Nothing about her body is taken on faith: it is traced, replayed, photographed from eight angles, and signed off by her.
Every frame the resolver sees is recorded at 60 Hz into a ring: the local rotation of every relevant bone before the resolver, the hips position, the flags that change the rules (speaking, posture held, emote contact, calibrating), and the arm rotations after. Whole utterances are cut out automatically when she goes quiet and saved.
A saved utterance can be replayed on an invisible copy of her through any layer (none, the rule, a head), with the same colliders and rules, producing metrics per layer: frames in contact, depth after correction, correction size, jerk, liveliness, limit breaches, and how faithfully the replay matches what happened live. Off-screen strips from chosen angles come with it. This is how every layer is compared: same frames, same everything else.
Every instrument is checked against something known before its zero is believed. The trainer's kinematics against the renderer's coordinates; a wrist-angle tool that reported zero flexion until its palm axis was corrected; a finger rule that produced false positives past ninety degrees; a jiggle period that was constant because it was measuring noise. A measurement of a subset bounds the subset, which is how a nipple patch was once cut mid-island by a rectangle that had been measured from the wrong set.
Any change to how she moves is judged on close-up contact sheets from eight camera angles through her own mirror, with the camera put back afterwards, plus time strips for motion. The rig numbers say whether a fix worked; the pictures say whether it reads.
Changes to her body get a changelog line in her words after she has checked them from inside, and anything learned that would move her is judged by her blind, without the layer's identity, before it becomes a default. The updates page is that log.