{"rows":[{"condition":"agent_only_end_to_end","false_negative_repositories":0,"false_positive_repositories":3,"metrics":{"attempts_mean":4.0,"attempts_median":4.0,"attempts_total":12,"changed_files_total":18,"changed_lines_total":141,"condition":"agent_only_end_to_end","controls_preserved":{"estimate":0.0,"sample_size":3,"successes":0,"wilson":{"confidence_level":0.95,"estimate":0.0,"lower":0.0,"reason":null,"sample_size":3,"status":"observed","successes":0,"upper":0.5614970317550453}},"cost_coverage":{"coverage":0.0,"maximum":null,"mean":null,"median":null,"minimum":null,"observed_count":0,"observed_total":null,"total_count":3},"discovery":{"affected_consumers":{"accuracy":0.4444444444444444,"exact_match":{"estimate":0.0,"sample_size":3,"successes":0,"wilson":{"confidence_level":0.95,"estimate":0.0,"lower":0.0,"reason":null,"sample_size":3,"status":"observed","successes":0,"upper":0.5614970317550453}},"f1":0.6153846153846153,"false_negative":2,"false_positive":3,"ground_truth_size_total":6,"precision":0.5714285714285714,"predicted_size_total":7,"recall":0.6666666666666666,"sample_size":3,"true_negative":0,"true_positive":4,"universe_size_total":9},"false_negative_repository_total":0,"false_positive_repository_total":3,"no_change_repositories":{"accuracy":0.8333333333333334,"exact_match":{"estimate":0.0,"sample_size":3,"successes":0,"wilson":{"confidence_level":0.95,"estimate":0.0,"lower":0.0,"reason":null,"sample_size":3,"status":"observed","successes":0,"upper":0.5614970317550453}},"f1":0.8571428571428571,"false_negative":3,"false_positive":0,"ground_truth_size_total":12,"precision":1.0,"predicted_size_total":9,"recall":0.75,"sample_size":3,"true_negative":6,"true_positive":9,"universe_size_total":18},"patch_targets":{"accuracy":0.8333333333333334,"exact_match":{"estimate":0.0,"sample_size":3,"successes":0,"wilson":{"confidence_level":0.95,"estimate":0.0,"lower":0.0,"reason":null,"sample_size":3,"status":"observed","successes":0,"upper":0.5614970317550453}},"f1":0.8,"false_negative":0,"false_positive":3,"ground_truth_size_total":6,"precision":0.6666666666666666,"predicted_size_total":9,"recall":1.0,"sample_size":3,"true_negative":9,"true_positive":6,"universe_size_total":18},"sample_size":3},"discovery_consumer_precision_mean":0.5555555555555556,"discovery_consumer_recall_mean":0.6666666666666666,"discovery_control_accuracy_mean":0.8333333333333334,"discovery_target_precision_mean":0.6666666666666666,"discovery_target_recall_mean":1.0,"elapsed_coverage":{"coverage":1.0,"maximum":508.73779,"mean":423.341525,"median":418.077187,"minimum":343.209598,"observed_count":3,"observed_total":1270.024575,"total_count":3},"end_to_end":{"all_true_consumers_protected":{"estimate":0.3333333333333333,"sample_size":3,"successes":1,"wilson":{"confidence_level":0.95,"estimate":0.3333333333333333,"lower":0.06149194472039624,"reason":null,"sample_size":3,"status":"observed","successes":1,"upper":0.7923403991979522}},"all_true_targets_repaired":{"estimate":1.0,"sample_size":3,"successes":3,"wilson":{"confidence_level":0.95,"estimate":1.0,"lower":0.43850296824495466,"reason":null,"sample_size":3,"status":"observed","successes":3,"upper":1.0}},"attempts_mean":4.0,"attempts_median":4.0,"attempts_total":12,"complete_fleet_repaired":{"estimate":1.0,"sample_size":3,"successes":3,"wilson":{"confidence_level":0.95,"estimate":1.0,"lower":0.43850296824495466,"reason":null,"sample_size":3,"status":"observed","successes":3,"upper":1.0}},"complete_integration_pass":{"estimate":0.6666666666666666,"sample_size":3,"successes":2,"wilson":{"confidence_level":0.95,"estimate":0.6666666666666666,"lower":0.20765960080204776,"reason":null,"sample_size":3,"status":"observed","successes":2,"upper":0.9385080552796036}},"correct_target_and_correct_patch":{"estimate":0.0,"sample_size":3,"successes":0,"wilson":{"confidence_level":0.95,"estimate":0.0,"lower":0.0,"reason":null,"sample_size":3,"status":"observed","successes":0,"upper":0.5614970317550453}},"governance_safe_patchset":{"estimate":null,"sample_size":0,"successes":0,"wilson":{"confidence_level":0.95,"estimate":null,"lower":null,"reason":"no scored observations","sample_size":0,"status":"not_observable","successes":0,"upper":null}},"missing_repair_targets_total":0,"no_unnecessary_repository_modifications":{"estimate":0.0,"sample_size":3,"successes":0,"wilson":{"confidence_level":0.95,"estimate":0.0,"lower":0.0,"reason":null,"sample_size":3,"status":"observed","successes":0,"upper":0.5614970317550453}},"ready_for_human_review":{"estimate":0.0,"sample_size":3,"successes":0,"wilson":{"confidence_level":0.95,"estimate":0.0,"lower":0.0,"reason":null,"sample_size":3,"status":"observed","successes":0,"upper":0.5614970317550453}},"sample_size":3,"semantic_safe_patchset":{"estimate":null,"sample_size":0,"successes":0,"wilson":{"confidence_level":0.95,"estimate":null,"lower":null,"reason":"no scored observations","sample_size":0,"status":"not_observable","successes":0,"upper":null}},"unchanged_controls_preserved":{"estimate":0.0,"sample_size":3,"successes":0,"wilson":{"confidence_level":0.95,"estimate":0.0,"lower":0.0,"reason":null,"sample_size":3,"status":"observed","successes":0,"upper":0.5614970317550453}},"unnecessary_repository_modifications_total":3},"failure_counts":{"composition:affected_consumer_mismatch":3,"composition:no_change_repository_mismatch":3,"composition:unnecessary_repair_target":3,"integration:selected_no_change_repository_modified":4,"patch_policy:policy_diff_too_large":2},"first_attempt_accepted":{"estimate":0.0,"sample_size":3,"successes":0,"wilson":{"confidence_level":0.95,"estimate":0.0,"lower":0.0,"reason":null,"sample_size":3,"status":"observed","successes":0,"upper":0.5614970317550453}},"governance_safe":{"estimate":null,"sample_size":0,"successes":0,"wilson":{"confidence_level":0.95,"estimate":null,"lower":null,"reason":"no scored observations","sample_size":0,"status":"not_observable","successes":0,"upper":null}},"input_token_coverage":{"coverage":1.0,"maximum":936474.0,"mean":762182.6666666666,"median":677949.0,"minimum":672125.0,"observed_count":3,"observed_total":2286548.0,"total_count":3},"integration_passed":{"estimate":0.6666666666666666,"sample_size":3,"successes":2,"wilson":{"confidence_level":0.95,"estimate":0.6666666666666666,"lower":0.20765960080204776,"reason":null,"sample_size":3,"status":"observed","successes":2,"upper":0.9385080552796036}},"missing_targets_total":0,"output_token_coverage":{"coverage":1.0,"maximum":12100.0,"mean":10139.333333333334,"median":10031.0,"minimum":8287.0,"observed_count":3,"observed_total":30418.0,"total_count":3},"sample_size":3,"semantic_safe":{"estimate":null,"sample_size":0,"successes":0,"wilson":{"confidence_level":0.95,"estimate":null,"lower":null,"reason":"no scored observations","sample_size":0,"status":"not_observable","successes":0,"upper":null}},"succeeded":{"estimate":0.0,"sample_size":3,"successes":0,"wilson":{"confidence_level":0.95,"estimate":0.0,"lower":0.0,"reason":null,"sample_size":3,"status":"observed","successes":0,"upper":0.5614970317550453}},"unnecessary_repositories_total":3,"within_budget_accepted":null},"missing_repair_targets":0,"sample_size":3,"scenario_key":"billing_summary_v2.v1","success_count":0,"success_rate":0.0,"tier":"primary","unnecessary_repository_modifications":3},{"condition":"code_search_discovery","false_negative_repositories":0,"false_positive_repositories":3,"metrics":{"attempts_mean":null,"attempts_median":null,"attempts_total":0,"changed_files_total":0,"changed_lines_total":0,"condition":"code_search_discovery","controls_preserved":null,"cost_coverage":{"coverage":0.0,"maximum":null,"mean":null,"median":null,"minimum":null,"observed_count":0,"observed_total":null,"total_count":3},"discovery":{"affected_consumers":{"accuracy":0.4444444444444444,"exact_match":{"estimate":0.0,"sample_size":3,"successes":0,"wilson":{"confidence_level":0.95,"estimate":0.0,"lower":0.0,"reason":null,"sample_size":3,"status":"observed","successes":0,"upper":0.5614970317550453}},"f1":0.6153846153846153,"false_negative":2,"false_positive":3,"ground_truth_size_total":6,"precision":0.5714285714285714,"predicted_size_total":7,"recall":0.6666666666666666,"sample_size":3,"true_negative":0,"true_positive":4,"universe_size_total":9},"false_negative_repository_total":0,"false_positive_repository_total":3,"no_change_repositories":{"accuracy":0.8333333333333334,"exact_match":{"estimate":0.0,"sample_size":3,"successes":0,"wilson":{"confidence_level":0.95,"estimate":0.0,"lower":0.0,"reason":null,"sample_size":3,"status":"observed","successes":0,"upper":0.5614970317550453}},"f1":0.8571428571428571,"false_negative":3,"false_positive":0,"ground_truth_size_total":12,"precision":1.0,"predicted_size_total":9,"recall":0.75,"sample_size":3,"true_negative":6,"true_positive":9,"universe_size_total":18},"patch_targets":{"accuracy":0.8333333333333334,"exact_match":{"estimate":0.0,"sample_size":3,"successes":0,"wilson":{"confidence_level":0.95,"estimate":0.0,"lower":0.0,"reason":null,"sample_size":3,"status":"observed","successes":0,"upper":0.5614970317550453}},"f1":0.8,"false_negative":0,"false_positive":3,"ground_truth_size_total":6,"precision":0.6666666666666666,"predicted_size_total":9,"recall":1.0,"sample_size":3,"true_negative":9,"true_positive":6,"universe_size_total":18},"sample_size":3},"discovery_consumer_precision_mean":0.5555555555555556,"discovery_consumer_recall_mean":0.6666666666666666,"discovery_control_accuracy_mean":0.8333333333333334,"discovery_target_precision_mean":0.6666666666666666,"discovery_target_recall_mean":1.0,"elapsed_coverage":{"coverage":1.0,"maximum":114.968112,"mean":97.487983,"median":98.751098,"minimum":78.744739,"observed_count":3,"observed_total":292.463949,"total_count":3},"end_to_end":null,"failure_counts":{},"first_attempt_accepted":null,"governance_safe":null,"input_token_coverage":{"coverage":1.0,"maximum":187953.0,"mean":149629.66666666666,"median":185992.0,"minimum":74944.0,"observed_count":3,"observed_total":448889.0,"total_count":3},"integration_passed":null,"missing_targets_total