The new local workbench authorizes exact native VLA requests and compares independently verified execution records. A 40-pair engineering study retains both the matching actions and the one camera-feedback difference.
This dated update retains the original native-v3 grid. Its 0.802 ms mean precedes copy/entry hardening; the latest update reports the revised writer and Launch Gate separately. Read the latest update
Direct submission and active authorization each succeed in 33/40 LIBERO-Spatial episodes. All 5,420 final action bytes and 40 paired outcomes match, and the seven failures in each lane remain visible. This fixed engineering grid measures integration compatibility; earlier observation-only cohorts stay separate.
Sixty synthetic invalid-authorization attempts cover six fault classes; all are blocked with zero forbidden writer calls. Authorization plus admission averages 0.802 ms, with 1.040 ms P95 over 5,420 actions. The timing excludes policy inference and environment execution.
The new comparison locates one camera-2 feedback fingerprint difference at task 0 / state 48 / step 4. Actions, cursors, outcomes and unexpected authorization differences agree. The camera difference remains in the report, with its source frames and the saved-pose gap explicitly identified.
The locally built macOS App imports all 40 formal episodes with an independently selected key, displays actual Panda task geometry, and retains signed export and reimport. Native replay and a downloadable comparison are available on the project homepage. The physical robot interface remains read-only; Panda physical safety has not been validated.
