{
  "audit_json": {
    "artifact_resolution": {
      "candidate_run_dirs": [
        "evidences"
      ],
      "manifest_path_mismatch": false,
      "primary_failure_code": null,
      "resolution_mode": "workspace_evidence_root",
      "scanned_scopes": [
        {
          "candidate_count": 1,
          "candidates": [
            {
              "actual_cluster_id": "10748116073341772391",
              "actual_end_ts": 1784603100,
              "actual_start_ts": 1784602800,
              "analysis_patch_path": "",
              "analysis_patch_present": false,
              "diagnosis_report_path": "evidences/workspace_evidence_report.json",
              "diagnosis_report_present": true,
              "evidence_dir": "evidences",
              "evidence_dir_present": true,
              "invalid_reasons": [],
              "logs_dir": "",
              "logs_dir_present": false,
              "manifest_path": "",
              "manifest_path_mismatch": false,
              "manifest_path_mismatch_fields": [],
              "manifest_present": false,
              "manifest_valid_json": true,
              "resolution_mode": "workspace_evidence_root",
              "run_dir": "evidences",
              "summary_json_path": "",
              "summary_json_present": false,
              "summary_txt_path": "",
              "summary_txt_present": false,
              "valid_bundle": true
            }
          ],
          "decision": "resolved",
          "selected": true,
          "source_kind": "audit_dir",
          "source_root": "/data/eval_audits/evalrun-dca9124193b04b2c873027cd9fb2ca6f/evalcase-1dd5f72ba4a944f09e48502ce55f1652",
          "telemetry_path": ".opencode/telemetry.log",
          "valid_candidate_count": 1
        },
        {
          "candidate_count": 1,
          "candidates": [
            {
              "actual_cluster_id": "10748116073341772391",
              "actual_end_ts": 1784603100,
              "actual_start_ts": 1784602800,
              "analysis_patch_path": "",
              "analysis_patch_present": false,
              "diagnosis_report_path": "evidences/workspace_evidence_report.json",
              "diagnosis_report_present": true,
              "evidence_dir": "evidences",
              "evidence_dir_present": true,
              "invalid_reasons": [],
              "logs_dir": "",
              "logs_dir_present": false,
              "manifest_path": "",
              "manifest_path_mismatch": false,
              "manifest_path_mismatch_fields": [],
              "manifest_present": false,
              "manifest_valid_json": true,
              "resolution_mode": "workspace_evidence_root",
              "run_dir": "evidences",
              "summary_json_path": "",
              "summary_json_present": false,
              "summary_txt_path": "",
              "summary_txt_present": false,
              "valid_bundle": true
            }
          ],
          "decision": "skipped_lower_priority",
          "selected": false,
          "source_kind": "workspace_dir",
          "source_root": "/data/workspaces/ws-eval-1784618267343-4a2d5ff3",
          "telemetry_path": ".opencode/telemetry.log",
          "valid_candidate_count": 1
        }
      ],
      "selected_candidate": {
        "actual_cluster_id": "10748116073341772391",
        "actual_end_ts": 1784603100,
        "actual_start_ts": 1784602800,
        "analysis_patch_path": "",
        "analysis_patch_present": false,
        "diagnosis_report_path": "evidences/workspace_evidence_report.json",
        "diagnosis_report_present": true,
        "evidence_dir": "evidences",
        "evidence_dir_present": true,
        "invalid_reasons": [],
        "logs_dir": "",
        "logs_dir_present": false,
        "manifest_path": "",
        "manifest_path_mismatch": false,
        "manifest_path_mismatch_fields": [],
        "manifest_present": false,
        "manifest_valid_json": true,
        "resolution_mode": "workspace_evidence_root",
        "run_dir": "evidences",
        "summary_json_path": "",
        "summary_json_present": false,
        "summary_txt_path": "",
        "summary_txt_present": false,
        "valid_bundle": true
      },
      "selected_scope": "audit_dir",
      "selected_trace_candidate": {
        "actual_cluster_id": "10748116073341772391",
        "actual_end_ts": 1784603100,
        "actual_start_ts": 1784602800,
        "analysis_patch_path": "",
        "analysis_patch_present": false,
        "diagnosis_report_path": "evidences/workspace_evidence_report.json",
        "diagnosis_report_present": true,
        "evidence_dir": "evidences",
        "evidence_dir_present": true,
        "invalid_reasons": [],
        "logs_dir": "",
        "logs_dir_present": false,
        "manifest_path": "",
        "manifest_path_mismatch": false,
        "manifest_path_mismatch_fields": [],
        "manifest_present": false,
        "manifest_valid_json": true,
        "resolution_mode": "workspace_evidence_root",
        "run_dir": "evidences",
        "summary_json_path": "",
        "summary_json_present": false,
        "summary_txt_path": "",
        "summary_txt_present": false,
        "valid_bundle": true
      },
      "status": "resolved"
    },
    "case_id": "top-cpu-overload-nodes-and-log-patterns-in-premium-cluster-during-latest-window",
    "contract": {
      "actual_cluster_id": "10748116073341772391",
      "cluster_scope_match": true,
      "expected_allowed_window_id": "bjt-2026-07-21-1100-5m",
      "expected_cluster_id": "10748116073341772391",
      "expected_parent_id": "2026948324023603200",
      "time_window_match": true
    },
    "critical_evidence_verdict": "partial",
    "evidence_collection_verdict": "pass",
    "evidence_routing_verdict": "pass",
    "execution_contract": {
      "artifact_facts": {
        "analysis_patch_present": false,
        "core_evidence_bundle_present": false,
        "diagnosis_report_present": true,
        "summary_artifact_present": false,
        "supplemental_loki_present": false
      },
      "channel_facts": {
        "events": true,
        "logs": true,
        "query_or_metric": true,
        "ticket_context": false
      },
      "checks": [
        {
          "description": "The minimal ad-hoc workflow is present.",
          "id": "required_workflow_present",
          "match": "all_of",
          "passed": true,
          "required": true,
          "rules": [
            {
              "matched": true,
              "matched_keys": [
                "final_answer_emitted"
              ],
