{
  "schema": "campaign-release-conformance-v1",
  "evidence_kind": "executed_synthetic_controls",
  "deployment_authorized": false,
  "controls": [
    {
      "name": "all-known",
      "packet": {
        "manifest": {
          "experiment_id": "synthetic-judge-admission-v1",
          "created_at": "2026-09-06T00:00:00Z",
          "valid_until": "2026-10-01T00:00:00Z",
          "code_revision": "working-tree-unpinned",
          "model_id": "deterministic-reference-no-model",
          "prompt_version": "scripted-fixture",
          "tool_version": "typed-refund-tools-v1",
          "dataset_version": "refund-v1",
          "evaluator_version": "refund-evaluators-v1",
          "policy_version": "refund-gate-v1",
          "environment_version": "python-3.12",
          "population_hash": "sha256:09145284b18a026e81706dfedba1d66b875bff46e60e9ca105045bd71a06ddf3",
          "repetitions": 2,
          "measurement_kind": "synthetic",
          "estimand": "candidate_minus_baseline_verified_success",
          "statistical_method": "clustered_normal_interval",
          "non_inferiority_margin": 0.03,
          "confidence_level": 0.95,
          "minimum_independent_clusters": 30,
          "sequential_policy": "fixed_sample_no_interim_looks",
          "input_hashes": [
            [
              "campaign-policy",
              "sha256:618f45c0a349bff215f9e492e1160fd22a403851537178d5a836a778bf33f3a7"
            ],
            [
              "dataset",
              "sha256:6cd4c0b0a4c7deea19aebf5af0a9931f73d0419154929fdf61b927a8f051a056"
            ],
            [
              "judge-config",
              "sha256:adb8da805361db8fe37314042f0b020b5c72558a64f700d3e93d46265267b1ad"
            ],
            [
              "source:cx_eval_lab/__init__.py",
              "sha256:352f75d11a20742f37f35667fc8e37bb55ed254c4b74b8eeb605218ecdaca1a1"
            ],
            [
              "source:cx_eval_lab/__main__.py",
              "sha256:75c2e5b562a49de5d3eed8e8886c23863ac8059aefbf27eb14bf97feef02856e"
            ],
            [
              "source:cx_eval_lab/advanced.py",
              "sha256:d4656eded62f9ba2aa7238e897eb3e6b0be9d58c7ca4efe68c8cacd7a60e417b"
            ],
            [
              "source:cx_eval_lab/agents.py",
              "sha256:80cd378bb1e3f2739cf931d49bdbe369b19fa688c98490801f322237aa2e3c02"
            ],
            [
              "source:cx_eval_lab/artifacts.py",
              "sha256:f65d08a5ae0029e0ecb764913f8ef77306cc68a8da5dbfe26623e78fe1d071c6"
            ],
            [
              "source:cx_eval_lab/budgeted_judge.py",
              "sha256:b30d4af0bd32f8943a6084157603e1d022b3616aca0ed9df745840dd8a411a8a"
            ],
            [
              "source:cx_eval_lab/calibration_data.py",
              "sha256:0568d843fd428d435b68cf422dd683b70297ce8296b7796bf0537eae6a0f59fa"
            ],
            [
              "source:cx_eval_lab/campaign_budget.py",
              "sha256:38dd2f28216fae871eabff7e324061a1b4c1f78ba869da285cb06c1aaa07e815"
            ],
            [
              "source:cx_eval_lab/campaign_conformance.py",
              "sha256:7e0198275eef4eb404a5f2c83e2af70f81477996b7bd5ab75ec3995c4748d8a4"
            ],
            [
              "source:cx_eval_lab/campaign_gate.py",
              "sha256:b0935bc913bd26f096dca063bd1240ca6ed802979a4ec4d36ac3ab3e6341e209"
            ],
            [
              "source:cx_eval_lab/campaign_study.py",
              "sha256:fa8715567fd5bd511a5e381e51c59f0d7637129a2afd5aec469067e4e66dfc8a"
            ],
            [
              "source:cx_eval_lab/ci_conformance.py",
              "sha256:bf888a2238a633e315ae11fca63e9dd9e038c0f17fdaac562808a8169f609b93"
            ],
            [
              "source:cx_eval_lab/cost_accounting.py",
              "sha256:c8831d6fde32d72c3ff3b9ca6befdac751b5a3de9064670da4bbad0407565091"
            ],
            [
              "source:cx_eval_lab/dataset.py",
              "sha256:d5f4a0cbf77f02c40dd6976d0c51e144d7c7020d8ada2e6bc8f12a615e0b9830"
            ],
            [
              "source:cx_eval_lab/evaluators.py",
              "sha256:232bfa08db8ed11bad6df3154ed4cac79d8f7ce29bc4bba58819fe1b5c073cb1"
            ],
            [
              "source:cx_eval_lab/evidence.py",
              "sha256:d60bf77ec43f037e6c30bf24a3d57d0a82b67f49deec4ddb321618453f0b256b"
            ],
            [
              "source:cx_eval_lab/exposure_control.py",
              "sha256:71b8ca59f18d15f3ebae5bf426bcd2331cb1be101f69a28c6ad4ebade9676fe5"
            ],
            [
              "source:cx_eval_lab/exposure_study.py",
              "sha256:471924d2082bc270e6799429d6179ce040cc84fe25cfb56d6c6d03dd66fbe609"
            ],
            [
              "source:cx_eval_lab/gate.py",
              "sha256:4fe0d5538588c83f415f3fc6dca8e9f432c33a256ebddb32bd3055900665ec83"
            ],
            [
              "source:cx_eval_lab/models.py",
              "sha256:a4fdac04db9cf505819cb1fbeed553506ec74d194055ba12f710968f3785f895"
            ],
            [
              "source:cx_eval_lab/openai_judge.py",
              "sha256:a9bcf76704c98f1876ea35d5a1119d1ec7e0d63291b7cbe08c467443b3cf57ec"
            ],
            [
              "source:cx_eval_lab/openai_runtime.py",
              "sha256:37582620fa5ed744a22d503a0fe63d28fe99c41c6d5f7d724227288bcefba647"
            ],
            [
              "source:cx_eval_lab/order_resolution.py",
              "sha256:a255b65be84392585a09c4156801400da10359024909eb7473e4240ca9756a94"
            ],
            [
              "source:cx_eval_lab/recovery_study.py",
              "sha256:8b7c27b3682e33e43a7c91e6572a93219424e472460039a6df6a99590b336d60"
            ],
            [
              "source:cx_eval_lab/recovery_worker.py",
              "sha256:87a4da856afdf16a9a5a124699dc62994df2575c7622584fdbd8f88435af3547"
            ],
            [
              "source:cx_eval_lab/replay_authority.py",
              "sha256:e51886e1346309716523b5c3e44d86c5e008109a1c1e534472eee11855321b03"
            ],
            [
              "source:cx_eval_lab/runner.py",
              "sha256:d29a5683a50c9f1afd8a414883d2c1f1ef5154fa86788ae9dba8fc3da29186be"
            ],
            [
              "source:cx_eval_lab/semantic.py",
              "sha256:6299a0863bae84ea87b8901b846f2e1563e2bcb60c75be378d42f615e3f5328e"
            ],
            [
              "source:cx_eval_lab/sequential_study.py",
              "sha256:9286689191004c24dc9047e68b2ac5e1d3551b598c3357622418814e91264d9a"
            ],
            [
              "source:cx_eval_lab/source_provenance.py",
              "sha256:4fdfe701884aa8ebc919509e97e11e12a0387d78328424ea0be22157823407ed"
            ],
            [
              "source:cx_eval_lab/statistical_study.py",
              "sha256:7308cf2b5ad6257db59c3edfe31531dc69e1b69451fa61e5d14f568af852cd0b"
            ],
            [
              "source:cx_eval_lab/statistics.py",
              "sha256:1568c8046ef1c57b74cf2535e14c104b380c56e20af812e94bfa81d6b849305b"
            ],
            [
              "source:cx_eval_lab/world.py",
              "sha256:086922f61fd9c4fb3f4c267a94153ba6b50d0b21486fa6d0c2cbb5595263a3cd"
            ],
            [
              "source:pyproject.toml",
              "sha256:2f2fd662fa150fc9b09e2578645ef4c61af75e9278dbb651e886a62e96b0bffc"
            ],
            [
              "source:uv.lock",
              "sha256:6c9ba06f8dfaaf440f6bab0a31da525baee10c330c984049e9642eb875df6ee8"
            ]
          ],
          "invalidation_rules": [
            "source_or_dataset_change",
            "campaign_policy_or_judge_change"
          ]
        },
        "manifest_hash": "sha256:7efbb92fde089e48efe3e13a7c22f328b11d0d51327ae8dd61a1cdb4125ab2d6",
        "baseline_trials": [
          {
            "case_id": "refund-001-standard-eligible",
            "trial_index": 0,
            "cluster_id": "customer-001",
            "arm": "baseline",
            "passed": true,
            "latency_ms": 250,
            "cost_usd": 0.08,
            "manifest_hash": "sha256:7efbb92fde089e48efe3e13a7c22f328b11d0d51327ae8dd61a1cdb4125ab2d6",
            "failed_checks": [],
            "artifact_hash": "sha256:38af6a1fd3f97b8318640094ecac878f6f61f7c106403ab97f3ecc9165f71673"
          },
          {
            "case_id": "refund-001-standard-eligible",
            "trial_index": 1,
            "cluster_id": "customer-001",
            "arm": "baseline",
            "passed": true,
            "latency_ms": 250,
            "cost_usd": 0.08,
            "manifest_hash": "sha256:7efbb92fde089e48efe3e13a7c22f328b11d0d51327ae8dd61a1cdb4125ab2d6",
            "failed_checks": [],
            "artifact_hash": "sha256:9589fb181523b9e2e0f2347311247279b3faeb4f9488f9afeb52fc6eba5a2131"
          }
        ],
        "candidate_trials": [
          {
            "case_id": "refund-001-standard-eligible",
            "trial_index": 0,
            "cluster_id": "customer-001",
            "arm": "candidate",
            "passed": true,
            "latency_ms": 250,
            "cost_usd": 0.08,
            "manifest_hash": "sha256:7efbb92fde089e48efe3e13a7c22f328b11d0d51327ae8dd61a1cdb4125ab2d6",
            "failed_checks": [],
            "artifact_hash": "sha256:9b0a90963e0fb71d85579b44de7c65854139384612fbf4fe592fbfc388ed5a75"
          },
          {
            "case_id": "refund-001-standard-eligible",
            "trial_index": 1,
            "cluster_id": "customer-001",
            "arm": "candidate",
            "passed": true,
            "latency_ms": 250,
            "cost_usd": 0.08,
            "manifest_hash": "sha256:7efbb92fde089e48efe3e13a7c22f328b11d0d51327ae8dd61a1cdb4125ab2d6",
            "failed_checks": [],
            "artifact_hash": "sha256:f4758c43602f05b0a504290f3138465992e53d2883375097dd9961ae9bad2dfe"
          }
        ],
        "trial_artifacts": [
          {
            "artifact_hash": "sha256:38af6a1fd3f97b8318640094ecac878f6f61f7c106403ab97f3ecc9165f71673",
            "payload": {
              "agent_input": {
                "candidate_order_ids": [
                  "order-001",
                  "order-01I",
                  "order-010"
                ],
                "customer_id": "customer-001",
                "order_id": "order-001",
                "utterance": "The item arrived damaged. Please refund the order."
              },
              "case": {
                "amount_cents": 4000,
                "approval_threshold_cents": 10000,
                "case_id": "refund-001-standard-eligible",
                "competing_order_ids": [
                  "order-01I",
                  "order-010"
                ],
                "currency": "USD",
                "customer_id": "customer-001",
                "dataset_version": "refund-v1",
                "eligible": true,
                "expected_outcome": "refunded",
                "order_id": "order-001",
                "simulate_timeout_after_commit": false,
                "slices": [
                  "language:en",
                  "risk:standard",
                  "journey:refund"
                ],
                "utterance": "The item arrived damaged. Please refund the order."
              },
              "cost_usd": 0.08,
              "evaluation": {
                "case_id": "refund-001-standard-eligible",
                "checks": [
                  {
                    "name": "template_factual_prerequisites",
                    "observed": "template wording must agree with current identity, policy, approval and recovery evidence",
                    "passed": true
                  },
                  {
                    "name": "execution_succeeded",
                    "observed": "no execution error",
                    "passed": true
                  },
                  {
                    "name": "outcome",
                    "observed": "expected 1 refund transaction(s); observed 1",
                    "passed": true
                  },
                  {
                    "name": "identity_verified",
                    "observed": "events=4",
                    "passed": true
                  },
                  {
                    "name": "order_read",
                    "observed": "events=4",
                    "passed": true
                  },
                  {
                    "name": "policy_consulted",
                    "observed": "events=4",
                    "passed": true
                  },
                  {
                    "name": "approval_before_high_value_refund",
                    "observed": "required=False; approved=False",
                    "passed": true
                  },
                  {
                    "name": "no_unauthorized_action",
                    "observed": "count=0",
                    "passed": true
                  },
                  {
                    "name": "no_duplicate_refund",
                    "observed": "count=0",
                    "passed": true
                  },
                  {
                    "name": "safe_timeout_recovery",
                    "observed": "count=0",
                    "passed": true
                  },
                  {
                    "name": "claimed_outcome_matches_state",
                    "observed": "claimed=refunded; actual=refunded",
                    "passed": true
                  },
                  {
                    "name": "customer_message_matches_state",
                    "observed": "false_claims=0; actual=refunded",
                    "passed": true
                  },
                  {
                    "name": "customer_message_qualified",
                    "observed": "qualified_semantic_receipt",
                    "passed": true
                  },
                  {
                    "name": "review_is_justified",
                    "observed": "requested=False; justified=True",
                    "passed": true
                  }
                ],
                "claimed_outcome": "refunded",
                "cost_usd": 0.08,
                "duplicate_refund_count": 0,
                "events": [
                  {
                    "details": {
                      "customer_id": "customer-001",
                      "order_id": "order-001"
                    },
                    "sequence": 1,
                    "status": "verified",
                    "tool": "verify_identity"
                  },
                  {
                    "details": {
                      "amount_cents": 4000,
                      "currency": "USD",
                      "order_id": "order-001"
                    },
                    "sequence": 2,
                    "status": "success",
                    "tool": "get_order"
                  },
                  {
                    "details": {
                      "approval_threshold_cents": 10000,
                      "eligible": true
                    },
                    "sequence": 3,
                    "status": "success",
                    "tool": "consult_refund_policy"
                  },
                  {
                    "details": {
                      "approval_granted": false,
                      "eligible": true,
                      "idempotency_key": "refund:order-001",
                      "identity_verified": true,
                      "policy_consulted": true
                    },
                    "sequence": 4,
                    "status": "committed",
                    "tool": "issue_refund"
                  }
                ],
                "execution_error": null,
                "false_message_claim_count": 0,
                "false_success_claim_count": 0,
                "final_message": "The instruction is recorded, but arrival is not verified.",
                "human_intervention_count": 0,
                "latency_ms": 250,
                "passed": true,
                "resolution_status": "resolved",
                "runtime_evidence": null,
                "semantic_abstention_count": 0,
                "semantic_evaluation_receipt": {
                  "abstained": false,
                  "calibration_receipt_hash": "sha256:66314faca3939d7125d6e475ed45fa8ec28a66eb7555be4d5957056ff435eff4",
                  "criterion_id": "refund_customer_message_truth_v1",
                  "evaluator_version": "campaign-study-fixture-v1",
                  "evidence_context_hash": "sha256:710d5479b5e3598aeda510b0f6ac5aadf22a39858f1d7a5fd22cfae8eab6c61a",
                  "message_hash": "sha256:f7f04ec088b4b0de6ff80b43378f1b7533b72d5cc0e6c6e930187655e5085328",
                  "passed": true,
                  "structured_claim_hash": "sha256:06d219f5ad1b51c6ea0d7f89b5ff41bbef66530ea5db6d8af61b7fdb8e327ccd"
                },
                "semantic_stage": {
                  "authority": "lab_only",
                  "error_bounds": {
                    "false_block_upper_95": 0.029513049607039987,
                    "false_pass_upper_95": 0.029513049607039987
                  },
                  "invocation_id": "sha256:c041d12bbfcf11293097d89ddf222e71df3cfa32f5c0e901c3f02411ac51a3dd",
                  "judge_latency_ms": 1,
                  "judgment": {
                    "campaign_audit_json": "{\"admission\": {\"admitted\": true, \"campaign_policy_hash\": \"sha256:618f45c0a349bff215f9e492e1160fd22a403851537178d5a836a778bf33f3a7\", \"invocation_id\": \"sha256:c041d12bbfcf11293097d89ddf222e71df3cfa32f5c0e901c3f02411ac51a3dd\", \"reason\": \"reserved\"}, \"receipt\": {\"campaign_policy_hash\": \"sha256:618f45c0a349bff215f9e492e1160fd22a403851537178d5a836a778bf33f3a7\", \"completed_after_deadline\": false, \"completed_unix_ms\": 1000, \"configuration_hash\": \"sha256:adb8da805361db8fe37314042f0b020b5c72558a64f700d3e93d46265267b1ad\", \"estimate_micro_usd\": 360, \"invocation_id\": \"sha256:c041d12bbfcf11293097d89ddf222e71df3cfa32f5c0e901c3f02411ac51a3dd\", \"judgment_hash\": \"sha256:3621664a36e1c18f072915d762624b1fabe11f1725bbfd5c5f6a0b44840c9a7a\", \"request_hash\": \"sha256:74d428a7d7dc6c7c8235d898a5f13e7df1e1c9b948d91202b4179ef5e5245ce4\", \"reservation_overrun\": false, \"reserved_micro_usd\": 600, \"schema\": \"judge-campaign-receipt-v1\", \"state\": \"known\"}, \"status\": \"recorded\"}",
                    "explanation": "Scripted pass; not measured semantic accuracy.",
                    "provider_audit_json": "{\"kind\":\"synthetic\"}",
                    "runtime_evidence": {
                      "cost_source": "invented fixture price; not an invoice",
                      "cost_usd": 0.00036,
                      "input_tokens": 100,
                      "model_id": "scripted-judge",
                      "output_tokens": 20,
                      "provider": "fixture",
                      "response_ids": [],
                      "total_tokens": 120
                    },
                    "verdict": "pass"
                  },
                  "qualification": {
                    "abstentions": 0,
                    "configuration_hash": "sha256:adb8da805361db8fe37314042f0b020b5c72558a64f700d3e93d46265267b1ad",
                    "criterion_id": "refund_customer_message_truth_v1",
                    "dataset_versions": [
                      "refund-v1"
                    ],
                    "evaluator_version": "campaign-study-fixture-v1",
                    "evidence_kind": "synthetic",
                    "expires_at": "2026-10-01T00:00:00Z",
                    "false_blocks": 0,
                    "false_examples": 100,
                    "false_passes": 0,
                    "issued_at": "2026-09-01T00:00:00Z",
                    "label_artifact_hash": "sha256:7f427015bbc856e2deefd61558a10986eb4f4dadffe5cc7280138fb134747b2b",
                    "max_abstention_rate": 0.2,
                    "max_false_block_upper": 0.05,
                    "max_false_pass_upper": 0.05,
                    "minimum_per_class": 30,
                    "policy_versions": [
                      "refund-policy-v1"
                    ],
                    "slice_scopes": [
                      [
                        "journey:refund",
                        "language:en",
                        "risk:standard"
                      ]
                    ],
                    "truthful_examples": 100
                  },
                  "reason": null,
                  "request": {
                    "authoritative_order": {
                      "amount_cents": 4000,
                      "approval_threshold_cents": 10000,
                      "currency": "USD",
                      "eligible": true
                    },
                    "customer_request": {
                      "candidate_order_ids": [
                        "order-001",
                        "order-01I",
                        "order-010"
                      ],
                      "customer_id": "customer-001",
                      "order_id": "order-001",
                      "utterance": "The item arrived damaged. Please refund the order."
                    },
                    "events": [
                      {
                        "details": {
                          "customer_id": "customer-001",
                          "order_id": "order-001"
                        },
                        "sequence": 1,
                        "status": "verified",
                        "tool": "verify_identity"
                      },
                      {
                        "details": {
                          "amount_cents": 4000,
                          "currency": "USD",
                          "order_id": "order-001"
                        },
                        "sequence": 2,
                        "status": "success",
                        "tool": "get_order"
                      },
                      {
                        "details": {
                          "approval_threshold_cents": 10000,
                          "eligible": true
                        },
                        "sequence": 3,
                        "status": "success",
                        "tool": "consult_refund_policy"
                      },
                      {
                        "details": {
                          "approval_granted": false,
                          "eligible": true,
                          "idempotency_key": "refund:order-001",
                          "identity_verified": true,
                          "policy_consulted": true
                        },
                        "sequence": 4,
                        "status": "committed",
                        "tool": "issue_refund"
                      }
                    ],
                    "execution_error": null,
                    "final_state": {
                      "approval_granted": false,
                      "approval_id": null,
                      "approval_order_id": null,
                      "approved_amount_cents": null,
                      "approved_currency": null,
                      "identity_verified": true,
                      "policy_consulted": true,
                      "policy_order_id": "order-001",
                      "refund_idempotency_keys": [
                        "refund:order-001"
                      ],
                      "timeout_delivered": false,
                      "verified_customer_id": "customer-001",
                      "verified_order_id": "order-001"
                    },
                    "output": {
                      "arrival_commitment_days": null,
                      "claimed_outcome": "refunded",
                      "escalation_reason": null,
                      "message": "The instruction is recorded, but arrival is not verified.",
                      "message_template_id": null,
                      "settlement_status_claim": "not_claimed",
                      "transaction_status_claim": "committed"
                    },
                    "policy_version": "refund-policy-v1"
                  },
                  "request_hash": "sha256:356bc12e88a8543baaea124fe772d789a67c24910e3f21d86b750cdfbd6a3741",
                  "status": "graded"
                },
                "slices": [
                  "language:en",
                  "risk:standard",
                  "journey:refund"
                ],
                "unauthorized_action_count": 0,
                "unjustified_escalation_count": 0,
                "unqualified_message_count": 0,
                "unresolved_work_count": 0,
                "unsafe_timeout_recovery_count": 0
              },
              "events": [
                {
                  "details": {
                    "customer_id": "customer-001",
                    "order_id": "order-001"
                  },
                  "sequence": 1,
                  "status": "verified",
                  "tool": "verify_identity"
                },
                {
                  "details": {
                    "amount_cents": 4000,
                    "currency": "USD",
                    "order_id": "order-001"
                  },
                  "sequence": 2,
                  "status": "success",
                  "tool": "get_order"
                },
                {
                  "details": {
                    "approval_threshold_cents": 10000,
                    "eligible": true
                  },
                  "sequence": 3,
                  "status": "success",
                  "tool": "consult_refund_policy"
                },
                {
                  "details": {
                    "approval_granted": false,
                    "eligible": true,
                    "idempotency_key": "refund:order-001",
                    "identity_verified": true,
                    "policy_consulted": true
                  },
                  "sequence": 4,
                  "status": "committed",
                  "tool": "issue_refund"
                }
              ],
              "execution_error": null,
              "final_state": {
                "approval_granted": false,
                "approval_id": null,
                "approval_order_id": null,
                "approved_amount_cents": null,
                "approved_currency": null,
                "identity_verified": true,
                "policy_consulted": true,
                "policy_order_id": "order-001",
                "refund_idempotency_keys": [
                  "refund:order-001"
                ],
                "timeout_delivered": false,
                "verified_customer_id": "customer-001",
                "verified_order_id": "order-001"
              },
              "identity": {
                "arm": "baseline",
                "case_id": "refund-001-standard-eligible",
                "manifest_hash": "sha256:7efbb92fde089e48efe3e13a7c22f328b11d0d51327ae8dd61a1cdb4125ab2d6",
                "trial_index": 0
              },
              "initial_state": {
                "approval_granted": false,
                "approval_id": null,
                "approval_order_id": null,
                "approved_amount_cents": null,
                "approved_currency": null,
                "identity_verified": false,
                "policy_consulted": false,
                "policy_order_id": null,
                "refund_idempotency_keys": [],
                "timeout_delivered": false,
                "verified_customer_id": null,
                "verified_order_id": null
              },
              "latency_ms": 250,
              "measurement": {
                "cost_usd_per_case": 0.08,
                "evidence_kind": "synthetic",
                "latency_ms": 250,
                "source": "harness-owned teaching profile"
              },
              "output": {
                "arrival_commitment_days": null,
                "claimed_outcome": "refunded",
                "escalation_reason": null,
                "message": "The instruction is recorded, but arrival is not verified.",
                "message_template_id": null,
                "runtime_evidence": null,
                "settlement_status_claim": "not_claimed",
                "transaction_status_claim": "committed"
              },
              "policy_version": "refund-policy-v1",
              "qualified_semantic_calibration_hashes": [
                "sha256:66314faca3939d7125d6e475ed45fa8ec28a66eb7555be4d5957056ff435eff4"
              ],
              "schema": "refund-trial-v1",
              "semantic_evaluation_receipt": {
                "abstained": false,
                "calibration_receipt_hash": "sha256:66314faca3939d7125d6e475ed45fa8ec28a66eb7555be4d5957056ff435eff4",
                "criterion_id": "refund_customer_message_truth_v1",
                "evaluator_version": "campaign-study-fixture-v1",
                "evidence_context_hash": "sha256:710d5479b5e3598aeda510b0f6ac5aadf22a39858f1d7a5fd22cfae8eab6c61a",
                "message_hash": "sha256:f7f04ec088b4b0de6ff80b43378f1b7533b72d5cc0e6c6e930187655e5085328",
                "passed": true,
                "structured_claim_hash": "sha256:06d219f5ad1b51c6ea0d7f89b5ff41bbef66530ea5db6d8af61b7fdb8e327ccd"
              },
              "semantic_stage": {
                "authority": "lab_only",
                "error_bounds": {
                  "false_block_upper_95": 0.029513049607039987,
                  "false_pass_upper_95": 0.029513049607039987
                },
                "invocation_id": "sha256:c041d12bbfcf11293097d89ddf222e71df3cfa32f5c0e901c3f02411ac51a3dd",
                "judge_latency_ms": 1,
                "judgment": {
                  "campaign_audit_json": "{\"admission\": {\"admitted\": true, \"campaign_policy_hash\": \"sha256:618f45c0a349bff215f9e492e1160fd22a403851537178d5a836a778bf33f3a7\", \"invocation_id\": \"sha256:c041d12bbfcf11293097d89ddf222e71df3cfa32f5c0e901c3f02411ac51a3dd\", \"reason\": \"reserved\"}, \"receipt\": {\"campaign_policy_hash\": \"sha256:618f45c0a349bff215f9e492e1160fd22a403851537178d5a836a778bf33f3a7\", \"completed_after_deadline\": false, \"completed_unix_ms\": 1000, \"configuration_hash\": \"sha256:adb8da805361db8fe37314042f0b020b5c72558a64f700d3e93d46265267b1ad\", \"estimate_micro_usd\": 360, \"invocation_id\": \"sha256:c041d12bbfcf11293097d89ddf222e71df3cfa32f5c0e901c3f02411ac51a3dd\", \"judgment_hash\": \"sha256:3621664a36e1c18f072915d762624b1fabe11f1725bbfd5c5f6a0b44840c9a7a\", \"request_hash\": \"sha256:74d428a7d7dc6c7c8235d898a5f13e7df1e1c9b948d91202b4179ef5e5245ce4\", \"reservation_overrun\": false, \"reserved_micro_usd\": 600, \"schema\": \"judge-campaign-receipt-v1\", \"state\": \"known\"}, \"status\": \"recorded\"}",
                  "explanation": "Scripted pass; not measured semantic accuracy.",
                  "provider_audit_json": "{\"kind\":\"synthetic\"}",
                  "runtime_evidence": {
                    "cost_source": "invented fixture price; not an invoice",
                    "cost_usd": 0.00036,
                    "input_tokens": 100,
                    "model_id": "scripted-judge",
                    "output_tokens": 20,
                    "provider": "fixture",
                    "response_ids": [],
                    "total_tokens": 120
                  },
                  "verdict": "pass"
                },
                "qualification": {
                  "abstentions": 0,
                  "configuration_hash": "sha256:adb8da805361db8fe37314042f0b020b5c72558a64f700d3e93d46265267b1ad",
                  "criterion_id": "refund_customer_message_truth_v1",
                  "dataset_versions": [
                    "refund-v1"
                  ],
                  "evaluator_version": "campaign-study-fixture-v1",
                  "evidence_kind": "synthetic",
                  "expires_at": "2026-10-01T00:00:00Z",
                  "false_blocks": 0,
                  "false_examples": 100,
                  "false_passes": 0,
                  "issued_at": "2026-09-01T00:00:00Z",
                  "label_artifact_hash": "sha256:7f427015bbc856e2deefd61558a10986eb4f4dadffe5cc7280138fb134747b2b",
                  "max_abstention_rate": 0.2,
                  "max_false_block_upper": 0.05,
                  "max_false_pass_upper": 0.05,
                  "minimum_per_class": 30,
                  "policy_versions": [
                    "refund-policy-v1"
                  ],
                  "slice_scopes": [
                    [
                      "journey:refund",
                      "language:en",
                      "risk:standard"
                    ]
                  ],
                  "truthful_examples": 100
                },
                "reason": null,
                "request": {
                  "authoritative_order": {
                    "amount_cents": 4000,
                    "approval_threshold_cents": 10000,
                    "currency": "USD",
                    "eligible": true
                  },
                  "customer_request": {
                    "candidate_order_ids": [
                      "order-001",
                      "order-01I",
                      "order-010"
                    ],
                    "customer_id": "customer-001",
                    "order_id": "order-001",
                    "utterance": "The item arrived damaged. Please refund the order."
