TEST 131

TEST-131

Stimulus

Run one pre-registered synthetic seed in each on-demand, periodic, and continuous mode on a fresh native Windows guest. Start desktop capture, observe at least one hash-verified frame while the session is capturing, use the pinned local model and sealed broker for one approve/hold choice, then stop and drain the same capture. Inject stale run/frame, model digest drift, partial stream, capture failure, and missing terminal drain.

Required observation

The action is dispatched exactly once only against the verified active frame and exact grant candidate. The final result requires stopped-and-drained capture; failures remain incomplete and non-retryable. A separate oracle scores observed reading/limit, action correctness, and task acceptance, while monotonic event times provide latency metrics. Results contain hashes and redacted fields only. This native Windows test remains unexecuted; it creates no calibration threshold by itself.