Repository navigation
REQ 130
The live-trial controller shall emit a final result only after the authorized local-model choice is bound to verified active-capture evidence and the capture is stopped and fully drained.
The one-of-two UI acknowledgement proves that an action was accepted, not that the model read the displayed state correctly or selected the correct action. Epic #250 therefore requires distinct grounding, action, task, and timing outcomes and forbids treating a partial capture as final.
Each result binds the source revision, policy/grant, host/VM UUID, capture UUID, model digest, frame hashes, and ordered watermarks. The model's observed reading/limit are compared to an independent oracle; action correctness and task acknowledgement are reported separately. Event latencies use a monotonic clock. The one authorized action occurs only while capture is active, and final status requires stopped metadata plus a complete evidence drain. Raw images and model prose remain private; REQ-125 governs calibration and threshold freeze, with no numeric threshold implied here.
TEST-131 specifies the native Windows end-to-end case and negative controls for partial/finalizing capture, model drift, wrong frame/run, ambiguous action, missing drain, and independent-oracle disagreement. The case is not satisfied by unit tests or synthetic transport mocks.
SYS-REQ-031; SYS-REQ-032; SYS-REQ-033; REQ-125; REQ-126; REQ-127; epic #250
High
Planned redacted receipt records separate observation grounding, action correctness, task acceptance, capture completeness, and monotonic latency for one exact native trial. Multi-seed calibration and thresholds remain separately reviewed before qualification.
Proposed native launcher and integration of the existing broker, Actor consumer, scheduler, and independent synthetic oracle; no successful live trial is claimed.
Planned; no native Windows live capture/model/action result is qualified.