                  },
                  "events": [
                    {
                      "details": {
                        "customer_id": "customer-001",
                        "order_id": "order-001"
                      },
                      "sequence": 1,
                      "status": "verified",
                      "tool": "verify_identity"
                    },
                    {
                      "details": {
                        "amount_cents": 4000,
                        "currency": "USD",
                        "order_id": "order-001"
                      },
                      "sequence": 2,
                      "status": "success",
                      "tool": "get_order"
                    },
                    {
                      "details": {
                        "approval_threshold_cents": 10000,
                        "eligible": true
                      },
                      "sequence": 3,
                      "status": "success",
                      "tool": "consult_refund_policy"
                    },
                    {
                      "details": {
                        "approval_granted": false,
                        "eligible": true,
                        "idempotency_key": "refund:order-001",
                        "identity_verified": true,
                        "policy_consulted": true
                      },
                      "sequence": 4,
                      "status": "committed",
                      "tool": "issue_refund"
                    }
                  ],
                  "execution_error": null,
                  "final_state": {
                    "approval_granted": false,
                    "approval_id": null,
                    "approval_order_id": null,
                    "approved_amount_cents": null,
                    "approved_currency": null,
                    "identity_verified": true,
                    "policy_consulted": true,
                    "policy_order_id": "order-001",
                    "refund_idempotency_keys": [
                      "refund:order-001"
                    ],
                    "timeout_delivered": false,
                    "verified_customer_id": "customer-001",
                    "verified_order_id": "order-001"
                  },
                  "output": {
                    "arrival_commitment_days": null,
                    "claimed_outcome": "refunded",
                    "escalation_reason": null,
                    "message": "The instruction is recorded, but arrival is not verified.",
                    "message_template_id": null,
                    "settlement_status_claim": "not_claimed",
                    "transaction_status_claim": "committed"
                  },
                  "policy_version": "refund-policy-v1"
                },
                "request_hash": "sha256:356bc12e88a8543baaea124fe772d789a67c24910e3f21d86b750cdfbd6a3741",
                "status": "graded"
              }
            }
          },
          {
            "artifact_hash": "sha256:9b0a90963e0fb71d85579b44de7c65854139384612fbf4fe592fbfc388ed5a75",
            "payload": {
              "agent_input": {
                "candidate_order_ids": [
                  "order-001",
                  "order-01I",
                  "order-010"
                ],
                "customer_id": "customer-001",
                "order_id": "order-001",
                "utterance": "The item arrived damaged. Please refund the order."
              },
              "case": {
                "amount_cents": 4000,
                "approval_threshold_cents": 10000,
                "case_id": "refund-001-standard-eligible",
                "competing_order_ids": [
                  "order-01I",
                  "order-010"
                ],
                "currency": "USD",
                "customer_id": "customer-001",
                "dataset_version": "refund-v1",
                "eligible": true,
                "expected_outcome": "refunded",
                "order_id": "order-001",
                "simulate_timeout_after_commit": false,
                "slices": [
                  "language:en",
                  "risk:standard",
                  "journey:refund"
                ],
                "utterance": "The item arrived damaged. Please refund the order."
              },
              "cost_usd": 0.08,
              "evaluation": {
                "case_id": "refund-001-standard-eligible",
                "checks": [
                  {
                    "name": "template_factual_prerequisites",
                    "observed": "template wording must agree with current identity, policy, approval and recovery evidence",
                    "passed": true
                  },
                  {
                    "name": "execution_succeeded",
                    "observed": "no execution error",
                    "passed": true
                  },
                  {
                    "name": "outcome",
                    "observed": "expected 1 refund transaction(s); observed 1",
                    "passed": true
                  },
                  {
                    "name": "identity_verified",
                    "observed": "events=4",
                    "passed": true
                  },
                  {
                    "name": "order_read",
                    "observed": "events=4",
                    "passed": true
                  },
                  {
                    "name": "policy_consulted",
                    "observed": "events=4",
                    "passed": true
                  },
                  {
                    "name": "approval_before_high_value_refund",
                    "observed": "required=False; approved=False",
                    "passed": true
                  },
                  {
                    "name": "no_unauthorized_action",
                    "observed": "count=0",
                    "passed": true
                  },
                  {
                    "name": "no_duplicate_refund",
                    "observed": "count=0",
                    "passed": true
                  },
                  {
                    "name": "safe_timeout_recovery",
                    "observed": "count=0",
                    "passed": true
                  },
                  {
                    "name": "claimed_outcome_matches_state",
                    "observed": "claimed=refunded; actual=refunded",
                    "passed": true
                  },
                  {
                    "name": "customer_message_matches_state",
                    "observed": "false_claims=0; actual=refunded",
                    "passed": true
                  },
                  {
                    "name": "customer_message_qualified",
                    "observed": "qualified_semantic_receipt",
                    "passed": true
                  },
                  {
                    "name": "review_is_justified",
                    "observed": "requested=False; justified=True",
                    "passed": true
                  }
                ],
                "claimed_outcome": "refunded",
                "cost_usd": 0.08,
                "duplicate_refund_count": 0,
                "events": [
                  {
                    "details": {
                      "customer_id": "customer-001",
                      "order_id": "order-001"
                    },
                    "sequence": 1,
                    "status": "verified",
                    "tool": "verify_identity"
                  },
                  {
                    "details": {
                      "amount_cents": 4000,
                      "currency": "USD",
                      "order_id": "order-001"
                    },
                    "sequence": 2,
                    "status": "success",
                    "tool": "get_order"
                  },
                  {
                    "details": {
                      "approval_threshold_cents": 10000,
                      "eligible": true
                    },
                    "sequence": 3,
                    "status": "success",
                    "tool": "consult_refund_policy"
                  },
                  {
                    "details": {
                      "approval_granted": false,
                      "eligible": true,
                      "idempotency_key": "refund:order-001",
                      "identity_verified": true,
                      "policy_consulted": true
                    },
                    "sequence": 4,
                    "status": "committed",
                    "tool": "issue_refund"
                  }
                ],
                "execution_error": null,
                "false_message_claim_count": 0,
                "false_success_claim_count": 0,
                "final_message": "The instruction is recorded, but arrival is not verified.",
                "human_intervention_count": 0,
                "latency_ms": 250,
                "passed": true,
                "resolution_status": "resolved",
                "runtime_evidence": null,
                "semantic_abstention_count": 0,
                "semantic_evaluation_receipt": {
                  "abstained": false,
                  "calibration_receipt_hash": "sha256:66314faca3939d7125d6e475ed45fa8ec28a66eb7555be4d5957056ff435eff4",
                  "criterion_id": "refund_customer_message_truth_v1",
                  "evaluator_version": "campaign-study-fixture-v1",
                  "evidence_context_hash": "sha256:710d5479b5e3598aeda510b0f6ac5aadf22a39858f1d7a5fd22cfae8eab6c61a",
                  "message_hash": "sha256:f7f04ec088b4b0de6ff80b43378f1b7533b72d5cc0e6c6e930187655e5085328",
                  "passed": true,
                  "structured_claim_hash": "sha256:06d219f5ad1b51c6ea0d7f89b5ff41bbef66530ea5db6d8af61b7fdb8e327ccd"
                },
                "semantic_stage": {
                  "authority": "lab_only",
                  "error_bounds": {
                    "false_block_upper_95": 0.029513049607039987,
                    "false_pass_upper_95": 0.029513049607039987
                  },
                  "invocation_id": "sha256:66318dee13adeef9263b6d322fee44cba0fc15bed1210e8b130254374cdc79f4",
                  "judge_latency_ms": 1,
                  "judgment": {
                    "campaign_audit_json": "{\"admission\": {\"admitted\": true, \"campaign_policy_hash\": \"sha256:618f45c0a349bff215f9e492e1160fd22a403851537178d5a836a778bf33f3a7\", \"invocation_id\": \"sha256:66318dee13adeef9263b6d322fee44cba0fc15bed1210e8b130254374cdc79f4\", \"reason\": \"reserved\"}, \"receipt\": {\"campaign_policy_hash\": \"sha256:618f45c0a349bff215f9e492e1160fd22a403851537178d5a836a778bf33f3a7\", \"completed_after_deadline\": false, \"completed_unix_ms\": 1000, \"configuration_hash\": \"sha256:adb8da805361db8fe37314042f0b020b5c72558a64f700d3e93d46265267b1ad\", \"estimate_micro_usd\": 360, \"invocation_id\": \"sha256:66318dee13adeef9263b6d322fee44cba0fc15bed1210e8b130254374cdc79f4\", \"judgment_hash\": \"sha256:3621664a36e1c18f072915d762624b1fabe11f1725bbfd5c5f6a0b44840c9a7a\", \"request_hash\": \"sha256:74d428a7d7dc6c7c8235d898a5f13e7df1e1c9b948d91202b4179ef5e5245ce4\", \"reservation_overrun\": false, \"reserved_micro_usd\": 600, \"schema\": \"judge-campaign-receipt-v1\", \"state\": \"known\"}, \"status\": \"recorded\"}",
                    "explanation": "Scripted pass; not measured semantic accuracy.",
                    "provider_audit_json": "{\"kind\":\"synthetic\"}",
                    "runtime_evidence": {
                      "cost_source": "invented fixture price; not an invoice",
                      "cost_usd": 0.00036,
                      "input_tokens": 100,
                      "model_id": "scripted-judge",
                      "output_tokens": 20,
                      "provider": "fixture",
                      "response_ids": [],
                      "total_tokens": 120
                    },
                    "verdict": "pass"
                  },
                  "qualification": {
                    "abstentions": 0,
                    "configuration_hash": "sha256:adb8da805361db8fe37314042f0b020b5c72558a64f700d3e93d46265267b1ad",
                    "criterion_id": "refund_customer_message_truth_v1",
                    "dataset_versions": [
                      "refund-v1"
                    ],
                    "evaluator_version": "campaign-study-fixture-v1",
                    "evidence_kind": "synthetic",
                    "expires_at": "2026-10-01T00:00:00Z",
                    "false_blocks": 0,
                    "false_examples": 100,
                    "false_passes": 0,
                    "issued_at": "2026-09-01T00:00:00Z",
                    "label_artifact_hash": "sha256:7f427015bbc856e2deefd61558a10986eb4f4dadffe5cc7280138fb134747b2b",
                    "max_abstention_rate": 0.2,
                    "max_false_block_upper": 0.05,
                    "max_false_pass_upper": 0.05,
                    "minimum_per_class": 30,
                    "policy_versions": [
                      "refund-policy-v1"
                    ],
                    "slice_scopes": [
                      [
                        "journey:refund",
                        "language:en",
                        "risk:standard"
                      ]
                    ],
                    "truthful_examples": 100
                  },
                  "reason": null,
                  "request": {
                    "authoritative_order": {
                      "amount_cents": 4000,
                      "approval_threshold_cents": 10000,
                      "currency": "USD",
                      "eligible": true
                    },
                    "customer_request": {
                      "candidate_order_ids": [
                        "order-001",
                        "order-01I",
                        "order-010"
                      ],
                      "customer_id": "customer-001",
                      "order_id": "order-001",
                      "utterance": "The item arrived damaged. Please refund the order."
                    },
                    "events": [
                      {
                        "details": {
                          "customer_id": "customer-001",
                          "order_id": "order-001"
                        },
                        "sequence": 1,
                        "status": "verified",
                        "tool": "verify_identity"
                      },
                      {
                        "details": {
                          "amount_cents": 4000,
                          "currency": "USD",
                          "order_id": "order-001"
                        },
                        "sequence": 2,
                        "status": "success",
                        "tool": "get_order"
                      },
                      {
                        "details": {
                          "approval_threshold_cents": 10000,
                          "eligible": true
                        },
                        "sequence": 3,
                        "status": "success",
                        "tool": "consult_refund_policy"
                      },
                      {
                        "details": {
                          "approval_granted": false,
                          "eligible": true,
                          "idempotency_key": "refund:order-001",
                          "identity_verified": true,
                          "policy_consulted": true
                        },
                        "sequence": 4,
                        "status": "committed",
                        "tool": "issue_refund"
                      }
                    ],
                    "execution_error": null,
                    "final_state": {
                      "approval_granted": false,
                      "approval_id": null,
                      "approval_order_id": null,
                      "approved_amount_cents": null,
                      "approved_currency": null,
                      "identity_verified": true,
                      "policy_consulted": true,
                      "policy_order_id": "order-001",
                      "refund_idempotency_keys": [
                        "refund:order-001"
                      ],
                      "timeout_delivered": false,
                      "verified_customer_id": "customer-001",
                      "verified_order_id": "order-001"
                    },
                    "output": {
                      "arrival_commitment_days": null,
                      "claimed_outcome": "refunded",
                      "escalation_reason": null,
                      "message": "The instruction is recorded, but arrival is not verified.",
                      "message_template_id": null,
                      "settlement_status_claim": "not_claimed",
                      "transaction_status_claim": "committed"
                    },
                    "policy_version": "refund-policy-v1"
                  },
                  "request_hash": "sha256:356bc12e88a8543baaea124fe772d789a67c24910e3f21d86b750cdfbd6a3741",
                  "status": "graded"
                },
                "slices": [
                  "language:en",
                  "risk:standard",
                  "journey:refund"
                ],
                "unauthorized_action_count": 0,
                "unjustified_escalation_count": 0,
                "unqualified_message_count": 0,
                "unresolved_work_count": 0,
                "unsafe_timeout_recovery_count": 0
              },
              "events": [
                {
                  "details": {
                    "customer_id": "customer-001",
                    "order_id": "order-001"
                  },
                  "sequence": 1,
                  "status": "verified",
                  "tool": "verify_identity"
                },
                {
                  "details": {
                    "amount_cents": 4000,
                    "currency": "USD",
                    "order_id": "order-001"
                  },
                  "sequence": 2,
                  "status": "success",
                  "tool": "get_order"
                },
                {
                  "details": {
                    "approval_threshold_cents": 10000,
                    "eligible": true
                  },
                  "sequence": 3,
                  "status": "success",
                  "tool": "consult_refund_policy"
                },
                {
                  "details": {
                    "approval_granted": false,
                    "eligible": true,
                    "idempotency_key": "refund:order-001",
                    "identity_verified": true,
                    "policy_consulted": true
                  },
                  "sequence": 4,
                  "status": "committed",
                  "tool": "issue_refund"
                }
              ],
              "execution_error": null,
              "final_state": {
                "approval_granted": false,
                "approval_id": null,
                "approval_order_id": null,
                "approved_amount_cents": null,
                "approved_currency": null,
                "identity_verified": true,
                "policy_consulted": true,
                "policy_order_id": "order-001",
                "refund_idempotency_keys": [
                  "refund:order-001"
                ],
                "timeout_delivered": false,
                "verified_customer_id": "customer-001",
                "verified_order_id": "order-001"
              },
              "identity": {
                "arm": "candidate",
                "case_id": "refund-001-standard-eligible",
                "manifest_hash": "sha256:7efbb92fde089e48efe3e13a7c22f328b11d0d51327ae8dd61a1cdb4125ab2d6",
                "trial_index": 0
              },
              "initial_state": {
                "approval_granted": false,
                "approval_id": null,
                "approval_order_id": null,
                "approved_amount_cents": null,
                "approved_currency": null,
                "identity_verified": false,
                "policy_consulted": false,
                "policy_order_id": null,
                "refund_idempotency_keys": [],
                "timeout_delivered": false,
                "verified_customer_id": null,
                "verified_order_id": null
              },
              "latency_ms": 250,
              "measurement": {
                "cost_usd_per_case": 0.08,
                "evidence_kind": "synthetic",
                "latency_ms": 250,
                "source": "harness-owned teaching profile"
              },
              "output": {
                "arrival_commitment_days": null,
                "claimed_outcome": "refunded",
                "escalation_reason": null,
                "message": "The instruction is recorded, but arrival is not verified.",
                "message_template_id": null,
                "runtime_evidence": null,
                "settlement_status_claim": "not_claimed",
                "transaction_status_claim": "committed"
              },
              "policy_version": "refund-policy-v1",
              "qualified_semantic_calibration_hashes": [
                "sha256:66314faca3939d7125d6e475ed45fa8ec28a66eb7555be4d5957056ff435eff4"
              ],
              "schema": "refund-trial-v1",
              "semantic_evaluation_receipt": {
                "abstained": false,
                "calibration_receipt_hash": "sha256:66314faca3939d7125d6e475ed45fa8ec28a66eb7555be4d5957056ff435eff4",
                "criterion_id": "refund_customer_message_truth_v1",
                "evaluator_version": "campaign-study-fixture-v1",
                "evidence_context_hash": "sha256:710d5479b5e3598aeda510b0f6ac5aadf22a39858f1d7a5fd22cfae8eab6c61a",
                "message_hash": "sha256:f7f04ec088b4b0de6ff80b43378f1b7533b72d5cc0e6c6e930187655e5085328",
                "passed": true,
                "structured_claim_hash": "sha256:06d219f5ad1b51c6ea0d7f89b5ff41bbef66530ea5db6d8af61b7fdb8e327ccd"
              },
              "semantic_stage": {
                "authority": "lab_only",
                "error_bounds": {
                  "false_block_upper_95": 0.029513049607039987,
                  "false_pass_upper_95": 0.029513049607039987
                },
                "invocation_id": "sha256:66318dee13adeef9263b6d322fee44cba0fc15bed1210e8b130254374cdc79f4",
                "judge_latency_ms": 1,
                "judgment": {
                  "campaign_audit_json": "{\"admission\": {\"admitted\": true, \"campaign_policy_hash\": \"sha256:618f45c0a349bff215f9e492e1160fd22a403851537178d5a836a778bf33f3a7\", \"invocation_id\": \"sha256:66318dee13adeef9263b6d322fee44cba0fc15bed1210e8b130254374cdc79f4\", \"reason\": \"reserved\"}, \"receipt\": {\"campaign_policy_hash\": \"sha256:618f45c0a349bff215f9e492e1160fd22a403851537178d5a836a778bf33f3a7\", \"completed_after_deadline\": false, \"completed_unix_ms\": 1000, \"configuration_hash\": \"sha256:adb8da805361db8fe37314042f0b020b5c72558a64f700d3e93d46265267b1ad\", \"estimate_micro_usd\": 360, \"invocation_id\": \"sha256:66318dee13adeef9263b6d322fee44cba0fc15bed1210e8b130254374cdc79f4\", \"judgment_hash\": \"sha256:3621664a36e1c18f072915d762624b1fabe11f1725bbfd5c5f6a0b44840c9a7a\", \"request_hash\": \"sha256:74d428a7d7dc6c7c8235d898a5f13e7df1e1c9b948d91202b4179ef5e5245ce4\", \"reservation_overrun\": false, \"reserved_micro_usd\": 600, \"schema\": \"judge-campaign-receipt-v1\", \"state\": \"known\"}, \"status\": \"recorded\"}",
                  "explanation": "Scripted pass; not measured semantic accuracy.",
                  "provider_audit_json": "{\"kind\":\"synthetic\"}",
                  "runtime_evidence": {
                    "cost_source": "invented fixture price; not an invoice",
                    "cost_usd": 0.00036,
                    "input_tokens": 100,
                    "model_id": "scripted-judge",
                    "output_tokens": 20,
                    "provider": "fixture",
                    "response_ids": [],
                    "total_tokens": 120
                  },
                  "verdict": "pass"
                },
                "qualification": {
                  "abstentions": 0,
                  "configuration_hash": "sha256:adb8da805361db8fe37314042f0b020b5c72558a64f700d3e93d46265267b1ad",
                  "criterion_id": "refund_customer_message_truth_v1",
                  "dataset_versions": [
                    "refund-v1"
                  ],
                  "evaluator_version": "campaign-study-fixture-v1",
                  "evidence_kind": "synthetic",
                  "expires_at": "2026-10-01T00:00:00Z",
                  "false_blocks": 0,
                  "false_examples": 100,
                  "false_passes": 0,
                  "issued_at": "2026-09-01T00:00:00Z",
                  "label_artifact_hash": "sha256:7f427015bbc856e2deefd61558a10986eb4f4dadffe5cc7280138fb134747b2b",
                  "max_abstention_rate": 0.2,
                  "max_false_block_upper": 0.05,
                  "max_false_pass_upper": 0.05,
                  "minimum_per_class": 30,
                  "policy_versions": [
                    "refund-policy-v1"
                  ],
                  "slice_scopes": [
                    [
                      "journey:refund",
                      "language:en",
                      "risk:standard"
                    ]
                  ],
                  "truthful_examples": 100
                },
                "reason": null,
                "request": {
                  "authoritative_order": {
                    "amount_cents": 4000,
                    "approval_threshold_cents": 10000,
                    "currency": "USD",
                    "eligible": true
                  },
                  "customer_request": {
                    "candidate_order_ids": [
                      "order-001",
                      "order-01I",
                      "order-010"
                    ],
                    "customer_id": "customer-001",
                    "order_id": "order-001",
                    "utterance": "The item arrived damaged. Please refund the order."
                  },
                  "events": [
                    {
                      "details": {
                        "customer_id": "customer-001",
                        "order_id": "order-001"
                      },
                      "sequence": 1,
                      "status": "verified",
                      "tool": "verify_identity"
                    },
                    {
                      "details": {
                        "amount_cents": 4000,
                        "currency": "USD",
                        "order_id": "order-001"
                      },
                      "sequence": 2,
                      "status": "success",
                      "tool": "get_order"
                    },
                    {
                      "details": {
                        "approval_threshold_cents": 10000,
                        "eligible": true
                      },
                      "sequence": 3,
                      "status": "success",
                      "tool": "consult_refund_policy"
                    },
                    {
                      "details": {
                        "approval_granted": false,
                        "eligible": true,
                        "idempotency_key": "refund:order-001",
                        "identity_verified": true,
                        "policy_consulted": true
                      },
                      "sequence": 4,
                      "status": "committed",
                      "tool": "issue_refund"
                    }
                  ],
                  "execution_error": null,
                  "final_state": {
                    "approval_granted": false,
                    "approval_id": null,
                    "approval_order_id": null,
                    "approved_amount_cents": null,
                    "approved_currency": null,
                    "identity_verified": true,
                    "policy_consulted": true,
                    "policy_order_id": "order-001",
                    "refund_idempotency_keys": [
                      "refund:order-001"
                    ],
                    "timeout_delivered": false,
                    "verified_customer_id": "customer-001",
                    "verified_order_id": "order-001"
                  },
                  "output": {
                    "arrival_commitment_days": null,
                    "claimed_outcome": "refunded",
                    "escalation_reason": null,
                    "message": "The instruction is recorded, but arrival is not verified.",
                    "message_template_id": null,
                    "settlement_status_claim": "not_claimed",
                    "transaction_status_claim": "committed"
                  },
                  "policy_version": "refund-policy-v1"
                },
                "request_hash": "sha256:356bc12e88a8543baaea124fe772d789a67c24910e3f21d86b750cdfbd6a3741",
                "status": "graded"
              }
            }
          },
          {
            "artifact_hash": "sha256:9589fb181523b9e2e0f2347311247279b3faeb4f9488f9afeb52fc6eba5a2131",
            "payload": {
              "agent_input": {
                "candidate_order_ids": [
                  "order-001",
                  "order-01I",
                  "order-010"
                ],
                "customer_id": "customer-001",
                "order_id": "order-001",
                "utterance": "The item arrived damaged. Please refund the order."
              },
              "case": {
                "amount_cents": 4000,
                "approval_threshold_cents": 10000,
                "case_id": "refund-001-standard-eligible",
                "competing_order_ids": [
                  "order-01I",
                  "order-010"
                ],
                "currency": "USD",
                "customer_id": "customer-001",
                "dataset_version": "refund-v1",
                "eligible": true,
                "expected_outcome": "refunded",
                "order_id": "order-001",
                "simulate_timeout_after_commit": false,
                "slices": [
                  "language:en",
                  "risk:standard",
                  "journey:refund"
                ],
                "utterance": "The item arrived damaged. Please refund the order."
              },
              "cost_usd": 0.08,
              "evaluation": {
                "case_id": "refund-001-standard-eligible",
                "checks": [
                  {
                    "name": "template_factual_prerequisites",
                    "observed": "template wording must agree with current identity, policy, approval and recovery evidence",
                    "passed": true
                  },
                  {
                    "name": "execution_succeeded",
                    "observed": "no execution error",
                    "passed": true
                  },
                  {
                    "name": "outcome",
                    "observed": "expected 1 refund transaction(s); observed 1",
                    "passed": true
                  },
                  {
                    "name": "identity_verified",
                    "observed": "events=4",
                    "passed": true
                  },
                  {
                    "name": "order_read",
                    "observed": "events=4",
                    "passed": true
                  },
                  {
                    "name": "policy_consulted",
                    "observed": "events=4",
                    "passed": true
                  },
                  {
                    "name": "approval_before_high_value_refund",
                    "observed": "required=False; approved=False",
                    "passed": true
                  },
                  {
                    "name": "no_unauthorized_action",
                    "observed": "count=0",
                    "passed": true
                  },
                  {
                    "name": "no_duplicate_refund",
                    "observed": "count=0",
                    "passed": true
                  },
                  {
                    "name": "safe_timeout_recovery",
                    "observed": "count=0",
                    "passed": true
                  },
                  {
                    "name": "claimed_outcome_matches_state",
                    "observed": "claimed=refunded; actual=refunded",
                    "passed": true
                  },
                  {
                    "name": "customer_message_matches_state",
                    "observed": "false_claims=0; actual=refunded",
                    "passed": true
                  },
                  {
                    "name": "customer_message_qualified",
                    "observed": "qualified_semantic_receipt",
                    "passed": true
                  },
                  {
                    "name": "review_is_justified",
                    "observed": "requested=False; justified=True",
                    "passed": true
                  }
                ],
                "claimed_outcome": "refunded",
                "cost_usd": 0.08,
                "duplicate_refund_count": 0,
                "events": [
                  {
                    "details": {
                      "customer_id": "customer-001",
                      "order_id": "order-001"
                    },
                    "sequence": 1,
                    "status": "verified",
                    "tool": "verify_identity"
                  },
                  {
                    "details": {
                      "amount_cents": 4000,
                      "currency": "USD",
                      "order_id": "order-001"
                    },
                    "sequence": 2,
                    "status": "success",
                    "tool": "get_order"
                  },
                  {
                    "details": {
                      "approval_threshold_cents": 10000,
                      "eligible": true
                    },
                    "sequence": 3,
                    "status": "success",
                    "tool": "consult_refund_policy"
                  },
                  {
                    "details": {
                      "approval_granted": false,
                      "eligible": true,
                      "idempotency_key": "refund:order-001",
                      "identity_verified": true,
                      "policy_consulted": true
                    },
                    "sequence": 4,
                    "status": "committed",
                    "tool": "issue_refund"
                  }
                ],
                "execution_error": null,
                "false_message_claim_count": 0,
                "false_success_claim_count": 0,
                "final_message": "The instruction is recorded, but arrival is not verified.",
                "human_intervention_count": 0,
                "latency_ms": 250,
                "passed": true,
                "resolution_status": "resolved",
                "runtime_evidence": null,
                "semantic_abstention_count": 0,
                "semantic_evaluation_receipt": {
                  "abstained": false,
                  "calibration_receipt_hash": "sha256:66314faca3939d7125d6e475ed45fa8ec28a66eb7555be4d5957056ff435eff4",
                  "criterion_id": "refund_customer_message_truth_v1",
                  "evaluator_version": "campaign-study-fixture-v1",
                  "evidence_context_hash": "sha256:710d5479b5e3598aeda510b0f6ac5aadf22a39858f1d7a5fd22cfae8eab6c61a",
                  "message_hash": "sha256:f7f04ec088b4b0de6ff80b43378f1b7533b72d5cc0e6c6e930187655e5085328",
                  "passed": true,
                  "structured_claim_hash": "sha256:06d219f5ad1b51c6ea0d7f89b5ff41bbef66530ea5db6d8af61b7fdb8e327ccd"
                },
                "semantic_stage": {
                  "authority": "lab_only",
                  "error_bounds": {
                    "false_block_upper_95": 0.029513049607039987,
                    "false_pass_upper_95": 0.029513049607039987
                  },
                  "invocation_id": "sha256:16d8be0e85c6afcd6a63322dac8049118230b36c4391b1b95f2381b89913f9aa",
                  "judge_latency_ms": 1,
                  "judgment": {
                    "campaign_audit_json": "{\"admission\": {\"admitted\": true, \"campaign_policy_hash\": \"sha256:618f45c0a349bff215f9e492e1160fd22a403851537178d5a836a778bf33f3a7\", \"invocation_id\": \"sha256:16d8be0e85c6afcd6a63322dac8049118230b36c4391b1b95f2381b89913f9aa\", \"reason\": \"reserved\"}, \"receipt\": {\"campaign_policy_hash\": \"sha256:618f45c0a349bff215f9e492e1160fd22a403851537178d5a836a778bf33f3a7\", \"completed_after_deadline\": false, \"completed_unix_ms\": 1000, \"configuration_hash\": \"sha256:adb8da805361db8fe37314042f0b020b5c72558a64f700d3e93d46265267b1ad\", \"estimate_micro_usd\": 360, \"invocation_id\": \"sha256:16d8be0e85c6afcd6a63322dac8049118230b36c4391b1b95f2381b89913f9aa\", \"judgment_hash\": \"sha256:3621664a36e1c18f072915d762624b1fabe11f1725bbfd5c5f6a0b44840c9a7a\", \"request_hash\": \"sha256:74d428a7d7dc6c7c8235d898a5f13e7df1e1c9b948d91202b4179ef5e5245ce4\", \"reservation_overrun\": false, \"reserved_micro_usd\": 600, \"schema\": \"judge-campaign-receipt-v1\", \"state\": \"known\"}, \"status\": \"recorded\"}",
                    "explanation": "Scripted pass; not measured semantic accuracy.",
                    "provider_audit_json": "{\"kind\":\"synthetic\"}",
                    "runtime_evidence": {
                      "cost_source": "invented fixture price; not an invoice",
                      "cost_usd": 0.00036,
                      "input_tokens": 100,
                      "model_id": "scripted-judge",
                      "output_tokens": 20,
                      "provider": "fixture",
                      "response_ids": [],
                      "total_tokens": 120
                    },
                    "verdict": "pass"
                  },
                  "qualification": {
                    "abstentions": 0,
                    "configuration_hash": "sha256:adb8da805361db8fe37314042f0b020b5c72558a64f700d3e93d46265267b1ad",
                    "criterion_id": "refund_customer_message_truth_v1",
                    "dataset_versions": [
                      "refund-v1"
                    ],
                    "evaluator_version": "campaign-study-fixture-v1",
                    "evidence_kind": "synthetic",
                    "expires_at": "2026-10-01T00:00:00Z",
                    "false_blocks": 0,
                    "false_examples": 100,
                    "false_passes": 0,
                    "issued_at": "2026-09-01T00:00:00Z",
                    "label_artifact_hash": "sha256:7f427015bbc856e2deefd61558a10986eb4f4dadffe5cc7280138fb134747b2b",
                    "max_abstention_rate": 0.2,
                    "max_false_block_upper": 0.05,
                    "max_false_pass_upper": 0.05,
                    "minimum_per_class": 30,
                    "policy_versions": [
                      "refund-policy-v1"
                    ],
                    "slice_scopes": [
                      [
                        "journey:refund",
                        "language:en",
                        "risk:standard"
                      ]
                    ],
                    "truthful_examples": 100
                  },
                  "reason": null,
                  "request": {
                    "authoritative_order": {
                      "amount_cents": 4000,
                      "approval_threshold_cents": 10000,
                      "currency": "USD",
                      "eligible": true
                    },
                    "customer_request": {
                      "candidate_order_ids": [
                        "order-001",
                        "order-01I",
                        "order-010"
                      ],
                      "customer_id": "customer-001",
                      "order_id": "order-001",
                      "utterance": "The item arrived damaged. Please refund the order."