              "matcher": "subset_observed",
              "path": "",
              "paths": [
                "final_answer_emitted"
              ],
              "source": "workflow_check",
              "threshold": null
            }
          ],
          "snippet": "matched execution contract facts: final_answer_emitted"
        },
        {
          "description": "This ad-hoc case must not invoke collect_evidence.py.",
          "id": "collect_evidence_not_used",
          "match": "all_of",
          "passed": true,
          "required": true,
          "rules": [
            {
              "matched": true,
              "matched_keys": [
                "collect_evidence_not_used"
              ],
              "matcher": "subset_observed",
              "path": "",
              "paths": [
                "collect_evidence_not_used"
              ],
              "source": "workflow_check",
              "threshold": null
            }
          ],
          "snippet": "matched execution contract facts: collect_evidence_not_used"
        },
        {
          "description": "Workspace diagnosis report is present.",
          "id": "required_report_present",
          "match": "all_of",
          "passed": true,
          "required": true,
          "rules": [
            {
              "matched": true,
              "matched_keys": [
                "diagnosis_report_present"
              ],
              "matcher": "subset_observed",
              "path": "",
              "paths": [
                "diagnosis_report_present"
              ],
              "source": "artifact_fact",
              "threshold": null
            }
          ],
          "snippet": "matched execution contract facts: diagnosis_report_present"
        },
        {
          "description": "At least one query or metric evidence channel is observed.",
          "id": "query_or_metric_channel_observed",
          "match": "all_of",
          "passed": true,
          "required": true,
          "rules": [
            {
              "matched": true,
              "matched_keys": [
                "query_or_metric"
              ],
              "matcher": "count_gte",
              "path": "",
              "paths": [
                "query_or_metric"
              ],
              "source": "channel_fact",
              "threshold": 1.0
            }
          ],
          "snippet": "matched execution contract facts: query_or_metric"
        }
      ],
      "status": "pass",
      "workflow_facts": {
        "collect_evidence_not_used": true,
        "final_answer_emitted": true
      }
    },
    "execution_contract_verdict": "pass",
    "overall_verdict": "fail",
    "quality_verdict": "pass",
    "semantic_shadow": {
      "available": false,
      "status": "not_available"
    },
    "source": "deterministic_judge",
    "strategy": "deterministic_trace_heavy_v1",
    "workflow_verdict": "pass"
  },
  "facts_json": {
    "artifact_observations": {
      "analysis_patch_present": false,
      "diagnosis_report_present": true,
      "summary_artifact_present": false,
      "supplemental_loki_present": false
    },
    "artifacts": {
      "analysis_patch_path": "",
      "diagnosis_report_path": "evidences/workspace_evidence_report.json",
      "evidence_dir": "evidences",
      "logs_dir": "",
      "manifest_path": "",
      "manifest_path_mismatch": false,
      "resolution_mode": "workspace_evidence_root",
      "run_dir": "evidences",
      "source_root": "/data/eval_audits/evalrun-dca9124193b04b2c873027cd9fb2ca6f/evalcase-1dd5f72ba4a944f09e48502ce55f1652",
      "source_scope": "audit_dir",
      "summary_txt_path": "",
      "telemetry_path": ".opencode/telemetry.log"
    },
    "case_id": "top-cpu-overload-nodes-and-log-patterns-in-premium-cluster-during-latest-window",
    "cluster": {
      "actual_cluster_id": "10748116073341772391",
      "allowed_cluster_ids": [
        "10748116073341772391",
        "2026948324023603200"
      ],
      "expected_cluster_id": "10748116073341772391",
      "expected_parent_id": "2026948324023603200",
      "matched": true
    },
    "contract": {
      "actual_cluster_id": "10748116073341772391",
      "cluster_scope_match": true,
      "expected_allowed_window_id": "bjt-2026-07-21-1100-5m",
      "expected_cluster_id": "10748116073341772391",
      "expected_parent_id": "2026948324023603200",
      "time_window_match": true
    },
    "critical_evidence": {
      "configured": true,
      "matched_check_ids": [
        "output_cluster_id_matches_reference",
        "output_window_start_matches_reference",
        "output_window_end_matches_reference"
      ]
    },
    "evidence": {
      "analysis_patch_path": "",
      "core_evidence_bundle_present": false,
      "decoded_plan_indexes_used": null,
      "events_attempted_implies_artifact_present": true,
      "events_channel_observed": true,
      "events_collection_attempted": false,
      "events_output_file_present": false,
      "hotspot_detected": false,
      "infra_clear": false,
      "kill_query_signal_present": false,
      "log_error_counts": {},
      "logs_channel_observed": true,
      "logs_dir_present": false,
      "memory_tracker_panic_present": false,
      "metric_evidence": [
        {
          "chart_data_ref": {
            "cluster_id": "10748116073341772391",
            "end": 1784603100,
            "kind": "promql_range",
            "promql": "sum by (pod, instance) (rate(process_cpu_seconds_total{component=\"tidb\"}[1m]))",
            "start": 1784602800,
            "step": "60"
          },
          "group": "tidb_cpu_rate",
          "id": "tidb_cpu_rate",
          "metric_group_id": "tidb_cpu_rate",
          "name": "tidb_cpu_rate",
          "promql": "sum by (pod, instance) (rate(process_cpu_seconds_total{component=\"tidb\"}[1m]))",
          "ref": {
            "metric_group_id": "tidb_cpu_rate",
            "type": "metric"
          },
          "series_summaries": [
            {
              "avg": 7.966333333340784,
              "labels": {
                "instance": "db-2027220408993841152-tidb-2j6dju",
                "pod": "db-2027220408993841152-tidb-2j6dju"
              },
              "max": 8.369666666661699,
              "min": 7.3980000000447035,
              "point_count": 6
            },
            {
              "avg": 8.774000000001656,
              "labels": {
                "instance": "db-2027220408993841152-tidb-4pf3vl",
                "pod": "db-2027220408993841152-tidb-4pf3vl"
              },
              "max": 9.318166666726272,
