{
  "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-b5720968534a4efd9cb8fd9e4f335402",
          "telemetry_path": ".opencode/telemetry.log",
          "valid_candidate_count": 1
        },
        {
          "candidate_count": 1,
          "candidates": [
            {
              "actual_cluster_id": "2026948324023603200",
              "actual_end_ts": 1784603100,
              "actual_start_ts": 1784603040,
              "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-1784617422348-71a05571",
          "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": "premium-keyspace-storage-query-semantics",
    "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": false,
        "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",
                "collect_evidence_not_used"
              ],
              "matcher": "subset_observed",
              "path": "",
              "paths": [
                "final_answer_emitted",
                "collect_evidence_not_used"
              ],
              "source": "workflow_check",
              "threshold": null
            }
          ],
          "snippet": "matched execution contract facts: final_answer_emitted, 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": "fail",
    "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-b5720968534a4efd9cb8fd9e4f335402",
      "source_scope": "audit_dir",
      "summary_txt_path": "",
      "telemetry_path": ".opencode/telemetry.log"
    },
    "case_id": "premium-keyspace-storage-query-semantics",
    "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_window_start_matches_reference",
        "output_window_end_matches_reference",
        "output_target_metric_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": false,
      "logs_dir_present": false,
      "memory_tracker_panic_present": false,
      "metric_evidence": [
        {
          "chart_data_ref": {
            "cluster_id": "2026948324023603200",
            "end": 1784603100,
            "kind": "promql_range",
            "promql": "tikv_store_size_bytes{type=\"used\", instance=\"db-125d9e7b-tikv-pmbces\"}",
            "start": 1784603040,
            "step": "60"
          },
          "group": "tikv_store_size_bytes_used_discovery",
          "id": "tikv_store_size_bytes_used_discovery",
          "metric_group_id": "tikv_store_size_bytes_used_discovery",
          "name": "tikv_store_size_bytes_used_label_discovery",
          "promql": "tikv_store_size_bytes{type=\"used\", instance=\"db-125d9e7b-tikv-pmbces\"}",
          "ref": {
            "metric_group_id": "tikv_store_size_bytes_used_discovery",
            "type": "metric"
          },
          "series_summaries": [],
          "summary": {
            "point_count": 0,
            "series_count": 0
          },
          "time_window": {
            "end": 1784603100,
            "start": 1784603040,
            "step": "60"
          },
          "unit": ""
        },
        {
          "chart_data_ref": {
            "cluster_id": "2026948324023603200",
            "end": 1784603100,
            "kind": "promql_range",
            "promql": "tikv_store_size_bytes{type=\"used\"}",
            "start": 1784603040,
            "step": "60"
          },
          "group": "tikv_store_size_bytes_used_all",
          "id": "tikv_store_size_bytes_used_all",
          "metric_group_id": "tikv_store_size_bytes_used_all",
          "name": "tikv_store_size_bytes_used_all",
          "promql": "tikv_store_size_bytes{type=\"used\"}",
          "ref": {
            "metric_group_id": "tikv_store_size_bytes_used_all",
            "type": "metric"
          },
          "series_summaries": [
            {
              "avg": 498482.0,
              "labels": {
                "__name__": "tikv_store_size_bytes",
                "component": "tikv",
                "container": "tikv",
                "endpoint": "status",
                "exported_sharedpool_id": "2026948324023603200",
                "instance": "db-125d9e7b-tikv-uinnfa",
                "job": "tidb2026948324023603200/db-125d9e7b-tikv",
                "keyspace_name": "SYSTEM",
                "namespace": "tidb2026948324023603200",
                "pod": "db-125d9e7b-tikv-uinnfa",
                "pod_name": "db-125d9e7b-tikv-uinnfa",
                "prometheus": "tidb2026948324023603200/o11y-vmagent",
                "region_id": "11",
                "sharedpool_id": "2026948324023603200",
                "storage_class": "standard",
                "tidb_cluster_id": "2026948324023603200",
                "type": "used",
                "vm_account_id": "612",