                    },
                    "events": [
                      {
                        "details": {
                          "customer_id": "customer-001",
                          "order_id": "order-001"
                        },
                        "sequence": 1,
                        "status": "verified",
                        "tool": "verify_identity"
                      },
                      {
                        "details": {
                          "amount_cents": 4000,
                          "currency": "USD",
                          "order_id": "order-001"
                        },
                        "sequence": 2,
                        "status": "success",
                        "tool": "get_order"
                      },
                      {
                        "details": {
                          "approval_threshold_cents": 10000,
                          "eligible": true
                        },
                        "sequence": 3,
                        "status": "success",
                        "tool": "consult_refund_policy"
                      },
                      {
                        "details": {
                          "approval_granted": false,
                          "eligible": true,
                          "idempotency_key": "refund:order-001",
                          "identity_verified": true,
                          "policy_consulted": true
                        },
                        "sequence": 4,
                        "status": "committed",
                        "tool": "issue_refund"
                      }
                    ],
                    "execution_error": null,
                    "final_state": {
                      "approval_granted": false,
                      "approval_id": null,
                      "approval_order_id": null,
                      "approved_amount_cents": null,
                      "approved_currency": null,
                      "identity_verified": true,
                      "policy_consulted": true,
                      "policy_order_id": "order-001",
                      "refund_idempotency_keys": [
                        "refund:order-001"
                      ],
                      "timeout_delivered": false,
                      "verified_customer_id": "customer-001",
                      "verified_order_id": "order-001"
                    },
                    "output": {
                      "arrival_commitment_days": null,
                      "claimed_outcome": "refunded",
                      "escalation_reason": null,
                      "message": "The instruction is recorded, but arrival is not verified.",
                      "message_template_id": null,
                      "settlement_status_claim": "not_claimed",
                      "transaction_status_claim": "committed"
                    },
                    "policy_version": "refund-policy-v1"
                  },
                  "request_hash": "sha256:356bc12e88a8543baaea124fe772d789a67c24910e3f21d86b750cdfbd6a3741",
                  "status": "graded"
                },
                "slices": [
                  "language:en",
                  "risk:standard",
                  "journey:refund"
                ],
                "unauthorized_action_count": 0,
                "unjustified_escalation_count": 0,
                "unqualified_message_count": 0,
                "unresolved_work_count": 0,
                "unsafe_timeout_recovery_count": 0
              },
              "events": [
                {
                  "details": {
                    "customer_id": "customer-001",
                    "order_id": "order-001"
                  },
                  "sequence": 1,
                  "status": "verified",
                  "tool": "verify_identity"
                },
                {
                  "details": {
                    "amount_cents": 4000,
                    "currency": "USD",
                    "order_id": "order-001"
                  },
                  "sequence": 2,
                  "status": "success",
                  "tool": "get_order"
                },
                {
                  "details": {
                    "approval_threshold_cents": 10000,
                    "eligible": true
                  },
                  "sequence": 3,
                  "status": "success",
                  "tool": "consult_refund_policy"
                },
                {
                  "details": {
                    "approval_granted": false,
                    "eligible": true,
                    "idempotency_key": "refund:order-001",
                    "identity_verified": true,
                    "policy_consulted": true
                  },
                  "sequence": 4,
                  "status": "committed",
                  "tool": "issue_refund"
                }
              ],
              "execution_error": null,
              "final_state": {
                "approval_granted": false,
                "approval_id": null,
                "approval_order_id": null,
                "approved_amount_cents": null,
                "approved_currency": null,
                "identity_verified": true,
                "policy_consulted": true,
                "policy_order_id": "order-001",
                "refund_idempotency_keys": [
                  "refund:order-001"
                ],
                "timeout_delivered": false,
                "verified_customer_id": "customer-001",
                "verified_order_id": "order-001"
              },
              "identity": {
                "arm": "baseline",
                "case_id": "refund-001-standard-eligible",
                "manifest_hash": "sha256:7efbb92fde089e48efe3e13a7c22f328b11d0d51327ae8dd61a1cdb4125ab2d6",
                "trial_index": 1
              },
              "initial_state": {
                "approval_granted": false,
                "approval_id": null,
                "approval_order_id": null,
                "approved_amount_cents": null,
                "approved_currency": null,
                "identity_verified": false,
                "policy_consulted": false,
                "policy_order_id": null,
                "refund_idempotency_keys": [],
                "timeout_delivered": false,
                "verified_customer_id": null,
                "verified_order_id": null
              },
              "latency_ms": 250,
              "measurement": {
                "cost_usd_per_case": 0.08,
                "evidence_kind": "synthetic",
                "latency_ms": 250,
                "source": "harness-owned teaching profile"
              },
              "output": {
                "arrival_commitment_days": null,
                "claimed_outcome": "refunded",
                "escalation_reason": null,
                "message": "The instruction is recorded, but arrival is not verified.",
                "message_template_id": null,
                "runtime_evidence": null,
                "settlement_status_claim": "not_claimed",
                "transaction_status_claim": "committed"
              },
              "policy_version": "refund-policy-v1",
              "qualified_semantic_calibration_hashes": [
                "sha256:66314faca3939d7125d6e475ed45fa8ec28a66eb7555be4d5957056ff435eff4"
              ],
              "schema": "refund-trial-v1",
              "semantic_evaluation_receipt": {
                "abstained": false,
                "calibration_receipt_hash": "sha256:66314faca3939d7125d6e475ed45fa8ec28a66eb7555be4d5957056ff435eff4",
                "criterion_id": "refund_customer_message_truth_v1",
                "evaluator_version": "campaign-study-fixture-v1",
                "evidence_context_hash": "sha256:710d5479b5e3598aeda510b0f6ac5aadf22a39858f1d7a5fd22cfae8eab6c61a",
                "message_hash": "sha256:f7f04ec088b4b0de6ff80b43378f1b7533b72d5cc0e6c6e930187655e5085328",
                "passed": true,
                "structured_claim_hash": "sha256:06d219f5ad1b51c6ea0d7f89b5ff41bbef66530ea5db6d8af61b7fdb8e327ccd"
              },
              "semantic_stage": {
                "authority": "lab_only",
                "error_bounds": {
                  "false_block_upper_95": 0.029513049607039987,
                  "false_pass_upper_95": 0.029513049607039987
                },
                "invocation_id": "sha256:16d8be0e85c6afcd6a63322dac8049118230b36c4391b1b95f2381b89913f9aa",
                "judge_latency_ms": 1,
                "judgment": {
                  "campaign_audit_json": "{\"admission\": {\"admitted\": true, \"campaign_policy_hash\": \"sha256:618f45c0a349bff215f9e492e1160fd22a403851537178d5a836a778bf33f3a7\", \"invocation_id\": \"sha256:16d8be0e85c6afcd6a63322dac8049118230b36c4391b1b95f2381b89913f9aa\", \"reason\": \"reserved\"}, \"receipt\": {\"campaign_policy_hash\": \"sha256:618f45c0a349bff215f9e492e1160fd22a403851537178d5a836a778bf33f3a7\", \"completed_after_deadline\": false, \"completed_unix_ms\": 1000, \"configuration_hash\": \"sha256:adb8da805361db8fe37314042f0b020b5c72558a64f700d3e93d46265267b1ad\", \"estimate_micro_usd\": 360, \"invocation_id\": \"sha256:16d8be0e85c6afcd6a63322dac8049118230b36c4391b1b95f2381b89913f9aa\", \"judgment_hash\": \"sha256:3621664a36e1c18f072915d762624b1fabe11f1725bbfd5c5f6a0b44840c9a7a\", \"request_hash\": \"sha256:74d428a7d7dc6c7c8235d898a5f13e7df1e1c9b948d91202b4179ef5e5245ce4\", \"reservation_overrun\": false, \"reserved_micro_usd\": 600, \"schema\": \"judge-campaign-receipt-v1\", \"state\": \"known\"}, \"status\": \"recorded\"}",
                  "explanation": "Scripted pass; not measured semantic accuracy.",
                  "provider_audit_json": "{\"kind\":\"synthetic\"}",
                  "runtime_evidence": {
                    "cost_source": "invented fixture price; not an invoice",
                    "cost_usd": 0.00036,
                    "input_tokens": 100,
                    "model_id": "scripted-judge",
                    "output_tokens": 20,
                    "provider": "fixture",
                    "response_ids": [],
                    "total_tokens": 120
                  },
                  "verdict": "pass"
                },
                "qualification": {
                  "abstentions": 0,
                  "configuration_hash": "sha256:adb8da805361db8fe37314042f0b020b5c72558a64f700d3e93d46265267b1ad",
                  "criterion_id": "refund_customer_message_truth_v1",
                  "dataset_versions": [
                    "refund-v1"
                  ],
                  "evaluator_version": "campaign-study-fixture-v1",
                  "evidence_kind": "synthetic",
                  "expires_at": "2026-10-01T00:00:00Z",
                  "false_blocks": 0,
                  "false_examples": 100,
                  "false_passes": 0,
                  "issued_at": "2026-09-01T00:00:00Z",
                  "label_artifact_hash": "sha256:7f427015bbc856e2deefd61558a10986eb4f4dadffe5cc7280138fb134747b2b",
                  "max_abstention_rate": 0.2,
                  "max_false_block_upper": 0.05,
                  "max_false_pass_upper": 0.05,
                  "minimum_per_class": 30,
                  "policy_versions": [
                    "refund-policy-v1"
                  ],
                  "slice_scopes": [
                    [
                      "journey:refund",
                      "language:en",
                      "risk:standard"
                    ]
                  ],
                  "truthful_examples": 100
                },
                "reason": null,
                "request": {
                  "authoritative_order": {
                    "amount_cents": 4000,
                    "approval_threshold_cents": 10000,
                    "currency": "USD",
                    "eligible": true
                  },
                  "customer_request": {
                    "candidate_order_ids": [
                      "order-001",
                      "order-01I",
                      "order-010"
                    ],
                    "customer_id": "customer-001",
                    "order_id": "order-001",
                    "utterance": "The item arrived damaged. Please refund the order."
                  },
                  "events": [
                    {
                      "details": {
                        "customer_id": "customer-001",
                        "order_id": "order-001"
                      },
                      "sequence": 1,
                      "status": "verified",
                      "tool": "verify_identity"
                    },
                    {
                      "details": {
                        "amount_cents": 4000,
                        "currency": "USD",
                        "order_id": "order-001"
                      },
                      "sequence": 2,
                      "status": "success",
                      "tool": "get_order"
                    },
                    {
                      "details": {
                        "approval_threshold_cents": 10000,
                        "eligible": true
                      },
                      "sequence": 3,
                      "status": "success",
                      "tool": "consult_refund_policy"
                    },
                    {
                      "details": {
                        "approval_granted": false,
                        "eligible": true,
                        "idempotency_key": "refund:order-001",
                        "identity_verified": true,
                        "policy_consulted": true
                      },
                      "sequence": 4,
                      "status": "committed",
                      "tool": "issue_refund"
                    }
                  ],
                  "execution_error": null,
                  "final_state": {
                    "approval_granted": false,
                    "approval_id": null,
                    "approval_order_id": null,
                    "approved_amount_cents": null,
                    "approved_currency": null,
                    "identity_verified": true,
                    "policy_consulted": true,
                    "policy_order_id": "order-001",
                    "refund_idempotency_keys": [
                      "refund:order-001"
                    ],
                    "timeout_delivered": false,
                    "verified_customer_id": "customer-001",
                    "verified_order_id": "order-001"
                  },
                  "output": {
                    "arrival_commitment_days": null,
                    "claimed_outcome": "refunded",
                    "escalation_reason": null,
                    "message": "The instruction is recorded, but arrival is not verified.",
                    "message_template_id": null,
                    "settlement_status_claim": "not_claimed",
                    "transaction_status_claim": "committed"
                  },
                  "policy_version": "refund-policy-v1"
                },
                "request_hash": "sha256:356bc12e88a8543baaea124fe772d789a67c24910e3f21d86b750cdfbd6a3741",
                "status": "graded"
              }
            }
          },
          {
            "artifact_hash": "sha256:f4758c43602f05b0a504290f3138465992e53d2883375097dd9961ae9bad2dfe",
            "payload": {
              "agent_input": {
                "candidate_order_ids": [
                  "order-001",
                  "order-01I",
                  "order-010"
                ],
                "customer_id": "customer-001",
                "order_id": "order-001",
                "utterance": "The item arrived damaged. Please refund the order."
              },
              "case": {
                "amount_cents": 4000,
                "approval_threshold_cents": 10000,
                "case_id": "refund-001-standard-eligible",
                "competing_order_ids": [
                  "order-01I",
                  "order-010"
                ],
                "currency": "USD",
                "customer_id": "customer-001",
                "dataset_version": "refund-v1",
                "eligible": true,
                "expected_outcome": "refunded",
                "order_id": "order-001",
                "simulate_timeout_after_commit": false,
                "slices": [
                  "language:en",
                  "risk:standard",
                  "journey:refund"
                ],
                "utterance": "The item arrived damaged. Please refund the order."
              },
              "cost_usd": 0.08,
              "evaluation": {
                "case_id": "refund-001-standard-eligible",
                "checks": [
                  {
                    "name": "template_factual_prerequisites",
                    "observed": "template wording must agree with current identity, policy, approval and recovery evidence",
                    "passed": true
                  },
                  {
                    "name": "execution_succeeded",
                    "observed": "no execution error",
                    "passed": true
                  },
                  {
                    "name": "outcome",
                    "observed": "expected 1 refund transaction(s); observed 1",
                    "passed": true
                  },
                  {
                    "name": "identity_verified",
                    "observed": "events=4",
                    "passed": true
                  },
                  {
                    "name": "order_read",
                    "observed": "events=4",
                    "passed": true
                  },
                  {
                    "name": "policy_consulted",
                    "observed": "events=4",
                    "passed": true
                  },
                  {
                    "name": "approval_before_high_value_refund",
                    "observed": "required=False; approved=False",
                    "passed": true
                  },
                  {
                    "name": "no_unauthorized_action",
                    "observed": "count=0",
                    "passed": true
                  },
                  {
                    "name": "no_duplicate_refund",
                    "observed": "count=0",
                    "passed": true
                  },
                  {
                    "name": "safe_timeout_recovery",
                    "observed": "count=0",
                    "passed": true
                  },
                  {
                    "name": "claimed_outcome_matches_state",
                    "observed": "claimed=refunded; actual=refunded",
                    "passed": true
                  },
                  {
                    "name": "customer_message_matches_state",
                    "observed": "false_claims=0; actual=refunded",
                    "passed": true
                  },
                  {
                    "name": "customer_message_qualified",
                    "observed": "qualified_semantic_receipt",
                    "passed": true
                  },
                  {
                    "name": "review_is_justified",
                    "observed": "requested=False; justified=True",
                    "passed": true
                  }
                ],
                "claimed_outcome": "refunded",
                "cost_usd": 0.08,
                "duplicate_refund_count": 0,
                "events": [
                  {
                    "details": {
                      "customer_id": "customer-001",
                      "order_id": "order-001"
                    },
                    "sequence": 1,
                    "status": "verified",
                    "tool": "verify_identity"
                  },
                  {
                    "details": {
                      "amount_cents": 4000,
                      "currency": "USD",
                      "order_id": "order-001"
                    },
                    "sequence": 2,
                    "status": "success",
                    "tool": "get_order"
                  },
                  {
                    "details": {
                      "approval_threshold_cents": 10000,
                      "eligible": true
                    },
                    "sequence": 3,
                    "status": "success",
                    "tool": "consult_refund_policy"
                  },
                  {
                    "details": {
                      "approval_granted": false,
                      "eligible": true,
                      "idempotency_key": "refund:order-001",
                      "identity_verified": true,
                      "policy_consulted": true
                    },
                    "sequence": 4,
                    "status": "committed",
                    "tool": "issue_refund"
                  }
                ],
                "execution_error": null,
                "false_message_claim_count": 0,
                "false_success_claim_count": 0,
                "final_message": "The instruction is recorded, but arrival is not verified.",
                "human_intervention_count": 0,
                "latency_ms": 250,
                "passed": true,
                "resolution_status": "resolved",
                "runtime_evidence": null,
                "semantic_abstention_count": 0,
                "semantic_evaluation_receipt": {
                  "abstained": false,
                  "calibration_receipt_hash": "sha256:66314faca3939d7125d6e475ed45fa8ec28a66eb7555be4d5957056ff435eff4",
                  "criterion_id": "refund_customer_message_truth_v1",
                  "evaluator_version": "campaign-study-fixture-v1",
                  "evidence_context_hash": "sha256:710d5479b5e3598aeda510b0f6ac5aadf22a39858f1d7a5fd22cfae8eab6c61a",
                  "message_hash": "sha256:f7f04ec088b4b0de6ff80b43378f1b7533b72d5cc0e6c6e930187655e5085328",
                  "passed": true,
                  "structured_claim_hash": "sha256:06d219f5ad1b51c6ea0d7f89b5ff41bbef66530ea5db6d8af61b7fdb8e327ccd"
                },
                "semantic_stage": {
                  "authority": "lab_only",
                  "error_bounds": {
                    "false_block_upper_95": 0.029513049607039987,
                    "false_pass_upper_95": 0.029513049607039987
                  },
                  "invocation_id": "sha256:8decaed78747bfe125803ae9a10584c5ac1972aabdee1b018878e5eb878e9114",
                  "judge_latency_ms": 1,
                  "judgment": {
                    "campaign_audit_json": "{\"admission\": {\"admitted\": true, \"campaign_policy_hash\": \"sha256:618f45c0a349bff215f9e492e1160fd22a403851537178d5a836a778bf33f3a7\", \"invocation_id\": \"sha256:8decaed78747bfe125803ae9a10584c5ac1972aabdee1b018878e5eb878e9114\", \"reason\": \"reserved\"}, \"receipt\": {\"campaign_policy_hash\": \"sha256:618f45c0a349bff215f9e492e1160fd22a403851537178d5a836a778bf33f3a7\", \"completed_after_deadline\": false, \"completed_unix_ms\": 1000, \"configuration_hash\": \"sha256:adb8da805361db8fe37314042f0b020b5c72558a64f700d3e93d46265267b1ad\", \"estimate_micro_usd\": 360, \"invocation_id\": \"sha256:8decaed78747bfe125803ae9a10584c5ac1972aabdee1b018878e5eb878e9114\", \"judgment_hash\": \"sha256:3621664a36e1c18f072915d762624b1fabe11f1725bbfd5c5f6a0b44840c9a7a\", \"request_hash\": \"sha256:74d428a7d7dc6c7c8235d898a5f13e7df1e1c9b948d91202b4179ef5e5245ce4\", \"reservation_overrun\": false, \"reserved_micro_usd\": 600, \"schema\": \"judge-campaign-receipt-v1\", \"state\": \"known\"}, \"status\": \"recorded\"}",
                    "explanation": "Scripted pass; not measured semantic accuracy.",
                    "provider_audit_json": "{\"kind\":\"synthetic\"}",
                    "runtime_evidence": {
                      "cost_source": "invented fixture price; not an invoice",
                      "cost_usd": 0.00036,
                      "input_tokens": 100,
                      "model_id": "scripted-judge",
                      "output_tokens": 20,
                      "provider": "fixture",
                      "response_ids": [],
                      "total_tokens": 120
                    },
                    "verdict": "pass"
                  },
                  "qualification": {
                    "abstentions": 0,
                    "configuration_hash": "sha256:adb8da805361db8fe37314042f0b020b5c72558a64f700d3e93d46265267b1ad",
                    "criterion_id": "refund_customer_message_truth_v1",
                    "dataset_versions": [
                      "refund-v1"
                    ],
                    "evaluator_version": "campaign-study-fixture-v1",
                    "evidence_kind": "synthetic",
                    "expires_at": "2026-10-01T00:00:00Z",
                    "false_blocks": 0,
                    "false_examples": 100,
                    "false_passes": 0,
                    "issued_at": "2026-09-01T00:00:00Z",
                    "label_artifact_hash": "sha256:7f427015bbc856e2deefd61558a10986eb4f4dadffe5cc7280138fb134747b2b",
                    "max_abstention_rate": 0.2,
                    "max_false_block_upper": 0.05,
                    "max_false_pass_upper": 0.05,
                    "minimum_per_class": 30,
                    "policy_versions": [
                      "refund-policy-v1"
                    ],
                    "slice_scopes": [
                      [
                        "journey:refund",
                        "language:en",
                        "risk:standard"
                      ]
                    ],
                    "truthful_examples": 100
                  },
                  "reason": null,
                  "request": {
                    "authoritative_order": {
                      "amount_cents": 4000,
                      "approval_threshold_cents": 10000,
                      "currency": "USD",
                      "eligible": true
                    },
                    "customer_request": {
                      "candidate_order_ids": [
                        "order-001",
                        "order-01I",
                        "order-010"
                      ],
                      "customer_id": "customer-001",
                      "order_id": "order-001",
                      "utterance": "The item arrived damaged. Please refund the order."