              "min": 7.972666666656733,
              "point_count": 6
            },
            {
              "avg": 10.021388888897167,
              "labels": {
                "instance": "db-2027220408993841152-tidb-79yuza",
                "pod": "db-2027220408993841152-tidb-79yuza"
              },
              "max": 10.44249999995033,
              "min": 9.745166666681568,
              "point_count": 6
            },
            {
              "avg": 8.719138888869848,
              "labels": {
                "instance": "db-2027220408993841152-tidb-7gleep",
                "pod": "db-2027220408993841152-tidb-7gleep"
              },
              "max": 9.053333333258827,
              "min": 7.973999999960264,
              "point_count": 6
            },
            {
              "avg": 7.464944444451895,
              "labels": {
                "instance": "db-2027220408993841152-tidb-7qc8pe",
                "pod": "db-2027220408993841152-tidb-7qc8pe"
              },
              "max": 8.640666666626931,
              "min": 6.771333333229025,
              "point_count": 6
            },
            {
              "avg": 9.44808333331926,
              "labels": {
                "instance": "db-2027220408993841152-tidb-7sixuh",
                "pod": "db-2027220408993841152-tidb-7sixuh"
              },
              "max": 10.76116666669647,
              "min": 7.906833333273728,
              "point_count": 6
            },
            {
              "avg": 9.567305555546449,
              "labels": {
                "instance": "db-2027220408993841152-tidb-glddbi",
                "pod": "db-2027220408993841152-tidb-glddbi"
              },
              "max": 11.122666666656732,
              "min": 8.679833333318433,
              "point_count": 6
            },
            {
              "avg": 10.150888888889716,
              "labels": {
                "instance": "db-2027220408993841152-tidb-jvlne6",
                "pod": "db-2027220408993841152-tidb-jvlne6"
              },
              "max": 10.511666666716337,
              "min": 9.587333333368102,
              "point_count": 6
            },
            {
              "avg": 8.669805555563007,
              "labels": {
                "instance": "db-2027220408993841152-tidb-krx4fy",
                "pod": "db-2027220408993841152-tidb-krx4fy"
              },
              "max": 9.755499999970198,
              "min": 7.7738333333283665,
              "point_count": 6
            },
            {
              "avg": 8.147833333330022,
              "labels": {
                "instance": "db-2027220408993841152-tidb-p1s0s2",
                "pod": "db-2027220408993841152-tidb-p1s0s2"
              },
              "max": 8.822000000004968,
              "min": 7.711999999980132,
              "point_count": 6
            }
          ],
          "summary": {
            "avg": 8.89297222222098,
            "max": 11.122666666656732,
            "min": 6.771333333229025,
            "point_count": 60,
            "series_count": 10
          },
          "time_window": {
            "end": 1784603100,
            "start": 1784602800,
            "step": "60"
          },
          "unit": "cores"
        },
        {
          "chart_data_ref": {
            "cluster_id": "10748116073341772391",
            "end": 1784603100,
            "kind": "promql_range",
            "promql": "max by (pod, instance) (tidb_server_maxprocs)",
            "start": 1784602800,
            "step": "60"
          },
          "group": "tidb_cpu_quota",
          "id": "tidb_cpu_quota",
          "metric_group_id": "tidb_cpu_quota",
          "name": "tidb_cpu_quota",
          "promql": "max by (pod, instance) (tidb_server_maxprocs)",
          "ref": {
            "metric_group_id": "tidb_cpu_quota",
            "type": "metric"
          },
          "series_summaries": [
            {
              "avg": 31.0,
              "labels": {
                "instance": "db-2027220408993841152-tidb-2j6dju",
                "pod": "db-2027220408993841152-tidb-2j6dju"
              },
              "max": 31.0,
              "min": 31.0,
              "point_count": 6
            },
            {
              "avg": 31.0,
              "labels": {
                "instance": "db-2027220408993841152-tidb-4pf3vl",
                "pod": "db-2027220408993841152-tidb-4pf3vl"
              },
              "max": 31.0,
              "min": 31.0,
              "point_count": 6
            },
            {
              "avg": 31.0,
              "labels": {
                "instance": "db-2027220408993841152-tidb-79yuza",
                "pod": "db-2027220408993841152-tidb-79yuza"
              },
              "max": 31.0,
              "min": 31.0,
              "point_count": 6
            },
            {
              "avg": 31.0,
              "labels": {
                "instance": "db-2027220408993841152-tidb-7gleep",
                "pod": "db-2027220408993841152-tidb-7gleep"
              },
              "max": 31.0,
              "min": 31.0,
              "point_count": 6
            },
            {
              "avg": 31.0,
              "labels": {
                "instance": "db-2027220408993841152-tidb-7qc8pe",
                "pod": "db-2027220408993841152-tidb-7qc8pe"
              },
              "max": 31.0,
              "min": 31.0,
              "point_count": 6
            },
            {
              "avg": 31.0,
              "labels": {
                "instance": "db-2027220408993841152-tidb-7sixuh",
                "pod": "db-2027220408993841152-tidb-7sixuh"
              },
              "max": 31.0,
              "min": 31.0,
              "point_count": 6
            },
            {
              "avg": 31.0,
              "labels": {
                "instance": "db-2027220408993841152-tidb-glddbi",
                "pod": "db-2027220408993841152-tidb-glddbi"
              },
              "max": 31.0,
              "min": 31.0,
              "point_count": 6
            },
            {
              "avg": 31.0,
              "labels": {
                "instance": "db-2027220408993841152-tidb-jvlne6",
                "pod": "db-2027220408993841152-tidb-jvlne6"
              },
              "max": 31.0,
              "min": 31.0,
              "point_count": 6
            },
            {
              "avg": 31.0,
              "labels": {
                "instance": "db-2027220408993841152-tidb-krx4fy",
                "pod": "db-2027220408993841152-tidb-krx4fy"
              },
              "max": 31.0,
