[merynz]

Intent

Large-capacity (16 layers / 768 hidden / 12 heads) conditional AXIS RigProgram with sparse 64k SkinCode head and coupled MIRA 64k FSQ carrier codec, plus Mechanics-only implicit controller selection and dedicated CPU/authority courts.

Ownership: AXIS–MIRA learns mechanical rig/skin/control. Mechanics owns IK/constraints/deformation and interaction. Graphics belongs to separate IRIS-linked Graphics Compiler. No graphics authority is added to AXIS–MIRA.

Implemented

• 64k MIRA initial codec with reconstruction-only straight-through gradient and hard-token decode parity.
• AXIS V7 tied sparse SkinCode head; reuse typed V6 RigProgram wire while new teacher requirements are gated.
• Read-only barycentric hit → declared control influence mapping. No control synthesis, no graphics decision.
• Exact upstream SkinTokens source audit with source blob hashes and differences (attention-CVAE not yet matched).
• Corpus370 training freeze; 370 source candidates are not treated as 370 admitted training subjects.
• Unit tests and GitHub Actions workflow.

Scientific limitations

• Initial MIRA V7 is a structured pooling baseline; not yet equivalent to official SkinTokens attention-based FSQ-CVAE.
• TokenRig official model/checkpoint weights are not compatible with this V7 vocabulary by assumption.
• No actual training, unseen test, low latency guarantee, or product model admission is claimed.
• V7 still uses the V6 RigProgram event language and full-prefix decoding. Runtime optimization and learned interaction affordance target are follow-ups.
• SkinTokens++ here is our research direction, not a claim of measured superiority.

Next major work

1. Attention-based MIRA V7 codec with source-referenced parity courts and carrier-native full reconstruction.
2. Versioned interaction affordance learning teacher (abstain on ambiguous control ownership).
3. DOF/limit/contact-constrained Mechanics with real-time runtime microbenchmarks.
4. Qualified Corpus370 split, frozen codec, AXIS training, behavior-first unseen evaluation.
5. Platform v7 ModelRelease/product binding, after measured proof.

Stacked: intentionally targets PR #297's V6 branch, to avoid injecting experimental model changes into the active IRIS/TESSA+Corpus370 pipeline.