                "vm_project_id": "187"
              },
              "max": 498482.0,
              "min": 498482.0,
              "point_count": 2
            }
          ],
          "summary": {
            "avg": 498482.0,
            "max": 498482.0,
            "min": 498482.0,
            "point_count": 2,
            "series_count": 1
          },
          "time_window": {
            "end": 1784603100,
            "start": 1784603040,
            "step": "60"
          },
          "unit": ""
        },
        {
          "chart_data_ref": {
            "cluster_id": "2026948324023603200",
            "end": 1784603100,
            "kind": "promql_range",
            "promql": "tikv_store_size_bytes{type=\"used\", pod=\"db-125d9e7b-tikv-pmbces\"}",
            "start": 1784602800,
            "step": "60"
          },
          "group": "tikv_store_size_bytes_target",
          "id": "tikv_store_size_bytes_target",
          "metric_group_id": "tikv_store_size_bytes_target",
          "name": "tikv_store_size_bytes_target_instance",
          "promql": "tikv_store_size_bytes{type=\"used\", pod=\"db-125d9e7b-tikv-pmbces\"}",
          "ref": {
            "metric_group_id": "tikv_store_size_bytes_target",
            "type": "metric"
          },
          "series_summaries": [],
          "summary": {
            "point_count": 0,
            "series_count": 0
          },
          "time_window": {
            "end": 1784603100,
            "start": 1784602800,
            "step": "60"
          },
          "unit": ""
        },
        {
          "chart_data_ref": {
            "cluster_id": "2026948324023603200",
            "end": 1784603100,
            "kind": "promql_range",
            "promql": "tikv_store_size_bytes{type=\"used\", instance=~\"db-125d9e7b-tikv-pmbces.*\"}",
            "start": 1784602800,
            "step": "60"
          },
          "group": "tikv_store_size_bytes_target_v2",
          "id": "tikv_store_size_bytes_target_v2",
          "metric_group_id": "tikv_store_size_bytes_target_v2",
          "name": "tikv_store_size_bytes_target_instance_v2",
          "promql": "tikv_store_size_bytes{type=\"used\", instance=~\"db-125d9e7b-tikv-pmbces.*\"}",
          "ref": {
            "metric_group_id": "tikv_store_size_bytes_target_v2",
            "type": "metric"
          },
          "series_summaries": [],
          "summary": {
            "point_count": 0,
            "series_count": 0
          },
          "time_window": {
            "end": 1784603100,
            "start": 1784602800,
            "step": "60"
          },
          "unit": ""
        },
        {
          "chart_data_ref": {
            "cluster_id": "2026948324023603200",
            "end": 1784603100,
            "kind": "promql_range",
            "promql": "tikv_store_size_bytes{type=\"used\"}",
            "start": 1784602800,
            "step": "30"
          },
          "group": "tikv_store_size_bytes_full_window",
          "id": "tikv_store_size_bytes_full_window",
          "metric_group_id": "tikv_store_size_bytes_full_window",
          "name": "tikv_store_size_bytes_full_window",
          "promql": "tikv_store_size_bytes{type=\"used\"}",
          "ref": {
            "metric_group_id": "tikv_store_size_bytes_full_window",
            "type": "metric"
          },
          "series_summaries": [
            {
              "avg": 498482.0,
              "labels": {
                "__name__": "tikv_store_size_bytes",
                "component": "tikv",
                "container": "tikv",
                "endpoint": "status",
                "exported_sharedpool_id": "2026948324023603200",
                "instance": "db-125d9e7b-tikv-uinnfa",
                "job": "tidb2026948324023603200/db-125d9e7b-tikv",
                "keyspace_name": "SYSTEM",
                "namespace": "tidb2026948324023603200",
                "pod": "db-125d9e7b-tikv-uinnfa",
                "pod_name": "db-125d9e7b-tikv-uinnfa",
                "prometheus": "tidb2026948324023603200/o11y-vmagent",
                "region_id": "11",
                "sharedpool_id": "2026948324023603200",
                "storage_class": "standard",
                "tidb_cluster_id": "2026948324023603200",
                "type": "used",
                "vm_account_id": "612",
                "vm_project_id": "187"
              },
              "max": 498482.0,
              "min": 498482.0,
              "point_count": 11
            }
          ],
          "summary": {
            "avg": 498482.0,
            "max": 498482.0,
            "min": 498482.0,
            "point_count": 11,
