# Current visual evidence audit — revision 13

- Physical footage uses original G1 pixels with existing face-only blur and contact annotations. Source 00:57–03:50 becomes one 173-second continuous, normal-speed excerpt. The montage contains six selected excerpts at 1×; the full take runs concurrently from film time zero.
- No baseline physical failure footage was supplied or synthesized. Hardware manuscript counts are shown as a separate fixed +40 ms benchmark.
- Simulation comparison identities: LUCID left, synchronized reference center, filtered-error PI right. Raw fixed/off capture keys must not be interpreted as fixed-DR/no-DR methods.
- Selected comparison recordings use +40 ms added delay and scheduled velocity impulses. They are separate from the unseen +60 ms aggregate. The film shows complete six-second walking and turning captures; the website retains other excerpts with selection limitations.
- Skeletons and measured plots come from captured G1 states and corresponding source references. XY pelvis translation is removed for skeleton display; height and orientation remain. Root error uses original world coordinates.
- Denoising, latent geometry, feedback arrows and channel icons are explanatory diagrams. They do not fabricate training logs or force measurements.
- The 1,024-environment parallel clip is frozen-policy evaluation replay with no resets, shown at 0.25× simulation time. It does not show PPO training updates. Training uses 4,096 environments.
- The +21.7 figure is a percentage-point gain, not percent relative change. Reported confidence intervals and the hybrid’s highest tested means remain visible.
- The closing formation is a choreographed visualization. Deployment includes the policy and low-level controller; the encoder and curriculum scheduler remain training components.

Current verification records: COMPLIANCE.json, qa/ENCODED_QA.json, qa/INTEGRATED_TIMELINE_QA.json, qa/HARDWARE_QA.json and qa/PAGES_QA.json. Earlier notes are preserved in the local archive.