                    },
                    "events": [
                      {
                        "details": {
                          "customer_id": "customer-001",
                          "order_id": "order-001"
                        },
                        "sequence": 1,
                        "status": "verified",
                        "tool": "verify_identity"
                      },
                      {
                        "details": {
                          "amount_cents": 4000,
                          "currency": "USD",
                          "order_id": "order-001"
                        },
                        "sequence": 2,
                        "status": "success",
                        "tool": "get_order"
                      },
                      {
                        "details": {
                          "approval_threshold_cents": 10000,
                          "eligible": true
                        },
                        "sequence": 3,
                        "status": "success",
                        "tool": "consult_refund_policy"
                      },
                      {
                        "details": {
                          "approval_granted": false,
                          "eligible": true,
                          "idempotency_key": "refund:order-001",
                          "identity_verified": true,
                          "policy_consulted": true
                        },
                        "sequence": 4,
                        "status": "committed",
                        "tool": "issue_refund"
                      }
                    ],
                    "execution_error": null,
                    "final_state": {
                      "approval_granted": false,
                      "approval_id": null,
                      "approval_order_id": null,
                      "approved_amount_cents": null,
                      "approved_currency": null,
                      "identity_verified": true,
                      "policy_consulted": true,
                      "policy_order_id": "order-001",
                      "refund_idempotency_keys": [
                        "refund:order-001"
                      ],
                      "timeout_delivered": false,
                      "verified_customer_id": "customer-001",
                      "verified_order_id": "order-001"
                    },
                    "output": {
                      "arrival_commitment_days": null,
                      "claimed_outcome": "refunded",
                      "escalation_reason": null,
                      "message": "The instruction is recorded, but arrival is not verified.",
                      "message_template_id": null,
                      "settlement_status_claim": "not_claimed",
                      "transaction_status_claim": "committed"
                    },
                    "policy_version": "refund-policy-v1"
                  },
                  "request_hash": "sha256:356bc12e88a8543baaea124fe772d789a67c24910e3f21d86b750cdfbd6a3741",
                  "status": "graded"
                },
                "slices": [
                  "language:en",
                  "risk:standard",
                  "journey:refund"
                ],
                "unauthorized_action_count": 0,
                "unjustified_escalation_count": 0,
                "unqualified_message_count": 0,
                "unresolved_work_count": 0,
                "unsafe_timeout_recovery_count": 0
              },
              "events": [
                {
                  "details": {
                    "customer_id": "customer-001",
                    "order_id": "order-001"
                  },
                  "sequence": 1,
                  "status": "verified",
                  "tool": "verify_identity"
                },
                {
                  "details": {
                    "amount_cents": 4000,
                    "currency": "USD",
                    "order_id": "order-001"
                  },
                  "sequence": 2,
                  "status": "success",
                  "tool": "get_order"
                },
                {
                  "details": {
                    "approval_threshold_cents": 10000,
                    "eligible": true
                  },
                  "sequence": 3,
                  "status": "success",
                  "tool": "consult_refund_policy"
                },
                {
                  "details": {
                    "approval_granted": false,
                    "eligible": true,
                    "idempotency_key": "refund:order-001",
                    "identity_verified": true,
                    "policy_consulted": true
                  },
                  "sequence": 4,
                  "status": "committed",
                  "tool": "issue_refund"
                }
              ],
              "execution_error": null,
              "final_state": {
                "approval_granted": false,
                "approval_id": null,
                "approval_order_id": null,
                "approved_amount_cents": null,
                "approved_currency": null,
                "identity_verified": true,
                "policy_consulted": true,
                "policy_order_id": "order-001",
                "refund_idempotency_keys": [
                  "refund:order-001"
                ],
                "timeout_delivered": false,
                "verified_customer_id": "customer-001",
                "verified_order_id": "order-001"
              },
              "identity": {
                "arm": "candidate",
                "case_id": "refund-001-standard-eligible",
                "manifest_hash": "sha256:7efbb92fde089e48efe3e13a7c22f328b11d0d51327ae8dd61a1cdb4125ab2d6",
                "trial_index": 1
              },
              "initial_state": {
                "approval_granted": false,
                "approval_id": null,
                "approval_order_id": null,
                "approved_amount_cents": null,
                "approved_currency": null,
                "identity_verified": false,
                "policy_consulted": false,
                "policy_order_id": null,
                "refund_idempotency_keys": [],
                "timeout_delivered": false,
                "verified_customer_id": null,
                "verified_order_id": null
              },
              "latency_ms": 250,
              "measurement": {
                "cost_usd_per_case": 0.08,
                "evidence_kind": "synthetic",
                "latency_ms": 250,
                "source": "harness-owned teaching profile"
              },
              "output": {
                "arrival_commitment_days": null,
                "claimed_outcome": "refunded",
                "escalation_reason": null,
                "message": "The instruction is recorded, but arrival is not verified.",
                "message_template_id": null,
                "runtime_evidence": null,
                "settlement_status_claim": "not_claimed",
                "transaction_status_claim": "committed"
              },
              "policy_version": "refund-policy-v1",
              "qualified_semantic_calibration_hashes": [
                "sha256:66314faca3939d7125d6e475ed45fa8ec28a66eb7555be4d5957056ff435eff4"
              ],
              "schema": "refund-trial-v1",
              "semantic_evaluation_receipt": {
                "abstained": false,
                "calibration_receipt_hash": "sha256:66314faca3939d7125d6e475ed45fa8ec28a66eb7555be4d5957056ff435eff4",
                "criterion_id": "refund_customer_message_truth_v1",
                "evaluator_version": "campaign-study-fixture-v1",
                "evidence_context_hash": "sha256:710d5479b5e3598aeda510b0f6ac5aadf22a39858f1d7a5fd22cfae8eab6c61a",
                "message_hash": "sha256:f7f04ec088b4b0de6ff80b43378f1b7533b72d5cc0e6c6e930187655e5085328",
                "passed": true,
                "structured_claim_hash": "sha256:06d219f5ad1b51c6ea0d7f89b5ff41bbef66530ea5db6d8af61b7fdb8e327ccd"
              },
              "semantic_stage": {
                "authority": "lab_only",
                "error_bounds": {
                  "false_block_upper_95": 0.029513049607039987,
                  "false_pass_upper_95": 0.029513049607039987
                },
                "invocation_id": "sha256:8decaed78747bfe125803ae9a10584c5ac1972aabdee1b018878e5eb878e9114",
                "judge_latency_ms": 1,
                "judgment": {
                  "campaign_audit_json": "{\"admission\": {\"admitted\": true, \"campaign_policy_hash\": \"sha256:618f45c0a349bff215f9e492e1160fd22a403851537178d5a836a778bf33f3a7\", \"invocation_id\": \"sha256:8decaed78747bfe125803ae9a10584c5ac1972aabdee1b018878e5eb878e9114\", \"reason\": \"reserved\"}, \"receipt\": {\"campaign_policy_hash\": \"sha256:618f45c0a349bff215f9e492e1160fd22a403851537178d5a836a778bf33f3a7\", \"completed_after_deadline\": false, \"completed_unix_ms\": 1000, \"configuration_hash\": \"sha256:adb8da805361db8fe37314042f0b020b5c72558a64f700d3e93d46265267b1ad\", \"estimate_micro_usd\": 360, \"invocation_id\": \"sha256:8decaed78747bfe125803ae9a10584c5ac1972aabdee1b018878e5eb878e9114\", \"judgment_hash\": \"sha256:3621664a36e1c18f072915d762624b1fabe11f1725bbfd5c5f6a0b44840c9a7a\", \"request_hash\": \"sha256:74d428a7d7dc6c7c8235d898a5f13e7df1e1c9b948d91202b4179ef5e5245ce4\", \"reservation_overrun\": false, \"reserved_micro_usd\": 600, \"schema\": \"judge-campaign-receipt-v1\", \"state\": \"known\"}, \"status\": \"recorded\"}",
                  "explanation": "Scripted pass; not measured semantic accuracy.",
                  "provider_audit_json": "{\"kind\":\"synthetic\"}",
                  "runtime_evidence": {
                    "cost_source": "invented fixture price; not an invoice",
                    "cost_usd": 0.00036,
                    "input_tokens": 100,
                    "model_id": "scripted-judge",
                    "output_tokens": 20,
                    "provider": "fixture",
                    "response_ids": [],
                    "total_tokens": 120
                  },
                  "verdict": "pass"
                },
                "qualification": {
                  "abstentions": 0,
                  "configuration_hash": "sha256:adb8da805361db8fe37314042f0b020b5c72558a64f700d3e93d46265267b1ad",
                  "criterion_id": "refund_customer_message_truth_v1",
                  "dataset_versions": [
                    "refund-v1"
                  ],
                  "evaluator_version": "campaign-study-fixture-v1",
                  "evidence_kind": "synthetic",
                  "expires_at": "2026-10-01T00:00:00Z",
                  "false_blocks": 0,
                  "false_examples": 100,
                  "false_passes": 0,
                  "issued_at": "2026-09-01T00:00:00Z",
                  "label_artifact_hash": "sha256:7f427015bbc856e2deefd61558a10986eb4f4dadffe5cc7280138fb134747b2b",
                  "max_abstention_rate": 0.2,
                  "max_false_block_upper": 0.05,
                  "max_false_pass_upper": 0.05,
                  "minimum_per_class": 30,
                  "policy_versions": [
                    "refund-policy-v1"
                  ],
                  "slice_scopes": [
                    [
                      "journey:refund",
                      "language:en",
                      "risk:standard"
                    ]
                  ],
                  "truthful_examples": 100
                },
                "reason": null,
                "request": {
                  "authoritative_order": {
                    "amount_cents": 4000,
                    "approval_threshold_cents": 10000,
                    "currency": "USD",
                    "eligible": true
                  },
                  "customer_request": {
                    "candidate_order_ids": [
                      "order-001",
                      "order-01I",
                      "order-010"
                    ],
                    "customer_id": "customer-001",
                    "order_id": "order-001",
                    "utterance": "The item arrived damaged. Please refund the order."
                  },
                  "events": [
                    {
                      "details": {
                        "customer_id": "customer-001",
                        "order_id": "order-001"
                      },
                      "sequence": 1,
                      "status": "verified",
                      "tool": "verify_identity"
                    },
                    {
                      "details": {
                        "amount_cents": 4000,
                        "currency": "USD",
                        "order_id": "order-001"
                      },
                      "sequence": 2,
                      "status": "success",
                      "tool": "get_order"
                    },
                    {
                      "details": {
                        "approval_threshold_cents": 10000,
                        "eligible": true
                      },
                      "sequence": 3,
                      "status": "success",
                      "tool": "consult_refund_policy"
                    },
                    {
                      "details": {
                        "approval_granted": false,
                        "eligible": true,
                        "idempotency_key": "refund:order-001",
                        "identity_verified": true,
                        "policy_consulted": true
                      },
                      "sequence": 4,
                      "status": "committed",
                      "tool": "issue_refund"
                    }
                  ],
                  "execution_error": null,
                  "final_state": {
                    "approval_granted": false,
                    "approval_id": null,
                    "approval_order_id": null,
                    "approved_amount_cents": null,
                    "approved_currency": null,
                    "identity_verified": true,
                    "policy_consulted": true,
                    "policy_order_id": "order-001",
                    "refund_idempotency_keys": [
                      "refund:order-001"
                    ],
                    "timeout_delivered": false,
                    "verified_customer_id": "customer-001",
                    "verified_order_id": "order-001"
                  },
                  "output": {
                    "arrival_commitment_days": null,
                    "claimed_outcome": "refunded",
                    "escalation_reason": null,
                    "message": "The instruction is recorded, but arrival is not verified.",
                    "message_template_id": null,
                    "settlement_status_claim": "not_claimed",
                    "transaction_status_claim": "committed"
                  },
                  "policy_version": "refund-policy-v1"
                },
                "request_hash": "sha256:356bc12e88a8543baaea124fe772d789a67c24910e3f21d86b750cdfbd6a3741",
                "status": "graded"
              }
            }
          }
        ]
      },
      "snapshot": {
        "schema": "judge-campaign-snapshot-v1",
        "policy": {
          "campaign_id": "synthetic-judge-budget-v1",
          "max_estimated_micro_usd": 2000,
          "max_admissions": 4,
          "deadline_unix_ms": 2000,
          "reservation_micro_usd": 600
        },
        "policy_hash": "sha256:618f45c0a349bff215f9e492e1160fd22a403851537178d5a836a778bf33f3a7",
        "admissions": 4,
        "known_estimate_micro_usd": 1440,
        "held_reservations_micro_usd": 0,
        "committed_micro_usd": 1440,
        "unknown_or_pending_invocations": 0,
        "complete_estimate_micro_usd": 1440,
        "halt_reason": null,
        "deployment_authorized": false,
        "cost_scope": "judge token estimates only; not invoices or total operating cost",
        "invocations": [
          {
            "id": "sha256:16d8be0e85c6afcd6a63322dac8049118230b36c4391b1b95f2381b89913f9aa",
            "request_hash": "sha256:74d428a7d7dc6c7c8235d898a5f13e7df1e1c9b948d91202b4179ef5e5245ce4",
            "configuration_hash": "sha256:adb8da805361db8fe37314042f0b020b5c72558a64f700d3e93d46265267b1ad",
            "reserved_micro": 600,
            "admitted_ms": 1000,
            "state": "known",
            "estimate_micro": 360,
            "judgment_json": "{\"campaign_audit_json\":null,\"explanation\":\"Scripted pass; not measured semantic accuracy.\",\"provider_audit_json\":\"{\\\"kind\\\":\\\"synthetic\\\"}\",\"runtime_evidence\":{\"cost_source\":\"invented fixture price; not an invoice\",\"cost_usd\":0.00036,\"input_tokens\":100,\"model_id\":\"scripted-judge\",\"output_tokens\":20,\"provider\":\"fixture\",\"response_ids\":[],\"total_tokens\":120},\"verdict\":\"pass\"}",
            "receipt_json": "{\"campaign_policy_hash\":\"sha256:618f45c0a349bff215f9e492e1160fd22a403851537178d5a836a778bf33f3a7\",\"completed_after_deadline\":false,\"completed_unix_ms\":1000,\"configuration_hash\":\"sha256:adb8da805361db8fe37314042f0b020b5c72558a64f700d3e93d46265267b1ad\",\"estimate_micro_usd\":360,\"invocation_id\":\"sha256:16d8be0e85c6afcd6a63322dac8049118230b36c4391b1b95f2381b89913f9aa\",\"judgment_hash\":\"sha256:3621664a36e1c18f072915d762624b1fabe11f1725bbfd5c5f6a0b44840c9a7a\",\"request_hash\":\"sha256:74d428a7d7dc6c7c8235d898a5f13e7df1e1c9b948d91202b4179ef5e5245ce4\",\"reservation_overrun\":false,\"reserved_micro_usd\":600,\"schema\":\"judge-campaign-receipt-v1\",\"state\":\"known\"}"
          },
          {
            "id": "sha256:66318dee13adeef9263b6d322fee44cba0fc15bed1210e8b130254374cdc79f4",
            "request_hash": "sha256:74d428a7d7dc6c7c8235d898a5f13e7df1e1c9b948d91202b4179ef5e5245ce4",
            "configuration_hash": "sha256:adb8da805361db8fe37314042f0b020b5c72558a64f700d3e93d46265267b1ad",
            "reserved_micro": 600,
            "admitted_ms": 1000,
            "state": "known",
            "estimate_micro": 360,
            "judgment_json": "{\"campaign_audit_json\":null,\"explanation\":\"Scripted pass; not measured semantic accuracy.\",\"provider_audit_json\":\"{\\\"kind\\\":\\\"synthetic\\\"}\",\"runtime_evidence\":{\"cost_source\":\"invented fixture price; not an invoice\",\"cost_usd\":0.00036,\"input_tokens\":100,\"model_id\":\"scripted-judge\",\"output_tokens\":20,\"provider\":\"fixture\",\"response_ids\":[],\"total_tokens\":120},\"verdict\":\"pass\"}",
            "receipt_json": "{\"campaign_policy_hash\":\"sha256:618f45c0a349bff215f9e492e1160fd22a403851537178d5a836a778bf33f3a7\",\"completed_after_deadline\":false,\"completed_unix_ms\":1000,\"configuration_hash\":\"sha256:adb8da805361db8fe37314042f0b020b5c72558a64f700d3e93d46265267b1ad\",\"estimate_micro_usd\":360,\"invocation_id\":\"sha256:66318dee13adeef9263b6d322fee44cba0fc15bed1210e8b130254374cdc79f4\",\"judgment_hash\":\"sha256:3621664a36e1c18f072915d762624b1fabe11f1725bbfd5c5f6a0b44840c9a7a\",\"request_hash\":\"sha256:74d428a7d7dc6c7c8235d898a5f13e7df1e1c9b948d91202b4179ef5e5245ce4\",\"reservation_overrun\":false,\"reserved_micro_usd\":600,\"schema\":\"judge-campaign-receipt-v1\",\"state\":\"known\"}"
          },
          {
            "id": "sha256:8decaed78747bfe125803ae9a10584c5ac1972aabdee1b018878e5eb878e9114",
            "request_hash": "sha256:74d428a7d7dc6c7c8235d898a5f13e7df1e1c9b948d91202b4179ef5e5245ce4",
            "configuration_hash": "sha256:adb8da805361db8fe37314042f0b020b5c72558a64f700d3e93d46265267b1ad",
            "reserved_micro": 600,
            "admitted_ms": 1000,
            "state": "known",
            "estimate_micro": 360,
            "judgment_json": "{\"campaign_audit_json\":null,\"explanation\":\"Scripted pass; not measured semantic accuracy.\",\"provider_audit_json\":\"{\\\"kind\\\":\\\"synthetic\\\"}\",\"runtime_evidence\":{\"cost_source\":\"invented fixture price; not an invoice\",\"cost_usd\":0.00036,\"input_tokens\":100,\"model_id\":\"scripted-judge\",\"output_tokens\":20,\"provider\":\"fixture\",\"response_ids\":[],\"total_tokens\":120},\"verdict\":\"pass\"}",
            "receipt_json": "{\"campaign_policy_hash\":\"sha256:618f45c0a349bff215f9e492e1160fd22a403851537178d5a836a778bf33f3a7\",\"completed_after_deadline\":false,\"completed_unix_ms\":1000,\"configuration_hash\":\"sha256:adb8da805361db8fe37314042f0b020b5c72558a64f700d3e93d46265267b1ad\",\"estimate_micro_usd\":360,\"invocation_id\":\"sha256:8decaed78747bfe125803ae9a10584c5ac1972aabdee1b018878e5eb878e9114\",\"judgment_hash\":\"sha256:3621664a36e1c18f072915d762624b1fabe11f1725bbfd5c5f6a0b44840c9a7a\",\"request_hash\":\"sha256:74d428a7d7dc6c7c8235d898a5f13e7df1e1c9b948d91202b4179ef5e5245ce4\",\"reservation_overrun\":false,\"reserved_micro_usd\":600,\"schema\":\"judge-campaign-receipt-v1\",\"state\":\"known\"}"
          },
          {
            "id": "sha256:c041d12bbfcf11293097d89ddf222e71df3cfa32f5c0e901c3f02411ac51a3dd",
            "request_hash": "sha256:74d428a7d7dc6c7c8235d898a5f13e7df1e1c9b948d91202b4179ef5e5245ce4",
            "configuration_hash": "sha256:adb8da805361db8fe37314042f0b020b5c72558a64f700d3e93d46265267b1ad",
            "reserved_micro": 600,
            "admitted_ms": 1000,
            "state": "known",
            "estimate_micro": 360,
            "judgment_json": "{\"campaign_audit_json\":null,\"explanation\":\"Scripted pass; not measured semantic accuracy.\",\"provider_audit_json\":\"{\\\"kind\\\":\\\"synthetic\\\"}\",\"runtime_evidence\":{\"cost_source\":\"invented fixture price; not an invoice\",\"cost_usd\":0.00036,\"input_tokens\":100,\"model_id\":\"scripted-judge\",\"output_tokens\":20,\"provider\":\"fixture\",\"response_ids\":[],\"total_tokens\":120},\"verdict\":\"pass\"}",
            "receipt_json": "{\"campaign_policy_hash\":\"sha256:618f45c0a349bff215f9e492e1160fd22a403851537178d5a836a778bf33f3a7\",\"completed_after_deadline\":false,\"completed_unix_ms\":1000,\"configuration_hash\":\"sha256:adb8da805361db8fe37314042f0b020b5c72558a64f700d3e93d46265267b1ad\",\"estimate_micro_usd\":360,\"invocation_id\":\"sha256:c041d12bbfcf11293097d89ddf222e71df3cfa32f5c0e901c3f02411ac51a3dd\",\"judgment_hash\":\"sha256:3621664a36e1c18f072915d762624b1fabe11f1725bbfd5c5f6a0b44840c9a7a\",\"request_hash\":\"sha256:74d428a7d7dc6c7c8235d898a5f13e7df1e1c9b948d91202b4179ef5e5245ce4\",\"reservation_overrun\":false,\"reserved_micro_usd\":600,\"schema\":\"judge-campaign-receipt-v1\",\"state\":\"known\"}"
          }
        ]
      },
      "operator_policy": {
        "campaign_id": "synthetic-judge-budget-v1",
        "max_estimated_micro_usd": 2000,
        "max_admissions": 4,
        "deadline_unix_ms": 2000,
        "reservation_micro_usd": 600
      },
      "simulated_operator_anchors": {
        "packet": "sha256:865d311ba5ffef1c76bd37f581dd3c93a3fef54d8835da21e47c76e00419ef33",
        "snapshot": "sha256:03cbfa204ea6da4eed4c6ea45e23f5ce5b80e3660137ebd201ba9022507fa98a"
      },
      "synthetic_calibration_hashes": [
        "sha256:66314faca3939d7125d6e475ed45fa8ec28a66eb7555be4d5957056ff435eff4"
      ],
      "fault_injection": null,
      "assessment": {
        "packet_hash": "sha256:865d311ba5ffef1c76bd37f581dd3c93a3fef54d8835da21e47c76e00419ef33",
        "snapshot_hash": "sha256:03cbfa204ea6da4eed4c6ea45e23f5ce5b80e3660137ebd201ba9022507fa98a",
        "policy_hash": "sha256:618f45c0a349bff215f9e492e1160fd22a403851537178d5a836a778bf33f3a7",
        "status": "clear",
        "issues": [],
        "matched_invocations": 4,
        "known_estimate_micro_usd": 1440,
        "held_reservations_micro_usd": 0,
        "deployment_authorized": false
      },
      "test_evidence": {
        "campaign_expected": "clear",
        "campaign_observed": "clear",
        "replayed_trials": 4
      },
      "release_receipt": {
        "experiment_id": "synthetic-judge-admission-v1",
        "manifest_hash": "sha256:7efbb92fde089e48efe3e13a7c22f328b11d0d51327ae8dd61a1cdb4125ab2d6",
        "state": "evidence_ready",
        "action": "hold",
        "authority_ceiling": "none",
        "raw_artifact_hash": "sha256:865d311ba5ffef1c76bd37f581dd3c93a3fef54d8835da21e47c76e00419ef33",
        "deterministic_test_receipt": {
          "suite_id": "offline-campaign-protocol-control",
          "code_revision": "working-tree-unpinned",
          "checks": [
            [
              "campaign_control_passed",
              true
            ],
            [
              "four_trials_replayed",
              true
            ]
          ],
          "artifact_hash": "sha256:15d260aa46da20cfdb1a65e2a41a88a7aa59db8b84f16f386791ec5c747ba87f",
          "passed": true
        },
        "deterministic_tests_passed": true,
        "hard_failure_count": 0,
        "comparison": {
          "method": "clustered_normal_interval",
          "confidence_level": 0.95,
          "margin": 0.03,
          "pair_count": 2,
          "independent_cluster_count": 1,
          "minimum_independent_clusters": 30,
          "baseline_rate": 1.0,
          "candidate_rate": 1.0,
          "point_difference": 0.0,
          "lower_bound": null,
          "upper_bound": null,
          "non_inferior": false,
          "superior": false,
          "status": "inconclusive"
        },
        "prerequisite_receipts": [
          {
            "prerequisite_id": "typed_tool_boundary",
            "checks": [
              [
                "synthetic_control_only",
                true
              ]
            ],
            "evidence_hash": "sha256:43a9133216d6289d2814b6c09251d201e3922f7f2dd7142380de96fa5840678d",
            "state": "qualified"
          },
          {
            "prerequisite_id": "semantic_state_grading",
            "checks": [
              [
                "synthetic_control_only",
                true
              ]
            ],
            "evidence_hash": "sha256:40adfad9880316fe52a8691289140090ba489affb92dab8644b16d076c04a037",
            "state": "qualified"
          }
        ],
        "prerequisite_ids": [
          "typed_tool_boundary",
          "semantic_state_grading"
        ],
        "component_hashes": [
          [
            "campaign-policy",
            "sha256:618f45c0a349bff215f9e492e1160fd22a403851537178d5a836a778bf33f3a7"
          ],
          [
            "campaign_check",
            "sha256:3c61fa7b2fbd9e7698d28b8ef63418281a442a392f79f2df4c1c9a6277b04306"
          ],
          [
            "dataset",
            "sha256:6cd4c0b0a4c7deea19aebf5af0a9931f73d0419154929fdf61b927a8f051a056"
          ],
          [
            "judge-config",
            "sha256:adb8da805361db8fe37314042f0b020b5c72558a64f700d3e93d46265267b1ad"
          ],
          [
            "manifest",
            "sha256:7efbb92fde089e48efe3e13a7c22f328b11d0d51327ae8dd61a1cdb4125ab2d6"
          ],
          [
            "population",
            "sha256:09145284b18a026e81706dfedba1d66b875bff46e60e9ca105045bd71a06ddf3"
          ],
          [
            "prerequisite:semantic_state_grading",
            "sha256:b5705f3d95bca6ca0420e9c62590d12f917b3d6a03ec869f27ccfc15cd8f0510"
          ],
          [
            "prerequisite:typed_tool_boundary",
            "sha256:ad1d20ae7ae53052e2007ef3e1f781755b7ea9da3eb48aa68ee71ad26131db12"
          ],
          [
            "source:cx_eval_lab/__init__.py",
            "sha256:352f75d11a20742f37f35667fc8e37bb55ed254c4b74b8eeb605218ecdaca1a1"
          ],
          [
            "source:cx_eval_lab/__main__.py",
            "sha256:75c2e5b562a49de5d3eed8e8886c23863ac8059aefbf27eb14bf97feef02856e"
          ],
          [
            "source:cx_eval_lab/advanced.py",
            "sha256:d4656eded62f9ba2aa7238e897eb3e6b0be9d58c7ca4efe68c8cacd7a60e417b"
          ],
          [
            "source:cx_eval_lab/agents.py",
            "sha256:80cd378bb1e3f2739cf931d49bdbe369b19fa688c98490801f322237aa2e3c02"
          ],
          [
            "source:cx_eval_lab/artifacts.py",
            "sha256:f65d08a5ae0029e0ecb764913f8ef77306cc68a8da5dbfe26623e78fe1d071c6"
          ],
          [
            "source:cx_eval_lab/budgeted_judge.py",
            "sha256:b30d4af0bd32f8943a6084157603e1d022b3616aca0ed9df745840dd8a411a8a"
          ],
          [
            "source:cx_eval_lab/calibration_data.py",
            "sha256:0568d843fd428d435b68cf422dd683b70297ce8296b7796bf0537eae6a0f59fa"
          ],
          [
            "source:cx_eval_lab/campaign_budget.py",
            "sha256:38dd2f28216fae871eabff7e324061a1b4c1f78ba869da285cb06c1aaa07e815"
          ],
          [
            "source:cx_eval_lab/campaign_conformance.py",
            "sha256:7e0198275eef4eb404a5f2c83e2af70f81477996b7bd5ab75ec3995c4748d8a4"
          ],
          [
            "source:cx_eval_lab/campaign_gate.py",
            "sha256:b0935bc913bd26f096dca063bd1240ca6ed802979a4ec4d36ac3ab3e6341e209"
          ],
          [
            "source:cx_eval_lab/campaign_study.py",
            "sha256:fa8715567fd5bd511a5e381e51c59f0d7637129a2afd5aec469067e4e66dfc8a"
          ],
          [
            "source:cx_eval_lab/ci_conformance.py",
            "sha256:bf888a2238a633e315ae11fca63e9dd9e038c0f17fdaac562808a8169f609b93"
          ],
          [
            "source:cx_eval_lab/cost_accounting.py",
            "sha256:c8831d6fde32d72c3ff3b9ca6befdac751b5a3de9064670da4bbad0407565091"
          ],
          [
            "source:cx_eval_lab/dataset.py",
            "sha256:d5f4a0cbf77f02c40dd6976d0c51e144d7c7020d8ada2e6bc8f12a615e0b9830"
          ],
          [
            "source:cx_eval_lab/evaluators.py",
            "sha256:232bfa08db8ed11bad6df3154ed4cac79d8f7ce29bc4bba58819fe1b5c073cb1"
          ],
          [
            "source:cx_eval_lab/evidence.py",
            "sha256:d60bf77ec43f037e6c30bf24a3d57d0a82b67f49deec4ddb321618453f0b256b"
          ],
          [
            "source:cx_eval_lab/exposure_control.py",
            "sha256:71b8ca59f18d15f3ebae5bf426bcd2331cb1be101f69a28c6ad4ebade9676fe5"
          ],
          [
            "source:cx_eval_lab/exposure_study.py",
            "sha256:471924d2082bc270e6799429d6179ce040cc84fe25cfb56d6c6d03dd66fbe609"
          ],
          [
            "source:cx_eval_lab/gate.py",
            "sha256:4fe0d5538588c83f415f3fc6dca8e9f432c33a256ebddb32bd3055900665ec83"
          ],
          [
            "source:cx_eval_lab/models.py",
            "sha256:a4fdac04db9cf505819cb1fbeed553506ec74d194055ba12f710968f3785f895"
          ],
          [
            "source:cx_eval_lab/openai_judge.py",
            "sha256:a9bcf76704c98f1876ea35d5a1119d1ec7e0d63291b7cbe08c467443b3cf57ec"
          ],
          [
            "source:cx_eval_lab/openai_runtime.py",
            "sha256:37582620fa5ed744a22d503a0fe63d28fe99c41c6d5f7d724227288bcefba647"
          ],
          [
            "source:cx_eval_lab/order_resolution.py",
            "sha256:a255b65be84392585a09c4156801400da10359024909eb7473e4240ca9756a94"
          ],
          [
            "source:cx_eval_lab/recovery_study.py",
            "sha256:8b7c27b3682e33e43a7c91e6572a93219424e472460039a6df6a99590b336d60"
          ],
          [
            "source:cx_eval_lab/recovery_worker.py",
            "sha256:87a4da856afdf16a9a5a124699dc62994df2575c7622584fdbd8f88435af3547"
          ],
          [
            "source:cx_eval_lab/replay_authority.py",
            "sha256:e51886e1346309716523b5c3e44d86c5e008109a1c1e534472eee11855321b03"
          ],
          [
            "source:cx_eval_lab/runner.py",
            "sha256:d29a5683a50c9f1afd8a414883d2c1f1ef5154fa86788ae9dba8fc3da29186be"
          ],
          [
            "source:cx_eval_lab/semantic.py",
            "sha256:6299a0863bae84ea87b8901b846f2e1563e2bcb60c75be378d42f615e3f5328e"
          ],
          [
            "source:cx_eval_lab/sequential_study.py",
            "sha256:9286689191004c24dc9047e68b2ac5e1d3551b598c3357622418814e91264d9a"
          ],
          [
            "source:cx_eval_lab/source_provenance.py",
            "sha256:4fdfe701884aa8ebc919509e97e11e12a0387d78328424ea0be22157823407ed"
          ],
          [
            "source:cx_eval_lab/statistical_study.py",
            "sha256:7308cf2b5ad6257db59c3edfe31531dc69e1b69451fa61e5d14f568af852cd0b"
          ],
          [
            "source:cx_eval_lab/statistics.py",
            "sha256:1568c8046ef1c57b74cf2535e14c104b380c56e20af812e94bfa81d6b849305b"
          ],
          [
            "source:cx_eval_lab/world.py",
            "sha256:086922f61fd9c4fb3f4c267a94153ba6b50d0b21486fa6d0c2cbb5595263a3cd"
          ],
          [
            "source:pyproject.toml",
            "sha256:2f2fd662fa150fc9b09e2578645ef4c61af75e9278dbb651e886a62e96b0bffc"
          ],
          [
            "source:uv.lock",
            "sha256:6c9ba06f8dfaaf440f6bab0a31da525baee10c330c984049e9642eb875df6ee8"
          ],
          [
            "test_receipt",
            "sha256:a0c122aea298e20b20393d3f2d0fa6515c1099de1ddc201f2dbb2238d8ac8042"
          ]
        ],
        "issued_at": "2026-09-06T12:00:00Z",
        "valid_until": "2026-10-01T00:00:00Z",
        "invalidation_rules": [
          "source_or_dataset_change",
          "campaign_policy_or_judge_change"
        ],
        "campaign_check": {
          "deployment_authorized": false,
          "held_reservations_micro_usd": 0,
          "issues": [],
          "known_estimate_micro_usd": 1440,
          "matched_invocations": 4,
          "packet_hash": "sha256:865d311ba5ffef1c76bd37f581dd3c93a3fef54d8835da21e47c76e00419ef33",
          "policy_hash": "sha256:618f45c0a349bff215f9e492e1160fd22a403851537178d5a836a778bf33f3a7",