              "min": 31.0,
              "point_count": 6
            },
            {
              "avg": 31.0,
              "labels": {
                "instance": "db-2027220408993841152-tidb-p1s0s2",
                "pod": "db-2027220408993841152-tidb-p1s0s2"
              },
              "max": 31.0,
              "min": 31.0,
              "point_count": 6
            }
          ],
          "summary": {
            "avg": 31.0,
            "max": 31.0,
            "min": 31.0,
            "point_count": 60,
            "series_count": 10
          },
          "time_window": {
            "end": 1784603100,
            "start": 1784602800,
            "step": "60"
          },
          "unit": "cores"
        },
        {
          "chart_data_ref": {
            "cluster_id": "10748116073341772391",
            "end": 1784603100,
            "kind": "promql_range",
            "promql": "sum by (pod, instance) (rate(process_cpu_seconds_total{component=\"tikv\"}[1m]))",
            "start": 1784602800,
            "step": "60"
          },
          "group": "tikv_cpu_rate",
          "id": "tikv_cpu_rate",
          "metric_group_id": "tikv_cpu_rate",
          "name": "tikv_cpu_rate",
          "promql": "sum by (pod, instance) (rate(process_cpu_seconds_total{component=\"tikv\"}[1m]))",
          "ref": {
            "metric_group_id": "tikv_cpu_rate",
            "type": "metric"
          },
          "series_summaries": [],
          "summary": {
            "point_count": 0,
            "series_count": 0
          },
          "time_window": {
            "end": 1784603100,
            "start": 1784602800,
            "step": "60"
          },
          "unit": "cores"
        },
        {
          "chart_data_ref": {
            "cluster_id": "10748116073341772391",
            "end": 1784603100,
            "kind": "promql_range",
            "promql": "max by (pod, instance) (tikv_server_cpu_cores_quota)",
            "start": 1784602800,
            "step": "60"
          },
          "group": "tikv_cpu_quota",
          "id": "tikv_cpu_quota",
          "metric_group_id": "tikv_cpu_quota",
          "name": "tikv_cpu_quota",
          "promql": "max by (pod, instance) (tikv_server_cpu_cores_quota)",
          "ref": {
            "metric_group_id": "tikv_cpu_quota",
            "type": "metric"
          },
          "series_summaries": [],
          "summary": {
            "point_count": 0,
            "series_count": 0
          },
          "time_window": {
            "end": 1784603100,
            "start": 1784602800,
            "step": "60"
          },
          "unit": "cores"
        },
        {
          "chart_data_ref": {
            "cluster_id": "10748116073341772391",
            "end": 1784603100,
            "kind": "promql_range",
            "promql": "sum by (pod, instance) (rate(process_cpu_seconds_total{component=\"pd\"}[1m]))",
            "start": 1784602800,
            "step": "60"
          },
          "group": "pd_cpu_rate",
          "id": "pd_cpu_rate",
          "metric_group_id": "pd_cpu_rate",
          "name": "pd_cpu_rate",
          "promql": "sum by (pod, instance) (rate(process_cpu_seconds_total{component=\"pd\"}[1m]))",
          "ref": {
            "metric_group_id": "pd_cpu_rate",
            "type": "metric"
          },
          "series_summaries": [],
          "summary": {
            "point_count": 0,
            "series_count": 0
          },
          "time_window": {
            "end": 1784603100,
            "start": 1784602800,
            "step": "60"
          },
          "unit": "cores"
        },
        {
          "chart_data_ref": {
            "cluster_id": "10748116073341772391",
            "end": 1784603100,
            "kind": "promql_range",
            "promql": "max by (pod, instance) (pd_service_maxprocs)",
            "start": 1784602800,
            "step": "60"
          },
          "group": "pd_cpu_quota",
          "id": "pd_cpu_quota",
          "metric_group_id": "pd_cpu_quota",
          "name": "pd_cpu_quota",
          "promql": "max by (pod, instance) (pd_service_maxprocs)",
          "ref": {
            "metric_group_id": "pd_cpu_quota",
            "type": "metric"
          },
          "series_summaries": [],
          "summary": {
            "point_count": 0,
            "series_count": 0
          },
          "time_window": {
            "end": 1784603100,
            "start": 1784602800,
            "step": "60"
          },
          "unit": "cores"
        }
      ],
      "panic_signal_artifact_present": false,
      "query_evidence": {
        "slow_query": [],
        "topsql": []
      },
      "query_or_metric_channel_observed": true,
      "slow_query_top_digest": "",
      "ticket_context_channel_observed": false,
      "topsql_collection_error_recorded": false
    },
    "execution_contract": {
      "artifact_facts": {
        "analysis_patch_present": false,
        "core_evidence_bundle_present": false,
        "diagnosis_report_present": true,
        "summary_artifact_present": false,
        "supplemental_loki_present": false
      },
      "channel_facts": {
        "events": true,
        "logs": true,
        "query_or_metric": true,
        "ticket_context": false
      },
      "checks": [
        {
          "description": "The minimal ad-hoc workflow is present.",
          "id": "required_workflow_present",
          "match": "all_of",
          "passed": true,
          "required": true,
          "rules": [
            {
              "matched": true,
              "matched_keys": [
                "final_answer_emitted"
              ],
              "matcher": "subset_observed",
              "path": "",
              "paths": [
                "final_answer_emitted"
              ],
              "source": "workflow_check",
              "threshold": null
            }
          ],
          "snippet": "matched execution contract facts: final_answer_emitted"
        },
        {
          "description": "This ad-hoc case must not invoke collect_evidence.py.",
          "id": "collect_evidence_not_used",
          "match": "all_of",
          "passed": true,
          "required": true,
          "rules": [
            {
              "matched": true,
              "matched_keys": [
                "collect_evidence_not_used"
              ],
              "matcher": "subset_observed",
              "path": "",
              "paths": [
                "collect_evidence_not_used"
              ],
              "source": "workflow_check",
