Claude Code Doc Review - 2026-05-27

The user's correction was right. Several generated docs treated intended design language as if it were verified runtime behavior. The source supports a strong architecture, but the language must stay precise: - "Diffusion" is a legacy service name for a conditioned one-step flow/token path. - SAN is wired and weight-backed locally, but not proven as the sole audio driver. - N'Ko ASR is anchored by trajectory-biased Transformer CTC, not MAOE routing. - MAOE is a post-ASR correction/admissibility/control layer. - The 128D vector is a family of related runtime buffers with known shims and mismatches, not a perfectly uniform contract everywhere yet. | Topic | Incorrect generated claim | Verified source evidence | Correct wording | |---|---|---|---| | Diffusion | The app uses a full multi-step diffusion sampler over the full 128D body state. | `MotionMixApp/Services/DiffusionService.swift` says the name is historical and the current path is ConditioningEncoder + FlowGenerator1Step. CoreML signatures are `dynamics [1,104] -> embedding [1,768]` and `noise [1,384,81] + conditioning [1,768] -> logits [1,384,81]`. | Conditioned one-step flow/token generation path with hub and rule-based fallbacks. | | 104D/128D | There is one universal 128D vector consumed identically everywhere. | `DiffusionService` builds 128D but truncates to 104D for `ConditioningEncoder`; Rust SAN defaults to 128D; some comments still say 104D; ClaimBridge comments say 104D while FFI reads 128 floats. | Treat dynamics as related source-specific vectors with explicit producer/consumer boundaries. | | SAN training | SAN was trained to understand BodyTruth modes and all modality masks. | `SANService` loads `san_weights.bin`/`san_manifest.json`, calls `san_step`, and logs flat inputs. No source inspected proves BodyTruth-mode training. | SAN is a Rust/Swift adaptive output pipeline with local weights and trajectory logging; training claims require named artifacts. | | SAN audible control | SAN drives music after calibration. | `SANService.mixFactor` defaults to `0.0`; Rust `SANConfig::default()` uses `mix_factor = 0.0`. | SAN affects audio only when weights load, the consumer blends SAN output, and mixFactor is above zero. | | FAN/NHA/TTT | These layers have proven learned selectivity, rhythm learning, and fixed-time calibration. | Source implements/configures FAN, MoE, NHA, TTT machinery, but behavior must be proven by weights and runtime logs. | Describe the implemented modules and require artifacts/logs before claiming learned performance. | | N'Ko ASR | MAOE is the trained ASR model or acoustic anchor. | `nko-brain-scanner/HANDOFF.md` identifies N'Ko Trajectory CTC, UnifiedCTCHead, Whisper large-v3 encoder features, 6-layer Transformer with trajectory bias injection, 20.57% CER. | MAOE