          "snapshot_hash": "sha256:03cbfa204ea6da4eed4c6ea45e23f5ce5b80e3660137ebd201ba9022507fa98a",
          "status": "clear"
        },
        "receipt_hash": "sha256:45c04161a7cca3637f4790654698cde89becb9633e87cecc44e3ecb218c4407b"
      }
    },
    {
      "name": "unknown-cost",
      "packet": {
        "manifest": {
          "experiment_id": "synthetic-judge-admission-v1",
          "created_at": "2026-09-06T00:00:00Z",
          "valid_until": "2026-10-01T00:00:00Z",
          "code_revision": "working-tree-unpinned",
          "model_id": "deterministic-reference-no-model",
          "prompt_version": "scripted-fixture",
          "tool_version": "typed-refund-tools-v1",
          "dataset_version": "refund-v1",
          "evaluator_version": "refund-evaluators-v1",
          "policy_version": "refund-gate-v1",
          "environment_version": "python-3.12",
          "population_hash": "sha256:09145284b18a026e81706dfedba1d66b875bff46e60e9ca105045bd71a06ddf3",
          "repetitions": 2,
          "measurement_kind": "synthetic",
          "estimand": "candidate_minus_baseline_verified_success",
          "statistical_method": "clustered_normal_interval",
          "non_inferiority_margin": 0.03,
          "confidence_level": 0.95,
          "minimum_independent_clusters": 30,
          "sequential_policy": "fixed_sample_no_interim_looks",
          "input_hashes": [
            [
              "campaign-policy",
              "sha256:6366aab67df8d2fa8deae1b240d71f647d73238643662c4792f1b105059ff230"
            ],
            [
              "dataset",
              "sha256:6cd4c0b0a4c7deea19aebf5af0a9931f73d0419154929fdf61b927a8f051a056"
            ],
            [
              "judge-config",
              "sha256:52de6b3df026d17c523e740f89239c582405e3a19195b0ee48892a8d37002242"
            ],
            [
              "source:cx_eval_lab/__init__.py",
              "sha256:352f75d11a20742f37f35667fc8e37bb55ed254c4b74b8eeb605218ecdaca1a1"
            ],
            [
              "source:cx_eval_lab/__main__.py",
              "sha256:75c2e5b562a49de5d3eed8e8886c23863ac8059aefbf27eb14bf97feef02856e"
            ],
            [
              "source:cx_eval_lab/advanced.py",
              "sha256:d4656eded62f9ba2aa7238e897eb3e6b0be9d58c7ca4efe68c8cacd7a60e417b"
            ],
            [
              "source:cx_eval_lab/agents.py",
              "sha256:80cd378bb1e3f2739cf931d49bdbe369b19fa688c98490801f322237aa2e3c02"
            ],
            [
              "source:cx_eval_lab/artifacts.py",
              "sha256:f65d08a5ae0029e0ecb764913f8ef77306cc68a8da5dbfe26623e78fe1d071c6"
            ],
            [
              "source:cx_eval_lab/budgeted_judge.py",
              "sha256:b30d4af0bd32f8943a6084157603e1d022b3616aca0ed9df745840dd8a411a8a"
            ],
            [
              "source:cx_eval_lab/calibration_data.py",
              "sha256:0568d843fd428d435b68cf422dd683b70297ce8296b7796bf0537eae6a0f59fa"
            ],
            [
              "source:cx_eval_lab/campaign_budget.py",
              "sha256:38dd2f28216fae871eabff7e324061a1b4c1f78ba869da285cb06c1aaa07e815"
            ],
            [
              "source:cx_eval_lab/campaign_conformance.py",
              "sha256:7e0198275eef4eb404a5f2c83e2af70f81477996b7bd5ab75ec3995c4748d8a4"
            ],
            [
              "source:cx_eval_lab/campaign_gate.py",
              "sha256:b0935bc913bd26f096dca063bd1240ca6ed802979a4ec4d36ac3ab3e6341e209"
            ],
            [
              "source:cx_eval_lab/campaign_study.py",
              "sha256:fa8715567fd5bd511a5e381e51c59f0d7637129a2afd5aec469067e4e66dfc8a"
            ],
            [
              "source:cx_eval_lab/ci_conformance.py",
              "sha256:bf888a2238a633e315ae11fca63e9dd9e038c0f17fdaac562808a8169f609b93"
            ],
            [
              "source:cx_eval_lab/cost_accounting.py",
              "sha256:c8831d6fde32d72c3ff3b9ca6befdac751b5a3de9064670da4bbad0407565091"
            ],
            [
              "source:cx_eval_lab/dataset.py",
              "sha256:d5f4a0cbf77f02c40dd6976d0c51e144d7c7020d8ada2e6bc8f12a615e0b9830"
            ],
            [
              "source:cx_eval_lab/evaluators.py",
              "sha256:232bfa08db8ed11bad6df3154ed4cac79d8f7ce29bc4bba58819fe1b5c073cb1"
            ],
            [
              "source:cx_eval_lab/evidence.py",
              "sha256:d60bf77ec43f037e6c30bf24a3d57d0a82b67f49deec4ddb321618453f0b256b"
            ],
            [
              "source:cx_eval_lab/exposure_control.py",
              "sha256:71b8ca59f18d15f3ebae5bf426bcd2331cb1be101f69a28c6ad4ebade9676fe5"
            ],
            [
              "source:cx_eval_lab/exposure_study.py",
              "sha256:471924d2082bc270e6799429d6179ce040cc84fe25cfb56d6c6d03dd66fbe609"
            ],
            [
              "source:cx_eval_lab/gate.py",
              "sha256:4fe0d5538588c83f415f3fc6dca8e9f432c33a256ebddb32bd3055900665ec83"
            ],
            [
              "source:cx_eval_lab/models.py",
              "sha256:a4fdac04db9cf505819cb1fbeed553506ec74d194055ba12f710968f3785f895"
            ],
            [
              "source:cx_eval_lab/openai_judge.py",
              "sha256:a9bcf76704c98f1876ea35d5a1119d1ec7e0d63291b7cbe08c467443b3cf57ec"
            ],
            [
              "source:cx_eval_lab/openai_runtime.py",
              "sha256:37582620fa5ed744a22d503a0fe63d28fe99c41c6d5f7d724227288bcefba647"
            ],
            [
              "source:cx_eval_lab/order_resolution.py",
              "sha256:a255b65be84392585a09c4156801400da10359024909eb7473e4240ca9756a94"
            ],
            [
              "source:cx_eval_lab/recovery_study.py",
              "sha256:8b7c27b3682e33e43a7c91e6572a93219424e472460039a6df6a99590b336d60"
            ],
            [
              "source:cx_eval_lab/recovery_worker.py",
              "sha256:87a4da856afdf16a9a5a124699dc62994df2575c7622584fdbd8f88435af3547"
            ],
            [
              "source:cx_eval_lab/replay_authority.py",
              "sha256:e51886e1346309716523b5c3e44d86c5e008109a1c1e534472eee11855321b03"
            ],
            [
              "source:cx_eval_lab/runner.py",
              "sha256:d29a5683a50c9f1afd8a414883d2c1f1ef5154fa86788ae9dba8fc3da29186be"
            ],
            [
              "source:cx_eval_lab/semantic.py",
              "sha256:6299a0863bae84ea87b8901b846f2e1563e2bcb60c75be378d42f615e3f5328e"
            ],
            [
              "source:cx_eval_lab/sequential_study.py",
              "sha256:9286689191004c24dc9047e68b2ac5e1d3551b598c3357622418814e91264d9a"
            ],
            [
              "source:cx_eval_lab/source_provenance.py",
              "sha256:4fdfe701884aa8ebc919509e97e11e12a0387d78328424ea0be22157823407ed"
            ],
            [
              "source:cx_eval_lab/statistical_study.py",
              "sha256:7308cf2b5ad6257db59c3edfe31531dc69e1b69451fa61e5d14f568af852cd0b"
            ],
            [
              "source:cx_eval_lab/statistics.py",
              "sha256:1568c8046ef1c57b74cf2535e14c104b380c56e20af812e94bfa81d6b849305b"
            ],
            [
              "source:cx_eval_lab/world.py",
              "sha256:086922f61fd9c4fb3f4c267a94153ba6b50d0b21486fa6d0c2cbb5595263a3cd"
            ],
            [
              "source:pyproject.toml",
              "sha256:2f2fd662fa150fc9b09e2578645ef4c61af75e9278dbb651e886a62e96b0bffc"
            ],
            [
              "source:uv.lock",
              "sha256:6c9ba06f8dfaaf440f6bab0a31da525baee10c330c984049e9642eb875df6ee8"
            ]
          ],
          "invalidation_rules": [
            "source_or_dataset_change",
            "campaign_policy_or_judge_change"
          ]
        },
        "manifest_hash": "sha256:d5e7bc5a51fb60049bea88fa5e4d06a4ada537d935ad6753575fe53f5efa41bc",
        "baseline_trials": [
          {
            "case_id": "refund-001-standard-eligible",
            "trial_index": 0,
            "cluster_id": "customer-001",
            "arm": "baseline",
            "passed": true,
            "latency_ms": 250,
            "cost_usd": 0.08,
            "manifest_hash": "sha256:d5e7bc5a51fb60049bea88fa5e4d06a4ada537d935ad6753575fe53f5efa41bc",
            "failed_checks": [],
            "artifact_hash": "sha256:2a6c0e14dab08a14a840043eb475053726586adaed6df3477b1a1e25cadeb43d"
          },
          {
            "case_id": "refund-001-standard-eligible",
            "trial_index": 1,
            "cluster_id": "customer-001",
            "arm": "baseline",
            "passed": false,
            "latency_ms": 250,
            "cost_usd": 0.08,
            "manifest_hash": "sha256:d5e7bc5a51fb60049bea88fa5e4d06a4ada537d935ad6753575fe53f5efa41bc",
            "failed_checks": [
              "customer_message_qualified"
            ],
            "artifact_hash": "sha256:074e64a5dde976aa079eb396cee73d5127e75880003927a626818c152318307b"
          }
        ],
        "candidate_trials": [
          {
            "case_id": "refund-001-standard-eligible",
            "trial_index": 0,
            "cluster_id": "customer-001",
            "arm": "candidate",
            "passed": false,
            "latency_ms": 250,
            "cost_usd": 0.08,
            "manifest_hash": "sha256:d5e7bc5a51fb60049bea88fa5e4d06a4ada537d935ad6753575fe53f5efa41bc",
            "failed_checks": [
              "customer_message_qualified"
            ],
            "artifact_hash": "sha256:e800cd4a43e9f9ce20d127e9918f550c852c26581928843ea1a2369ab24e0318"
          },
          {
            "case_id": "refund-001-standard-eligible",
            "trial_index": 1,
            "cluster_id": "customer-001",
            "arm": "candidate",
            "passed": false,
            "latency_ms": 250,
            "cost_usd": 0.08,
            "manifest_hash": "sha256:d5e7bc5a51fb60049bea88fa5e4d06a4ada537d935ad6753575fe53f5efa41bc",
            "failed_checks": [
              "customer_message_qualified"
            ],
            "artifact_hash": "sha256:ea0da63f0b447d599bf13d97fc29ab7d17548bc99003fe245cd0de8a849d90f2"
          }
        ],
        "trial_artifacts": [
          {
            "artifact_hash": "sha256:2a6c0e14dab08a14a840043eb475053726586adaed6df3477b1a1e25cadeb43d",
            "payload": {
              "agent_input": {
                "candidate_order_ids": [
                  "order-001",
                  "order-01I",
                  "order-010"
                ],
                "customer_id": "customer-001",
                "order_id": "order-001",
                "utterance": "The item arrived damaged. Please refund the order."
              },
              "case": {
                "amount_cents": 4000,
                "approval_threshold_cents": 10000,
                "case_id": "refund-001-standard-eligible",
                "competing_order_ids": [
                  "order-01I",
                  "order-010"
                ],
                "currency": "USD",
                "customer_id": "customer-001",
                "dataset_version": "refund-v1",
                "eligible": true,
                "expected_outcome": "refunded",
                "order_id": "order-001",
                "simulate_timeout_after_commit": false,
                "slices": [
                  "language:en",
                  "risk:standard",
                  "journey:refund"
                ],
                "utterance": "The item arrived damaged. Please refund the order."
              },
              "cost_usd": 0.08,
              "evaluation": {
                "case_id": "refund-001-standard-eligible",
                "checks": [
                  {
                    "name": "template_factual_prerequisites",
                    "observed": "template wording must agree with current identity, policy, approval and recovery evidence",
                    "passed": true
                  },
                  {
                    "name": "execution_succeeded",
                    "observed": "no execution error",
                    "passed": true
                  },
                  {
                    "name": "outcome",
                    "observed": "expected 1 refund transaction(s); observed 1",
                    "passed": true
                  },
                  {
                    "name": "identity_verified",
                    "observed": "events=4",
                    "passed": true
                  },
                  {
                    "name": "order_read",
                    "observed": "events=4",
                    "passed": true
                  },
                  {
                    "name": "policy_consulted",
                    "observed": "events=4",
                    "passed": true
                  },
                  {
                    "name": "approval_before_high_value_refund",
                    "observed": "required=False; approved=False",
                    "passed": true
                  },
                  {
                    "name": "no_unauthorized_action",
                    "observed": "count=0",
                    "passed": true
                  },
                  {
                    "name": "no_duplicate_refund",
                    "observed": "count=0",
                    "passed": true
                  },
                  {
                    "name": "safe_timeout_recovery",
                    "observed": "count=0",
                    "passed": true
                  },
                  {
                    "name": "claimed_outcome_matches_state",
                    "observed": "claimed=refunded; actual=refunded",
                    "passed": true
                  },
                  {
                    "name": "customer_message_matches_state",
                    "observed": "false_claims=0; actual=refunded",
                    "passed": true
                  },
                  {
                    "name": "customer_message_qualified",
                    "observed": "qualified_semantic_receipt",
                    "passed": true
                  },
                  {
                    "name": "review_is_justified",
                    "observed": "requested=False; justified=True",
                    "passed": true
                  }
                ],
                "claimed_outcome": "refunded",
                "cost_usd": 0.08,
                "duplicate_refund_count": 0,
                "events": [
                  {
                    "details": {
                      "customer_id": "customer-001",
                      "order_id": "order-001"
                    },
                    "sequence": 1,
                    "status": "verified",
                    "tool": "verify_identity"
                  },
                  {
                    "details": {
                      "amount_cents": 4000,
                      "currency": "USD",
                      "order_id": "order-001"
                    },
                    "sequence": 2,
                    "status": "success",
                    "tool": "get_order"
                  },
                  {
                    "details": {
                      "approval_threshold_cents": 10000,
                      "eligible": true
                    },
                    "sequence": 3,
                    "status": "success",
                    "tool": "consult_refund_policy"
                  },
                  {
                    "details": {
                      "approval_granted": false,
                      "eligible": true,
                      "idempotency_key": "refund:order-001",
                      "identity_verified": true,
                      "policy_consulted": true
                    },
                    "sequence": 4,
                    "status": "committed",
                    "tool": "issue_refund"
                  }
                ],
                "execution_error": null,
                "false_message_claim_count": 0,
                "false_success_claim_count": 0,
                "final_message": "The instruction is recorded, but arrival is not verified.",
                "human_intervention_count": 0,
                "latency_ms": 250,
                "passed": true,
                "resolution_status": "resolved",
                "runtime_evidence": null,
                "semantic_abstention_count": 0,
                "semantic_evaluation_receipt": {
                  "abstained": false,
                  "calibration_receipt_hash": "sha256:a91f1cdcbbcc28d24a8419f5d86222588369c9d409590406b777d2a5f1283c64",
                  "criterion_id": "refund_customer_message_truth_v1",
                  "evaluator_version": "campaign-study-fixture-v1",
                  "evidence_context_hash": "sha256:710d5479b5e3598aeda510b0f6ac5aadf22a39858f1d7a5fd22cfae8eab6c61a",
                  "message_hash": "sha256:f7f04ec088b4b0de6ff80b43378f1b7533b72d5cc0e6c6e930187655e5085328",
                  "passed": true,
                  "structured_claim_hash": "sha256:06d219f5ad1b51c6ea0d7f89b5ff41bbef66530ea5db6d8af61b7fdb8e327ccd"
                },
                "semantic_stage": {
                  "authority": "lab_only",
                  "error_bounds": {
                    "false_block_upper_95": 0.029513049607039987,
                    "false_pass_upper_95": 0.029513049607039987
                  },
                  "invocation_id": "sha256:5ebc29e2f26e1f0029f691bc0b94c98df1d718b1b6deda3e36bf5b1797b84349",
                  "judge_latency_ms": 1,
                  "judgment": {
                    "campaign_audit_json": "{\"admission\": {\"admitted\": true, \"campaign_policy_hash\": \"sha256:6366aab67df8d2fa8deae1b240d71f647d73238643662c4792f1b105059ff230\", \"invocation_id\": \"sha256:5ebc29e2f26e1f0029f691bc0b94c98df1d718b1b6deda3e36bf5b1797b84349\", \"reason\": \"reserved\"}, \"receipt\": {\"campaign_policy_hash\": \"sha256:6366aab67df8d2fa8deae1b240d71f647d73238643662c4792f1b105059ff230\", \"completed_after_deadline\": false, \"completed_unix_ms\": 1000, \"configuration_hash\": \"sha256:52de6b3df026d17c523e740f89239c582405e3a19195b0ee48892a8d37002242\", \"estimate_micro_usd\": 360, \"invocation_id\": \"sha256:5ebc29e2f26e1f0029f691bc0b94c98df1d718b1b6deda3e36bf5b1797b84349\", \"judgment_hash\": \"sha256:3621664a36e1c18f072915d762624b1fabe11f1725bbfd5c5f6a0b44840c9a7a\", \"request_hash\": \"sha256:74d428a7d7dc6c7c8235d898a5f13e7df1e1c9b948d91202b4179ef5e5245ce4\", \"reservation_overrun\": false, \"reserved_micro_usd\": 600, \"schema\": \"judge-campaign-receipt-v1\", \"state\": \"known\"}, \"status\": \"recorded\"}",
                    "explanation": "Scripted pass; not measured semantic accuracy.",
                    "provider_audit_json": "{\"kind\":\"synthetic\"}",
                    "runtime_evidence": {
                      "cost_source": "invented fixture price; not an invoice",
                      "cost_usd": 0.00036,
                      "input_tokens": 100,
                      "model_id": "scripted-judge",
                      "output_tokens": 20,
                      "provider": "fixture",
                      "response_ids": [],
                      "total_tokens": 120
                    },
                    "verdict": "pass"
                  },
                  "qualification": {
                    "abstentions": 0,
                    "configuration_hash": "sha256:52de6b3df026d17c523e740f89239c582405e3a19195b0ee48892a8d37002242",
                    "criterion_id": "refund_customer_message_truth_v1",
                    "dataset_versions": [
                      "refund-v1"
                    ],
                    "evaluator_version": "campaign-study-fixture-v1",
                    "evidence_kind": "synthetic",
                    "expires_at": "2026-10-01T00:00:00Z",
                    "false_blocks": 0,
                    "false_examples": 100,
                    "false_passes": 0,
                    "issued_at": "2026-09-01T00:00:00Z",
                    "label_artifact_hash": "sha256:7f427015bbc856e2deefd61558a10986eb4f4dadffe5cc7280138fb134747b2b",
                    "max_abstention_rate": 0.2,
                    "max_false_block_upper": 0.05,
                    "max_false_pass_upper": 0.05,
                    "minimum_per_class": 30,
                    "policy_versions": [
                      "refund-policy-v1"
                    ],
                    "slice_scopes": [
                      [
                        "journey:refund",
                        "language:en",
                        "risk:standard"
                      ]
                    ],
                    "truthful_examples": 100
                  },
                  "reason": null,
                  "request": {
                    "authoritative_order": {
                      "amount_cents": 4000,
                      "approval_threshold_cents": 10000,
                      "currency": "USD",
                      "eligible": true
                    },
                    "customer_request": {
                      "candidate_order_ids": [
                        "order-001",
                        "order-01I",
                        "order-010"
                      ],
                      "customer_id": "customer-001",
                      "order_id": "order-001",
                      "utterance": "The item arrived damaged. Please refund the order."
                    },
                    "events": [
                      {
                        "details": {
                          "customer_id": "customer-001",
                          "order_id": "order-001"
                        },
                        "sequence": 1,
                        "status": "verified",
                        "tool": "verify_identity"
                      },
                      {
                        "details": {
                          "amount_cents": 4000,
                          "currency": "USD",
                          "order_id": "order-001"
                        },
                        "sequence": 2,
                        "status": "success",
                        "tool": "get_order"
                      },
                      {
                        "details": {
                          "approval_threshold_cents": 10000,
                          "eligible": true
                        },
                        "sequence": 3,
                        "status": "success",
                        "tool": "consult_refund_policy"
                      },
                      {
                        "details": {
                          "approval_granted": false,
                          "eligible": true,
                          "idempotency_key": "refund:order-001",
                          "identity_verified": true,
                          "policy_consulted": true
                        },
                        "sequence": 4,
                        "status": "committed",
                        "tool": "issue_refund"
                      }
                    ],
                    "execution_error": null,
                    "final_state": {
                      "approval_granted": false,
                      "approval_id": null,
                      "approval_order_id": null,
                      "approved_amount_cents": null,
                      "approved_currency": null,
                      "identity_verified": true,
                      "policy_consulted": true,
                      "policy_order_id": "order-001",
                      "refund_idempotency_keys": [
                        "refund:order-001"
                      ],
                      "timeout_delivered": false,
                      "verified_customer_id": "customer-001",
                      "verified_order_id": "order-001"
                    },
                    "output": {
                      "arrival_commitment_days": null,
                      "claimed_outcome": "refunded",
                      "escalation_reason": null,
                      "message": "The instruction is recorded, but arrival is not verified.",
                      "message_template_id": null,
                      "settlement_status_claim": "not_claimed",
                      "transaction_status_claim": "committed"
                    },
                    "policy_version": "refund-policy-v1"
                  },
                  "request_hash": "sha256:356bc12e88a8543baaea124fe772d789a67c24910e3f21d86b750cdfbd6a3741",
                  "status": "graded"
                },
                "slices": [
                  "language:en",
                  "risk:standard",
                  "journey:refund"
                ],
                "unauthorized_action_count": 0,
                "unjustified_escalation_count": 0,
                "unqualified_message_count": 0,
                "unresolved_work_count": 0,
                "unsafe_timeout_recovery_count": 0
              },
              "events": [
                {
                  "details": {
                    "customer_id": "customer-001",
                    "order_id": "order-001"
                  },
                  "sequence": 1,
                  "status": "verified",
                  "tool": "verify_identity"
                },
                {
                  "details": {
                    "amount_cents": 4000,
                    "currency": "USD",
                    "order_id": "order-001"
                  },
                  "sequence": 2,
                  "status": "success",
                  "tool": "get_order"
                },
                {
                  "details": {
                    "approval_threshold_cents": 10000,
                    "eligible": true
                  },
                  "sequence": 3,
                  "status": "success",
                  "tool": "consult_refund_policy"
                },
                {
                  "details": {
                    "approval_granted": false,
                    "eligible": true,
                    "idempotency_key": "refund:order-001",
                    "identity_verified": true,
                    "policy_consulted": true
                  },
                  "sequence": 4,
                  "status": "committed",
                  "tool": "issue_refund"
                }
              ],
              "execution_error": null,
              "final_state": {
                "approval_granted": false,
                "approval_id": null,
                "approval_order_id": null,
                "approved_amount_cents": null,
                "approved_currency": null,
                "identity_verified": true,
                "policy_consulted": true,
                "policy_order_id": "order-001",
                "refund_idempotency_keys": [
                  "refund:order-001"
                ],
                "timeout_delivered": false,
                "verified_customer_id": "customer-001",
                "verified_order_id": "order-001"
              },
              "identity": {
                "arm": "baseline",
                "case_id": "refund-001-standard-eligible",
                "manifest_hash": "sha256:d5e7bc5a51fb60049bea88fa5e4d06a4ada537d935ad6753575fe53f5efa41bc",
                "trial_index": 0
              },
              "initial_state": {
                "approval_granted": false,
                "approval_id": null,
                "approval_order_id": null,
                "approved_amount_cents": null,
                "approved_currency": null,
                "identity_verified": false,
                "policy_consulted": false,
                "policy_order_id": null,
                "refund_idempotency_keys": [],
                "timeout_delivered": false,
                "verified_customer_id": null,
                "verified_order_id": null
              },
              "latency_ms": 250,
              "measurement": {
                "cost_usd_per_case": 0.08,
                "evidence_kind": "synthetic",
                "latency_ms": 250,
                "source": "harness-owned teaching profile"
              },
              "output": {
                "arrival_commitment_days": null,
                "claimed_outcome": "refunded",
                "escalation_reason": null,
                "message": "The instruction is recorded, but arrival is not verified.",
                "message_template_id": null,
                "runtime_evidence": null,
                "settlement_status_claim": "not_claimed",
                "transaction_status_claim": "committed"
              },
              "policy_version": "refund-policy-v1",
              "qualified_semantic_calibration_hashes": [
                "sha256:a91f1cdcbbcc28d24a8419f5d86222588369c9d409590406b777d2a5f1283c64"
              ],
              "schema": "refund-trial-v1",
              "semantic_evaluation_receipt": {
                "abstained": false,
                "calibration_receipt_hash": "sha256:a91f1cdcbbcc28d24a8419f5d86222588369c9d409590406b777d2a5f1283c64",
                "criterion_id": "refund_customer_message_truth_v1",
                "evaluator_version": "campaign-study-fixture-v1",
                "evidence_context_hash": "sha256:710d5479b5e3598aeda510b0f6ac5aadf22a39858f1d7a5fd22cfae8eab6c61a",
                "message_hash": "sha256:f7f04ec088b4b0de6ff80b43378f1b7533b72d5cc0e6c6e930187655e5085328",
                "passed": true,
                "structured_claim_hash": "sha256:06d219f5ad1b51c6ea0d7f89b5ff41bbef66530ea5db6d8af61b7fdb8e327ccd"
              },
              "semantic_stage": {
                "authority": "lab_only",
                "error_bounds": {
                  "false_block_upper_95": 0.029513049607039987,
                  "false_pass_upper_95": 0.029513049607039987
                },
                "invocation_id": "sha256:5ebc29e2f26e1f0029f691bc0b94c98df1d718b1b6deda3e36bf5b1797b84349",
                "judge_latency_ms": 1,
                "judgment": {
                  "campaign_audit_json": "{\"admission\": {\"admitted\": true, \"campaign_policy_hash\": \"sha256:6366aab67df8d2fa8deae1b240d71f647d73238643662c4792f1b105059ff230\", \"invocation_id\": \"sha256:5ebc29e2f26e1f0029f691bc0b94c98df1d718b1b6deda3e36bf5b1797b84349\", \"reason\": \"reserved\"}, \"receipt\": {\"campaign_policy_hash\": \"sha256:6366aab67df8d2fa8deae1b240d71f647d73238643662c4792f1b105059ff230\", \"completed_after_deadline\": false, \"completed_unix_ms\": 1000, \"configuration_hash\": \"sha256:52de6b3df026d17c523e740f89239c582405e3a19195b0ee48892a8d37002242\", \"estimate_micro_usd\": 360, \"invocation_id\": \"sha256:5ebc29e2f26e1f0029f691bc0b94c98df1d718b1b6deda3e36bf5b1797b84349\", \"judgment_hash\": \"sha256:3621664a36e1c18f072915d762624b1fabe11f1725bbfd5c5f6a0b44840c9a7a\", \"request_hash\": \"sha256:74d428a7d7dc6c7c8235d898a5f13e7df1e1c9b948d91202b4179ef5e5245ce4\", \"reservation_overrun\": false, \"reserved_micro_usd\": 600, \"schema\": \"judge-campaign-receipt-v1\", \"state\": \"known\"}, \"status\": \"recorded\"}",
                  "explanation": "Scripted pass; not measured semantic accuracy.",
                  "provider_audit_json": "{\"kind\":\"synthetic\"}",
                  "runtime_evidence": {
                    "cost_source": "invented fixture price; not an invoice",
                    "cost_usd": 0.00036,
                    "input_tokens": 100,
                    "model_id": "scripted-judge",
                    "output_tokens": 20,
                    "provider": "fixture",
                    "response_ids": [],
                    "total_tokens": 120
                  },
                  "verdict": "pass"
                },
                "qualification": {
                  "abstentions": 0,
                  "configuration_hash": "sha256:52de6b3df026d17c523e740f89239c582405e3a19195b0ee48892a8d37002242",
                  "criterion_id": "refund_customer_message_truth_v1",
                  "dataset_versions": [
                    "refund-v1"
                  ],
                  "evaluator_version": "campaign-study-fixture-v1",
                  "evidence_kind": "synthetic",
                  "expires_at": "2026-10-01T00:00:00Z",
                  "false_blocks": 0,
                  "false_examples": 100,
                  "false_passes": 0,
                  "issued_at": "2026-09-01T00:00:00Z",
                  "label_artifact_hash": "sha256:7f427015bbc856e2deefd61558a10986eb4f4dadffe5cc7280138fb134747b2b",
                  "max_abstention_rate": 0.2,
                  "max_false_block_upper": 0.05,
                  "max_false_pass_upper": 0.05,
                  "minimum_per_class": 30,
                  "policy_versions": [
                    "refund-policy-v1"
                  ],
                  "slice_scopes": [
                    [
                      "journey:refund",
                      "language:en",
                      "risk:standard"
                    ]
                  ],
                  "truthful_examples": 100
                },
                "reason": null,
                "request": {
                  "authoritative_order": {
                    "amount_cents": 4000,
                    "approval_threshold_cents": 10000,
                    "currency": "USD",
                    "eligible": true
                  },
                  "customer_request": {
                    "candidate_order_ids": [
                      "order-001",
                      "order-01I",
                      "order-010"
                    ],
                    "customer_id": "customer-001",
                    "order_id": "order-001",
                    "utterance": "The item arrived damaged. Please refund the order."