              "threshold": null
            }
          ],
          "snippet": "matched execution contract facts: collect_evidence_not_used"
        },
        {
          "description": "Workspace diagnosis report is present.",
          "id": "required_report_present",
          "match": "all_of",
          "passed": true,
          "required": true,
          "rules": [
            {
              "matched": true,
              "matched_keys": [
                "diagnosis_report_present"
              ],
              "matcher": "subset_observed",
              "path": "",
              "paths": [
                "diagnosis_report_present"
              ],
              "source": "artifact_fact",
              "threshold": null
            }
          ],
          "snippet": "matched execution contract facts: diagnosis_report_present"
        },
        {
          "description": "At least one query or metric evidence channel is observed.",
          "id": "query_or_metric_channel_observed",
          "match": "all_of",
          "passed": true,
          "required": true,
          "rules": [
            {
              "matched": true,
              "matched_keys": [
                "query_or_metric"
              ],
              "matcher": "count_gte",
              "path": "",
              "paths": [
                "query_or_metric"
              ],
              "source": "channel_fact",
              "threshold": 1.0
            }
          ],
          "snippet": "matched execution contract facts: query_or_metric"
        }
      ],
      "status": "pass",
      "workflow_facts": {
        "collect_evidence_not_used": true,
        "final_answer_emitted": true
      }
    },
    "output": {
      "analyze_primary_culprit_asserted": false,
      "claims_panic_not_confirmable": false,
      "diagnosis_cpu_hotspot_asserted": false,
      "diagnosis_kill_query_chain_asserted": false,
      "diagnosis_memory_tracker_panic_asserted": false,
      "diagnosis_plan_stats_issue_asserted": false,
      "diagnosis_single_instance_panic_restart_asserted": false,
      "fake_alarm_primary_asserted": false,
      "focuses_on_missing_index": false,
      "mentions_cpu_saturation": false,
      "mentions_est_act_divergence": false,
      "mentions_idle_transaction": false,
      "mentions_intermediate_result_explosion": false,
      "mentions_kill_query": false,
      "mentions_memory_tracker_consume": false,
      "mentions_p99_latency_spike": false,
      "mentions_rollback_cleanup": false,
      "mentions_statistics_uninitialized": false,
      "mentions_v8_5_2": false,
      "mentions_v8_5_3_or_fixed_release": false,
      "mentions_without_fix": false,
      "missing_index_primary_root_cause_asserted": false,
      "puller_or_source_side_primary_root_cause_asserted": false,
      "query_only_primary_root_cause_asserted": false,
      "wide_scan_primary_root_cause_asserted": false
    },
    "routing": {
      "analysis_event_items_present": false,
      "analysis_has_key_evidence": false,
      "analysis_has_supporting_evidence": false,
      "analysis_items_shape_valid": false,
      "analysis_log_items_present": false,
      "analysis_log_or_event_channel_resolvable": true,
      "analysis_metric_refs_required": false,
      "analysis_metric_refs_resolvable": true,
      "analysis_patch_present": false,
      "analysis_patch_valid_json": false,
      "analysis_query_refs_required": false,
      "analysis_query_refs_resolvable": true,
      "classification_contract_valid": false,
      "diagnosis_analyst_contract_present": false,
      "fault_classification_contract_valid": false,
      "primary_fault_class_component_restart_or_panic": false,
      "primary_fault_class_resource_control_or_capacity": false,
      "primary_fault_class_service_unavailability_or_worker_failure": false,
      "primary_fault_class_sql_plan_or_query": false,
      "primary_fault_class_storage_or_infra_io": false,
      "primary_fault_class_tidb_cpu_gc_or_memory": false,
      "primary_fault_class_tikv_read_hotspot": false,
      "primary_fault_class_unknown_insufficient_evidence": false,
      "selected_diagnosis_skills_present": false,
      "signal_tags_contract_valid": false
    },
    "runtime_context_present": true,
    "window": {
      "actual_end_ts": 1784603100,
      "actual_start_ts": 1784602800,
      "expected_allowed_window_id": "bjt-2026-07-21-1100-5m",
      "expected_end_ts": 1784603100,
      "expected_start_ts": 1784602800,
      "matched_allowed_window": true
    },
    "workflow": {
      "architecture_context_loaded": false,
      "collect_evidence_invoked": false,
      "diagnosis_run_initialized": true,
      "direct_conclusion_by_investigator": false,
      "evidence_collection_completed": true,
      "evidence_collection_started": false,
      "final_answer_emitted": true,
      "misdiagnosis_guardrails_loaded": false,
      "premise_verification_recorded": false,
      "summary_read_completed": false,
      "targeted_panic_log_query_completed": false
    }
  },
  "invalid_subtype": null,
  "judgment_verdict": "fail",
  "packet_json": {
    "artifact_resolution": {
      "candidate_run_dirs": [
        "evidences"
      ],
      "manifest_path_mismatch": false,
      "primary_failure_code": null,
      "resolution_mode": "workspace_evidence_root",
      "scanned_scopes": [
        {
          "candidate_count": 1,
          "candidates": [
            {
              "actual_cluster_id": "10748116073341772391",
              "actual_end_ts": 1784603100,
              "actual_start_ts": 1784602800,
              "analysis_patch_path": "",
              "analysis_patch_present": false,
              "diagnosis_report_path": "evidences/workspace_evidence_report.json",
              "diagnosis_report_present": true,
              "evidence_dir": "evidences",
              "evidence_dir_present": true,
              "invalid_reasons": [],
              "logs_dir": "",
              "logs_dir_present": false,
              "manifest_path": "",
              "manifest_path_mismatch": false,
              "manifest_path_mismatch_fields": [],
              "manifest_present": false,
              "manifest_valid_json": true,
              "resolution_mode": "workspace_evidence_root",
              "run_dir": "evidences",
              "summary_json_path": "",
              "summary_json_present": false,
              "summary_txt_path": "",
              "summary_txt_present": false,
              "valid_bundle": true
            }
          ],
          "decision": "resolved",
          "selected": true,
          "source_kind": "audit_dir",