            "series_count": 1
          },
          "time_window": {
            "end": 1784603100,
            "start": 1784602800,
            "step": "30"
          },
          "unit": ""
        },
        {
          "chart_data_ref": {
            "cluster_id": "2026948324023603200",
            "end": 1784603100,
            "kind": "promql_range",
            "promql": "tikv_store_size_bytes{type=\"used\", pod_name=~\"db-125d9e7b-tikv-pmbces.*\"}",
            "start": 1784602800,
            "step": "30"
          },
          "group": "tikv_store_size_pmbces",
          "id": "tikv_store_size_pmbces",
          "metric_group_id": "tikv_store_size_pmbces",
          "name": "tikv_store_size_pmbces_start",
          "promql": "tikv_store_size_bytes{type=\"used\", pod_name=~\"db-125d9e7b-tikv-pmbces.*\"}",
          "ref": {
            "metric_group_id": "tikv_store_size_pmbces",
            "type": "metric"
          },
          "series_summaries": [],
          "summary": {
            "point_count": 0,
            "series_count": 0
          },
          "time_window": {
            "end": 1784603100,
            "start": 1784602800,
            "step": "30"
          },
          "unit": ""
        },
        {
          "chart_data_ref": {
            "cluster_id": "2026948324023603200",
            "end": 1784603100,
            "kind": "promql_range",
            "promql": "count by (keyspace_name) (tikv_store_size_bytes{type=\"used\", pod_name=~\".*pmbces.*\"})",
            "start": 1784603040,
            "step": "60"
          },
          "group": "tikv_store_size_pmbces_kscount",
          "id": "tikv_store_size_pmbces_kscount",
          "metric_group_id": "tikv_store_size_pmbces_kscount",
          "name": "tikv_store_size_pmbces_count_ksname",
          "promql": "count by (keyspace_name) (tikv_store_size_bytes{type=\"used\", pod_name=~\".*pmbces.*\"})",
          "ref": {
            "metric_group_id": "tikv_store_size_pmbces_kscount",
            "type": "metric"
          },
          "series_summaries": [],
          "summary": {
            "point_count": 0,
            "series_count": 0
          },
          "time_window": {
            "end": 1784603100,
            "start": 1784603040,
            "step": "60"
          },
          "unit": ""
        },
        {
          "chart_data_ref": {
            "cluster_id": "2026948324023603200",
            "end": 1784603100,
            "kind": "promql_range",
            "promql": "count by (keyspace_name) (tikv_store_size_bytes{type=\"used\"})",
            "start": 1784603040,
            "step": "60"
          },
          "group": "tikv_store_size_kscount_global",
          "id": "tikv_store_size_kscount_global",
          "metric_group_id": "tikv_store_size_kscount_global",
          "name": "tikv_store_size_count_ks_global",
          "promql": "count by (keyspace_name) (tikv_store_size_bytes{type=\"used\"})",
          "ref": {
            "metric_group_id": "tikv_store_size_kscount_global",
            "type": "metric"
          },
          "series_summaries": [
            {
              "avg": 1.0,
              "labels": {
                "keyspace_name": "SYSTEM"
              },
              "max": 1.0,
              "min": 1.0,
              "point_count": 2
            }
          ],
          "summary": {
            "avg": 1.0,
            "max": 1.0,
            "min": 1.0,
            "point_count": 2,
            "series_count": 1
          },
          "time_window": {
            "end": 1784603100,
            "start": 1784603040,
            "step": "60"
          },
          "unit": ""
        }
      ],
      "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": false,
        "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",
                "collect_evidence_not_used"
              ],
              "matcher": "subset_observed",
              "path": "",
              "paths": [
                "final_answer_emitted",
                "collect_evidence_not_used"
              ],
              "source": "workflow_check",
              "threshold": null
            }
          ],
          "snippet": "matched execution contract facts: final_answer_emitted, 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-b5720968534a4efd9cb8fd9e4f335402",
          "telemetry_path": ".opencode/telemetry.log",
          "valid_candidate_count": 1
        },
        {
          "candidate_count": 1,
          "candidates": [
            {
              "actual_cluster_id": "2026948324023603200",
              "actual_end_ts": 1784603100,
              "actual_start_ts": 1784603040,
              "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-1784617422348-71a05571",
          "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 time window matches one allowed case window.",
          "id": "time_window_match",
          "passed": true,
          "required": true,
          "snippet": "1784602800..1784603100"
        },
        {
          "description": "Actual diagnosis cluster scope matches expected cluster id.",
          "id": "cluster_scope_match",
          "passed": true,
          "required": true,
          "snippet": "10748116073341772391"
        }
      ],
      "status": "pass"
    },
    "critical_evidence": {
      "checks": [
        {
          "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 target_metric matches the runtime reference.",
          "earned_score": 6,
          "id": "output_target_metric_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": "\"tikv_store_size_bytes\"",
              "source": "output_json"
            }
          ],
          "snippet": "\"tikv_store_size_bytes\""
        },
        {
          "artifact_paths": [],
          "description": "Output JSON evidence_sufficient matches the runtime reference.",
          "earned_score": 0,
          "id": "output_evidence_sufficient_matches_reference",
          "match": "all_of",
          "max_score": 6,
          "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 cluster cardinality matches the runtime reference ordering and values.",
          "earned_score": 0,
          "id": "output_clusters_match_reference",
          "match": "all_of",
          "max_score": 20,
          "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": 42,
      "raw_score": 16,
      "score": 30,
      "status": "partial"
    },
    "critical_evidence_verdict": "partial",
    "diagnosis": {
      "checks": [
        {
          "description": "Output JSON window.start matches the runtime reference.",
          "id": "output_window_start_matches_reference",
          "match": "all_of",
          "passed": true,
          "rules": [
            {
              "key_fields": [],
              "keywords": [],
              "matched": true,
              "matched_keys": [
                "window.start"
              ],
              "matcher": "json_path_equals_reference",
              "numeric_fields": [],
              "path": "window.start",
              "paths": [],
              "patterns": [],
              "reference_output_id": "reference_answer",
              "reference_path": "window.start",
              "source": "output_json",
              "threshold": null
            }
          ],
          "snippet": "matched output facts: window.start"
        },
        {
          "description": "Output JSON window.end matches the runtime reference.",
          "id": "output_window_end_matches_reference",
          "match": "all_of",
          "passed": true,
          "rules": [
            {
              "key_fields": [],
              "keywords": [],
              "matched": true,
              "matched_keys": [
                "window.end"
              ],
              "matcher": "json_path_equals_reference",
              "numeric_fields": [],
              "path": "window.end",
              "paths": [],
              "patterns": [],
              "reference_output_id": "reference_answer",
              "reference_path": "window.end",
              "source": "output_json",
              "threshold": null
            }
          ],
          "snippet": "matched output facts: window.end"
        },
        {
          "description": "Output JSON target_metric matches the runtime reference.",
          "id": "output_target_metric_matches_reference",
          "match": "all_of",
          "passed": true,
          "rules": [
            {
              "key_fields": [],
              "keywords": [],
              "matched": true,
              "matched_keys": [
                "target_metric"
              ],
              "matcher": "json_path_equals_reference",
              "numeric_fields": [],
              "path": "target_metric",
              "paths": [],
              "patterns": [],
              "reference_output_id": "reference_answer",
              "reference_path": "target_metric",
              "source": "output_json",
              "threshold": null
            }
          ],
          "snippet": "matched output facts: target_metric"
        },
        {
          "description": "Output JSON evidence_sufficient matches the runtime reference.",
          "id": "output_evidence_sufficient_matches_reference",
          "match": "all_of",