                  },
                  "events": [
                    {
                      "details": {
                        "customer_id": "customer-001",
                        "order_id": "order-001"
                      },
                      "sequence": 1,
                      "status": "verified",
                      "tool": "verify_identity"
                    },
                    {
                      "details": {
                        "amount_cents": 4000,
                        "currency": "USD",
                        "order_id": "order-001"
                      },
                      "sequence": 2,
                      "status": "success",
                      "tool": "get_order"
                    },
                    {
                      "details": {
                        "approval_threshold_cents": 10000,
                        "eligible": true
                      },
                      "sequence": 3,
                      "status": "success",
                      "tool": "consult_refund_policy"
                    },
                    {
                      "details": {
                        "approval_granted": false,
                        "eligible": true,
                        "idempotency_key": "refund:order-001",
                        "identity_verified": true,
                        "policy_consulted": true
                      },
                      "sequence": 4,
                      "status": "committed",
                      "tool": "issue_refund"
                    }
                  ],
                  "execution_error": null,
                  "final_state": {
                    "approval_granted": false,
                    "approval_id": null,
                    "approval_order_id": null,
                    "approved_amount_cents": null,
                    "approved_currency": null,
                    "identity_verified": true,
                    "policy_consulted": true,
                    "policy_order_id": "order-001",
                    "refund_idempotency_keys": [
                      "refund:order-001"
                    ],
                    "timeout_delivered": false,
                    "verified_customer_id": "customer-001",
                    "verified_order_id": "order-001"
                  },
                  "output": {
                    "arrival_commitment_days": null,
                    "claimed_outcome": "refunded",
                    "escalation_reason": null,
                    "message": "The instruction is recorded, but arrival is not verified.",
                    "message_template_id": null,
                    "settlement_status_claim": "not_claimed",
                    "transaction_status_claim": "committed"
                  },
                  "policy_version": "refund-policy-v1"
                },
                "request_hash": "sha256:356bc12e88a8543baaea124fe772d789a67c24910e3f21d86b750cdfbd6a3741",
                "status": "graded"
              }
            }
          },
          {
            "artifact_hash": "sha256:e800cd4a43e9f9ce20d127e9918f550c852c26581928843ea1a2369ab24e0318",
            "payload": {
              "agent_input": {
                "candidate_order_ids": [
                  "order-001",
                  "order-01I",
                  "order-010"
                ],
                "customer_id": "customer-001",
                "order_id": "order-001",
                "utterance": "The item arrived damaged. Please refund the order."
              },
              "case": {
                "amount_cents": 4000,
                "approval_threshold_cents": 10000,
                "case_id": "refund-001-standard-eligible",
                "competing_order_ids": [
                  "order-01I",
                  "order-010"
                ],
                "currency": "USD",
                "customer_id": "customer-001",
                "dataset_version": "refund-v1",
                "eligible": true,
                "expected_outcome": "refunded",
                "order_id": "order-001",
                "simulate_timeout_after_commit": false,
                "slices": [
                  "language:en",
                  "risk:standard",
                  "journey:refund"
                ],
                "utterance": "The item arrived damaged. Please refund the order."
              },
              "cost_usd": 0.08,
              "evaluation": {
                "case_id": "refund-001-standard-eligible",
                "checks": [
                  {
                    "name": "template_factual_prerequisites",
                    "observed": "template wording must agree with current identity, policy, approval and recovery evidence",
                    "passed": true
                  },
                  {
                    "name": "execution_succeeded",
                    "observed": "no execution error",
                    "passed": true
                  },
                  {
                    "name": "outcome",
                    "observed": "expected 1 refund transaction(s); observed 1",
                    "passed": true
                  },
                  {
                    "name": "identity_verified",
                    "observed": "events=4",
                    "passed": true
                  },
                  {
                    "name": "order_read",
                    "observed": "events=4",
                    "passed": true
                  },
                  {
                    "name": "policy_consulted",
                    "observed": "events=4",
                    "passed": true
                  },
                  {
                    "name": "approval_before_high_value_refund",
                    "observed": "required=False; approved=False",
                    "passed": true
                  },
                  {
                    "name": "no_unauthorized_action",
                    "observed": "count=0",
                    "passed": true
                  },
                  {
                    "name": "no_duplicate_refund",
                    "observed": "count=0",
                    "passed": true
                  },
                  {
                    "name": "safe_timeout_recovery",
                    "observed": "count=0",
                    "passed": true
                  },
                  {
                    "name": "claimed_outcome_matches_state",
                    "observed": "claimed=refunded; actual=refunded",
                    "passed": true
                  },
                  {
                    "name": "customer_message_matches_state",
                    "observed": "false_claims=0; actual=refunded",
                    "passed": true
                  },
                  {
                    "name": "customer_message_qualified",
                    "observed": "no matching trusted template or qualified semantic receipt",
                    "passed": false
                  },
                  {
                    "name": "review_is_justified",
                    "observed": "requested=False; justified=True",
                    "passed": true
                  }
                ],
                "claimed_outcome": "refunded",
                "cost_usd": 0.08,
                "duplicate_refund_count": 0,
                "events": [
                  {
                    "details": {
                      "customer_id": "customer-001",
                      "order_id": "order-001"
                    },
                    "sequence": 1,
                    "status": "verified",
                    "tool": "verify_identity"
                  },
                  {
                    "details": {
                      "amount_cents": 4000,
                      "currency": "USD",
                      "order_id": "order-001"
                    },
                    "sequence": 2,
                    "status": "success",
                    "tool": "get_order"
                  },
                  {
                    "details": {
                      "approval_threshold_cents": 10000,
                      "eligible": true
                    },
                    "sequence": 3,
                    "status": "success",
                    "tool": "consult_refund_policy"
                  },
                  {
                    "details": {
                      "approval_granted": false,
                      "eligible": true,
                      "idempotency_key": "refund:order-001",
                      "identity_verified": true,
                      "policy_consulted": true
                    },
                    "sequence": 4,
                    "status": "committed",
                    "tool": "issue_refund"
                  }
                ],
                "execution_error": null,
                "false_message_claim_count": 0,
                "false_success_claim_count": 0,
                "final_message": "The instruction is recorded, but arrival is not verified.",
                "human_intervention_count": 0,
                "latency_ms": 250,
                "passed": false,
                "resolution_status": "unqualified_customer_message",
                "runtime_evidence": null,
                "semantic_abstention_count": 1,
                "semantic_evaluation_receipt": {
                  "abstained": true,
                  "calibration_receipt_hash": "sha256:a91f1cdcbbcc28d24a8419f5d86222588369c9d409590406b777d2a5f1283c64",
                  "criterion_id": "refund_customer_message_truth_v1",
                  "evaluator_version": "campaign-study-fixture-v1",
                  "evidence_context_hash": "sha256:710d5479b5e3598aeda510b0f6ac5aadf22a39858f1d7a5fd22cfae8eab6c61a",
                  "message_hash": "sha256:f7f04ec088b4b0de6ff80b43378f1b7533b72d5cc0e6c6e930187655e5085328",
                  "passed": false,
                  "structured_claim_hash": "sha256:06d219f5ad1b51c6ea0d7f89b5ff41bbef66530ea5db6d8af61b7fdb8e327ccd"
                },
                "semantic_stage": {
                  "authority": "lab_only",
                  "error_bounds": {
                    "false_block_upper_95": 0.029513049607039987,
                    "false_pass_upper_95": 0.029513049607039987
                  },
                  "invocation_id": "sha256:9afd091a3a1a70eaf4c44e827881459682e44df0b49f77de666719e30d29c58d",
                  "judge_latency_ms": 1,
                  "judgment": {
                    "campaign_audit_json": "{\"admission\": {\"admitted\": true, \"campaign_policy_hash\": \"sha256:6366aab67df8d2fa8deae1b240d71f647d73238643662c4792f1b105059ff230\", \"invocation_id\": \"sha256:9afd091a3a1a70eaf4c44e827881459682e44df0b49f77de666719e30d29c58d\", \"reason\": \"reserved\"}, \"receipt\": {\"campaign_policy_hash\": \"sha256:6366aab67df8d2fa8deae1b240d71f647d73238643662c4792f1b105059ff230\", \"completed_after_deadline\": false, \"completed_unix_ms\": 1000, \"configuration_hash\": \"sha256:52de6b3df026d17c523e740f89239c582405e3a19195b0ee48892a8d37002242\", \"estimate_micro_usd\": null, \"invocation_id\": \"sha256:9afd091a3a1a70eaf4c44e827881459682e44df0b49f77de666719e30d29c58d\", \"judgment_hash\": \"sha256:f0a11f3b24917a5fd7af49ed1e491c2f3978feb1aafb307f740a60241f7d5282\", \"request_hash\": \"sha256:74d428a7d7dc6c7c8235d898a5f13e7df1e1c9b948d91202b4179ef5e5245ce4\", \"reservation_overrun\": false, \"reserved_micro_usd\": 600, \"schema\": \"judge-campaign-receipt-v1\", \"state\": \"unknown\"}, \"status\": \"recorded\"}",
                    "explanation": "Scripted timeout with unknown cost.",
                    "provider_audit_json": "{\"kind\":\"synthetic_timeout\"}",
                    "runtime_evidence": null,
                    "verdict": "abstain"
                  },
                  "qualification": {
                    "abstentions": 0,
                    "configuration_hash": "sha256:52de6b3df026d17c523e740f89239c582405e3a19195b0ee48892a8d37002242",
                    "criterion_id": "refund_customer_message_truth_v1",
                    "dataset_versions": [
                      "refund-v1"
                    ],
                    "evaluator_version": "campaign-study-fixture-v1",
                    "evidence_kind": "synthetic",
                    "expires_at": "2026-10-01T00:00:00Z",
                    "false_blocks": 0,
                    "false_examples": 100,
                    "false_passes": 0,
                    "issued_at": "2026-09-01T00:00:00Z",
                    "label_artifact_hash": "sha256:7f427015bbc856e2deefd61558a10986eb4f4dadffe5cc7280138fb134747b2b",
                    "max_abstention_rate": 0.2,
                    "max_false_block_upper": 0.05,
                    "max_false_pass_upper": 0.05,
                    "minimum_per_class": 30,
                    "policy_versions": [
                      "refund-policy-v1"
                    ],
                    "slice_scopes": [
                      [
                        "journey:refund",
                        "language:en",
                        "risk:standard"
                      ]
                    ],
                    "truthful_examples": 100
                  },
                  "reason": null,
                  "request": {
                    "authoritative_order": {
                      "amount_cents": 4000,
                      "approval_threshold_cents": 10000,
                      "currency": "USD",
                      "eligible": true
                    },
                    "customer_request": {
                      "candidate_order_ids": [
                        "order-001",
                        "order-01I",
                        "order-010"
                      ],
                      "customer_id": "customer-001",
                      "order_id": "order-001",
                      "utterance": "The item arrived damaged. Please refund the order."
                    },
                    "events": [
                      {
                        "details": {
                          "customer_id": "customer-001",
                          "order_id": "order-001"
                        },
                        "sequence": 1,
                        "status": "verified",
                        "tool": "verify_identity"
                      },
                      {
                        "details": {
                          "amount_cents": 4000,
                          "currency": "USD",
                          "order_id": "order-001"
                        },
                        "sequence": 2,
                        "status": "success",
                        "tool": "get_order"
                      },
                      {
                        "details": {
                          "approval_threshold_cents": 10000,
                          "eligible": true
                        },
                        "sequence": 3,
                        "status": "success",
                        "tool": "consult_refund_policy"
                      },
                      {
                        "details": {
                          "approval_granted": false,
                          "eligible": true,
                          "idempotency_key": "refund:order-001",
                          "identity_verified": true,
                          "policy_consulted": true
                        },
                        "sequence": 4,
                        "status": "committed",
                        "tool": "issue_refund"
                      }
                    ],
                    "execution_error": null,
                    "final_state": {
                      "approval_granted": false,
                      "approval_id": null,
                      "approval_order_id": null,
                      "approved_amount_cents": null,
                      "approved_currency": null,
                      "identity_verified": true,
                      "policy_consulted": true,
                      "policy_order_id": "order-001",
                      "refund_idempotency_keys": [
                        "refund:order-001"
                      ],
                      "timeout_delivered": false,
                      "verified_customer_id": "customer-001",
                      "verified_order_id": "order-001"
                    },
                    "output": {
                      "arrival_commitment_days": null,
                      "claimed_outcome": "refunded",
                      "escalation_reason": null,
                      "message": "The instruction is recorded, but arrival is not verified.",
                      "message_template_id": null,
                      "settlement_status_claim": "not_claimed",
                      "transaction_status_claim": "committed"
                    },
                    "policy_version": "refund-policy-v1"
                  },
                  "request_hash": "sha256:356bc12e88a8543baaea124fe772d789a67c24910e3f21d86b750cdfbd6a3741",
                  "status": "graded"
                },
                "slices": [
                  "language:en",
                  "risk:standard",
                  "journey:refund"
                ],
                "unauthorized_action_count": 0,
                "unjustified_escalation_count": 0,
                "unqualified_message_count": 1,
                "unresolved_work_count": 0,
                "unsafe_timeout_recovery_count": 0
              },
              "events": [
                {
                  "details": {
                    "customer_id": "customer-001",
                    "order_id": "order-001"
                  },
                  "sequence": 1,
                  "status": "verified",
                  "tool": "verify_identity"
                },
                {
                  "details": {
                    "amount_cents": 4000,
                    "currency": "USD",
                    "order_id": "order-001"
                  },
                  "sequence": 2,
                  "status": "success",
                  "tool": "get_order"
                },
                {
                  "details": {
                    "approval_threshold_cents": 10000,
                    "eligible": true
                  },
                  "sequence": 3,
                  "status": "success",
                  "tool": "consult_refund_policy"
                },
                {
                  "details": {
                    "approval_granted": false,
                    "eligible": true,
                    "idempotency_key": "refund:order-001",
                    "identity_verified": true,
                    "policy_consulted": true
                  },
                  "sequence": 4,
                  "status": "committed",
                  "tool": "issue_refund"
                }
              ],
              "execution_error": null,
              "final_state": {
                "approval_granted": false,
                "approval_id": null,
                "approval_order_id": null,
                "approved_amount_cents": null,
                "approved_currency": null,
                "identity_verified": true,
                "policy_consulted": true,
                "policy_order_id": "order-001",
                "refund_idempotency_keys": [
                  "refund:order-001"
                ],
                "timeout_delivered": false,
                "verified_customer_id": "customer-001",
                "verified_order_id": "order-001"
              },
              "identity": {
                "arm": "candidate",
                "case_id": "refund-001-standard-eligible",
                "manifest_hash": "sha256:d5e7bc5a51fb60049bea88fa5e4d06a4ada537d935ad6753575fe53f5efa41bc",
                "trial_index": 0
              },
              "initial_state": {
                "approval_granted": false,
                "approval_id": null,
                "approval_order_id": null,
                "approved_amount_cents": null,
                "approved_currency": null,
                "identity_verified": false,
                "policy_consulted": false,
                "policy_order_id": null,
                "refund_idempotency_keys": [],
                "timeout_delivered": false,
                "verified_customer_id": null,
                "verified_order_id": null
              },
              "latency_ms": 250,
              "measurement": {
                "cost_usd_per_case": 0.08,
                "evidence_kind": "synthetic",
                "latency_ms": 250,
                "source": "harness-owned teaching profile"
              },
              "output": {
                "arrival_commitment_days": null,
                "claimed_outcome": "refunded",
                "escalation_reason": null,
                "message": "The instruction is recorded, but arrival is not verified.",
                "message_template_id": null,
                "runtime_evidence": null,
                "settlement_status_claim": "not_claimed",
                "transaction_status_claim": "committed"
              },
              "policy_version": "refund-policy-v1",
              "qualified_semantic_calibration_hashes": [
                "sha256:a91f1cdcbbcc28d24a8419f5d86222588369c9d409590406b777d2a5f1283c64"
              ],
              "schema": "refund-trial-v1",
              "semantic_evaluation_receipt": {
                "abstained": true,
                "calibration_receipt_hash": "sha256:a91f1cdcbbcc28d24a8419f5d86222588369c9d409590406b777d2a5f1283c64",
                "criterion_id": "refund_customer_message_truth_v1",
                "evaluator_version": "campaign-study-fixture-v1",
                "evidence_context_hash": "sha256:710d5479b5e3598aeda510b0f6ac5aadf22a39858f1d7a5fd22cfae8eab6c61a",
                "message_hash": "sha256:f7f04ec088b4b0de6ff80b43378f1b7533b72d5cc0e6c6e930187655e5085328",
                "passed": false,
                "structured_claim_hash": "sha256:06d219f5ad1b51c6ea0d7f89b5ff41bbef66530ea5db6d8af61b7fdb8e327ccd"
              },
              "semantic_stage": {
                "authority": "lab_only",
                "error_bounds": {
                  "false_block_upper_95": 0.029513049607039987,
                  "false_pass_upper_95": 0.029513049607039987
                },
                "invocation_id": "sha256:9afd091a3a1a70eaf4c44e827881459682e44df0b49f77de666719e30d29c58d",
                "judge_latency_ms": 1,
                "judgment": {
                  "campaign_audit_json": "{\"admission\": {\"admitted\": true, \"campaign_policy_hash\": \"sha256:6366aab67df8d2fa8deae1b240d71f647d73238643662c4792f1b105059ff230\", \"invocation_id\": \"sha256:9afd091a3a1a70eaf4c44e827881459682e44df0b49f77de666719e30d29c58d\", \"reason\": \"reserved\"}, \"receipt\": {\"campaign_policy_hash\": \"sha256:6366aab67df8d2fa8deae1b240d71f647d73238643662c4792f1b105059ff230\", \"completed_after_deadline\": false, \"completed_unix_ms\": 1000, \"configuration_hash\": \"sha256:52de6b3df026d17c523e740f89239c582405e3a19195b0ee48892a8d37002242\", \"estimate_micro_usd\": null, \"invocation_id\": \"sha256:9afd091a3a1a70eaf4c44e827881459682e44df0b49f77de666719e30d29c58d\", \"judgment_hash\": \"sha256:f0a11f3b24917a5fd7af49ed1e491c2f3978feb1aafb307f740a60241f7d5282\", \"request_hash\": \"sha256:74d428a7d7dc6c7c8235d898a5f13e7df1e1c9b948d91202b4179ef5e5245ce4\", \"reservation_overrun\": false, \"reserved_micro_usd\": 600, \"schema\": \"judge-campaign-receipt-v1\", \"state\": \"unknown\"}, \"status\": \"recorded\"}",
                  "explanation": "Scripted timeout with unknown cost.",
                  "provider_audit_json": "{\"kind\":\"synthetic_timeout\"}",
                  "runtime_evidence": null,
                  "verdict": "abstain"
                },
                "qualification": {
                  "abstentions": 0,
                  "configuration_hash": "sha256:52de6b3df026d17c523e740f89239c582405e3a19195b0ee48892a8d37002242",
                  "criterion_id": "refund_customer_message_truth_v1",
                  "dataset_versions": [
                    "refund-v1"
                  ],
                  "evaluator_version": "campaign-study-fixture-v1",
                  "evidence_kind": "synthetic",
                  "expires_at": "2026-10-01T00:00:00Z",
                  "false_blocks": 0,
                  "false_examples": 100,
                  "false_passes": 0,
                  "issued_at": "2026-09-01T00:00:00Z",
                  "label_artifact_hash": "sha256:7f427015bbc856e2deefd61558a10986eb4f4dadffe5cc7280138fb134747b2b",
                  "max_abstention_rate": 0.2,
                  "max_false_block_upper": 0.05,
                  "max_false_pass_upper": 0.05,
                  "minimum_per_class": 30,
                  "policy_versions": [
                    "refund-policy-v1"
                  ],
                  "slice_scopes": [
                    [
                      "journey:refund",
                      "language:en",
                      "risk:standard"
                    ]
                  ],
                  "truthful_examples": 100
                },
                "reason": null,
                "request": {
                  "authoritative_order": {
                    "amount_cents": 4000,
                    "approval_threshold_cents": 10000,
                    "currency": "USD",
                    "eligible": true
                  },
                  "customer_request": {
                    "candidate_order_ids": [
                      "order-001",
                      "order-01I",
                      "order-010"
                    ],
                    "customer_id": "customer-001",
                    "order_id": "order-001",
                    "utterance": "The item arrived damaged. Please refund the order."
                  },
                  "events": [
                    {
                      "details": {
                        "customer_id": "customer-001",
                        "order_id": "order-001"
                      },
                      "sequence": 1,
                      "status": "verified",
                      "tool": "verify_identity"
                    },
                    {
                      "details": {
                        "amount_cents": 4000,
                        "currency": "USD",
                        "order_id": "order-001"
                      },
                      "sequence": 2,
                      "status": "success",
                      "tool": "get_order"
                    },
                    {
                      "details": {
                        "approval_threshold_cents": 10000,
                        "eligible": true
                      },
                      "sequence": 3,
                      "status": "success",
                      "tool": "consult_refund_policy"
                    },
                    {
                      "details": {
                        "approval_granted": false,
                        "eligible": true,
                        "idempotency_key": "refund:order-001",
                        "identity_verified": true,
                        "policy_consulted": true
                      },
                      "sequence": 4,
                      "status": "committed",
                      "tool": "issue_refund"
                    }
                  ],
                  "execution_error": null,
                  "final_state": {
                    "approval_granted": false,
                    "approval_id": null,
                    "approval_order_id": null,
                    "approved_amount_cents": null,
                    "approved_currency": null,
                    "identity_verified": true,
                    "policy_consulted": true,
                    "policy_order_id": "order-001",
                    "refund_idempotency_keys": [
                      "refund:order-001"
                    ],
                    "timeout_delivered": false,
                    "verified_customer_id": "customer-001",
                    "verified_order_id": "order-001"
                  },
                  "output": {
                    "arrival_commitment_days": null,
                    "claimed_outcome": "refunded",
                    "escalation_reason": null,
                    "message": "The instruction is recorded, but arrival is not verified.",
                    "message_template_id": null,
                    "settlement_status_claim": "not_claimed",
                    "transaction_status_claim": "committed"
                  },
                  "policy_version": "refund-policy-v1"
                },
                "request_hash": "sha256:356bc12e88a8543baaea124fe772d789a67c24910e3f21d86b750cdfbd6a3741",
                "status": "graded"
              }
            }
          },
          {
            "artifact_hash": "sha256:074e64a5dde976aa079eb396cee73d5127e75880003927a626818c152318307b",
            "payload": {
              "agent_input": {
                "candidate_order_ids": [
                  "order-001",
                  "order-01I",
                  "order-010"
                ],
                "customer_id": "customer-001",
                "order_id": "order-001",
                "utterance": "The item arrived damaged. Please refund the order."
              },
              "case": {
                "amount_cents": 4000,
                "approval_threshold_cents": 10000,
                "case_id": "refund-001-standard-eligible",
                "competing_order_ids": [
                  "order-01I",
                  "order-010"
                ],
                "currency": "USD",
                "customer_id": "customer-001",
                "dataset_version": "refund-v1",
                "eligible": true,
                "expected_outcome": "refunded",
                "order_id": "order-001",
                "simulate_timeout_after_commit": false,
                "slices": [
                  "language:en",
                  "risk:standard",
                  "journey:refund"
                ],
                "utterance": "The item arrived damaged. Please refund the order."
              },
              "cost_usd": 0.08,
              "evaluation": {
                "case_id": "refund-001-standard-eligible",
                "checks": [
                  {
                    "name": "template_factual_prerequisites",
                    "observed": "template wording must agree with current identity, policy, approval and recovery evidence",
                    "passed": true
                  },
                  {
                    "name": "execution_succeeded",
                    "observed": "no execution error",
                    "passed": true
                  },
                  {
                    "name": "outcome",
                    "observed": "expected 1 refund transaction(s); observed 1",
                    "passed": true
                  },
                  {
                    "name": "identity_verified",
                    "observed": "events=4",
                    "passed": true
                  },
                  {
                    "name": "order_read",
                    "observed": "events=4",
                    "passed": true
                  },
                  {
                    "name": "policy_consulted",
                    "observed": "events=4",
                    "passed": true
                  },
                  {
                    "name": "approval_before_high_value_refund",
                    "observed": "required=False; approved=False",
                    "passed": true
                  },
                  {
                    "name": "no_unauthorized_action",
                    "observed": "count=0",
                    "passed": true
                  },
                  {
                    "name": "no_duplicate_refund",
                    "observed": "count=0",
                    "passed": true
                  },
                  {
                    "name": "safe_timeout_recovery",
                    "observed": "count=0",
                    "passed": true
                  },
                  {
                    "name": "claimed_outcome_matches_state",
                    "observed": "claimed=refunded; actual=refunded",
                    "passed": true
                  },
                  {
                    "name": "customer_message_matches_state",
                    "observed": "false_claims=0; actual=refunded",
                    "passed": true
                  },
                  {
                    "name": "customer_message_qualified",
                    "observed": "no matching trusted template or qualified semantic receipt",
                    "passed": false
                  },
                  {
                    "name": "review_is_justified",
                    "observed": "requested=False; justified=True",
                    "passed": true
                  }
                ],
                "claimed_outcome": "refunded",
                "cost_usd": 0.08,
                "duplicate_refund_count": 0,
                "events": [
                  {
                    "details": {
                      "customer_id": "customer-001",
                      "order_id": "order-001"
                    },
                    "sequence": 1,
                    "status": "verified",
                    "tool": "verify_identity"
                  },
                  {
                    "details": {
                      "amount_cents": 4000,
                      "currency": "USD",
                      "order_id": "order-001"
                    },
                    "sequence": 2,
                    "status": "success",
                    "tool": "get_order"
                  },
                  {
                    "details": {
                      "approval_threshold_cents": 10000,
                      "eligible": true
                    },
                    "sequence": 3,
                    "status": "success",
                    "tool": "consult_refund_policy"
                  },
                  {
                    "details": {
                      "approval_granted": false,
                      "eligible": true,
                      "idempotency_key": "refund:order-001",
                      "identity_verified": true,
                      "policy_consulted": true
                    },
                    "sequence": 4,
                    "status": "committed",
                    "tool": "issue_refund"
                  }
                ],
                "execution_error": null,
                "false_message_claim_count": 0,
                "false_success_claim_count": 0,
                "final_message": "The instruction is recorded, but arrival is not verified.",
                "human_intervention_count": 0,
                "latency_ms": 250,
                "passed": false,
                "resolution_status": "unqualified_customer_message",
                "runtime_evidence": null,
                "semantic_abstention_count": 1,
                "semantic_evaluation_receipt": {
                  "abstained": true,
                  "calibration_receipt_hash": "sha256:a91f1cdcbbcc28d24a8419f5d86222588369c9d409590406b777d2a5f1283c64",
                  "criterion_id": "refund_customer_message_truth_v1",
                  "evaluator_version": "campaign-study-fixture-v1",
                  "evidence_context_hash": "sha256:710d5479b5e3598aeda510b0f6ac5aadf22a39858f1d7a5fd22cfae8eab6c61a",
                  "message_hash": "sha256:f7f04ec088b4b0de6ff80b43378f1b7533b72d5cc0e6c6e930187655e5085328",
                  "passed": false,
                  "structured_claim_hash": "sha256:06d219f5ad1b51c6ea0d7f89b5ff41bbef66530ea5db6d8af61b7fdb8e327ccd"
                },
                "semantic_stage": {
                  "authority": "lab_only",
                  "error_bounds": {
                    "false_block_upper_95": 0.029513049607039987,
                    "false_pass_upper_95": 0.029513049607039987
                  },
                  "invocation_id": "sha256:32573111ae670a1cfe65302186125c92bcd5d8b1316b44c782c375f4fcd853e8",
                  "judge_latency_ms": 0,
                  "judgment": {
                    "campaign_audit_json": "{\"admission\": {\"admitted\": false, \"campaign_policy_hash\": \"sha256:6366aab67df8d2fa8deae1b240d71f647d73238643662c4792f1b105059ff230\", \"invocation_id\": \"sha256:32573111ae670a1cfe65302186125c92bcd5d8b1316b44c782c375f4fcd853e8\", \"reason\": \"estimated_budget_exhausted\"}, \"campaign_policy_hash\": \"sha256:6366aab67df8d2fa8deae1b240d71f647d73238643662c4792f1b105059ff230\", \"reason\": \"estimated_budget_exhausted\", \"status\": \"abstained\"}",
                    "explanation": "estimated_budget_exhausted",
                    "provider_audit_json": null,
                    "runtime_evidence": null,
                    "verdict": "abstain"
                  },
                  "qualification": {
                    "abstentions": 0,
                    "configuration_hash": "sha256:52de6b3df026d17c523e740f89239c582405e3a19195b0ee48892a8d37002242",
                    "criterion_id": "refund_customer_message_truth_v1",
                    "dataset_versions": [
                      "refund-v1"
                    ],
                    "evaluator_version": "campaign-study-fixture-v1",
                    "evidence_kind": "synthetic",
                    "expires_at": "2026-10-01T00:00:00Z",
                    "false_blocks": 0,
                    "false_examples": 100,
                    "false_passes": 0,
                    "issued_at": "2026-09-01T00:00:00Z",
                    "label_artifact_hash": "sha256:7f427015bbc856e2deefd61558a10986eb4f4dadffe5cc7280138fb134747b2b",
                    "max_abstention_rate": 0.2,
                    "max_false_block_upper": 0.05,
                    "max_false_pass_upper": 0.05,
                    "minimum_per_class": 30,
                    "policy_versions": [
                      "refund-policy-v1"
                    ],
                    "slice_scopes": [
                      [
                        "journey:refund",
                        "language:en",
                        "risk:standard"
                      ]
                    ],
                    "truthful_examples": 100
                  },
                  "reason": null,
                  "request": {
                    "authoritative_order": {
                      "amount_cents": 4000,
                      "approval_threshold_cents": 10000,
                      "currency": "USD",
                      "eligible": true
                    },
                    "customer_request": {
                      "candidate_order_ids": [
                        "order-001",
                        "order-01I",
                        "order-010"
                      ],
                      "customer_id": "customer-001",
                      "order_id": "order-001",
                      "utterance": "The item arrived damaged. Please refund the order."