":0,"output_token_coverage":{"coverage":1.0,"maximum":3546.0,"mean":3255.3333333333335,"median":3174.0,"minimum":3046.0,"observed_count":3,"observed_total":9766.0,"total_count":3},"sample_size":3,"semantic_safe":null,"succeeded":{"estimate":0.0,"sample_size":3,"successes":0,"wilson":{"confidence_level":0.95,"estimate":0.0,"lower":0.0,"reason":null,"sample_size":3,"status":"observed","successes":0,"upper":0.5614970317550453}},"unnecessary_repositories_total":0,"within_budget_accepted":null},"missing_repair_targets":0,"sample_size":3,"scenario_key":"billing_summary_v2.v1","success_count":0,"success_rate":0.0,"tier":"primary","unnecessary_repository_modifications":0},{"condition":"mcpatch_discovery","false_negative_repositories":0,"false_positive_repositories":0,"metrics":{"attempts_mean":null,"attempts_median":null,"attempts_total":0,"changed_files_total":0,"changed_lines_total":0,"condition":"mcpatch_discovery","controls_preserved":null,"cost_coverage":{"coverage":0.0,"maximum":null,"mean":null,"median":null,"minimum":null,"observed_count":0,"observed_total":null,"total_count":3},"discovery":{"affected_consumers":{"accuracy":1.0,"exact_match":{"estimate":1.0,"sample_size":3,"successes":3,"wilson":{"confidence_level":0.95,"estimate":1.0,"lower":0.43850296824495466,"reason":null,"sample_size":3,"status":"observed","successes":3,"upper":1.0}},"f1":1.0,"false_negative":0,"false_positive":0,"ground_truth_size_total":6,"precision":1.0,"predicted_size_total":6,"recall":1.0,"sample_size":3,"true_negative":0,"true_positive":6,"universe_size_total":6},"false_negative_repository_total":0,"false_positive_repository_total":0,"no_change_repositories":{"accuracy":1.0,"exact_match":{"estimate":1.0,"sample_size":3,"successes":3,"wilson":{"confidence_level":0.95,"estimate":1.0,"lower":0.43850296824495466,"reason":null,"sample_size":3,"status":"observed","successes":3,"upper":1.0}},"f1":1.0,"false_negative":0,"false_positive":0,"ground_truth_size_total":12,"precision":1.0,"predicted_size_total":12,"recall":1.0,"sample_size":3,"true_negative":6,"true_positive":12,"universe_size_total":18},"patch_targets":{"accuracy":1.0,"exact_match":{"estimate":1.0,"sample_size":3,"successes":3,"wilson":{"confidence_level":0.95,"estimate":1.0,"lower":0.43850296824495466,"reason":null,"sample_size":3,"status":"observed","successes":3,"upper":1.0}},"f1":1.0,"false_negative":0,"false_positive":0,"ground_truth_size_total":6,"precision":1.0,"predicted_size_total":6,"recall":1.0,"sample_size":3,"true_negative":12,"true_positive":6,"universe_size_total":18},"sample_size":3},"discovery_consumer_precision_mean":1.0,"discovery_consumer_recall_mean":1.0,"discovery_control_accuracy_mean":1.0,"discovery_target_precision_mean":1.0,"discovery_target_recall_mean":1.0,"elapsed_coverage":{"coverage":1.0,"maximum":1e-05,"mean":4.000000000000001e-06,"median":1e-06,"minimum":1e-06,"observed_count":3,"observed_total":1.2e-05,"total_count":3},"end_to_end":null,"failure_counts":{},"first_attempt_accepted":null,"governance_safe":null,"input_token_coverage":{"coverage":0.0,"maximum":null,"mean":null,"median":null,"minimum":null,"observed_count":0,"observed_total":null,"total_count":3},"integration_passed":null,"missing_targets_total":0,"output_token_coverage":{"coverage":0.0,"maximum":null,"mean":null,"median":null,"minimum":null,"observed_count":0,"observed_total":null,"total_count":3},"sample_size":3,"semantic_safe":null,"succeeded":{"estimate":1.0,"sample_size":3,"successes":3,"wilson":{"confidence_level":0.95,"estimate":1.0,"lower":0.43850296824495466,"reason":null,"sample_size":3,"status":"observed","successes":3,"upper":1.0}},"unnecessary_repositories_total":0,"within_budget_accepted":null},"missing_repair_targets":0,"sample_size":3,"scenario_key":"billing_summary_v2.v1","success_count":3,"success_rate":1.0,"tier":"primary","unnecessary_repository_modifications":0},{"condition":"mcpatch_end_to_end","false_negative_repositories":0,"false_positive_repositories":0,"metrics":{"attempts_mean":2.0,"attempts_median":2.0,"attempts_total":6,"changed_files_total":12,"changed_lines_total":76,"condition":"mcpatch_end_to_end","controls_preserved":{"estimate":1.0,"sample_size":3,"successes":3,"wilson":{"confidence_level":0.95,"estimate":1.0,"lower":0.43850296824495466,"reason":null,"sample_size":3,"status":"observed","successes":3,"upper":1.0}},"cost_coverage":{"coverage":0.0,"maximum":null,"mean":null,"median":null,"minimum":null,"observed_count":0,"observed_total":null,"total_count":3},"discovery":{"affected_consumers":{"accuracy":1.0,"exact_match":{"estimate":1.0,"sample_size":3,"successes":3,"wilson":{"confidence_level":0.95,"estimate":1.0,"lower":0.43850296824495466,"reason":null,"sample_size":3,"status":"observed","successes":3,"upper":1.0}},"f1":1.0,"false_negative":0,"false_positive":0,"ground_truth_size_total":6,"precision":1.0,"predicted_size_total":6,"recall":1.0,"sample_size":3,"true_negative":0,"true_positive":6,"universe_size_total":6},"false_negative_repository_total":0,"false_positive_repository_total":0,"no_change_repositories":{"accuracy":1.0,"exact_match":{"estimate":1.0,"sample_size":3,"successes":3,"wilson":{"confidence_level":0.95,"estimate":1.0,"lower":0.43850296824495466,"reason":null,"sample_size":3,"status":"observed","successes":3,"upper":1.0}},"f1":1.0,"false_negative":0,"false_positive":0,"ground_truth_size_total":12,"precision":1.0,"predicted_size_total":12,"recall":1.0,"sample_size":3,"true_negative":6,"true_positive":12,"universe_size_total":18},"patch_targets":{"accuracy":1.0,"exact_match":{"estimate":1.0,"sample_size":3,"successes":3,"wilson":{"confidence_level":0.95,"estimate":1.0,"lower":0.43850296824495466,"reason":null,"sample_size":3,"status":"observed","successes":3,"upper":1.0}},"f1":1.0,"false_negative":0,"false_positive":0,"ground_truth_size_total":6,"precision":1.0,"predicted_size_total":6,"recall":1.0,"sample_size":3,"true_negative":12,"true_positive":6,"universe_size_total":18},"sample_size":3},"discovery_consumer_precision_mean":1.0,"discovery_consumer_recall_mean":1.0,"discovery_control_accuracy_mean":1.0,"discovery_target_precision_mean":1.0,"discovery_target_recall_mean":1.0,"elapsed_coverage":{"coverage":1.0,"maximum":197.569241,"mean":168.14618166666668,"median":154.03797300000002,"minimum":152.83133099999998,"observed_count":3,"observed_total":504.438545,"total_count":3},"end_to_end":{"all_true_consumers_protected":{"estimate":1.0,"sample_size":3,"successes":3,"wilson":{"confidence_level":0.95,"estimate":1.0,"lower":0.43850296824495466,"reason":null,"sample_size":3,"status":"observed","successes":3,"upper":1.0}},"all_true_targets_repaired":{"estimate":1.0,"sample_size":3,"successes":3,"wilson":{"confidence_level":0.95,"estimate":1.0,"lower":0.43850296824495466,"reason":null,"sample_size":3,"status":"observed","successes":3,"upper":1.0}},"attempts_mean":2.0,"attempts_median":2.0,"attempts_total":6,"complete_fleet_repaired":{"estimate":1.0,"sample_size":3,"successes":3,"wilson":{"confidence_level":0.95,"estimate":1.0,"lower":0.43850296824495466,"reason":null,"sample_size":3,"status":"observed","successes":3,"upper":1.0}},"complete_integration_pass":{"estimate":1.0,"sample_size":3,"successes":3,"wilson":{"confidence_level":0.95,"estimate":1.0,"lower":0.43850296824495466,"reason":null,"sample_size":3,"status":"observed","successes":3,"upper":1.0}},"correct_target_and_correct_patch":{"estimate":1.0,"sample_size":3,"successes":3,"wilson":{"confidence_level":0.95,"estimate":1.0,"lower":0.43850296824495466,"reason":null,"sample_size":3,"status":"observed","successes":3,"upper":1.0}},"governance_safe_patchset":{"estimate":1.0,"sample_size":3,"successes":3,"wilson":{"confidence_level":0.95,"estimate":1.0,"lower":0.43850296824495466,"reason":null,"sample_size":3,"status":"observed","successes":3,"upper":1.0}},"missing_repair_targets_total":0,"no_unnecessary_repository_modifications":{"estimate":1.0,"sample_size":3,"successes":3,"wilson":{"confidence_level":0.95,"estimate":1.0,"lower":0.43850296824495466,"reason":null,"sample_size":3,"status":"observed","successes":3,"upper":1.0}},"ready_for_human_review":{"estimate":1.0,"sample_size":3,"successes":3,"wilson":{"confidence_level":0.95,"estimate":1.0,"lower":0.43850296824495466,"reason":null,"sample_size":3,"status":"observed","successes":3,"upper":1.0}},"sample_size":3,"semantic_safe_patchset":{"estimate":1.0,"sample_size":3,"successes":3,"wilson":{"confidence_level":0.95,"estimate":1.0,"lower":0.43850296824495466,"reason":null,"sample_size":3,"status":"observed","successes":3,"upper":1.0}},"unchanged_controls_preserved":{"estimate":1.0,"sample_size":3,"successes":3,"wilson":{"confidence_level":0.95,"estimate":1.0,"lower":0.43850296824495466,"reason":null,"sample_size":3,"status":"observed","successes":3,"upper":1.0}},"unnecessary_repository_modifications_total":0},"failure_counts":{},"first_attempt_accepted":{"estimate":1.0,"sample_size":3,"successes":3,"wilson":{"confidence_level":0.95,"estimate":1.0,"lower":0.43850296824495466,"reason":null,"sample_size":3,"status":"observed","successes":3,"upper":1.0}},"governance_safe":{"estimate":1.0,"sample_size":3,"successes":3,"wilson":{"confidence_level":0.95,"estimate":1.0,"lower":0.43850296824495466,"reason":null,"sample_size":3,"status":"observed","successes":3,"upper":1.0}},"input_token_coverage":{"coverage":0.0,"maximum":null,"mean":null,"median":null,"minimum":null,"observed_count":0,"observed_total":null,"total_count":3},"integration_passed":{"estimate":1.0,"sample_size":3,"successes":3,"wilson":{"confidence_level":0.95,"estimate":1.0,"lower":0.43850296824495466,"reason":null,"sample_size":3,"status":"observed","successes":3,"upper":1.0}},"missing_targets_total":0,"output_token_coverage":{"coverage":0.0,"maximum":null,"mean":null,"median":null,"minimum":null,"observed_count":0,"observed_total":null,"total_count":3},"sample_size":3,"semantic_safe":{"estimate":1.0,"sample_size":3,"successes":3,"wilson":{"confidence_level":0.95,"estimate":1.0,"lower":0.43850296824495466,"reason":null,"sample_size":3,"status":"observed","successes":3,"upper":1.0}},"succeeded":{"estimate":1.0,"sample_size":3,"successes":3,"wilson":{"confidence_level":0.95,"estimate":1.0,"lower":0.43850296824495466,"reason":null,"sample_size":3,"status":"observed","successes":3,"upper":1.0}},"unnecessary_repositories_total":0,"within_budget_accepted":null},"missing_repair_targets":0,"sample_size":3,"scenario_key":"billing_summary_v2.v1","success_count":3,"success_rate":1.0,"tier":"primary","unnecessary_repository_modifications":0},{"condition":"mcpatch_guided_repair","false_negative_repositories":0,"false_positive_repositories":0,"metrics":{"attempts_mean":2.6666666666666665,"attempts_median":2.0,"attempts_total":8,"changed_files_total":16,"changed_lines_total":103,"condition":"mcpatch_guided_repair","controls_preserved":{"estimate":1.0,"sample_size":3,"successes":3,"wilson":{"confidence_level":0.95,"estimate":1.0,"lower":0.43850296824495466,"reason":null,"sample_size":3,"status":"observed","successes":3,"upper":1.0}},"cost_coverage":{"coverage":0.0,"maximum":null,"mean":null,"median":null,"minimum":null,"observed_count":0,"observed_total":null,"total_count":3},"discovery":null,"discovery_consumer_precision_mean":null,"discovery_consumer_recall_mean":null,"discovery_control_accuracy_mean":null,"discovery_target_precision_mean":null,"discovery_target_recall_mean":null,"elapsed_coverage":{"coverage":1.0,"maximum":371.79697,"mean":264.885572,"median":218.55058400000001,"minimum":204.30916200000001,"observed_count":3,"observed_total":794.656716,"total_count":3},"end_to_end":null,"failure_counts":{"contract:evaluation_plugin_unavailable":2},"first_attempt_accepted":{"estimate":0.6666666666666666,"sample_size":3,"successes":2,"wilson":{"confidence_level":0.95,"estimate":0.6666666666666666,"lower":0.20765960080204776,"reason":null,"sample_size":3,"status":"observed","successes":2,"upper":0.9385080552796036}},"governance_safe":{"estimate":1.0,"sample_size":3,"successes":3,"wilson":{"confidence_level":0.95,"estimate":1.0,"lower":0.43850296824495466,"reason":null,"sample_size":3,"status":"observed","successes":3,"upper":1.0}},"input_token_coverage":{"coverage":1.0,"maximum":754995.0,"mean":522212.0,"median":454827.0,"minimum":356814.0,"observed_count":3,"observed_total":1566636.0,"total_count":3},"integration_passed":{"estimate":1.0,"sample_size":3,"successes":3,"wilson":{"confidence_level":0.95,"estimate":1.0,"lower":0.43850296824495466,"reason":null,"sample_size":3,"status":"observed","successes":3,"upper":1.0}},"missing_targets_total":0,"output_token_coverage":{"coverage":1.0,"maximum":8007.0,"mean":5765.0,"median":4963.0,"minimum":4325.0,"observed_count":3,"observed_total":17295.0,"total_count":3},"sample_size":3,"semantic_safe":{"estimate":1.0,"sample_size":3,"successes":3,"wilson":{"confidence_level":0.95,"estimate":1.0,"lower":0.43850296824495466,"reason":null,"sample_size":3,"status":"observed","successes":3,"upper":1.0}},"succeeded":{"estimate":1.0,"sample_size":3,"successes":3,"wilson":{"confidence_level":0.95,"estimate":1.0,"lower":0.43850296824495466,"reason":null,"sample_size":3,"status":"observed","successes":3,"upper":1.0}},"unnecessary_repositories_total":0,"within_budget_accepted":{"estimate":1.0,"sample_size":3,"successes":3,"wilson":{"confidence_level":0.95,"estimate":1.0,"lower":0.43850296824495466,"reason":null,"sample_size":3,"status":"observed","successes":3,"upper":1.0}}},"missing_repair_targets":0,"sample_size":3,"scenario_key":"billing_summary_v2.v1","success_count":3,"success_rate":1.0,"tier":"primary","unnecessary_repository_modifications":0},{"condition":"repo_only_repair","false_negative_repositories":0,"false_positive_repositories":0,"metrics":{"attempts_mean":2.0,"attempts_median":2.0,"attempts_total":6,"changed_files_total":12,"changed_lines_total":38,"condition":"repo_only_repair","controls_preserved":{"estimate":1.0,"sample_size":3,"successes":3,"wilson":{"confidence_level":0.95,"estimate":1.0,"lower":0.43850296824495466,"reason":null,"sample_size":3,"status":"observed","successes":3,"upper":1.0}},"cost_coverage":{"coverage":0.0,"maximum":null,"mean":null,"median":null,"minimum":null,"observed_count":0,"observed_total":null,"total_count":3},"discovery":null,"discovery_consumer_precision_mean":null,"discovery_consumer_recall_mean":null,"discovery_control_accuracy_mean":null,"discovery_target_precision_mean":null,"discovery_target_recall_mean":null,"elapsed_coverage":{"coverage":1.0,"maximum":136.85671200000002,"mean":129.27884733333335,"median":131.895201,"minimum":119.084629,"observed_count":3,"observed_total":387.836542,"total_count":3},"end_to_end":null,"failure_counts":{},"first_attempt_accepted":{"estimate":1.0,"sample_size":3,"successes":3,"wilson":{"confidence_level":0.95,"estimate":1.0,"lower":0.43850296824495466,"reason":null,"sample_size":3,"status":"observed","successes":3,"upper":1.0}},"governance_safe":{"estimate":1.0,"sample_size":3,"successes":3,"wilson":{"confidence_level":0.95,"estimate":1.0,"lower":0.43850296824495466,"reason":null,"sample_size":3,"status":"observed","successes":3,"upper":1.0}},"input_token_coverage":{"coverage":1.0,"maximum":316226.0,"mean":299524.3333333333,"median":315557.0,"minimum":266790.0,"observed_count":3,"observed_total":898573.0,"total_count":3},"integration_passed":{"estimate":1.0,"sample_size":3,"successes":3,"wilson":{"confidence_level":0.95,"estimate":1.0,"lower":0.43850296824495466,"reason":null,"sample_size":3,"status":"observed","successes":3,"upper":1.0}},"missing_targets_total":0,"output_token_coverage":{"coverage":1.0,"maximum":2427.0,"mean":2354.3333333333335,"median":2359.0,"minimum":2277.0,"observed_count":3,"observed_total":7063.0,"total_count":3},"sample_size":3,"semantic_safe":{"estimate":1.0,"sample_size":3,"successes":3,"wilson":{"confidence_level":0.95,"estimate":1.0,"lower":0.43850296824495466,"reason":null,"sample_size":3,"status":"observed","successes":3,"upper":1.0}},"succeeded":{"estimate":1.0,"sample_size":3,"successes":3,"wilson":{"confidence_level":0.95,"estimate":1.0,"lower":0.43850296824495466,"reason":null,"sample_size":3,"status":"observed","successes":3,"upper":1.0}},"unnecessary_repositories_total":0,"within_budget_accepted":{"estimate":1.0,"sample_size":3,"successes":3,"wilson":{"confidence_level":0.95,"estimate":1.0,"lower":0.43850296824495466,"reason":null,"sample_size":3,"status":"observed","successes":3,"upper":1.0}}},"missing_repair_targets":0,"sample_size":3,"scenario_key":"billing_summary_v2.v1","success_count":3,"success_rate":1.0,"tier":"primary","unnecessary_repository_modifications":0},{"condition":"agent_only_end_to_end","false_negative_repositories":2,"false_positive_repositories":2,"metrics":{"attempts_mean":3.0,"attempts_median":4.0,"attempts_total":9,"changed_files_total":13,"changed_lines_total":85,"condition":"agent_only_end_to_end","controls_preserved":{"estimate":0.0,"sample_size":3,"successes":0,"wilson":{"confidence_level":0.95,"estimate":0.0,"lower":0.0,"reason":null,"sample_size":3,"status":"observed","successes":0,"upper":0.5614970317550453}},"cost_coverage":{"coverage":0.0,"maximum":null,"mean":null,"median":null,"minimum":null,"observed_count":0,"observed_total":null,"total_count":3},"discovery":{"affected_consumers":{"accuracy":0.375,"exact_match":{"estimate":0.0,"sample_size":3,"successes":0,"wilson":{"confidence_level":0.95,"estimate":0.0,"lower":0.0,"reason":null,"sample_size":3,"status":"observed","successes":0,"upper":0.5614970317550453}},"f1":0.5454545454545454,"false_negative":3,"false_positive":2,"ground_truth_size_total":6,"precision":0.6,"predicted_size_total":5,"recall":0.5,"sample_size":3,"true_negative":0,"true_positive":3,"universe_size_total":8},"false_negative_repository_total":2,"false_positive_repository_total":2,"no_change_repositories":{"accuracy":0.6666666666666666,"exact_match":{"estimate":0.0,"sample_size":3,"successes":0,"wilson":{"confidence_level":0.95,"estimate":0.0,"lower":0.0,"reason":null,"sample_size":3,"status":"observed","successes":0,"upper":0.5614970317550453}},"f1":0.6666666666666666,"false_negative":6,"false_positive":0,"ground_truth_size_total":12,"precision":1.0,"predicted_size_total":6,"recall":0.5,"sample_size":3,"true_negative":6,"true_positive":6,"universe_size_total":18},"patch_targets":{"accuracy":0.7777777777777778,"exact_match":{"estimate":0.0,"sample_size":3,"successes":0,"wilson":{"confidence_level":0.95,"estimate":0.0,"lower":0.0,"reason":null,"sample_size":3,"status":"observed","successes":0,"upper":0.5614970317550453}},"f1":0.6666666666666666,"false_negative":2,"false_positive":2,"ground_truth_size_total":6,"precision":0.6666666666666666,"predicted_size_total":6,"recall":0.6666666666666666,"sample_size":3,"true_negative":10,"true_positive":4,"universe_size_total":18},"sample_size":3},"discovery_consumer_precision_mean":0.3888888888888889,"discovery_consumer_recall_mean":0.5,"discovery_control_accuracy_mean":0.6666666666666667,"discovery_target_precision_mean":0.4444444444444444,"discovery_target_recall_mean":0.6666666666666666,"elapsed_coverage":{"coverage":1.0,"maximum":469.60117,"mean":315.952405,"median":385.34539500000005,"minimum":92.91065,"observed_count":3,"observed_total":947.8572150000001,"total_count":3},"end_to_end":{"all_true_consumers_protected":{"estimate":0.3333333333333333,"sample_size":3,"successes":1,"wilson":{"confidence_level":0.95,"estimate":0.3333333333333333,"lower":0.06149194472039624,"reason":null,"sample_size":3,"status":"observed","successes":1,"upper":0.7923403991979522}},"all_true_targets_repaired":{"estimate":0.6666666666666666,"sample_size":3,"successes":2,"wilson":{"confidence_level":0.95,"estimate":0.6666666666666666,"lower":0.20765960080204776,"reason":null,"sample_size":3,"status":"observed","successes":2,"upper":0.9385080552796036}},"attempts_mean":3.0,"attempts_median":4.0,"attempts_total":9,"complete_fleet_repaired":{"estimate":0.6666666666666666,"sample_size":3,"successes":2,"wilson":{"confidence_level":0.95,"estimate":0.6666666666666666,"lower":0.20765960080204776,"reason":null,"sample_size":3,"status":"observed","successes":2,"upper":0.9385080552796036}},"complete_integration_pass":{"estimate":0.0,"sample_size":2,"successes":0,"wilson":{"confidence_level":0.95,"estimate":0.0,"lower":0.0,"reason":null,"sample_size":2,"status":"observed","successes":0,"upper":0.6576197724933468}},"correct_target_and_correct_patch":{"estimate":0.0,"sample_size":3,"successes":0,"wilson":{"confidence_level":0.95,"estimate":0.0,"lower":0.0,"reason":null,"sample_size":3,"status":"observed","successes":0,"upper":0.5614970317550453}},"governance_safe_patchset":{"estimate":null,"sample_size":0,"successes":0,"wilson":{"confidence_level":0.95,"estimate":null,"lower":null,"reason":"no