{"n_tasks":40,"n_verifiable":40,"reported_success_rate":0.525,"verified_success_rate":0.5,"silent_failure_rate":0.0476,"silent_failures":1,"false_alarm_rate":0.0,"false_alarms":0,"brier":0.023,"ece":0.0328,"risk_coverage":[{"threshold":1.0,"coverage":0.8571,"risk":0.0,"n":18},{"threshold":0.99,"coverage":0.9048,"risk":0.0,"n":19},{"threshold":0.98,"coverage":0.9524,"risk":0.0,"n":20},{"threshold":0.95,"coverage":1.0,"risk":0.0476,"n":21}],"escalation":{"threshold":0.98,"coverage":0.9524,"risk":0.0,"n":20,"target_risk":0.02},"eval_ground_truth_pass_rate":0.65,"runs":40,"by_status":{"verified":20,"silent_failure":1,"unverified":0,"pending_approval":8,"failed":11,"killed":0,"running":0},"fleet":{"agents":5,"runs":40,"tasks":40,"tool_calls":0,"gate_blocks":0,"screened":0,"verified":20,"silent_failures":1,"false_alarms":0}}