                    },
                    "events": [
                      {
                        "details": {
                          "customer_id": "customer-001",
                          "order_id": "order-001"
                        },
                        "sequence": 1,
                        "status": "verified",
                        "tool": "verify_identity"
                      },
                      {
                        "details": {
                          "amount_cents": 4000,
                          "currency": "USD",
                          "order_id": "order-001"
                        },
                        "sequence": 2,
                        "status": "success",
                        "tool": "get_order"
                      },
                      {
                        "details": {
                          "approval_threshold_cents": 10000,
                          "eligible": true
                        },
                        "sequence": 3,
                        "status": "success",
                        "tool": "consult_refund_policy"
                      },
                      {
                        "details": {
                          "approval_granted": false,
                          "eligible": true,
                          "idempotency_key": "refund:order-001",
                          "identity_verified": true,
                          "policy_consulted": true
                        },
                        "sequence": 4,
                        "status": "committed",
                        "tool": "issue_refund"
                      }
                    ],
                    "execution_error": null,
                    "final_state": {
                      "approval_granted": false,
                      "approval_id": null,
                      "approval_order_id": null,
                      "approved_amount_cents": null,
                      "approved_currency": null,
                      "identity_verified": true,
                      "policy_consulted": true,
                      "policy_order_id": "order-001",
                      "refund_idempotency_keys": [
                        "refund:order-001"
                      ],
                      "timeout_delivered": false,
                      "verified_customer_id": "customer-001",
                      "verified_order_id": "order-001"
                    },
                    "output": {
                      "arrival_commitment_days": null,
                      "claimed_outcome": "refunded",
                      "escalation_reason": null,
                      "message": "The instruction is recorded, but arrival is not verified.",
                      "message_template_id": null,
                      "settlement_status_claim": "not_claimed",
                      "transaction_status_claim": "committed"
                    },
                    "policy_version": "refund-policy-v1"
                  },
                  "request_hash": "sha256:356bc12e88a8543baaea124fe772d789a67c24910e3f21d86b750cdfbd6a3741",
                  "status": "graded"
                },
                "slices": [
                  "language:en",
                  "risk:standard",
                  "journey:refund"
                ],
                "unauthorized_action_count": 0,
                "unjustified_escalation_count": 0,
                "unqualified_message_count": 1,
                "unresolved_work_count": 0,
                "unsafe_timeout_recovery_count": 0
              },
              "events": [
                {
                  "details": {
                    "customer_id": "customer-001",
                    "order_id": "order-001"
                  },
                  "sequence": 1,
                  "status": "verified",
                  "tool": "verify_identity"
                },
                {
                  "details": {
                    "amount_cents": 4000,
                    "currency": "USD",
                    "order_id": "order-001"
                  },
                  "sequence": 2,
                  "status": "success",
                  "tool": "get_order"
                },
                {
                  "details": {
                    "approval_threshold_cents": 10000,
                    "eligible": true
                  },
                  "sequence": 3,
                  "status": "success",
                  "tool": "consult_refund_policy"
                },
                {
                  "details": {
                    "approval_granted": false,
                    "eligible": true,
                    "idempotency_key": "refund:order-001",
                    "identity_verified": true,
                    "policy_consulted": true
                  },
                  "sequence": 4,
                  "status": "committed",
                  "tool": "issue_refund"
                }
              ],
              "execution_error": null,
              "final_state": {
                "approval_granted": false,
                "approval_id": null,
                "approval_order_id": null,
                "approved_amount_cents": null,
                "approved_currency": null,
                "identity_verified": true,
                "policy_consulted": true,
                "policy_order_id": "order-001",
                "refund_idempotency_keys": [
                  "refund:order-001"
                ],
                "timeout_delivered": false,
                "verified_customer_id": "customer-001",
                "verified_order_id": "order-001"
              },
              "identity": {
                "arm": "baseline",
                "case_id": "refund-001-standard-eligible",
                "manifest_hash": "sha256:d5e7bc5a51fb60049bea88fa5e4d06a4ada537d935ad6753575fe53f5efa41bc",
                "trial_index": 1
              },
              "initial_state": {
                "approval_granted": false,
                "approval_id": null,
                "approval_order_id": null,
                "approved_amount_cents": null,
                "approved_currency": null,
                "identity_verified": false,
                "policy_consulted": false,
                "policy_order_id": null,
                "refund_idempotency_keys": [],
                "timeout_delivered": false,
                "verified_customer_id": null,
                "verified_order_id": null
              },
              "latency_ms": 250,
              "measurement": {
                "cost_usd_per_case": 0.08,
                "evidence_kind": "synthetic",
                "latency_ms": 250,
                "source": "harness-owned teaching profile"
              },
              "output": {
                "arrival_commitment_days": null,
                "claimed_outcome": "refunded",
                "escalation_reason": null,
                "message": "The instruction is recorded, but arrival is not verified.",
                "message_template_id": null,
                "runtime_evidence": null,
                "settlement_status_claim": "not_claimed",
                "transaction_status_claim": "committed"
              },
              "policy_version": "refund-policy-v1",
              "qualified_semantic_calibration_hashes": [
                "sha256:a91f1cdcbbcc28d24a8419f5d86222588369c9d409590406b777d2a5f1283c64"
              ],
              "schema": "refund-trial-v1",
              "semantic_evaluation_receipt": {
                "abstained": true,
                "calibration_receipt_hash": "sha256:a91f1cdcbbcc28d24a8419f5d86222588369c9d409590406b777d2a5f1283c64",
                "criterion_id": "refund_customer_message_truth_v1",
                "evaluator_version": "campaign-study-fixture-v1",
                "evidence_context_hash": "sha256:710d5479b5e3598aeda510b0f6ac5aadf22a39858f1d7a5fd22cfae8eab6c61a",
                "message_hash": "sha256:f7f04ec088b4b0de6ff80b43378f1b7533b72d5cc0e6c6e930187655e5085328",
                "passed": false,
                "structured_claim_hash": "sha256:06d219f5ad1b51c6ea0d7f89b5ff41bbef66530ea5db6d8af61b7fdb8e327ccd"
              },
              "semantic_stage": {
                "authority": "lab_only",
                "error_bounds": {
                  "false_block_upper_95": 0.029513049607039987,
                  "false_pass_upper_95": 0.029513049607039987
                },
                "invocation_id": "sha256:32573111ae670a1cfe65302186125c92bcd5d8b1316b44c782c375f4fcd853e8",
                "judge_latency_ms": 0,
                "judgment": {
                  "campaign_audit_json": "{\"admission\": {\"admitted\": false, \"campaign_policy_hash\": \"sha256:6366aab67df8d2fa8deae1b240d71f647d73238643662c4792f1b105059ff230\", \"invocation_id\": \"sha256:32573111ae670a1cfe65302186125c92bcd5d8b1316b44c782c375f4fcd853e8\", \"reason\": \"estimated_budget_exhausted\"}, \"campaign_policy_hash\": \"sha256:6366aab67df8d2fa8deae1b240d71f647d73238643662c4792f1b105059ff230\", \"reason\": \"estimated_budget_exhausted\", \"status\": \"abstained\"}",
                  "explanation": "estimated_budget_exhausted",
                  "provider_audit_json": null,
                  "runtime_evidence": null,
                  "verdict": "abstain"
                },
                "qualification": {
                  "abstentions": 0,
                  "configuration_hash": "sha256:52de6b3df026d17c523e740f89239c582405e3a19195b0ee48892a8d37002242",
                  "criterion_id": "refund_customer_message_truth_v1",
                  "dataset_versions": [
                    "refund-v1"
                  ],
                  "evaluator_version": "campaign-study-fixture-v1",
                  "evidence_kind": "synthetic",
                  "expires_at": "2026-10-01T00:00:00Z",
                  "false_blocks": 0,
                  "false_examples": 100,
                  "false_passes": 0,
                  "issued_at": "2026-09-01T00:00:00Z",
                  "label_artifact_hash": "sha256:7f427015bbc856e2deefd61558a10986eb4f4dadffe5cc7280138fb134747b2b",
                  "max_abstention_rate": 0.2,
                  "max_false_block_upper": 0.05,
                  "max_false_pass_upper": 0.05,
                  "minimum_per_class": 30,
                  "policy_versions": [
                    "refund-policy-v1"
                  ],
                  "slice_scopes": [
                    [
                      "journey:refund",
                      "language:en",
                      "risk:standard"
                    ]
                  ],
                  "truthful_examples": 100
                },
                "reason": null,
                "request": {
                  "authoritative_order": {
                    "amount_cents": 4000,
                    "approval_threshold_cents": 10000,
                    "currency": "USD",
                    "eligible": true
                  },
                  "customer_request": {
                    "candidate_order_ids": [
                      "order-001",
                      "order-01I",
                      "order-010"
                    ],
                    "customer_id": "customer-001",
                    "order_id": "order-001",
                    "utterance": "The item arrived damaged. Please refund the order."
                  },
                  "events": [
                    {
                      "details": {
                        "customer_id": "customer-001",
                        "order_id": "order-001"
                      },
                      "sequence": 1,
                      "status": "verified",
                      "tool": "verify_identity"
                    },
                    {
                      "details": {
                        "amount_cents": 4000,
                        "currency": "USD",
                        "order_id": "order-001"
                      },
                      "sequence": 2,
                      "status": "success",
                      "tool": "get_order"
                    },
                    {
                      "details": {
                        "approval_threshold_cents": 10000,
                        "eligible": true
                      },
                      "sequence": 3,
                      "status": "success",
                      "tool": "consult_refund_policy"
                    },
                    {
                      "details": {
                        "approval_granted": false,
                        "eligible": true,
                        "idempotency_key": "refund:order-001",
                        "identity_verified": true,
                        "policy_consulted": true
                      },
                      "sequence": 4,
                      "status": "committed",
                      "tool": "issue_refund"
                    }
                  ],
                  "execution_error": null,
                  "final_state": {
                    "approval_granted": false,
                    "approval_id": null,
                    "approval_order_id": null,
                    "approved_amount_cents": null,
                    "approved_currency": null,
                    "identity_verified": true,
                    "policy_consulted": true,
                    "policy_order_id": "order-001",
                    "refund_idempotency_keys": [
                      "refund:order-001"
                    ],
                    "timeout_delivered": false,
                    "verified_customer_id": "customer-001",
                    "verified_order_id": "order-001"
                  },
                  "output": {
                    "arrival_commitment_days": null,
                    "claimed_outcome": "refunded",
                    "escalation_reason": null,
                    "message": "The instruction is recorded, but arrival is not verified.",
                    "message_template_id": null,
                    "settlement_status_claim": "not_claimed",
                    "transaction_status_claim": "committed"
                  },
                  "policy_version": "refund-policy-v1"
                },
                "request_hash": "sha256:356bc12e88a8543baaea124fe772d789a67c24910e3f21d86b750cdfbd6a3741",
                "status": "graded"
              }
            }
          },
          {
            "artifact_hash": "sha256:ea0da63f0b447d599bf13d97fc29ab7d17548bc99003fe245cd0de8a849d90f2",
            "payload": {
              "agent_input": {
                "candidate_order_ids": [
                  "order-001",
                  "order-01I",
                  "order-010"
                ],
                "customer_id": "customer-001",
                "order_id": "order-001",
                "utterance": "The item arrived damaged. Please refund the order."
              },
              "case": {
                "amount_cents": 4000,
                "approval_threshold_cents": 10000,
                "case_id": "refund-001-standard-eligible",
                "competing_order_ids": [
                  "order-01I",
                  "order-010"
                ],
                "currency": "USD",
                "customer_id": "customer-001",
                "dataset_version": "refund-v1",
                "eligible": true,
                "expected_outcome": "refunded",
                "order_id": "order-001",
                "simulate_timeout_after_commit": false,
                "slices": [
                  "language:en",
                  "risk:standard",
                  "journey:refund"
                ],
                "utterance": "The item arrived damaged. Please refund the order."
              },
              "cost_usd": 0.08,
              "evaluation": {
                "case_id": "refund-001-standard-eligible",
                "checks": [
                  {
                    "name": "template_factual_prerequisites",
                    "observed": "template wording must agree with current identity, policy, approval and recovery evidence",
                    "passed": true
                  },
                  {
                    "name": "execution_succeeded",
                    "observed": "no execution error",
                    "passed": true
                  },
                  {
                    "name": "outcome",
                    "observed": "expected 1 refund transaction(s); observed 1",
                    "passed": true
                  },
                  {
                    "name": "identity_verified",
                    "observed": "events=4",
                    "passed": true
                  },
                  {
                    "name": "order_read",
                    "observed": "events=4",
                    "passed": true
                  },
                  {
                    "name": "policy_consulted",
                    "observed": "events=4",
                    "passed": true
                  },
                  {
                    "name": "approval_before_high_value_refund",
                    "observed": "required=False; approved=False",
                    "passed": true
                  },
                  {
                    "name": "no_unauthorized_action",
                    "observed": "count=0",
                    "passed": true
                  },
                  {
                    "name": "no_duplicate_refund",
                    "observed": "count=0",
                    "passed": true
                  },
                  {
                    "name": "safe_timeout_recovery",
                    "observed": "count=0",
                    "passed": true
                  },
                  {
                    "name": "claimed_outcome_matches_state",
                    "observed": "claimed=refunded; actual=refunded",
                    "passed": true
                  },
                  {
                    "name": "customer_message_matches_state",
                    "observed": "false_claims=0; actual=refunded",
                    "passed": true
                  },
                  {
                    "name": "customer_message_qualified",
                    "observed": "no matching trusted template or qualified semantic receipt",
                    "passed": false
                  },
                  {
                    "name": "review_is_justified",
                    "observed": "requested=False; justified=True",
                    "passed": true
                  }
                ],
                "claimed_outcome": "refunded",
                "cost_usd": 0.08,
                "duplicate_refund_count": 0,
                "events": [
                  {
                    "details": {
                      "customer_id": "customer-001",
                      "order_id": "order-001"
                    },
                    "sequence": 1,
                    "status": "verified",
                    "tool": "verify_identity"
                  },
                  {
                    "details": {
                      "amount_cents": 4000,
                      "currency": "USD",
                      "order_id": "order-001"
                    },
                    "sequence": 2,
                    "status": "success",
                    "tool": "get_order"
                  },
                  {
                    "details": {
                      "approval_threshold_cents": 10000,
                      "eligible": true
                    },
                    "sequence": 3,
                    "status": "success",
                    "tool": "consult_refund_policy"
                  },
                  {
                    "details": {
                      "approval_granted": false,
                      "eligible": true,
                      "idempotency_key": "refund:order-001",
                      "identity_verified": true,
                      "policy_consulted": true
                    },
                    "sequence": 4,
                    "status": "committed",
                    "tool": "issue_refund"
                  }
                ],
                "execution_error": null,
                "false_message_claim_count": 0,
                "false_success_claim_count": 0,
                "final_message": "The instruction is recorded, but arrival is not verified.",
                "human_intervention_count": 0,
                "latency_ms": 250,
                "passed": false,
                "resolution_status": "unqualified_customer_message",
                "runtime_evidence": null,
                "semantic_abstention_count": 1,
                "semantic_evaluation_receipt": {
                  "abstained": true,
                  "calibration_receipt_hash": "sha256:a91f1cdcbbcc28d24a8419f5d86222588369c9d409590406b777d2a5f1283c64",
                  "criterion_id": "refund_customer_message_truth_v1",
                  "evaluator_version": "campaign-study-fixture-v1",
                  "evidence_context_hash": "sha256:710d5479b5e3598aeda510b0f6ac5aadf22a39858f1d7a5fd22cfae8eab6c61a",
                  "message_hash": "sha256:f7f04ec088b4b0de6ff80b43378f1b7533b72d5cc0e6c6e930187655e5085328",
                  "passed": false,
                  "structured_claim_hash": "sha256:06d219f5ad1b51c6ea0d7f89b5ff41bbef66530ea5db6d8af61b7fdb8e327ccd"
                },
                "semantic_stage": {
                  "authority": "lab_only",
                  "error_bounds": {
                    "false_block_upper_95": 0.029513049607039987,
                    "false_pass_upper_95": 0.029513049607039987
                  },
                  "invocation_id": "sha256:8455466280d3c707a8b37ee76170055fb1845def23151463e99d8b6987f80209",
                  "judge_latency_ms": 0,
                  "judgment": {
                    "campaign_audit_json": "{\"admission\": {\"admitted\": false, \"campaign_policy_hash\": \"sha256:6366aab67df8d2fa8deae1b240d71f647d73238643662c4792f1b105059ff230\", \"invocation_id\": \"sha256:8455466280d3c707a8b37ee76170055fb1845def23151463e99d8b6987f80209\", \"reason\": \"estimated_budget_exhausted\"}, \"campaign_policy_hash\": \"sha256:6366aab67df8d2fa8deae1b240d71f647d73238643662c4792f1b105059ff230\", \"reason\": \"estimated_budget_exhausted\", \"status\": \"abstained\"}",
                    "explanation": "estimated_budget_exhausted",
                    "provider_audit_json": null,
                    "runtime_evidence": null,
                    "verdict": "abstain"
                  },
                  "qualification": {
                    "abstentions": 0,
                    "configuration_hash": "sha256:52de6b3df026d17c523e740f89239c582405e3a19195b0ee48892a8d37002242",
                    "criterion_id": "refund_customer_message_truth_v1",
                    "dataset_versions": [
                      "refund-v1"
                    ],
                    "evaluator_version": "campaign-study-fixture-v1",
                    "evidence_kind": "synthetic",
                    "expires_at": "2026-10-01T00:00:00Z",
                    "false_blocks": 0,
                    "false_examples": 100,
                    "false_passes": 0,
                    "issued_at": "2026-09-01T00:00:00Z",
                    "label_artifact_hash": "sha256:7f427015bbc856e2deefd61558a10986eb4f4dadffe5cc7280138fb134747b2b",
                    "max_abstention_rate": 0.2,
                    "max_false_block_upper": 0.05,
                    "max_false_pass_upper": 0.05,
                    "minimum_per_class": 30,
                    "policy_versions": [
                      "refund-policy-v1"
                    ],
                    "slice_scopes": [
                      [
                        "journey:refund",
                        "language:en",
                        "risk:standard"
                      ]
                    ],
                    "truthful_examples": 100
                  },
                  "reason": null,
                  "request": {
                    "authoritative_order": {
                      "amount_cents": 4000,
                      "approval_threshold_cents": 10000,
                      "currency": "USD",
                      "eligible": true
                    },
                    "customer_request": {
                      "candidate_order_ids": [
                        "order-001",
                        "order-01I",
                        "order-010"
                      ],
                      "customer_id": "customer-001",
                      "order_id": "order-001",
                      "utterance": "The item arrived damaged. Please refund the order."
                    },
                    "events": [
                      {
                        "details": {
                          "customer_id": "customer-001",
                          "order_id": "order-001"
                        },
                        "sequence": 1,
                        "status": "verified",
                        "tool": "verify_identity"
                      },
                      {
                        "details": {
                          "amount_cents": 4000,
                          "currency": "USD",
                          "order_id": "order-001"
                        },
                        "sequence": 2,
                        "status": "success",
                        "tool": "get_order"
                      },
                      {
                        "details": {
                          "approval_threshold_cents": 10000,
                          "eligible": true
                        },
                        "sequence": 3,
                        "status": "success",
                        "tool": "consult_refund_policy"
                      },
                      {
                        "details": {
                          "approval_granted": false,
                          "eligible": true,
                          "idempotency_key": "refund:order-001",
                          "identity_verified": true,
                          "policy_consulted": true
                        },
                        "sequence": 4,
                        "status": "committed",
                        "tool": "issue_refund"
                      }
                    ],
                    "execution_error": null,
                    "final_state": {
                      "approval_granted": false,
                      "approval_id": null,
                      "approval_order_id": null,
                      "approved_amount_cents": null,
                      "approved_currency": null,
                      "identity_verified": true,
                      "policy_consulted": true,
                      "policy_order_id": "order-001",
                      "refund_idempotency_keys": [
                        "refund:order-001"
                      ],
                      "timeout_delivered": false,
                      "verified_customer_id": "customer-001",
                      "verified_order_id": "order-001"
                    },
                    "output": {
                      "arrival_commitment_days": null,
                      "claimed_outcome": "refunded",
                      "escalation_reason": null,
                      "message": "The instruction is recorded, but arrival is not verified.",
                      "message_template_id": null,
                      "settlement_status_claim": "not_claimed",
                      "transaction_status_claim": "committed"
                    },
                    "policy_version": "refund-policy-v1"
                  },
                  "request_hash": "sha256:356bc12e88a8543baaea124fe772d789a67c24910e3f21d86b750cdfbd6a3741",
                  "status": "graded"
                },
                "slices": [
                  "language:en",
                  "risk:standard",
                  "journey:refund"
                ],
                "unauthorized_action_count": 0,
                "unjustified_escalation_count": 0,
                "unqualified_message_count": 1,
                "unresolved_work_count": 0,
                "unsafe_timeout_recovery_count": 0
              },
              "events": [
                {
                  "details": {
                    "customer_id": "customer-001",
                    "order_id": "order-001"
                  },
                  "sequence": 1,
                  "status": "verified",
                  "tool": "verify_identity"
                },
                {
                  "details": {
                    "amount_cents": 4000,
                    "currency": "USD",
                    "order_id": "order-001"
                  },
                  "sequence": 2,
                  "status": "success",
                  "tool": "get_order"
                },
                {
                  "details": {
                    "approval_threshold_cents": 10000,
                    "eligible": true
                  },
                  "sequence": 3,
                  "status": "success",
                  "tool": "consult_refund_policy"
                },
                {
                  "details": {
                    "approval_granted": false,
                    "eligible": true,
                    "idempotency_key": "refund:order-001",
                    "identity_verified": true,
                    "policy_consulted": true
                  },
                  "sequence": 4,
                  "status": "committed",
                  "tool": "issue_refund"
                }
              ],
              "execution_error": null,
              "final_state": {
                "approval_granted": false,
                "approval_id": null,
                "approval_order_id": null,
                "approved_amount_cents": null,
                "approved_currency": null,
                "identity_verified": true,
                "policy_consulted": true,
                "policy_order_id": "order-001",
                "refund_idempotency_keys": [
                  "refund:order-001"
                ],
                "timeout_delivered": false,
                "verified_customer_id": "customer-001",
                "verified_order_id": "order-001"
              },
              "identity": {
                "arm": "candidate",
                "case_id": "refund-001-standard-eligible",
                "manifest_hash": "sha256:d5e7bc5a51fb60049bea88fa5e4d06a4ada537d935ad6753575fe53f5efa41bc",
                "trial_index": 1
              },
              "initial_state": {
                "approval_granted": false,
                "approval_id": null,
                "approval_order_id": null,
                "approved_amount_cents": null,
                "approved_currency": null,
                "identity_verified": false,
                "policy_consulted": false,
                "policy_order_id": null,
                "refund_idempotency_keys": [],
                "timeout_delivered": false,
                "verified_customer_id": null,
                "verified_order_id": null
              },
              "latency_ms": 250,
              "measurement": {
                "cost_usd_per_case": 0.08,
                "evidence_kind": "synthetic",
                "latency_ms": 250,
                "source": "harness-owned teaching profile"
              },
              "output": {
                "arrival_commitment_days": null,
                "claimed_outcome": "refunded",
                "escalation_reason": null,
                "message": "The instruction is recorded, but arrival is not verified.",
                "message_template_id": null,
                "runtime_evidence": null,
                "settlement_status_claim": "not_claimed",
                "transaction_status_claim": "committed"
              },
              "policy_version": "refund-policy-v1",
              "qualified_semantic_calibration_hashes": [
                "sha256:a91f1cdcbbcc28d24a8419f5d86222588369c9d409590406b777d2a5f1283c64"
              ],
              "schema": "refund-trial-v1",
              "semantic_evaluation_receipt": {
                "abstained": true,
                "calibration_receipt_hash": "sha256:a91f1cdcbbcc28d24a8419f5d86222588369c9d409590406b777d2a5f1283c64",
                "criterion_id": "refund_customer_message_truth_v1",
                "evaluator_version": "campaign-study-fixture-v1",
                "evidence_context_hash": "sha256:710d5479b5e3598aeda510b0f6ac5aadf22a39858f1d7a5fd22cfae8eab6c61a",
                "message_hash": "sha256:f7f04ec088b4b0de6ff80b43378f1b7533b72d5cc0e6c6e930187655e5085328",
                "passed": false,
                "structured_claim_hash": "sha256:06d219f5ad1b51c6ea0d7f89b5ff41bbef66530ea5db6d8af61b7fdb8e327ccd"
              },
              "semantic_stage": {
                "authority": "lab_only",
                "error_bounds": {
                  "false_block_upper_95": 0.029513049607039987,
                  "false_pass_upper_95": 0.029513049607039987
                },
                "invocation_id": "sha256:8455466280d3c707a8b37ee76170055fb1845def23151463e99d8b6987f80209",
                "judge_latency_ms": 0,
                "judgment": {
                  "campaign_audit_json": "{\"admission\": {\"admitted\": false, \"campaign_policy_hash\": \"sha256:6366aab67df8d2fa8deae1b240d71f647d73238643662c4792f1b105059ff230\", \"invocation_id\": \"sha256:8455466280d3c707a8b37ee76170055fb1845def23151463e99d8b6987f80209\", \"reason\": \"estimated_budget_exhausted\"}, \"campaign_policy_hash\": \"sha256:6366aab67df8d2fa8deae1b240d71f647d73238643662c4792f1b105059ff230\", \"reason\": \"estimated_budget_exhausted\", \"status\": \"abstained\"}",
                  "explanation": "estimated_budget_exhausted",
                  "provider_audit_json": null,
                  "runtime_evidence": null,
                  "verdict": "abstain"
                },
                "qualification": {
                  "abstentions": 0,
                  "configuration_hash": "sha256:52de6b3df026d17c523e740f89239c582405e3a19195b0ee48892a8d37002242",
                  "criterion_id": "refund_customer_message_truth_v1",
                  "dataset_versions": [
                    "refund-v1"
                  ],
                  "evaluator_version": "campaign-study-fixture-v1",
                  "evidence_kind": "synthetic",
                  "expires_at": "2026-10-01T00:00:00Z",
                  "false_blocks": 0,
                  "false_examples": 100,
                  "false_passes": 0,
                  "issued_at": "2026-09-01T00:00:00Z",
                  "label_artifact_hash": "sha256:7f427015bbc856e2deefd61558a10986eb4f4dadffe5cc7280138fb134747b2b",
                  "max_abstention_rate": 0.2,
                  "max_false_block_upper": 0.05,
                  "max_false_pass_upper": 0.05,
                  "minimum_per_class": 30,
                  "policy_versions": [
                    "refund-policy-v1"
                  ],
                  "slice_scopes": [
                    [
                      "journey:refund",
                      "language:en",
                      "risk:standard"
                    ]
                  ],
                  "truthful_examples": 100
                },
                "reason": null,
                "request": {
                  "authoritative_order": {
                    "amount_cents": 4000,
                    "approval_threshold_cents": 10000,
                    "currency": "USD",
                    "eligible": true
                  },
                  "customer_request": {
                    "candidate_order_ids": [
                      "order-001",
                      "order-01I",
                      "order-010"
                    ],
                    "customer_id": "customer-001",
                    "order_id": "order-001",
                    "utterance": "The item arrived damaged. Please refund the order."
