real-time audio handle and render path

  crates/motion-bridge
    LUMM / pose / motion bridge code
    includes lim_rps.rs as a real but non-canonical legacy/research term

128D Dynamics Reality

There are multiple related 128D paths:

  • Rust `echelon_get_dynamics_128`.
  • Swift `EchelonBridge.getDynamics128()` overlay helper.
  • SAN flat input from Rust SAN FFI.
  • `DiffusionService.buildDynamicsVector()`.

These are not yet one perfectly invariant contract. The docs should state where
each consumer gets its vector.

Important caveat:

text
EchelonBridge.step()
  -> echelon_get_dynamics_128(...)
  -> claimBridge.detect(...)

That path does not use the Swift overlay helper that injects pose/mocopi/watch
slots. The overlay helper is still important, but it is not automatically the
input to every downstream consumer.

SAN Lane

text
SANService
  -> san_create_default()
  -> san_load_weights(data, manifest)
  -> san_step(core)
  -> san_get_output()
  -> SANTrajectoryLogger.logFrameDirect(...)

Verified current local artifact:

  • `san_manifest.json`: 76 tensors, 164,248 total parameters.
  • `san_weights.bin`: binary weights corresponding to that manifest.

Unverified in current local source:

  • exact training data count;
  • exact validation loss;
  • claim that every heuristic has been replaced;
  • claim that full 128D is trained end-to-end.

Diffusion / Flow Lane

text
DiffusionService
  buildDynamicsVector()
    -> uses EchelonBridge.getDynamics128() when available
    -> overlays activations and expressions

  CoreML path:
    dynamics prefix [0:104]
      -> ConditioningEncoder.mlmodelc
      -> embedding [768]
      -> optional TTTAdaptationLayer
      -> FlowGenerator1Step.mlmodelc
      -> logits
      -> TokenGrid

  fallback:
    rule-based TokenGrid generation

This is an implemented conditioned one-step flow/token generation path. It should
not be documented as a verified multi-step diffusion sampler.

Mac4 / K11 Live Split

text
Sony mocopi app/sensors
  -> Mac4 UDP :12351
  -> com.lume.mocopi-sidecar-mac4
  -> Unity LUMM UDP :9702
  -> MotionMix POST /mocopi/state at low rate

K11
  -> camera / pose viewer
  -> AirDeck gesture detector
  -> Rekordbox bridge
  -> keyboard/MIDI command safety

K11 must remain the final safety gate for Rekordbox. Mac4 and Unity can visualize
and publish state, but they should not directly spam Rekordbox.

N'Ko / Inscription Lane

text
Echelon dynamics
  -> ClaimBridge
  -> controlled motion claim signal
  -> Convex inscription mutation
  -> cc-inscription / cc-semantic-language ecosystem

The source verifies a motion inscription lane. It does not prove full natural
language phrasing from arbitrary movement yet.

Training And Capture

Current verified capture:

text
SANTrajectoryLogger
  -> Documents/san-training/<session-id>.jsonl
  -> schema_version = 2
  -> dims = 128
  -> subsampleRate = 6

Current desired session structure for new recordings:

text
recordings/<session-id>/
  manifest.json
  raw-video/
  frames/
  pose/
  sensor/
  san/
  labels/
  reports/

Training claims must point to an actual dataset manifest, command, commit, and
artifact hash before they appear as facts.

Promotion Decision

Promote into a technical note or architecture paper with implementation anchors.

Source Anchor

computational-choreography/01-system-overview/full-stack-map.md

Detected Structure

Code Anchors · Architecture