          "source_root": "/data/eval_audits/evalrun-dca9124193b04b2c873027cd9fb2ca6f/evalcase-1dd5f72ba4a944f09e48502ce55f1652",
          "telemetry_path": ".opencode/telemetry.log",
          "valid_candidate_count": 1
        },
        {
          "candidate_count": 1,
          "candidates": [
            {
              "actual_cluster_id": "10748116073341772391",
              "actual_end_ts": 1784603100,
              "actual_start_ts": 1784602800,
              "analysis_patch_path": "",
              "analysis_patch_present": false,
              "diagnosis_report_path": "evidences/workspace_evidence_report.json",
              "diagnosis_report_present": true,
              "evidence_dir": "evidences",
              "evidence_dir_present": true,
              "invalid_reasons": [],
              "logs_dir": "",
              "logs_dir_present": false,
              "manifest_path": "",
              "manifest_path_mismatch": false,
              "manifest_path_mismatch_fields": [],
              "manifest_present": false,
              "manifest_valid_json": true,
              "resolution_mode": "workspace_evidence_root",
              "run_dir": "evidences",
              "summary_json_path": "",
              "summary_json_present": false,
              "summary_txt_path": "",
              "summary_txt_present": false,
              "valid_bundle": true
            }
          ],
          "decision": "skipped_lower_priority",
          "selected": false,
          "source_kind": "workspace_dir",
          "source_root": "/data/workspaces/ws-eval-1784618267343-4a2d5ff3",
          "telemetry_path": ".opencode/telemetry.log",
          "valid_candidate_count": 1
        }
      ],
      "selected_candidate": {
        "actual_cluster_id": "10748116073341772391",
        "actual_end_ts": 1784603100,
        "actual_start_ts": 1784602800,
        "analysis_patch_path": "",
        "analysis_patch_present": false,
        "diagnosis_report_path": "evidences/workspace_evidence_report.json",
        "diagnosis_report_present": true,
        "evidence_dir": "evidences",
        "evidence_dir_present": true,
        "invalid_reasons": [],
        "logs_dir": "",
        "logs_dir_present": false,
        "manifest_path": "",
        "manifest_path_mismatch": false,
        "manifest_path_mismatch_fields": [],
        "manifest_present": false,
        "manifest_valid_json": true,
        "resolution_mode": "workspace_evidence_root",
        "run_dir": "evidences",
        "summary_json_path": "",
        "summary_json_present": false,
        "summary_txt_path": "",
        "summary_txt_present": false,
        "valid_bundle": true
      },
      "selected_scope": "audit_dir",
      "selected_trace_candidate": {
        "actual_cluster_id": "10748116073341772391",
        "actual_end_ts": 1784603100,
        "actual_start_ts": 1784602800,
        "analysis_patch_path": "",
        "analysis_patch_present": false,
        "diagnosis_report_path": "evidences/workspace_evidence_report.json",
        "diagnosis_report_present": true,
        "evidence_dir": "evidences",
        "evidence_dir_present": true,
        "invalid_reasons": [],
        "logs_dir": "",
        "logs_dir_present": false,
        "manifest_path": "",
        "manifest_path_mismatch": false,
        "manifest_path_mismatch_fields": [],
        "manifest_present": false,
        "manifest_valid_json": true,
        "resolution_mode": "workspace_evidence_root",
        "run_dir": "evidences",
        "summary_json_path": "",
        "summary_json_present": false,
        "summary_txt_path": "",
        "summary_txt_present": false,
        "valid_bundle": true
      },
      "status": "resolved"
    },
    "contract": {
      "checks": [
        {
          "description": "Actual diagnosis cluster scope matches the expected cluster id.",
          "id": "cluster_scope_match",
          "match": "all_of",
          "passed": true,
          "required": true,
          "rules": [
            {
              "allowed": [],
              "matched": true,
              "matched_keys": [
                "cluster_scope_match"
              ],
              "matcher": "bool_true",
              "path": "cluster_scope_match",
              "source": "trace_fact",
              "value": ""
            }
          ],
          "snippet": "matched trace facts: cluster_scope_match"
        }
      ],
      "status": "pass"
    },
    "critical_evidence": {
      "checks": [
        {
          "artifact_paths": [],
          "description": "Output JSON cluster_id matches the runtime reference.",
          "earned_score": 6,
          "id": "output_cluster_id_matches_reference",
          "match": "all_of",
          "max_score": 6,
          "passed": true,
          "required": true,
          "rules": [
            {
              "artifact_path": "",
              "artifacts": [],
              "matched": true,
              "matched_keyword": null,
              "matcher": "json_path_equals_reference",
              "snippet": "\"10748116073341772391\"",
              "source": "output_json"
            }
          ],
          "snippet": "\"10748116073341772391\""
        },
        {
          "artifact_paths": [],
          "description": "Output JSON window.start matches the runtime reference.",
          "earned_score": 5,
          "id": "output_window_start_matches_reference",
          "match": "all_of",
          "max_score": 5,
          "passed": true,
          "required": true,
          "rules": [
            {
              "artifact_path": "",
              "artifacts": [],
              "matched": true,
              "matched_keyword": null,
              "matcher": "json_path_equals_reference",
              "snippet": "\"2026-07-21T03:00:00Z\"",
              "source": "output_json"
            }
          ],
          "snippet": "\"2026-07-21T03:00:00Z\""
        },
        {
          "artifact_paths": [],
          "description": "Output JSON window.end matches the runtime reference.",
          "earned_score": 5,
          "id": "output_window_end_matches_reference",
          "match": "all_of",
          "max_score": 5,
          "passed": true,
          "required": true,
          "rules": [
            {
              "artifact_path": "",
              "artifacts": [],
              "matched": true,
              "matched_keyword": null,
              "matcher": "json_path_equals_reference",
              "snippet": "\"2026-07-21T03:05:00Z\"",
              "source": "output_json"
            }
