An image sequence and a gripper-drive history corroborate a completion claim. Compare the original verifier with the combined decision on the same physical scene, including the failures that survive every check.
| Physical reference | Baseline v0.3 | Corroborated |
|---|
Original successes that become incomplete or unknown. All physically complete cases remain counted.
Axial yaw is modulo 180°. Bounds come from the declared color, shape and camera model; they are not guarantees under arbitrary lighting or occlusion.
Jaw gap · mm
Gripper drive effort · N
Command, encoder and effort consistency can corroborate an open, settled jaw. It cannot establish table support or absence of contact with every arm link.
Requires 200 ms of observation history. Repeated pipeline calls measure compute variation; they are not additional physical trials or a hard real-time guarantee. These measurements are not a matched speedup comparison with a hosted Jev request or the previous parsed-state-only verifier.
| Stage | Calls | Median · ms | p95 · ms | Max · ms |
|---|
Physical reference, evidence sufficiency and algorithmic agreement are different measurements. An unknown decision can block an unsafe success claim without proving that the task failed. Shared replay, consistent fabricated inputs and unobservable physical properties remain explicit controls.