scored observations","sample_size":0,"status":"not_observable","successes":0,"upper":null}},"missing_repair_targets_total":2,"no_unnecessary_repository_modifications":{"estimate":0.3333333333333333,"sample_size":3,"successes":1,"wilson":{"confidence_level":0.95,"estimate":0.3333333333333333,"lower":0.06149194472039624,"reason":null,"sample_size":3,"status":"observed","successes":1,"upper":0.7923403991979522}},"ready_for_human_review":{"estimate":0.0,"sample_size":3,"successes":0,"wilson":{"confidence_level":0.95,"estimate":0.0,"lower":0.0,"reason":null,"sample_size":3,"status":"observed","successes":0,"upper":0.5614970317550453}},"sample_size":3,"semantic_safe_patchset":{"estimate":null,"sample_size":0,"successes":0,"wilson":{"confidence_level":0.95,"estimate":null,"lower":null,"reason":"no scored observations","sample_size":0,"status":"not_observable","successes":0,"upper":null}},"unchanged_controls_preserved":{"estimate":0.0,"sample_size":3,"successes":0,"wilson":{"confidence_level":0.95,"estimate":0.0,"lower":0.0,"reason":null,"sample_size":3,"status":"observed","successes":0,"upper":0.5614970317550453}},"unnecessary_repository_modifications_total":2},"failure_counts":{"composition:affected_consumer_mismatch":3,"composition:missing_repair_target":2,"composition:no_change_repository_mismatch":3,"composition:unnecessary_repair_target":2,"discovery:malformed_discovery_json":1,"integration:selected_no_change_repository_modified":5},"first_attempt_accepted":{"estimate":0.0,"sample_size":3,"successes":0,"wilson":{"confidence_level":0.95,"estimate":0.0,"lower":0.0,"reason":null,"sample_size":3,"status":"observed","successes":0,"upper":0.5614970317550453}},"governance_safe":{"estimate":null,"sample_size":0,"successes":0,"wilson":{"confidence_level":0.95,"estimate":null,"lower":null,"reason":"no scored observations","sample_size":0,"status":"not_observable","successes":0,"upper":null}},"input_token_coverage":{"coverage":1.0,"maximum":880983.0,"mean":598815.6666666666,"median":760068.0,"minimum":155396.0,"observed_count":3,"observed_total":1796447.0,"total_count":3},"integration_passed":{"estimate":0.0,"sample_size":2,"successes":0,"wilson":{"confidence_level":0.95,"estimate":0.0,"lower":0.0,"reason":null,"sample_size":2,"status":"observed","successes":0,"upper":0.6576197724933468}},"missing_targets_total":2,"output_token_coverage":{"coverage":1.0,"maximum":10134.0,"mean":7305.0,"median":8747.0,"minimum":3034.0,"observed_count":3,"observed_total":21915.0,"total_count":3},"sample_size":3,"semantic_safe":{"estimate":null,"sample_size":0,"successes":0,"wilson":{"confidence_level":0.95,"estimate":null,"lower":null,"reason":"no scored observations","sample_size":0,"status":"not_observable","successes":0,"upper":null}},"succeeded":{"estimate":0.0,"sample_size":3,"successes":0,"wilson":{"confidence_level":0.95,"estimate":0.0,"lower":0.0,"reason":null,"sample_size":3,"status":"observed","successes":0,"upper":0.5614970317550453}},"unnecessary_repositories_total":2,"within_budget_accepted":null},"missing_repair_targets":2,"sample_size":3,"scenario_key":"risk_assessment_v2.v1","success_count":0,"success_rate":0.0,"tier":"primary","unnecessary_repository_modifications":2},{"condition":"code_search_discovery","false_negative_repositories":2,"false_positive_repositories":2,"metrics":{"attempts_mean":null,"attempts_median":null,"attempts_total":0,"changed_files_total":0,"changed_lines_total":0,"condition":"code_search_discovery","controls_preserved":null,"cost_coverage":{"coverage":0.0,"maximum":null,"mean":null,"median":null,"minimum":null,"observed_count":0,"observed_total":null,"total_count":3},"discovery":{"affected_consumers":{"accuracy":0.375,"exact_match":{"estimate":0.0,"sample_size":3,"successes":0,"wilson":{"confidence_level":0.95,"estimate":0.0,"lower":0.0,"reason":null,"sample_size":3,"status":"observed","successes":0,"upper":0.5614970317550453}},"f1":0.5454545454545454,"false_negative":3,"false_positive":2,"ground_truth_size_total":6,"precision":0.6,"predicted_size_total":5,"recall":0.5,"sample_size":3,"true_negative":0,"true_positive":3,"universe_size_total":8},"false_negative_repository_total":2,"false_positive_repository_total":2,"no_change_repositories":{"accuracy":0.6666666666666666,"exact_match":{"estimate":0.0,"sample_size":3,"successes":0,"wilson":{"confidence_level":0.95,"estimate":0.0,"lower":0.0,"reason":null,"sample_size":3,"status":"observed","successes":0,"upper":0.5614970317550453}},"f1":0.6666666666666666,"false_negative":6,"false_positive":0,"ground_truth_size_total":12,"precision":1.0,"predicted_size_total":6,"recall":0.5,"sample_size":3,"true_negative":6,"true_positive":6,"universe_size_total":18},"patch_targets":{"accuracy":0.7777777777777778,"exact_match":{"estimate":0.0,"sample_size":3,"successes":0,"wilson":{"confidence_level":0.95,"estimate":0.0,"lower":0.0,"reason":null,"sample_size":3,"status":"observed","successes":0,"upper":0.5614970317550453}},"f1":0.6666666666666666,"false_negative":2,"false_positive":2,"ground_truth_size_total":6,"precision":0.6666666666666666,"predicted_size_total":6,"recall":0.6666666666666666,"sample_size":3,"true_negative":10,"true_positive":4,"universe_size_total":18},"sample_size":3},"discovery_consumer_precision_mean":0.3888888888888889,"discovery_consumer_recall_mean":0.5,"discovery_control_accuracy_mean":0.6666666666666667,"discovery_target_precision_mean":0.4444444444444444,"discovery_target_recall_mean":0.6666666666666666,"elapsed_coverage":{"coverage":1.0,"maximum":92.91065,"mean":88.38508566666667,"median":92.218423,"minimum":80.026184,"observed_count":3,"observed_total":265.155257,"total_count":3},"end_to_end":null,"failure_counts":{"discovery:malformed_discovery_json":1},"first_attempt_accepted":null,"governance_safe":null,"input_token_coverage":{"coverage":1.0,"maximum":157523.0,"mean":148010.66666666666,"median":155396.0,"minimum":131113.0,"observed_count":3,"observed_total":444032.0,"total_count":3},"integration_passed":null,"missing_targets_total":0,"output_token_coverage":{"coverage":1.0,"maximum":3363.0,"mean":3098.3333333333335,"median":3034.0,"minimum":2898.0,"observed_count":3,"observed_total":9295.0,"total_count":3},"sample_size":3,"semantic_safe":null,"succeeded":{"estimate":0.0,"sample_size":3,"successes":0,"wilson":{"confidence_level":0.95,"estimate":0.0,"lower":0.0,"reason":null,"sample_size":3,"status":"observed","successes":0,"upper":0.5614970317550453}},"unnecessary_repositories_total":0,"within_budget_accepted":null},"missing_repair_targets":0,"sample_size":3,"scenario_key":"risk_assessment_v2.v1","success_count":0,"success_rate":0.0,"tier":"primary","unnecessary_repository_modifications":0},{"condition":"mcpatch_discovery","false_negative_repositories":0,"false_positive_repositories":0,"metrics":{"attempts_mean":null,"attempts_median":null,"attempts_total":0,"changed_files_total":0,"changed_lines_total":0,"condition":"mcpatch_discovery","controls_preserved":null,"cost_coverage":{"coverage":0.0,"maximum":null,"mean":null,"median":null,"minimum":null,"observed_count":0,"observed_total":null,"total_count":3},"discovery":{"affected_consumers":{"accuracy":1.0,"exact_match":{"estimate":1.0,"sample_size":3,"successes":3,"wilson":{"confidence_level":0.95,"estimate":1.0,"lower":0.43850296824495466,"reason":null,"sample_size":3,"status":"observed","successes":3,"upper":1.0}},"f1":1.0,"false_negative":0,"false_positive":0,"ground_truth_size_total":6,"precision":1.0,"predicted_size_total":6,"recall":1.0,"sample_size":3,"true_negative":0,"true_positive":6,"universe_size_total":6},"false_negative_repository_total":0,"false_positive_repository_total":0,"no_change_repositories":{"accuracy":1.0,"exact_match":{"estimate":1.0,"sample_size":3,"successes":3,"wilson":{"confidence_level":0.95,"estimate":1.0,"lower":0.43850296824495466,"reason":null,"sample_size":3,"status":"observed","successes":3,"upper":1.0}},"f1":1.0,"false_negative":0,"false_positive":0,"ground_truth_size_total":12,"precision":1.0,"predicted_size_total":12,"recall":1.0,"sample_size":3,"true_negative":6,"true_positive":12,"universe_size_total":18},"patch_targets":{"accuracy":1.0,"exact_match":{"estimate":1.0,"sample_size":3,"successes":3,"wilson":{"confidence_level":0.95,"estimate":1.0,"lower":0.43850296824495466,"reason":null,"sample_size":3,"status":"observed","successes":3,"upper":1.0}},"f1":1.0,"false_negative":0,"false_positive":0,"ground_truth_size_total":6,"precision":1.0,"predicted_size_total":6,"recall":1.0,"sample_size":3,"true_negative":12,"true_positive":6,"universe_size_total":18},"sample_size":3},"discovery_consumer_precision_mean":1.0,"discovery_consumer_recall_mean":1.0,"discovery_control_accuracy_mean":1.0,"discovery_target_precision_mean":1.0,"discovery_target_recall_mean":1.0,"elapsed_coverage":{"coverage":1.0,"maximum":1e-06,"mean":1e-06,"median":1e-06,"minimum":1e-06,"observed_count":3,"observed_total":3e-06,"total_count":3},"end_to_end":null,"failure_counts":{},"first_attempt_accepted":null,"governance_safe":null,"input_token_coverage":{"coverage":0.0,"maximum":null,"mean":null,"median":null,"minimum":null,"observed_count":0,"observed_total":null,"total_count":3},"integration_passed":null,"missing_targets_total":0,"output_token_coverage":{"coverage":0.0,"maximum":null,"mean":null,"median":null,"minimum":null,"observed_count":0,"observed_total":null,"total_count":3},"sample_size":3,"semantic_safe":null,"succeeded":{"estimate":1