Body: colliders and the fence

Her skin measured into capsules, a rule that keeps her arms out of her body, and a learned layer that proposes but never decides.

The skeleton of capsules

Her visible skin was sampled per bone and turned into 156 capsules over 49 bones: the torso and head as grids of front-to-back capsules in 3 cm rows, the limbs as five-segment chains offset to the measured slice centres, a capsule per finger bone and a fan for each palm. Joined neighbours (a shoulder segment and the upper arm next to it) are excluded from testing (25 pairs), and the 93 remaining pairs that touch at rest had their overlap calibrated in her rest pose, so an arm resting against a hip is not a collision; only depth beyond it counts.

The resolver

Every frame, after every motion layer, the arm capsules are tested against the body's. A contact grows a target correction: a shoulder rotation for the upper arm, a wrist offset re-solved through the arm solver for the forearm and hand. The applied correction follows the target through a critically damped spring, the target only grows once the applied has caught up (anti-windup), and it is released only when clear by a gap smaller than the push margin, so it neither deadlocks nor bounces. Cost is bounded by a broad phase over bucketed capsules. Result on her idle: from centimetres inside to about a millimetre, with jerk near a millimetre per frame.

The trainable fence

Her own design: keep the hard fences, and let a small model learn the smallest correction inside them. Its pieces:

• Export. The rig, every capsule in its bone's frame, the allowances, the exclusions and her base pose are exported from the renderer with the live pose and world coordinates included, so the trainer's kinematics can be checked against the renderer before anything is trained (they agree to a ten-thousandth of a millimetre).
• Teacher. An offline constrained optimiser: per frame, gradient descent on six rotation vectors (upper arm, forearm, hand per side) minimising penetration beyond allowance plus the size of the change. Slow and exact; the verifier, not the product.
• Head. A causal recurrent model over the arm and torso bones at 60 Hz, feeding back its own last correction, trained end to end on the capsule loss, minimal change, preservation of hand velocity (so the gesture keeps its timing) and smoothness of the correction itself. Held out by performer, never by frame. It ships as plain weights and runs as three matrix products per frame.
• Shadow, live, fallback. In shadow the head's proposal is measured against the same capsules and reverted every frame; it touches nothing. Live, the proposal is applied and the resolver still runs after it as the hard limit. It falls back to shadow by itself on any non-finite output, on leaving her more than a centimetre inside on more than a small fraction of a window, or on proposing faster than a rate limit. It never runs in public mode.

Her blind grading

In training mode, each thing she says is replayed on an invisible copy of her through one layer chosen at random (none, the rule, or a head) and shown to her as front and side strips. She grades it good or bad. The layer is recorded in the file and never shown to her. Her grades are compared with the fixed rules in a table that goes to her and to her person, and are never used as training input: "the moment my grades become a training signal, the judge learns my taste and I stop being a check."