          "passed": false,
          "rules": [
            {
              "key_fields": [],
              "keywords": [],
              "matched": false,
              "matched_keys": [],
              "matcher": "json_path_equals_reference",
              "numeric_fields": [],
              "path": "evidence_sufficient",
              "paths": [],
              "patterns": [],
              "reference_output_id": "reference_answer",
              "reference_path": "evidence_sufficient",
              "source": "output_json",
              "threshold": null
            }
          ],
          "snippet": null
        },
        {
          "description": "Output JSON cluster cardinality matches the runtime reference.",
          "id": "output_clusters_match_reference",
          "match": "all_of",
          "passed": false,
          "rules": [
            {
              "key_fields": [
                "cluster_id",
                "target_tikv_instance",
                "keyspace_label",
                "keyspace_value_count"
              ],
              "keywords": [],
              "matched": false,
              "matched_keys": [],
              "matcher": "json_array_matches_reference",
              "numeric_fields": [],
              "path": "clusters",
              "paths": [],
              "patterns": [],
              "reference_output_id": "reference_answer",
              "reference_path": "clusters",
              "source": "output_json",
              "threshold": null
            }
          ],
          "snippet": null
        },
        {
          "description": "The answer includes the expected structured JSON payload.",
          "earned_score": 1,
          "id": "storage_json_output_present",
          "match": "all_of",
          "max_score": 1,
          "passed": true,
          "rules": [
            {
              "key_fields": [],
              "keywords": [
                "target_metric",
                "keyspace_label",
                "keyspace_value_count"
              ],
              "matched": true,
              "matched_keys": [
                "target_metric"
              ],
              "matcher": "keyword_any",
              "numeric_fields": [],
              "path": "",
              "paths": [],
              "patterns": [],
              "reference_output_id": "",
              "reference_path": "",
              "source": "output_text",
              "threshold": null
            }
          ],
          "snippet": "matched output facts: target_metric"
        }
      ],
      "hard_gate_passed": false,
      "max_score": 50,
      "raw_score": 4,
      "score": 21,
      "semantic_shadow": {
        "available": false,
        "checks": [],
        "packet_path": "/data/eval_audits/evalrun-dca9124193b04b2c873027cd9fb2ca6f/evalcase-b5720968534a4efd9cb8fd9e4f335402/semantic_packet_v1.json",
        "reason": "semantic_packet_missing",
        "status": "not_available",
        "summary": {
          "disagreement_flags": [],
          "shadow_max_score": 0,
          "shadow_score": 0
        }
      },
      "status": "fail"
    },
    "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": 2,
          "id": "query_or_metric_channel_observed",
          "match": "all_of",
          "max_score": 2,
          "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": 6,
      "raw_score": 6,
      "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": false,
        "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",
                "collect_evidence_not_used"
              ],
              "matcher": "subset_observed",
              "path": "",
              "paths": [
                "final_answer_emitted",
                "collect_evidence_not_used"
              ],
              "source": "workflow_check",
              "threshold": null
            }
          ],
          "snippet": "matched execution contract facts: final_answer_emitted, 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": "Correct premium keyspace-storage diagnosis must discover the live keyspace label on tikv_store_size_bytes and count its distinct values at the target snapshot time instead of assuming a fixed label name.",
    "overall_verdict": "fail",
    "quality_verdict": "fail",
    "scoring": {
      "failing_hard_gate_score_cap": 69,
      "final_score": 56,
      "hard_gate_passed": false,
      "passing_hard_gate_score_floor": 70,
      "raw_total_score": 56
    },
    "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": 3,
  "primary_failure_code": "critical_evidence_partial",
  "score": 56,
  "total_checks": 5
}