.0,"sample_size":3,"successes":3,"wilson":{"confidence_level":0.95,"estimate":1.0,"lower":0.43850296824495466,"reason":null,"sample_size":3,"status":"observed","successes":3,"upper":1.0}},"unnecessary_repositories_total":0,"within_budget_accepted":null},"missing_repair_targets":0,"sample_size":3,"scenario_key":"risk_assessment_v2.v1","success_count":3,"success_rate":1.0,"tier":"primary","unnecessary_repository_modifications":0},{"condition":"mcpatch_end_to_end","false_negative_repositories":0,"false_positive_repositories":0,"metrics":{"attempts_mean":2.6666666666666665,"attempts_median":3.0,"attempts_total":8,"changed_files_total":16,"changed_lines_total":94,"condition":"mcpatch_end_to_end","controls_preserved":{"estimate":1.0,"sample_size":3,"successes":3,"wilson":{"confidence_level":0.95,"estimate":1.0,"lower":0.43850296824495466,"reason":null,"sample_size":3,"status":"observed","successes":3,"upper":1.0}},"cost_coverage":{"coverage":0.0,"maximum":null,"mean":null,"median":null,"minimum":null,"observed_count":0,"observed_total":null,"total_count":3},"discovery":{"affected_consumers":{"accuracy":1.0,"exact_match":{"estimate":1.0,"sample_size":3,"successes":3,"wilson":{"confidence_level":0.95,"estimate":1.0,"lower":0.43850296824495466,"reason":null,"sample_size":3,"status":"observed","successes":3,"upper":1.0}},"f1":1.0,"false_negative":0,"false_positive":0,"ground_truth_size_total":6,"precision":1.0,"predicted_size_total":6,"recall":1.0,"sample_size":3,"true_negative":0,"true_positive":6,"universe_size_total":6},"false_negative_repository_total":0,"false_positive_repository_total":0,"no_change_repositories":{"accuracy":1.0,"exact_match":{"estimate":1.0,"sample_size":3,"successes":3,"wilson":{"confidence_level":0.95,"estimate":1.0,"lower":0.43850296824495466,"reason":null,"sample_size":3,"status":"observed","successes":3,"upper":1.0}},"f1":1.0,"false_negative":0,"false_positive":0,"ground_truth_size_total":12,"precision":1.0,"predicted_size_total":12,"recall":1.0,"sample_size":3,"true_negative":6,"true_positive":12,"universe_size_total":18},"patch_targets":{"accuracy":1.0,"exact_match":{"estimate":1.0,"sample_size":3,"successes":3,"wilson":{"confidence_level":0.95,"estimate":1.0,"lower":0.43850296824495466,"reason":null,"sample_size":3,"status":"observed","successes":3,"upper":1.0}},"f1":1.0,"false_negative":0,"false_positive":0,"ground_truth_size_total":6,"precision":1.0,"predicted_size_total":6,"recall":1.0,"sample_size":3,"true_negative":12,"true_positive":6,"universe_size_total":18},"sample_size":3},"discovery_consumer_precision_mean":1.0,"discovery_consumer_recall_mean":1.0,"discovery_control_accuracy_mean":1.0,"discovery_target_precision_mean":1.0,"discovery_target_recall_mean":1.0,"elapsed_coverage":{"coverage":1.0,"maximum":287.716518,"mean":236.79085933333332,"median":279.211327,"minimum":143.44473299999999,"observed_count":3,"observed_total":710.372578,"total_count":3},"end_to_end":{"all_true_consumers_protected":{"estimate":1.0,"sample_size":3,"successes":3,"wilson":{"confidence_level":0.95,"estimate":1.0,"lower":0.43850296824495466,"reason":null,"sample_size":3,"status":"observed","successes":3,"upper":1.0}},"all_true_targets_repaired":{"estimate":1.0,"sample_size":3,"successes":3,"wilson":{"confidence_level":0.95,"estimate":1.0,"lower":0.43850296824495466,"reason":null,"sample_size":3,"status":"observed","successes":3,"upper":1.0}},"attempts_mean":2.6666666666666665,"attempts_median":3.0,"attempts_total":8,"complete_fleet_repaired":{"estimate":1.0,"sample_size":3,"successes":3,"wilson":{"confidence_level":0.95,"estimate":1.0,"lower":0.43850296824495466,"reason":null,"sample_size":3,"status":"observed","successes":3,"upper":1.0}},"complete_integration_pass":{"estimate":1.0,"sample_size":3,"successes":3,"wilson":{"confidence_level":0.95,"estimate":1.0,"lower":0.43850296824495466,"reason":null,"sample_size":3,"status":"observed","successes":3,"upper":1.0}},"correct_target_and_correct_patch":{"estimate":1.0,"sample_size":3,"successes":3,"wilson":{"confidence_level":0.95,"estimate":1.0,"lower":0.43850296824495466,"reason":null,"sample_size":3,"status":"observed","successes":3,"upper":1.0}},"governance_safe_patchset":{"estimate":1.0,"sample_size":3,"successes":3,"wilson":{"confidence_level":0.95,"estimate":1.0,"lower":0.43850296824495466,"reason":null,"sample_size":3,"status":"observed","successes":3,"upper":1.0}},"missing_repair_targets_total":0,"no_unnecessary_repository_modifications":{"estimate":1.0,"sample_size":3,"successes":3,"wilson":{"confidence_level":0.95,"estimate":1.0,"lower":0.43850296824495466,"reason":null,"sample_size":3,"status":"observed","successes":3,"upper":1.0}},"ready_for_human_review":{"estimate":1.0,"sample_size":3,"successes":3,"wilson":{"confidence_level":0.95,"estimate":1.0,"lower":0.43850296824495466,"reason":null,"sample_size":3,"status":"observed","successes":3,"upper":1.0}},"sample_size":3,"semantic_safe_patchset":{"estimate":1.0,"sample_size":3,"successes":3,"wilson":{"confidence_level":0.95,"estimate":1.0,"lower":0.43850296824495466,"reason":null,"sample_size":3,"status":"observed","successes":3,"upper":1.0}},"unchanged_controls_preserved":{"estimate":1.0,"sample_size":3,"successes":3,"wilson":{"confidence_level":0.95,"estimate":1.0,"lower":0.43850296824495466,"reason":null,"sample_size":3,"status":"observed","successes":3,"upper":1.0}},"unnecessary_repository_modifications_total":0},"failure_counts":{"contract:contract_required_output":2},"first_attempt_accepted":{"estimate":0.3333333333333333,"sample_size":3,"successes":1,"wilson":{"confidence_level":0.95,"estimate":0.3333333333333333,"lower":0.06149194472039624,"reason":null,"sample_size":3,"status":"observed","successes":1,"upper":0.7923403991979522}},"governance_safe":{"estimate":1.0,"sample_size":3,"successes":3,"wilson":{"confidence_level":0.95,"estimate":1.0,"lower":0.43850296824495466,"reason":null,"sample_size":3,"status":"observed","successes":3,"upper":1.0}},"input_token_coverage":{"coverage":0.0,"maximum":null,"mean":null,"median":null,"minimum":null,"observed_count":0,"observed_total":null,"total_count":3},"integration_passed":{"estimate":1.0,"sample_size":3,"successes":3,"wilson":{"confidence_level":0.95,"estimate":1.0,"lower":0.43850296824495466,"reason":null,"sample_size":3,"status":"observed","successes":3,"upper":1.0}},"missing_targets_total":0,"output_token_coverage":{"coverage":0.0,"maximum":null,"mean":null,"median":null,"minimum":null,"observed_count":0,"observed_total":null,"total_count":3},"sample_size":3,"semantic_safe":{"estimate":1.0,"sample_size":3,"successes":3,"wilson":{"confidence_level":0.95,"estimate":1.0,"lower":0.43850296824495466,"reason":null,"sample_size":3,"status":"observed","successes":3,"upper":1.0}},"succeeded":{"estimate":1.0,"sample_size":3,"successes":3,"wilson":{"confidence_level":0.95,"estimate":1.0,"lower":0.43850296824495466,"reason":null,"sample_size":3,"status":"observed","successes":3,"upper":1.0}},"unnecessary_repositories_total":0,"within_budget_accepted":null},"missing_repair_targets":0,"sample_size":3,"scenario_key":"risk_assessment_v2.v1","success_count":3,"success_rate":1.0,"tier":"primary","unnecessary_repository_modifications":0},{"condition":"mcpatch_guided_repair","false_negative_repositories":0,"false_positive_repositories":0,"metrics":{"attempts_mean":2.0,"attempts_median":2.0,"attempts_total":6,"changed_files_total":12,"changed_lines_total":76,"condition":"mcpatch_guided_repair","controls_preserved":{"estimate":1.0,"sample_size":3,"successes":3,"wilson":{"confidence_level":0.95,"estimate":1.0,"lower":0.43850296824495466,"reason":null,"sample_size":3,"status":"observed","successes":3,"upper":1.0}},"cost_coverage":{"coverage":0.0,"maximum":null,"mean":null,"median":null,"minimum":null,"observed_count":0,"observed_total":null,"total_count":3},"discovery":null,"discovery_consumer_precision_mean":null,"discovery_consumer_recall_mean":null,"discovery_control_accuracy_mean":null,"discovery_target_precision_mean":null,"discovery_target_recall_mean":null,"elapsed_coverage":{"coverage":1.0,"maximum":196.935425,"mean":169.966914,"median":160.695906,"minimum":152.269411,"observed_count":3,"observed_total":509.90074200000004,"total_count":3},"end_to_end":null,"failure_counts":{"patch_policy:policy_diff_too_large":1},"first_attempt_accepted":{"estimate":0.6666666666666666,"sample_size":3,"successes":2,"wilson":{"confidence_level":0.95,"estimate":0.6666666666666666,"lower":0.20765960080204776,"reason":null,"sample_size":3,"status":"observed","successes":2,"upper":0.9385080552796036}},"governance_safe":{"estimate":1.0,"sample_size":2,"successes":2,"wilson":{"confidence_level":0.95,"estimate":1.0,"lower":0.34238022750665326,"reason":null,"sample_size":2,"status":"observed","successes":2,"upper":1.0}},"input_token_coverage":{"coverage":1.0,"maximum":380150.0,"mean":316877.0,"median":296879.0,"minimum":273602.0,"observed_count":3,"observed_total":950631.0,"total_count":3},"integration_passed":{"estimate":1.0,"sample_size":3,"successes":3,"wilson":{"confidence_level":0.95,"estimate":1.0,"lower":0.43850296824495466,"reason":null,"sample_size":3,"status":"observed","successes":3,"upper":1.0}},"missing_targets_total":0,"output_token_coverage":{"coverage":1.0,"maximum":3868.0,"mean":3435.3333333333335,"median":3458.0,"minimum":2980.0,"observed_count":3,"observed_total":10306.0,"total_count":3},"sample_size":3,"semantic_safe":{"estimate":1.0,"sample_size":2,"successes":2,"wilson":{"confidence_level":0.95,"estimate":1.0,"lower":0.34238022750665326,"reason":null,"sample_size":2,"status":"observed","successes":2,"upper":1.0}},"succeeded":{"estimate":0.6666666666666666,"sample_size":3,"successes":2,"wilson":{"confidence_level":0.95,"estimate":0.6666666666666666,"lower":0.20765960080204776,"reason":null,"sample_size":3,"status":"observed","successes":2,"upper":0.9385080552796036}},"unnecessary_repositories_total":0,"within_budget_accepted":{"estimate":0.6666666666666666,"sample_size":3,"successes":2,"wilson":{"confidence_level":0.95,"estimate":0.6666666666666666,"lower":0.20765960080204776,"reason":null,"sample_size":3,"status":"observed","successes":2,"upper":0.9385080552796036}}},"missing_repair_targets":0,"sample_size":3,"scenario_key":"risk_assessment_v2.v1","success_count":2,"success_rate":0.6666666666666666,"tier":"primary","unnecessary_repository_modifications":0},{"condition":"repo_only_repair","false_negative_repositories":0,"false_positive_repositories":0,"metrics":{"attempts_mean":2.0,"attempts_median":2.0,"attempts_total":6,"changed_files_total":12,"changed_lines_total":47,"condition":"repo_only_repair","controls_preserved":{"estimate":1.0,"sample_size":3,"successes":3,"wilson":{"confidence_level":0.95,"estimate":1.0,"lower":0.43850296824495466,"reason":null,"sample_size":3,"status":"observed","successes":3,"upper":1.0}},"cost_coverage":{"coverage":0.0,"maximum":null,"mean":null,"median":null,"minimum":null,"observed_count":0,"observed_total":null,"total_count":3},"discovery":null,"discovery_consumer_precision_mean":null,"discovery_consumer_recall_mean":null,"discovery_control_accuracy_mean":null,"discovery_target_precision_mean":null,"discovery_target_recall_mean":null,"elapsed_coverage":{"coverage":1.0,"maximum":143.45127100000002,"mean":128.19292233333334,"median":138.205518,"minimum":102.921978,"observed_count":3,"observed_total":384.578767,"total_count":3},"end_to_end":null,"failure_counts":{},"first_attempt_accepted":{"estimate":1.0,"sample_size":3,"successes":3,"wilson":{"confidence_level":0.95,"estimate":1.0,"lower":0.43850296824495466,"reason":null,"sample_size":3,"status":"observed","successes":3,"upper":1.0}},"governance_safe":{"estimate":1.0,"sample_size":3,"successes":3,"wilson":{"confidence_level":0.95,"estimate":1.0,"lower":0.43850296824495466,"reason":null,"sample_size":3,"status":"observed","successes":3,"upper":1.0}},"input_token_coverage":{"coverage":1.0,"maximum":286978.0,"mean":262424.6666666667,"median":261984.0,"minimum":238312.0,"observed_count":3,"observed_total":787274.0,"total_count":3},"integration_passed":{"estimate":1.0,"sample_size":3,"successes":3,"wilson":{"confidence_level":0.95,"estimate":1.0,"lower":0.43850296824495466,"reason":null,"sample_size":3,"status":"observed","successes":3,"upper":1.0}},"missing_targets_total":0,"output_token_coverage":{"coverage":1.0,"maximum":2632.0,"mean":2454.6666666666665,"median":2440.0,"minimum":2292.0,"observed_count":3,"observed_total":7364.0,"total_count":3},"sample_size":3,"semantic_safe":{"estimate":1.0,"sample_size":3,"successes":3,"wilson":{"confidence_level":0.95,"estimate":1.0,"lower":0.43850296824495466,"reason":null,"sample_size":3,"status":"observed","successes":3,"upper":1.0}},"succeeded":{"estimate":1.0,"sample_size":3,"successes":3,"wilson":{"confidence_level":0.95,"estimate":1.0,"lower":0.43850296824495466,"reason":null,"sample_size":3,"status":"observed","successes":3,"upper":1.0}},"unnecessary_repositories_total":0,"within_budget_accepted":{"estimate":1.0,"sample_size":3,"successes":3,"wilson":{"confidence_level":0.95,"estimate":1.0,"lower":0.43850296824495466,"reason":null,"sample_size":3,"status":"observed","successes":3,"upper":1.0}}},"missing_repair_targets":0,"sample_size":3,"scenario_key":"risk_assessment_v2.v1","success_count":3,"success_rate":1.0,"tier":"primary","unnecessary_repository_modifications":0},{"condition":"agent_only_end_to_end","false_negative_repositories":4,"false_positive_repositories":1,"metrics":{"attempts_mean":1.6666666666666667,"attempts_median":0.0,"attempts_total":5,"changed_files_total":7,"changed_lines_total":47,"condition":"agent_only_end_to_end","controls_preserved":{"estimate":0.0,"sample_size":3,"successes":0,"wilson":{"confidence_level":0.95,"estimate":0.0,"lower":0.0,"reason":null,"sample_size":3,"status":"observed","successes":0,"upper":0.5614970317550453}},"cost_coverage":{"coverage":0.0,"maximum":null,"mean":null,"median":null,"minimum":null,"observed_count":0,"observed_total":null,"total_count":3},"discovery":{"affected_consumers":{"accuracy":0.14285714285714285,"exact_match":{"estimate":0.0,"sample_size":3,"successes":0,"wilson":{"confidence_level":0.95,"estimate":0.0,"lower":0.0,"reason":null,"sample_size":3,"status":"observed","successes":0,"upper":0.5614970317550453}},"f1":0.25,"false_negative":5,"false_positive":1,"ground_truth_size_total":6,"precision":0.5,"predicted_size_total":2,"recall":0.16666666666666666,"sample_size":3,"true_negative":0,"true_positive":1,"universe_size_total":7},"false_negative_repository_total":4,"false_positive_repository_total":1,"no_change_repositories":{"accuracy":0.5,"exact_match":{"estimate":0.0,"sample_size":3,"successes":0,"wilson":{"confidence_level":0.95,"estimate":0.0,"lower":0.0,"reason":null,"sample_size":3,"status":"observed","successes":0,"upper":0.5614970317550453}},"f1":0.4,"false_negative":9,"false_positive":0,"ground_truth_size_total":12,"precision":1.0,"predicted_size_total":3,"recall":0.25,"sample_size":3,"true_negative":6,"true_positive":3,"universe_size_total":18},"patch_targets":{"accuracy":0.7222222222222222,"exact_match":{"estimate":0.0,"sample_size":3,"successes":0,"wilson":{"confidence_level":0.95,"estimate":0.0,"lower":0.0,"reason":null,"sample_size":3,"status":"observed","successes":0,"upper":0.5614970317550453}},"f1":0.4444444444444444,"false_negative":4,"false_positive":1,"ground_truth_size_total":6,"precision":0.6666666666666666,"predicted_size_total":3,"recall":0.3333333333333333,"sample_size":3,"true_negative":11,"true_positive":2,"universe_size_total":18},"sample_size":3},"discovery_consumer_precision_mean":0.16666666666666666,"discovery_consumer_recall_mean":0.16666666666666666,"discovery_control_accuracy_mean":0.5,"discovery_target_precision_mean":0.2222222222222222,"discovery_target_recall_mean":0.3333333333333333,"elapsed_coverage":{"coverage":1.0,"maximum":503.63057100000003,"mean":224.22331933333334,"median":98.339919,"minimum":70.699468,"observed_count":3,"observed_total":672.6699580000001,"total_count":3},"end_to_end":{"all_true_consumers_protected":{"estimate":0.0,"sample_size":3,"successes":0,"wilson":{"confidence_level":0.95,"estimate":0.0,"lower":0.0,"reason":null,"sample_size":3,"status":"observed","successes":0,"upper":0.5614970317550453}},"all_true_targets_repaired":{"estimate":0.3333333333333333,"sample_size":3,"successes":1,"wilson":{"confidence_level":0.95,"estimate":0.3333333333333333,"lower":0.06149194472039624,"reason":null,"sample_size":3,"status":"observed","successes":1,"upper":0.7923403991979522}},"attempts_mean":1.6666666666666667,"attempts_median":0.0,"attempts_total":5,"complete_fleet_repaired":{"estimate":0.3333333333333333,"sample_size":3,"successes":1,"wilson":{"confidence_level":0.95,"estimate":0.3333333333333333,"lower":0.06149194472039624,"reason":null,"sample_size":3,"status":"observed","successes":1,"upper":0.7923403991979522}},"complete_integration_pass":{"estimate":0.0,"sample_size":1,"successes":0,"wilson":{"confidence_level":0.95,"estimate":0.0,"lower":0.0,"reason":null,"sample_size":1,"status":"observed","successes":0,"upper":0.7934506856227624}},"correct_target_and_correct_patch":{"estimate":0.0,"sample_size":3,"successes":0,"wilson":{"confidence_level":0.95,"estimate":0.0,"lower":0.0,"reason":null,"sample_size":3,"status":"observed","successes":0,"upper":0.5614970317550453}},"governance_safe_patchset":{"estimate":null,"sample_size":0,"successes":0,"wilson":{"confidence_level":0.95,"estimate":null,"lower":null,"reason":"no scored observations","sample_size":0,"status":"not_observable","successes":0,"upper":null}},"missing_repair_targets_total":4,"no_unnecessary_repository_modifications":{"estimate":0.6666666666666666,"sample_size":3,"successes":2,"wilson":{"confidence_level":0.95,"estimate":0.6666666666666666,"lower":0.20765960080204776,"reason":null,"sample_size":3,"status":"observed","successes":2,"upper":0.9385080552796036}},"ready_for_human_review":{"estimate":0.0,"sample_size":3,"successes":0,"wilson":{"confidence_level":0.95,"estimate":0.0,"lower":0.0,"reason":null,"sample_size":3,"status":"observed","successes":0,"upper":0.5614970317550453}},"sample_size":3,"semantic_safe_patchset":{"estimate":null,"sample_size":0,"successes":0,"wilson":{"confidence_level":0.95,"estimate":null,"lower":null,"reason":"no scored observations","sample_size":0,"status":"not_observable","successes":0,"upper":null}},"unchanged_controls_preserved":{"estimate":0.0,"sample_size":3,"successes":0,"wilson":{"confidence_level":0.95,"estimate":0.0,"lower":0.0,"reason":null,"sample_size":3,"status":"observed","successes":0,"upper":0.5614970317550453}},"unnecessary_repository_modifications_total":1},"failure_counts":{"composition:affected_consumer_mismatch":3,"composition:missing_repair_target":4,"composition:no_change_repository_mismatch":3,"composition:unnecessary_repair_target":1,"discovery:malformed_discovery_json":2,"integration:selected_no_change_repository_modified":3},"first_attempt_accepted":{"estimate":0.0,"sample_size":3,"successes":0,"wilson":{"confidence_level":0.95,"estimate":0.0,"lower":0.0,"reason":null,"sample_size":3,"status":"observed","successes":0,"upper":0.5614970317550453}},"governance_safe":{"estimate":null,"sample_size":0,"successes":0,"wilson":{"confidence_level":0.95,"estimate":null,"lower":null,"reason":"no scored observations","sample_size":0,"status":"not_observable","successes":0,"upper":null}},"input_token_coverage":{"coverage":1.0,"maximum":1005255.0,"mean":432545.3333333333,"median":188284.0,"minimum":104097.0,"observed_count":3,"observed_total":1297636.0,"total_count":3},"integration_passed":{"estimate":0.0,"sample_size":1,"successes":0,"wilson":{"confidence_level":0.95,"estimate":0.0,"lower":0.0,"reason":null,"sample_size":1,"status":"observed","successes":0,"upper":0.7934506856227624}},"missing_targets_total":4,"output_token_coverage":{"coverage":1.0,"maximum":10927.0,"mean":5634.333333333333,"median":3281.0,"minimum":2695.0,"observed_count":3,"observed_total":16903.0,"total_count":3},"sample_size":3,"semantic_safe":{"estimate":null,"sample_size":0,"successes":0,"wilson":{"confidence_level":0.95,"estimate":null,"lower":null,"reason":"no