          ],
          "snippet": "\"2026-07-21T03:05:00Z\""
        },
        {
          "artifact_paths": [],
          "description": "Output JSON evidence_sufficient matches the runtime reference.",
          "earned_score": 0,
          "id": "evidence_sufficient_matches_reference",
          "match": "all_of",
          "max_score": 4,
          "passed": false,
          "required": true,
          "rules": [
            {
              "artifact_path": "",
              "artifacts": [],
              "matched": false,
              "matched_keyword": null,
              "matcher": "json_path_equals_reference",
              "snippet": null,
              "source": "output_json"
            }
          ],
          "snippet": null
        },
        {
          "artifact_paths": [],
          "description": "Output JSON ranked_nodes ordering and identity fields match the runtime reference.",
          "earned_score": 0,
          "id": "ranked_nodes_identity_matches_reference",
          "match": "all_of",
          "max_score": 30,
          "passed": false,
          "required": true,
          "rules": [
            {
              "artifact_path": "",
              "artifacts": [],
              "matched": false,
              "matched_keyword": null,
              "matcher": "json_array_matches_reference",
              "snippet": null,
              "source": "output_json"
            }
          ],
          "snippet": null
        },
        {
          "artifact_paths": [],
          "description": "Output JSON ranked_nodes numeric fields match the runtime reference.",
          "earned_score": 0,
          "id": "ranked_nodes_numeric_fields_match_reference",
          "match": "all_of",
          "max_score": 30,
          "passed": false,
          "required": true,
          "rules": [
            {
              "artifact_path": "",
              "artifacts": [],
              "matched": false,
              "matched_keyword": null,
              "matcher": "json_array_matches_reference",
              "snippet": null,
              "source": "output_json"
            }
          ],
          "snippet": null
        }
      ],
      "max_score": 80,
      "raw_max_score": 80,
      "raw_score": 16,
      "score": 16,
      "status": "partial"
    },
    "critical_evidence_verdict": "partial",
    "diagnosis": {
      "checks": [
        {
          "description": "identify_known_bug_63956",
          "id": "identify_known_bug_63956",
          "passed": false,
          "snippet": null
        },
        {
          "description": "identify_kill_query_idle_txn_chain",
          "id": "identify_kill_query_idle_txn_chain",
          "passed": false,
          "snippet": null
        },
        {
          "description": "identify_version_without_fix",
          "id": "identify_version_without_fix",
          "passed": false,
          "snippet": null
        },
        {
          "description": "recommend_fixed_version_upgrade",
          "id": "recommend_fixed_version_upgrade",
          "passed": false,
          "snippet": null
        }
      ],
      "hard_gate_passed": true,
      "max_score": 50,
      "raw_score": 0,
      "score": 0,
      "semantic_shadow": {
        "available": false,
        "checks": [],
        "packet_path": "/data/eval_audits/evalrun-dca9124193b04b2c873027cd9fb2ca6f/evalcase-1dd5f72ba4a944f09e48502ce55f1652/semantic_packet_v1.json",
        "reason": "semantic_packet_missing",
        "status": "not_available",
        "summary": {
          "disagreement_flags": [],
          "shadow_max_score": 0,
          "shadow_score": 0
        }
      },
      "status": "pass"
    },
    "evidence_collection": {
      "checks": [
        {
          "description": "Workspace diagnosis report is present and readable.",
          "earned_score": 2,
          "id": "required_report_present",
          "match": "all_of",
          "max_score": 2,
          "passed": true,
          "required": true,
          "rules": [
            {
              "matched": true,
              "matched_keys": [
                "diagnosis_report_present"
              ],
              "matcher": "subset_observed",
              "path": "",
              "paths": [
                "diagnosis_report_present"
              ],
              "source": "artifact_fact",
              "threshold": null
            }
          ],
          "snippet": "matched collection facts: diagnosis_report_present"
        },
        {
          "description": "Query or metric evidence is observed for this ad-hoc request.",
          "earned_score": 3,
          "id": "query_or_metric_channel_observed",
          "match": "all_of",
          "max_score": 3,
          "passed": true,
          "required": true,
          "rules": [
            {
              "matched": true,
              "matched_keys": [
                "query_or_metric"
              ],
              "matcher": "count_gte",
              "path": "",
              "paths": [
                "query_or_metric"
              ],
              "source": "channel_fact",
              "threshold": 1.0
            }
          ],
          "snippet": "matched collection facts: query_or_metric"
        },
        {
          "description": "If an events collection query is attempted, an events output file is persisted even when the result set is empty.",
          "earned_score": 2,
          "id": "attempted_events_collection_persisted",
          "match": "all_of",
          "max_score": 2,
          "passed": true,
          "required": false,
          "rules": [
            {
              "matched": true,
              "matched_keys": [
                "events_attempted_implies_artifact_present"
              ],
              "matcher": "bool_true",
              "path": "events_attempted_implies_artifact_present",
              "paths": [],
              "source": "evidence_fact",
              "threshold": null
            }
          ],
          "snippet": "matched collection facts: events_attempted_implies_artifact_present"
        }
      ],
      "max_score": 5,
      "raw_max_score": 7,
      "raw_score": 7,
      "score": 5,
      "status": "pass"
    },
    "evidence_collection_verdict": "pass",
    "evidence_routing": {
      "checks": [
        {
          "description": "Analysis patch is only required to be valid when it exists.",
          "earned_score": 0,
          "id": "analysis_patch_valid_when_present",
          "match": "any_of",
          "max_score": 0,
          "passed": true,
          "required": true,
          "rules": [
