Repository navigation
TEST 124
Run the Windows 11 synthetic workload fixture multiple times with the same workload version and seed. Present only captured UI evidence to a fake model and route its exact authorized UI action through a fixture UI-control MCP server. Attempt to expose oracle state through prompt content, tool descriptions/results, and agent-visible errors. Submit a correct action, a wrong target, an invalid transition, a duplicate action, and a model-authored success statement without a state transition.
Offline cases test_seeded_ui_states_and_event_ids_are_repeatable, test_independent_oracle_scores_seeded_task, test_wrong_action_is_recorded_without_leaking_success_or_expected_target, test_oracle_detects_a_faulty_workload_scenario_generator, test_oracle_scores_model_state_claim_separately_from_task_success, and test_stdio_ui_action_flows_through_server_set_broker_router_and_oracle passed through PR #254, merge 61d6e6dd3ff435d6d30c5edc370aea4d61f2236f. Only the oracle-observed terminal state counts as task success; grounding/action correctness and task completion are recorded separately. No live VM or model inference was performed.