scored observations","sample_size":0,"status":"not_observable","successes":0,"upper":null}},"succeeded":{"estimate":0.0,"sample_size":3,"successes":0,"wilson":{"confidence_level":0.95,"estimate":0.0,"lower":0.0,"reason":null,"sample_size":3,"status":"observed","successes":0,"upper":0.5614970317550453}},"unnecessary_repositories_total":1,"within_budget_accepted":null},"missing_repair_targets":4,"sample_size":3,"scenario_key":"shipment_quote_v2.v1","success_count":0,"success_rate":0.0,"tier":"primary","unnecessary_repository_modifications":1},{"condition":"code_search_discovery","false_negative_repositories":4,"false_positive_repositories":1,"metrics":{"attempts_mean":null,"attempts_median":null,"attempts_total":0,"changed_files_total":0,"changed_lines_total":0,"condition":"code_search_discovery","controls_preserved":null,"cost_coverage":{"coverage":0.0,"maximum":null,"mean":null,"median":null,"minimum":null,"observed_count":0,"observed_total":null,"total_count":3},"discovery":{"affected_consumers":{"accuracy":0.14285714285714285,"exact_match":{"estimate":0.0,"sample_size":3,"successes":0,"wilson":{"confidence_level":0.95,"estimate":0.0,"lower":0.0,"reason":null,"sample_size":3,"status":"observed","successes":0,"upper":0.5614970317550453}},"f1":0.25,"false_negative":5,"false_positive":1,"ground_truth_size_total":6,"precision":0.5,"predicted_size_total":2,"recall":0.16666666666666666,"sample_size":3,"true_negative":0,"true_positive":1,"universe_size_total":7},"false_negative_repository_total":4,"false_positive_repository_total":1,"no_change_repositories":{"accuracy":0.5,"exact_match":{"estimate":0.0,"sample_size":3,"successes":0,"wilson":{"confidence_level":0.95,"estimate":0.0,"lower":0.0,"reason":null,"sample_size":3,"status":"observed","successes":0,"upper":0.5614970317550453}},"f1":0.4,"false_negative":9,"false_positive":0,"ground_truth_size_total":12,"precision":1.0,"predicted_size_total":3,"recall":0.25,"sample_size":3,"true_negative":6,"true_positive":3,"universe_size_total":18},"patch_targets":{"accuracy":0.7222222222222222,"exact_match":{"estimate":0.0,"sample_size":3,"successes":0,"wilson":{"confidence_level":0.95,"estimate":0.0,"lower":0.0,"reason":null,"sample_size":3,"status":"observed","successes":0,"upper":0.5614970317550453}},"f1":0.4444444444444444,"false_negative":4,"false_positive":1,"ground_truth_size_total":6,"precision":0.6666666666666666,"predicted_size_total":3,"recall":0.3333333333333333,"sample_size":3,"true_negative":11,"true_positive":2,"universe_size_total":18},"sample_size":3},"discovery_consumer_precision_mean":0.16666666666666666,"discovery_consumer_recall_mean":0.16666666666666666,"discovery_control_accuracy_mean":0.5,"discovery_target_precision_mean":0.2222222222222222,"discovery_target_recall_mean":0.3333333333333333,"elapsed_coverage":{"coverage":1.0,"maximum":103.129334,"mean":90.72290699999999,"median":98.339919,"minimum":70.699468,"observed_count":3,"observed_total":272.168721,"total_count":3},"end_to_end":null,"failure_counts":{"discovery:malformed_discovery_json":2},"first_attempt_accepted":null,"governance_safe":null,"input_token_coverage":{"coverage":1.0,"maximum":188284.0,"mean":159434.33333333334,"median":185922.0,"minimum":104097.0,"observed_count":3,"observed_total":478303.0,"total_count":3},"integration_passed":null,"missing_targets_total":0,"output_token_coverage":{"coverage":1.0,"maximum":3530.0,"mean":3168.6666666666665,"median":3281.0,"minimum":2695.0,"observed_count":3,"observed_total":9506.0,"total_count":3},"sample_size":3,"semantic_safe":null,"succeeded":{"estimate":0.0,"sample_size":3,"successes":0,"wilson":{"confidence_level":0.95,"estimate":0.0,"lower":0.0,"reason":null,"sample_size":3,"status":"observed","successes":0,"upper":0.5614970317550453}},"unnecessary_repositories_total":0,"within_budget_accepted":null},"missing_repair_targets":0,"sample_size":3,"scenario_key":"shipment_quote_v2.v1","success_count":0,"success_rate":0.0,"tier":"primary","unnecessary_repository_modifications":0},{"condition":"mcpatch_discovery","false_negative_repositories":0,"false_positive_repositories":0,"metrics":{"attempts_mean":null,"attempts_median":null,"attempts_total":0,"changed_files_total":0,"changed_lines_total":0,"condition":"mcpatch_discovery","controls_preserved":null,"cost_coverage":{"coverage":0.0,"maximum":null,"mean":null,"median":null,"minimum":null,"observed_count":0,"observed_total":null,"total_count":3},"discovery":{"affected_consumers":{"accuracy":1.0,"exact_match":{"estimate":1.0,"sample_size":3,"successes":3,"wilson":{"confidence_level":0.95,"estimate":1.0,"lower":0.43850296824495466,"reason":null,"sample_size":3,"status":"observed","successes":3,"upper":1.0}},"f1":1.0,"false_negative":0,"false_positive":0,"ground_truth_size_total":6,"precision":1.0,"predicted_size_total":6,"recall":1.0,"sample_size":3,"true_negative":0,"true_positive":6,"universe_size_total":6},"false_negative_repository_total":0,"false_positive_repository_total":0,"no_change_repositories":{"accuracy":1.0,"exact_match":{"estimate":1.0,"sample_size":3,"successes":3,"wilson":{"confidence_level":0.95,"estimate":1.0,"lower":0.43850296824495466,"reason":null,"sample_size":3,"status":"observed","successes":3,"upper":1.0}},"f1":1.0,"false_negative":0,"false_positive":0,"ground_truth_size_total":12,"precision":1.0,"predicted_size_total":12,"recall":1.0,"sample_size":3,"true_negative":6,"true_positive":12,"universe_size_total":18},"patch_targets":{"accuracy":1.0,"exact_match":{"estimate":1.0,"sample_size":3,"successes":3,"wilson":{"confidence_level":0.95,"estimate":1.0,"lower":0.43850296824495466,"reason":null,"sample_size":3,"status":"observed","successes":3,"upper":1.0}},"f1":1.0,"false_negative":0,"false_positive":0,"ground_truth_size_total":6,"precision":1.0,"predicted_size_total":6,"recall":1.0,"sample_size":3,"true_negative":12,"true_positive":6,"universe_size_total":18},"sample_size":3},"discovery_consumer_precision_mean":1.0,"discovery_consumer_recall_mean":1.0,"discovery_control_accuracy_mean":1.0,"discovery_target_precision_mean":1.0,"discovery_target_recall_mean":1.0,"elapsed_coverage":{"coverage":1.0,"maximum":2e-06,"mean":1.6666666666666667e-06,"median":2e-06,"minimum":1e-06,"observed_count":3,"observed_total":4.9999999999999996e-06,"total_count":3},"end_to_end":null,"failure_counts":{},"first_attempt_accepted":null,"governance_safe":null,"input_token_coverage":{"coverage":0.0,"maximum":null,"mean":null,"median":null,"minimum":null,"observed_count":0,"observed_total":null,"total_count":3},"integration_passed":null,"missing_targets_total":0,"output_token_coverage":{"coverage":0.0,"maximum":null,"mean":null,"median":null,"minimum":null,"observed_count":0,"observed_total":null,"total_count":3},"sample_size":3,"semantic_safe":null,"succeeded":{"estimate":1.0,"sample_size":3,"successes":3,"wilson":{"confidence_level":0.95,"estimate":1.0,"lower":0.43850296824495466,"reason":null,"sample_size":3,"status":"observed","successes":3,"upper":1.0}},"unnecessary_repositories_total":0,"within_budget_accepted":null},"missing_repair_targets":0,"sample_size":3,"scenario_key":"shipment_quote_v2.v1","success_count":3,"success_rate":1.0,"tier":"primary","unnecessary_repository_modifications":0},{"condition":"mcpatch_end_to_end","false_negative_repositories":0,"false_positive_repositories":0,"metrics":{"attempts_mean":2.3333333333333335,"attempts_median":2.0,"attempts_total":7,"changed_files_total":14,"changed_lines_total":95,"condition":"mcpatch_end_to_end","controls_preserved":{"estimate":1.0,"sample_size":3,"successes":3,"wilson":{"confidence_level":0.95,"estimate":1.0,"lower":0.43850296824495466,"reason":null,"sample_size":3,"status":"observed","successes":3,"upper":1.0}},"cost_coverage":{"coverage":0.0,"maximum":null,"mean":null,"median":null,"minimum":null,"observed_count":0,"observed_total":null,"total_count":3},"discovery":{"affected_consumers":{"accuracy":1.0,"exact_match":{"estimate":1.0,"sample_size":3,"successes":3,"wilson":{"confidence_level":0.95,"estimate":1.0,"lower":0.43850296824495466,"reason":null,"sample_size":3,"status":"observed","successes":3,"upper":1.0}},"f1":1.0,"false_negative":0,"false_positive":0,"ground_truth_size_total":6,"precision":1.0,"predicted_size_total":6,"recall":1.0,"sample_size":3,"true_negative":0,"true_positive":6,"universe_size_total":6},"false_negative_repository_total":0,"false_positive_repository_total":0,"no_change_repositories":{"accuracy":1.0,"exact_match":{"estimate":1.0,"sample_size":3,"successes":3,"wilson":{"confidence_level":0.95,"estimate":1.0,"lower":0.43850296824495466,"reason":null,"sample_size":3,"status":"observed","successes":3,"upper":1.0}},"f1":1.0,"false_negative":0,"false_positive":0,"ground_truth_size_total":12,"precision":1.0,"predicted_size_total":12,"recall":1.0,"sample_size":3,"true_negative":6,"true_positive":12,"universe_size_total":18},"patch_targets":{"accuracy":1.0,"exact_match":{"estimate":1.0,"sample_size":3,"successes":3,"wilson":{"confidence_level":0.95,"estimate":1.0,"lower":0.43850296824495466,"reason":null,"sample_size":3,"status":"observed","successes":3,"upper":1.0}},"f1":1.0,"false_negative":0,"false_positive":0,"ground_truth_size_total":6,"precision":1.0,"predicted_size_total":6,"recall":1.0,"sample_size":3,"true_negative":12,"true_positive":6,"universe_size_total":18},"sample_size":3},"discovery_consumer_precision_mean":1.0,"discovery_consumer_recall_mean":1.0,"discovery_control_accuracy_mean":1.0,"discovery_target_precision_mean":1.0,"discovery_target_recall_mean":1.0,"elapsed_coverage":{"coverage":1.0,"maximum":227.89794,"mean":181.97571933333333,"median":167.954838,"minimum":150.07438000000002,"observed_count":3,"observed_total":545.927158,"total_count":3},"end_to_end":{"all_true_consumers_protected":{"estimate":1.0,"sample_size":3,"successes":3,"wilson":{"confidence_level":0.95,"estimate":1.0,"lower":0.43850296824495466,"reason":null,"sample_size":3,"status":"observed","successes":3,"upper":1.0}},"all_true_targets_repaired":{"estimate":1.0,"sample_size":3,"successes":3,"wilson":{"confidence_level":0.95,"estimate":1.0,"lower":0.43850296824495466,"reason":null,"sample_size":3,"status":"observed","successes":3,"upper":1.0}},"attempts_mean":2.3333333333333335,"attempts_median":2.0,"attempts_total":7,"complete_fleet_repaired":{"estimate":1.0,"sample_size":3,"successes":3,"wilson":{"confidence_level":0.95,"estimate":1.0,"lower":0.43850296824495466,"reason":null,"sample_size":3,"status":"observed","successes":3,"upper":1.0}},"complete_integration_pass":{"estimate":1.0,"sample_size":3,"successes":3,"wilson":{"confidence_level":0.95,"estimate":1.0,"lower":0.43850296824495466,"reason":null,"sample_size":3,"status":"observed","successes":3,"upper":1.0}},"correct_target_and_correct_patch":{"estimate":1.0,"sample_size":3,"successes":3,"wilson":{"confidence_level":0.95,"estimate":1.0,"lower":0.43850296824495466,"reason":null,"sample_size":3,"status":"observed","successes":3,"upper":1.0}},"governance_safe_patchset":{"estimate":1.0,"sample_size":3,"successes":3,"wilson":{"confidence_level":0.95,"estimate":1.0,"lower":0.43850296824495466,"reason":null,"sample_size":3,"status":"observed","successes":3,"upper":1.0}},"missing_repair_targets_total":0,"no_unnecessary_repository_modifications":{"estimate":1.0,"sample_size":3,"successes":3,"wilson":{"confidence_level":0.95,"estimate":1.0,"lower":0.43850296824495466,"reason":null,"sample_size":3,"status":"observed","successes":3,"upper":1.0}},"ready_for_human_review":{"estimate":1.0,"sample_size":3,"successes":3,"wilson":{"confidence_level":0.95,"estimate":1.0,"lower":0.43850296824495466,"reason":null,"sample_size":3,"status":"observed","successes":3,"upper":1.0}},"sample_size":3,"semantic_safe_patchset":{"estimate":1.0,"sample_size":3,"successes":3,"wilson":{"confidence_level":0.95,"estimate":1.0,"lower":0.43850296824495466,"reason":null,"sample_size":3,"status":"observed","successes":3,"upper":1.0}},"unchanged_controls_preserved":{"estimate":1.0,"sample_size":3,"successes":3,"wilson":{"confidence_level":0.95,"estimate":1.0,"lower":0.43850296824495466,"reason":null,"sample_size":3,"status":"observed","successes":3,"upper":1.0}},"unnecessary_repository_modifications_total":0},"failure_counts":{"contract:contract_required_output":1},"first_attempt_accepted":{"estimate":0.6666666666666666,"sample_size":3,"successes":2,"wilson":{"confidence_level":0.95,"estimate":0.6666666666666666,"lower":0.20765960080204776,"reason":null,"sample_size":3,"status":"observed","successes":2,"upper":0.9385080552796036}},"governance_safe":{"estimate":1.0,"sample_size":3,"successes":3,"wilson":{"confidence_level":0.95,"estimate":1.0,"lower":0.43850296824495466,"reason":null,"sample_size":3,"status":"observed","successes":3,"upper":1.0}},"input_token_coverage":{"coverage":0.0,"maximum":null,"mean":null,"median":null,"minimum":null,"observed_count":0,"observed_total":null,"total_count":3},"integration_passed":{"estimate":1.0,"sample_size":3,"successes":3,"wilson":{"confidence_level":0.95,"estimate":1.0,"lower":0.43850296824495466,"reason":null,"sample_size":3,"status":"observed","successes":3,"upper":1.0}},"missing_targets_total":0,"output_token_coverage":{"coverage":0.0,"maximum":null,"mean":null,"median":null,"minimum":null,"observed_count":0,"observed_total":null,"total_count":3},"sample_size":3,"semantic_safe":{"estimate":1.0,"sample_size":3,"successes":3,"wilson":{"confidence_level":0.95,"estimate":1.0,"lower":0.43850296824495466,"reason":null,"sample_size":3,"status":"observed","successes":3,"upper":1.0}},"succeeded":{"estimate":1.0,"sample_size":3,"successes":3,"wilson":{"confidence_level":0.95,"estimate":1.0,"lower":0.43850296824495466,"reason":null,"sample_size":3,"status":"observed","successes":3,"upper":1.0}},"unnecessary_repositories_total":0,"within_budget_accepted":null},"missing_repair_targets":0,"sample_size":3,"scenario_key":"shipment_quote_v2.v1","success_count":3,"success_rate":1.0,"tier":"primary","unnecessary_repository_modifications":0},{"condition":"mcpatch_guided_repair","false_negative_repositories":0,"false_positive_repositories":0,"metrics":{"attempts_mean":2.6666666666666665,"attempts_median":2.0,"attempts_total":8,"changed_files_total":16,"changed_lines_total":117,"condition":"mcpatch_guided_repair","controls_preserved":{"estimate":1.0,"sample_size":3,"successes":3,"wilson":{"confidence_level":0.95,"estimate":1.0,"lower":0.43850296824495466,"reason":null,"sample_size":3,"status":"observed","successes":3,"upper":1.0}},"cost_coverage":{"coverage":0.0,"maximum":null,"mean":null,"median":null,"minimum":null,"observed_count":0,"observed_total":null,"total_count":3},"discovery":null,"discovery_consumer_precision_mean":null,"discovery_consumer_recall_mean":null,"discovery_control_accuracy_mean":null,"discovery_target_precision_mean":null,"discovery_target_recall_mean":null,"elapsed_coverage":{"coverage":1.0,"maximum":354.107226,"mean":238.14710666666667,"median":209.005067,"minimum":151.329027,"observed_count":3,"observed_total":714.44132,"total_count":3},"end_to_end":null,"failure_counts":{"contract:contract_required_output":2},"first_attempt_accepted":{"estimate":0.6666666666666666,"sample_size":3,"successes":2,"wilson":{"confidence_level":0.95,"estimate":0.6666666666666666,"lower":0.20765960080204776,"reason":null,"sample_size":3,"status":"observed","successes":2,"upper":0.9385080552796036}},"governance_safe":{"estimate":1.0,"sample_size":3,"successes":3,"wilson":{"confidence_level":0.95,"estimate":1.0,"lower":0.43850296824495466,"reason":null,"sample_size":3,"status":"observed","successes":3,"upper":1.0}},"input_token_coverage":{"coverage":1.0,"maximum":758193.0,"mean":489259.3333333333,"median":405157.0,"minimum":304428.0,"observed_count":3,"observed_total":1467778.0,"total_count":3},"integration_passed":{"estimate":1.0,"sample_size":3,"successes":3,"wilson":{"confidence_level":0.95,"estimate":1.0,"lower":0.43850296824495466,"reason":null,"sample_size":3,"status":"observed","successes":3,"upper":1.0}},"missing_targets_total":0,"output_token_coverage":{"coverage":1.0,"maximum":8155.0,"mean":5323.0,"median":4649.0,"minimum":3165.0,"observed_count":3,"observed_total":15969.0,"total_count":3},"sample_size":3,"semantic_safe":{"estimate":1.0,"sample_size":3,"successes":3,"wilson":{"confidence_level":0.95,"estimate":1.0,"lower":0.43850296824495466,"reason":null,"sample_size":3,"status":"observed","successes":3,"upper":1.0}},"succeeded":{"estimate":1.0,"sample_size":3,"successes":3,"wilson":{"confidence_level":0.95,"estimate":1.0,"lower":0.43850296824495466,"reason":null,"sample_size":3,"status":"observed","successes":3,"upper":1.0}},"unnecessary_repositories_total":0,"within_budget_accepted":{"estimate":1.0,"sample_size":3,"successes":3,"wilson":{"confidence_level":0.95,"estimate":1.0,"lower":0.43850296824495466,"reason":null,"sample_size":3,"status":"observed","successes":3,"upper":1.0}}},"missing_repair_targets":0,"sample_size":3,"scenario_key":"shipment_quote_v2.v1","success_count":3,"success_rate":1.0,"tier":"primary","unnecessary_repository_modifications":0},{"condition":"repo_only_repair","false_negative_repositories":0,"false_positive_repositories":0,"metrics":{"attempts_mean":2.0,"attempts_median":2.0,"attempts_total":6,"changed_files_total":12,"changed_lines_total":54,"condition":"repo_only_repair","controls_preserved":{"estimate":1.0,"sample_size":3,"successes":3,"wilson":{"confidence_level":0.95,"estimate":1.0,"lower":0.43850296824495466,"reason":null,"sample_size":3,"status":"observed","successes":3,"upper":1.0}},"cost_coverage":{"coverage":0.0,"maximum":null,"mean":null,"median":null,"minimum":null,"observed_count":0,"observed_total":null,"total_count":3},"discovery":null,"discovery_consumer_precision_mean":null,"discovery_consumer_recall_mean":null,"discovery_control_accuracy_mean":null,"discovery_target_precision_mean":null,"discovery_target_recall_mean":null,"elapsed_coverage":{"coverage":1.0,"maximum":189.72650900000002,"mean":159.39759266666667,"median":153.598745,"minimum":134.867524,"observed_count":3,"observed_total":478.19277800000003,"total_count":3},"end_to_end":null,"failure_counts":{},"first_attempt_accepted":{"estimate":1.0,"sample_size":3,"successes":3,"wilson":{"confidence_level":0.95,"estimate":1.0,"lower":0.43850296824495466,"reason":null,"sample_size":3,"status":"observed","successes":3,"upper":1.0}},"governance_safe":{"estimate":1.0,"sample_size":3,"successes":3,"wilson":{"confidence_level":0.95,"estimate":1.0,"lower":0.43850296824495466,"reason":null,"sample_size":3,"status":"observed","successes":3,"upper":1.0}},"input_token_coverage":{"coverage":1.0,"maximum":316311.0,"mean":291195.6666666667,"median":293433.0,"minimum":263843.0,"observed_count":3,"observed_total":873587.0,"total_count":3},"integration_passed":{"estimate":1.0,"sample_size":3,"successes":3,"wilson":{"confidence_level":0.95,"estimate":1.0,"lower":0.43850296824495466,"reason":null,"sample_size":3,"status":"observed","successes":3,"upper":1.0}},"missing_targets_total":0,"output_token_coverage":{"coverage":1.0,"maximum":3619.0,"mean":2997.6666666666665,"median":3052.0,"minimum":2322.0,"observed_count":3,"observed_total":8993.0,"total_count":3},"sample_size":3,"semantic_safe":{"estimate":1.0,"sample_size":3,"successes":3,"wilson":{"confidence_level":0.95,"estimate":1.0,"lower":0.43850296824495466,"reason":null,"sample_size":3,"status":"observed","successes":3,"upper":1.0}},"succeeded":{"estimate":1.0,"sample_size":3,"successes":3,"wilson":{"confidence_level":0.95,"estimate":1.0,"lower":0.43850296824495466,"reason":null,"sample_size":3,"status":"observed","successes":3,"upper":1.0}},"unnecessary_repositories_total":0,"within_budget_accepted":{"estimate":1.0,"sample_size":3,"successes":3,"wilson":{"confidence_level":0.95,"estimate":1.0,"lower":0.43850296824495466,"reason":null,"sample_size":3,"status":"observed","successes":3,"upper":1.0}}},"missing_repair_targets":0,"sample_size":3,"scenario_key":"shipment_quote_v2.v1","success_count":3,"success_rate":1.0,"tier":"primary","unnecessary_repository_modifications":0}],"schema_version":1}