            {
              "matched": true,
              "matched_keys": [
                "analysis_patch_present"
              ],
              "matcher": "bool_false",
              "path": "analysis_patch_present",
              "paths": [],
              "source": "routing_fact"
            },
            {
              "matched": false,
              "matched_keys": [],
              "matcher": "bool_true",
              "path": "analysis_patch_valid_json",
              "paths": [],
              "source": "routing_fact"
            }
          ],
          "snippet": "matched routing facts: analysis_patch_present"
        }
      ],
      "max_score": 5,
      "raw_max_score": 0,
      "raw_score": 0,
      "score": 0,
      "status": "pass"
    },
    "evidence_routing_verdict": "pass",
    "execution_contract": {
      "artifact_facts": {
        "analysis_patch_present": false,
        "core_evidence_bundle_present": false,
        "diagnosis_report_present": true,
        "summary_artifact_present": false,
        "supplemental_loki_present": false
      },
      "channel_facts": {
        "events": true,
        "logs": true,
        "query_or_metric": true,
        "ticket_context": false
      },
      "checks": [
        {
          "description": "The minimal ad-hoc workflow is present.",
          "id": "required_workflow_present",
          "match": "all_of",
          "passed": true,
          "required": true,
          "rules": [
            {
              "matched": true,
              "matched_keys": [
                "final_answer_emitted"
              ],
              "matcher": "subset_observed",
              "path": "",
              "paths": [
                "final_answer_emitted"
              ],
              "source": "workflow_check",
              "threshold": null
            }
          ],
          "snippet": "matched execution contract facts: final_answer_emitted"
        },
        {
          "description": "This ad-hoc case must not invoke collect_evidence.py.",
          "id": "collect_evidence_not_used",
          "match": "all_of",
          "passed": true,
          "required": true,
          "rules": [
            {
              "matched": true,
              "matched_keys": [
                "collect_evidence_not_used"
              ],
              "matcher": "subset_observed",
              "path": "",
              "paths": [
                "collect_evidence_not_used"
              ],
              "source": "workflow_check",
              "threshold": null
            }
          ],
          "snippet": "matched execution contract facts: collect_evidence_not_used"
        },
        {
          "description": "Workspace diagnosis report is present.",
          "id": "required_report_present",
          "match": "all_of",
          "passed": true,
          "required": true,
          "rules": [
            {
              "matched": true,
              "matched_keys": [
                "diagnosis_report_present"
              ],
              "matcher": "subset_observed",
              "path": "",
              "paths": [
                "diagnosis_report_present"
              ],
              "source": "artifact_fact",
              "threshold": null
            }
          ],
          "snippet": "matched execution contract facts: diagnosis_report_present"
        },
        {
          "description": "At least one query or metric evidence channel is observed.",
          "id": "query_or_metric_channel_observed",
          "match": "all_of",
          "passed": true,
          "required": true,
          "rules": [
            {
              "matched": true,
              "matched_keys": [
                "query_or_metric"
              ],
              "matcher": "count_gte",
              "path": "",
              "paths": [
                "query_or_metric"
              ],
              "source": "channel_fact",
              "threshold": 1.0
            }
          ],
          "snippet": "matched execution contract facts: query_or_metric"
        }
      ],
      "status": "pass",
      "workflow_facts": {
        "collect_evidence_not_used": true,
        "final_answer_emitted": true
      }
    },
    "execution_contract_verdict": "pass",
    "ground_truth_root_cause": "Correctly answering this ad-hoc case requires ranking the top three CPU-overload nodes per component in the runtime-resolved Beijing 11:00-11:05 window and counting the total ERROR/WARN log lines for each selected node in the same 5-minute window.",
    "overall_verdict": "fail",
    "quality_verdict": "pass",
    "scoring": {
      "failing_hard_gate_score_cap": 69,
      "final_score": 21,
      "hard_gate_passed": false,
      "passing_hard_gate_score_floor": 70,
      "raw_total_score": 21
    },
    "strategy": "deterministic_trace_heavy_v1",
    "workflow": {
      "checks": [
        {
          "description": "Final answer is emitted for the ad-hoc query.",
          "earned_score": 0,
          "id": "final_answer_emitted",
          "match": "all_of",
          "max_score": 0,
          "passed": true,
          "required": true,
          "rules": [
            {
              "matched": true,
              "matched_keys": [
                "final_answer_emitted"
              ],
              "matcher": "bool_true",
              "path": "final_answer_emitted",
              "paths": [],
              "source": "workflow_fact",
              "threshold": null
            }
          ],
          "snippet": "matched workflow facts: final_answer_emitted"
        },
        {
          "description": "This ad-hoc case must not invoke collect_evidence.py.",
          "earned_score": 0,
          "id": "collect_evidence_not_used",
          "match": "all_of",
          "max_score": 0,
          "passed": true,
          "required": true,
          "rules": [
            {
              "matched": true,
              "matched_keys": [
                "collect_evidence_invoked"
              ],
              "matcher": "bool_false",
              "path": "collect_evidence_invoked",
              "paths": [],
              "source": "workflow_fact",
              "threshold": null
            }
          ],
          "snippet": "matched workflow facts: collect_evidence_invoked"
        }
      ],
      "max_score": 10,
      "raw_max_score": 0,
      "raw_score": 0,
      "score": 0,
      "status": "pass"
    },
    "workflow_verdict": "pass"
  },
  "passed_checks": 0,
  "primary_failure_code": "critical_evidence_partial",
  "score": 21,
  "total_checks": 0
}