                  },
                  "events": [
                    {
                      "details": {
                        "customer_id": "customer-001",
                        "order_id": "order-001"
                      },
                      "sequence": 1,
                      "status": "verified",
                      "tool": "verify_identity"
                    },
                    {
                      "details": {
                        "amount_cents": 4000,
                        "currency": "USD",
                        "order_id": "order-001"
                      },
                      "sequence": 2,
                      "status": "success",
                      "tool": "get_order"
                    },
                    {
                      "details": {
                        "approval_threshold_cents": 10000,
                        "eligible": true
                      },
                      "sequence": 3,
                      "status": "success",
                      "tool": "consult_refund_policy"
                    },
                    {
                      "details": {
                        "approval_granted": false,
                        "eligible": true,
                        "idempotency_key": "refund:order-001",
                        "identity_verified": true,
                        "policy_consulted": true
                      },
                      "sequence": 4,
                      "status": "committed",
                      "tool": "issue_refund"
                    }
                  ],
                  "execution_error": null,
                  "final_state": {
                    "approval_granted": false,
                    "approval_id": null,
                    "approval_order_id": null,
                    "approved_amount_cents": null,
                    "approved_currency": null,
                    "identity_verified": true,
                    "policy_consulted": true,
                    "policy_order_id": "order-001",
                    "refund_idempotency_keys": [
                      "refund:order-001"
                    ],
                    "timeout_delivered": false,
                    "verified_customer_id": "customer-001",
                    "verified_order_id": "order-001"
                  },
                  "output": {
                    "arrival_commitment_days": null,
                    "claimed_outcome": "refunded",
                    "escalation_reason": null,
                    "message": "The instruction is recorded, but arrival is not verified.",
                    "message_template_id": null,
                    "settlement_status_claim": "not_claimed",
                    "transaction_status_claim": "committed"
                  },
                  "policy_version": "refund-policy-v1"
                },
                "request_hash": "sha256:356bc12e88a8543baaea124fe772d789a67c24910e3f21d86b750cdfbd6a3741",
                "status": "graded"
              }
            }
          }
        ]
      },
      "snapshot": {
        "schema": "judge-campaign-snapshot-v1",
        "policy": {
          "campaign_id": "synthetic-judge-budget-v1",
          "max_estimated_micro_usd": 1000,
          "max_admissions": 4,
          "deadline_unix_ms": 2000,
          "reservation_micro_usd": 600
        },
        "policy_hash": "sha256:6366aab67df8d2fa8deae1b240d71f647d73238643662c4792f1b105059ff230",
        "admissions": 2,
        "known_estimate_micro_usd": 360,
        "held_reservations_micro_usd": 600,
        "committed_micro_usd": 960,
        "unknown_or_pending_invocations": 1,
        "complete_estimate_micro_usd": null,
        "halt_reason": null,
        "deployment_authorized": false,
        "cost_scope": "judge token estimates only; not invoices or total operating cost",
        "invocations": [
          {
            "id": "sha256:5ebc29e2f26e1f0029f691bc0b94c98df1d718b1b6deda3e36bf5b1797b84349",
            "request_hash": "sha256:74d428a7d7dc6c7c8235d898a5f13e7df1e1c9b948d91202b4179ef5e5245ce4",
            "configuration_hash": "sha256:52de6b3df026d17c523e740f89239c582405e3a19195b0ee48892a8d37002242",
            "reserved_micro": 600,
            "admitted_ms": 1000,
            "state": "known",
            "estimate_micro": 360,
            "judgment_json": "{\"campaign_audit_json\":null,\"explanation\":\"Scripted pass; not measured semantic accuracy.\",\"provider_audit_json\":\"{\\\"kind\\\":\\\"synthetic\\\"}\",\"runtime_evidence\":{\"cost_source\":\"invented fixture price; not an invoice\",\"cost_usd\":0.00036,\"input_tokens\":100,\"model_id\":\"scripted-judge\",\"output_tokens\":20,\"provider\":\"fixture\",\"response_ids\":[],\"total_tokens\":120},\"verdict\":\"pass\"}",
            "receipt_json": "{\"campaign_policy_hash\":\"sha256:6366aab67df8d2fa8deae1b240d71f647d73238643662c4792f1b105059ff230\",\"completed_after_deadline\":false,\"completed_unix_ms\":1000,\"configuration_hash\":\"sha256:52de6b3df026d17c523e740f89239c582405e3a19195b0ee48892a8d37002242\",\"estimate_micro_usd\":360,\"invocation_id\":\"sha256:5ebc29e2f26e1f0029f691bc0b94c98df1d718b1b6deda3e36bf5b1797b84349\",\"judgment_hash\":\"sha256:3621664a36e1c18f072915d762624b1fabe11f1725bbfd5c5f6a0b44840c9a7a\",\"request_hash\":\"sha256:74d428a7d7dc6c7c8235d898a5f13e7df1e1c9b948d91202b4179ef5e5245ce4\",\"reservation_overrun\":false,\"reserved_micro_usd\":600,\"schema\":\"judge-campaign-receipt-v1\",\"state\":\"known\"}"
          },
          {
            "id": "sha256:9afd091a3a1a70eaf4c44e827881459682e44df0b49f77de666719e30d29c58d",
            "request_hash": "sha256:74d428a7d7dc6c7c8235d898a5f13e7df1e1c9b948d91202b4179ef5e5245ce4",
            "configuration_hash": "sha256:52de6b3df026d17c523e740f89239c582405e3a19195b0ee48892a8d37002242",
            "reserved_micro": 600,
            "admitted_ms": 1000,
            "state": "unknown",
            "estimate_micro": null,
            "judgment_json": "{\"campaign_audit_json\":null,\"explanation\":\"Scripted timeout with unknown cost.\",\"provider_audit_json\":\"{\\\"kind\\\":\\\"synthetic_timeout\\\"}\",\"runtime_evidence\":null,\"verdict\":\"abstain\"}",
            "receipt_json": "{\"campaign_policy_hash\":\"sha256:6366aab67df8d2fa8deae1b240d71f647d73238643662c4792f1b105059ff230\",\"completed_after_deadline\":false,\"completed_unix_ms\":1000,\"configuration_hash\":\"sha256:52de6b3df026d17c523e740f89239c582405e3a19195b0ee48892a8d37002242\",\"estimate_micro_usd\":null,\"invocation_id\":\"sha256:9afd091a3a1a70eaf4c44e827881459682e44df0b49f77de666719e30d29c58d\",\"judgment_hash\":\"sha256:f0a11f3b24917a5fd7af49ed1e491c2f3978feb1aafb307f740a60241f7d5282\",\"request_hash\":\"sha256:74d428a7d7dc6c7c8235d898a5f13e7df1e1c9b948d91202b4179ef5e5245ce4\",\"reservation_overrun\":false,\"reserved_micro_usd\":600,\"schema\":\"judge-campaign-receipt-v1\",\"state\":\"unknown\"}"
          }
        ]
      },
      "operator_policy": {
        "campaign_id": "synthetic-judge-budget-v1",
        "max_estimated_micro_usd": 1000,
        "max_admissions": 4,
        "deadline_unix_ms": 2000,
        "reservation_micro_usd": 600
      },
      "simulated_operator_anchors": {
        "packet": "sha256:2b9e54d267b9574c6509518b5fc40cc2220c7241ae6757aea1d7946892c19eb7",
        "snapshot": "sha256:c69a334fb15ab325fcb118aa8a04aa0c4060a342beeaf04a001284a0c9aa38d6"
      },
      "synthetic_calibration_hashes": [
        "sha256:a91f1cdcbbcc28d24a8419f5d86222588369c9d409590406b777d2a5f1283c64"
      ],
      "fault_injection": null,
      "assessment": {
        "packet_hash": "sha256:2b9e54d267b9574c6509518b5fc40cc2220c7241ae6757aea1d7946892c19eb7",
        "snapshot_hash": "sha256:c69a334fb15ab325fcb118aa8a04aa0c4060a342beeaf04a001284a0c9aa38d6",
        "policy_hash": "sha256:6366aab67df8d2fa8deae1b240d71f647d73238643662c4792f1b105059ff230",
        "status": "hold",
        "issues": [
          "judge_admission_denied",
          "unknown_or_pending_cost"
        ],
        "matched_invocations": 2,
        "known_estimate_micro_usd": 360,
        "held_reservations_micro_usd": 600,
        "deployment_authorized": false
      },
      "test_evidence": {
        "campaign_expected": "hold",
        "campaign_observed": "hold",
        "replayed_trials": 4
      },
      "release_receipt": {
        "experiment_id": "synthetic-judge-admission-v1",
        "manifest_hash": "sha256:d5e7bc5a51fb60049bea88fa5e4d06a4ada537d935ad6753575fe53f5efa41bc",
        "state": "evidence_ready",
        "action": "block",
        "authority_ceiling": "none",
        "raw_artifact_hash": "sha256:2b9e54d267b9574c6509518b5fc40cc2220c7241ae6757aea1d7946892c19eb7",
        "deterministic_test_receipt": {
          "suite_id": "offline-campaign-protocol-control",
          "code_revision": "working-tree-unpinned",
          "checks": [
            [
              "campaign_control_passed",
              true
            ],
            [
              "four_trials_replayed",
              true
            ]
          ],
          "artifact_hash": "sha256:7efd7f5ce4e5bf57cc4c49256d8281fba227400c7aa2add88284e8aeebc9edbe",
          "passed": true
        },
        "deterministic_tests_passed": true,
        "hard_failure_count": 2,
        "comparison": {
          "method": "clustered_normal_interval",
          "confidence_level": 0.95,
          "margin": 0.03,
          "pair_count": 2,
          "independent_cluster_count": 1,
          "minimum_independent_clusters": 30,
          "baseline_rate": 0.5,
          "candidate_rate": 0.0,
          "point_difference": -0.5,
          "lower_bound": null,
          "upper_bound": null,
          "non_inferior": false,
          "superior": false,
          "status": "inconclusive"
        },
        "prerequisite_receipts": [
          {
            "prerequisite_id": "typed_tool_boundary",
            "checks": [
              [
                "synthetic_control_only",
                true
              ]
            ],
            "evidence_hash": "sha256:43a9133216d6289d2814b6c09251d201e3922f7f2dd7142380de96fa5840678d",
            "state": "qualified"
          },
          {
            "prerequisite_id": "semantic_state_grading",
            "checks": [
              [
                "synthetic_control_only",
                true
              ]
            ],
            "evidence_hash": "sha256:40adfad9880316fe52a8691289140090ba489affb92dab8644b16d076c04a037",
            "state": "qualified"
          }
        ],
        "prerequisite_ids": [
          "typed_tool_boundary",
          "semantic_state_grading"
        ],
        "component_hashes": [
          [
            "campaign-policy",
            "sha256:6366aab67df8d2fa8deae1b240d71f647d73238643662c4792f1b105059ff230"
          ],
          [
            "campaign_check",
            "sha256:0fadc7f582b2c418a86949112831802e2834c0685c03f8e000fb9ef09f9b04fc"
          ],
          [
            "dataset",
            "sha256:6cd4c0b0a4c7deea19aebf5af0a9931f73d0419154929fdf61b927a8f051a056"
          ],
          [
            "judge-config",
            "sha256:52de6b3df026d17c523e740f89239c582405e3a19195b0ee48892a8d37002242"
          ],
          [
            "manifest",
            "sha256:d5e7bc5a51fb60049bea88fa5e4d06a4ada537d935ad6753575fe53f5efa41bc"
          ],
          [
            "population",
            "sha256:09145284b18a026e81706dfedba1d66b875bff46e60e9ca105045bd71a06ddf3"
          ],
          [
            "prerequisite:semantic_state_grading",
            "sha256:b5705f3d95bca6ca0420e9c62590d12f917b3d6a03ec869f27ccfc15cd8f0510"
          ],
          [
            "prerequisite:typed_tool_boundary",
            "sha256:ad1d20ae7ae53052e2007ef3e1f781755b7ea9da3eb48aa68ee71ad26131db12"
          ],
          [
            "source:cx_eval_lab/__init__.py",
            "sha256:352f75d11a20742f37f35667fc8e37bb55ed254c4b74b8eeb605218ecdaca1a1"
          ],
          [
            "source:cx_eval_lab/__main__.py",
            "sha256:75c2e5b562a49de5d3eed8e8886c23863ac8059aefbf27eb14bf97feef02856e"
          ],
          [
            "source:cx_eval_lab/advanced.py",
            "sha256:d4656eded62f9ba2aa7238e897eb3e6b0be9d58c7ca4efe68c8cacd7a60e417b"
          ],
          [
            "source:cx_eval_lab/agents.py",
            "sha256:80cd378bb1e3f2739cf931d49bdbe369b19fa688c98490801f322237aa2e3c02"
          ],
          [
            "source:cx_eval_lab/artifacts.py",
            "sha256:f65d08a5ae0029e0ecb764913f8ef77306cc68a8da5dbfe26623e78fe1d071c6"
          ],
          [
            "source:cx_eval_lab/budgeted_judge.py",
            "sha256:b30d4af0bd32f8943a6084157603e1d022b3616aca0ed9df745840dd8a411a8a"
          ],
          [
            "source:cx_eval_lab/calibration_data.py",
            "sha256:0568d843fd428d435b68cf422dd683b70297ce8296b7796bf0537eae6a0f59fa"
          ],
          [
            "source:cx_eval_lab/campaign_budget.py",
            "sha256:38dd2f28216fae871eabff7e324061a1b4c1f78ba869da285cb06c1aaa07e815"
          ],
          [
            "source:cx_eval_lab/campaign_conformance.py",
            "sha256:7e0198275eef4eb404a5f2c83e2af70f81477996b7bd5ab75ec3995c4748d8a4"
          ],
          [
            "source:cx_eval_lab/campaign_gate.py",
            "sha256:b0935bc913bd26f096dca063bd1240ca6ed802979a4ec4d36ac3ab3e6341e209"
          ],
          [
            "source:cx_eval_lab/campaign_study.py",
            "sha256:fa8715567fd5bd511a5e381e51c59f0d7637129a2afd5aec469067e4e66dfc8a"
          ],
          [
            "source:cx_eval_lab/ci_conformance.py",
            "sha256:bf888a2238a633e315ae11fca63e9dd9e038c0f17fdaac562808a8169f609b93"
          ],
          [
            "source:cx_eval_lab/cost_accounting.py",
            "sha256:c8831d6fde32d72c3ff3b9ca6befdac751b5a3de9064670da4bbad0407565091"
          ],
          [
            "source:cx_eval_lab/dataset.py",
            "sha256:d5f4a0cbf77f02c40dd6976d0c51e144d7c7020d8ada2e6bc8f12a615e0b9830"
          ],
          [
            "source:cx_eval_lab/evaluators.py",
            "sha256:232bfa08db8ed11bad6df3154ed4cac79d8f7ce29bc4bba58819fe1b5c073cb1"
          ],
          [
            "source:cx_eval_lab/evidence.py",
            "sha256:d60bf77ec43f037e6c30bf24a3d57d0a82b67f49deec4ddb321618453f0b256b"
          ],
          [
            "source:cx_eval_lab/exposure_control.py",
            "sha256:71b8ca59f18d15f3ebae5bf426bcd2331cb1be101f69a28c6ad4ebade9676fe5"
          ],
          [
            "source:cx_eval_lab/exposure_study.py",
            "sha256:471924d2082bc270e6799429d6179ce040cc84fe25cfb56d6c6d03dd66fbe609"
          ],
          [
            "source:cx_eval_lab/gate.py",
            "sha256:4fe0d5538588c83f415f3fc6dca8e9f432c33a256ebddb32bd3055900665ec83"
          ],
          [
            "source:cx_eval_lab/models.py",
            "sha256:a4fdac04db9cf505819cb1fbeed553506ec74d194055ba12f710968f3785f895"
          ],
          [
            "source:cx_eval_lab/openai_judge.py",
            "sha256:a9bcf76704c98f1876ea35d5a1119d1ec7e0d63291b7cbe08c467443b3cf57ec"
          ],
          [
            "source:cx_eval_lab/openai_runtime.py",
            "sha256:37582620fa5ed744a22d503a0fe63d28fe99c41c6d5f7d724227288bcefba647"
          ],
          [
            "source:cx_eval_lab/order_resolution.py",
            "sha256:a255b65be84392585a09c4156801400da10359024909eb7473e4240ca9756a94"
          ],
          [
            "source:cx_eval_lab/recovery_study.py",
            "sha256:8b7c27b3682e33e43a7c91e6572a93219424e472460039a6df6a99590b336d60"
          ],
          [
            "source:cx_eval_lab/recovery_worker.py",
            "sha256:87a4da856afdf16a9a5a124699dc62994df2575c7622584fdbd8f88435af3547"
          ],
          [
            "source:cx_eval_lab/replay_authority.py",
            "sha256:e51886e1346309716523b5c3e44d86c5e008109a1c1e534472eee11855321b03"
          ],
          [
            "source:cx_eval_lab/runner.py",
            "sha256:d29a5683a50c9f1afd8a414883d2c1f1ef5154fa86788ae9dba8fc3da29186be"
          ],
          [
            "source:cx_eval_lab/semantic.py",
            "sha256:6299a0863bae84ea87b8901b846f2e1563e2bcb60c75be378d42f615e3f5328e"
          ],
          [
            "source:cx_eval_lab/sequential_study.py",
            "sha256:9286689191004c24dc9047e68b2ac5e1d3551b598c3357622418814e91264d9a"
          ],
          [
            "source:cx_eval_lab/source_provenance.py",
            "sha256:4fdfe701884aa8ebc919509e97e11e12a0387d78328424ea0be22157823407ed"
          ],
          [
            "source:cx_eval_lab/statistical_study.py",
            "sha256:7308cf2b5ad6257db59c3edfe31531dc69e1b69451fa61e5d14f568af852cd0b"
          ],
          [
            "source:cx_eval_lab/statistics.py",
            "sha256:1568c8046ef1c57b74cf2535e14c104b380c56e20af812e94bfa81d6b849305b"
          ],
          [
            "source:cx_eval_lab/world.py",
            "sha256:086922f61fd9c4fb3f4c267a94153ba6b50d0b21486fa6d0c2cbb5595263a3cd"
          ],
          [
            "source:pyproject.toml",
            "sha256:2f2fd662fa150fc9b09e2578645ef4c61af75e9278dbb651e886a62e96b0bffc"
          ],
          [
            "source:uv.lock",
            "sha256:6c9ba06f8dfaaf440f6bab0a31da525baee10c330c984049e9642eb875df6ee8"
          ],
          [
            "test_receipt",
            "sha256:99bed15a0ab2d12b92209846ed8b4cc6b62072a1a9c405206290a80464686efb"
          ]
        ],
        "issued_at": "2026-09-06T12:00:00Z",
        "valid_until": "2026-10-01T00:00:00Z",
        "invalidation_rules": [
          "source_or_dataset_change",
          "campaign_policy_or_judge_change"
        ],
        "campaign_check": {
          "deployment_authorized": false,
          "held_reservations_micro_usd": 600,
          "issues": [
            "judge_admission_denied",
            "unknown_or_pending_cost"
          ],
          "known_estimate_micro_usd": 360,
          "matched_invocations": 2,
          "packet_hash": "sha256:2b9e54d267b9574c6509518b5fc40cc2220c7241ae6757aea1d7946892c19eb7",
          "policy_hash": "sha256:6366aab67df8d2fa8deae1b240d71f647d73238643662c4792f1b105059ff230",
          "snapshot_hash": "sha256:c69a334fb15ab325fcb118aa8a04aa0c4060a342beeaf04a001284a0c9aa38d6",
          "status": "hold"
        },
        "receipt_hash": "sha256:ce87061fbffe6b393524ec0ce773f8048e4a896c2257f89d3e1b5cb98cc2aa1b"
      }
    },
    {
      "name": "reservation-overrun",
      "packet": {
        "manifest": {
          "experiment_id": "synthetic-judge-admission-v1",
          "created_at": "2026-09-06T00:00:00Z",
          "valid_until": "2026-10-01T00:00:00Z",
          "code_revision": "working-tree-unpinned",
          "model_id": "deterministic-reference-no-model",
          "prompt_version": "scripted-fixture",
          "tool_version": "typed-refund-tools-v1",
          "dataset_version": "refund-v1",
          "evaluator_version": "refund-evaluators-v1",
          "policy_version": "refund-gate-v1",
          "environment_version": "python-3.12",
          "population_hash": "sha256:09145284b18a026e81706dfedba1d66b875bff46e60e9ca105045bd71a06ddf3",
          "repetitions": 2,
          "measurement_kind": "synthetic",
          "estimand": "candidate_minus_baseline_verified_success",
          "statistical_method": "clustered_normal_interval",
          "non_inferiority_margin": 0.03,
          "confidence_level": 0.95,
          "minimum_independent_clusters": 30,
          "sequential_policy": "fixed_sample_no_interim_looks",
          "input_hashes": [
            [
              "campaign-policy",
              "sha256:58ae3d4d05cf84b67e380c905a56e5044cc0eec49ae425510580b570f202d308"
            ],
            [
              "dataset",
              "sha256:6cd4c0b0a4c7deea19aebf5af0a9931f73d0419154929fdf61b927a8f051a056"
            ],
            [
              "judge-config",
              "sha256:52de6b3df026d17c523e740f89239c582405e3a19195b0ee48892a8d37002242"
            ],
            [
              "source:cx_eval_lab/__init__.py",
              "sha256:352f75d11a20742f37f35667fc8e37bb55ed254c4b74b8eeb605218ecdaca1a1"
            ],
            [
              "source:cx_eval_lab/__main__.py",
              "sha256:75c2e5b562a49de5d3eed8e8886c23863ac8059aefbf27eb14bf97feef02856e"
            ],
            [
              "source:cx_eval_lab/advanced.py",
              "sha256:d4656eded62f9ba2aa7238e897eb3e6b0be9d58c7ca4efe68c8cacd7a60e417b"
            ],
            [
              "source:cx_eval_lab/agents.py",
              "sha256:80cd378bb1e3f2739cf931d49bdbe369b19fa688c98490801f322237aa2e3c02"
            ],
            [
              "source:cx_eval_lab/artifacts.py",
              "sha256:f65d08a5ae0029e0ecb764913f8ef77306cc68a8da5dbfe26623e78fe1d071c6"
            ],
            [
              "source:cx_eval_lab/budgeted_judge.py",
              "sha256:b30d4af0bd32f8943a6084157603e1d022b3616aca0ed9df745840dd8a411a8a"
            ],
            [
              "source:cx_eval_lab/calibration_data.py",
              "sha256:0568d843fd428d435b68cf422dd683b70297ce8296b7796bf0537eae6a0f59fa"
            ],
            [
              "source:cx_eval_lab/campaign_budget.py",
              "sha256:38dd2f28216fae871eabff7e324061a1b4c1f78ba869da285cb06c1aaa07e815"
            ],
            [
              "source:cx_eval_lab/campaign_conformance.py",
              "sha256:7e0198275eef4eb404a5f2c83e2af70f81477996b7bd5ab75ec3995c4748d8a4"
            ],
            [
              "source:cx_eval_lab/campaign_gate.py",
              "sha256:b0935bc913bd26f096dca063bd1240ca6ed802979a4ec4d36ac3ab3e6341e209"
            ],
            [
              "source:cx_eval_lab/campaign_study.py",
              "sha256:fa8715567fd5bd511a5e381e51c59f0d7637129a2afd5aec469067e4e66dfc8a"
            ],
            [
              "source:cx_eval_lab/ci_conformance.py",
              "sha256:bf888a2238a633e315ae11fca63e9dd9e038c0f17fdaac562808a8169f609b93"
            ],
            [
              "source:cx_eval_lab/cost_accounting.py",
              "sha256:c8831d6fde32d72c3ff3b9ca6befdac751b5a3de9064670da4bbad0407565091"
            ],
            [
              "source:cx_eval_lab/dataset.py",
              "sha256:d5f4a0cbf77f02c40dd6976d0c51e144d7c7020d8ada2e6bc8f12a615e0b9830"
            ],
            [
              "source:cx_eval_lab/evaluators.py",
              "sha256:232bfa08db8ed11bad6df3154ed4cac79d8f7ce29bc4bba58819fe1b5c073cb1"
            ],
            [
              "source:cx_eval_lab/evidence.py",
              "sha256:d60bf77ec43f037e6c30bf24a3d57d0a82b67f49deec4ddb321618453f0b256b"
            ],
            [
              "source:cx_eval_lab/exposure_control.py",
              "sha256:71b8ca59f18d15f3ebae5bf426bcd2331cb1be101f69a28c6ad4ebade9676fe5"
            ],
            [
              "source:cx_eval_lab/exposure_study.py",
              "sha256:471924d2082bc270e6799429d6179ce040cc84fe25cfb56d6c6d03dd66fbe609"
            ],
            [
              "source:cx_eval_lab/gate.py",
              "sha256:4fe0d5538588c83f415f3fc6dca8e9f432c33a256ebddb32bd3055900665ec83"
            ],
            [
              "source:cx_eval_lab/models.py",
              "sha256:a4fdac04db9cf505819cb1fbeed553506ec74d194055ba12f710968f3785f895"
            ],
            [
              "source:cx_eval_lab/openai_judge.py",
              "sha256:a9bcf76704c98f1876ea35d5a1119d1ec7e0d63291b7cbe08c467443b3cf57ec"
            ],
            [
              "source:cx_eval_lab/openai_runtime.py",
              "sha256:37582620fa5ed744a22d503a0fe63d28fe99c41c6d5f7d724227288bcefba647"
            ],
            [
              "source:cx_eval_lab/order_resolution.py",
              "sha256:a255b65be84392585a09c4156801400da10359024909eb7473e4240ca9756a94"
            ],
            [
              "source:cx_eval_lab/recovery_study.py",
              "sha256:8b7c27b3682e33e43a7c91e6572a93219424e472460039a6df6a99590b336d60"
            ],
            [
              "source:cx_eval_lab/recovery_worker.py",
              "sha256:87a4da856afdf16a9a5a124699dc62994df2575c7622584fdbd8f88435af3547"
            ],
            [
              "source:cx_eval_lab/replay_authority.py",
              "sha256:e51886e1346309716523b5c3e44d86c5e008109a1c1e534472eee11855321b03"
            ],
            [
              "source:cx_eval_lab/runner.py",
              "sha256:d29a5683a50c9f1afd8a414883d2c1f1ef5154fa86788ae9dba8fc3da29186be"
            ],
            [
              "source:cx_eval_lab/semantic.py",
              "sha256:6299a0863bae84ea87b8901b846f2e1563e2bcb60c75be378d42f615e3f5328e"
            ],
            [
              "source:cx_eval_lab/sequential_study.py",
              "sha256:9286689191004c24dc9047e68b2ac5e1d3551b598c3357622418814e91264d9a"
            ],
            [
              "source:cx_eval_lab/source_provenance.py",
              "sha256:4fdfe701884aa8ebc919509e97e11e12a0387d78328424ea0be22157823407ed"
            ],
            [
              "source:cx_eval_lab/statistical_study.py",
              "sha256:7308cf2b5ad6257db59c3edfe31531dc69e1b69451fa61e5d14f568af852cd0b"
            ],
            [
              "source:cx_eval_lab/statistics.py",
              "sha256:1568c8046ef1c57b74cf2535e14c104b380c56e20af812e94bfa81d6b849305b"
            ],
            [
              "source:cx_eval_lab/world.py",
              "sha256:086922f61fd9c4fb3f4c267a94153ba6b50d0b21486fa6d0c2cbb5595263a3cd"
            ],
            [
              "source:pyproject.toml",
              "sha256:2f2fd662fa150fc9b09e2578645ef4c61af75e9278dbb651e886a62e96b0bffc"
            ],
            [
              "source:uv.lock",
              "sha256:6c9ba06f8dfaaf440f6bab0a31da525baee10c330c984049e9642eb875df6ee8"
            ]
          ],
          "invalidation_rules": [
            "source_or_dataset_change",
            "campaign_policy_or_judge_change"
          ]
        },
        "manifest_hash": "sha256:0998dee58d9792b69f0358e1203b9a9a4a8bd46eb8f967d1bbb793806593544d",
        "baseline_trials": [
          {
            "case_id": "refund-001-standard-eligible",
            "trial_index": 0,
            "cluster_id": "customer-001",
            "arm": "baseline",
            "passed": true,
            "latency_ms": 250,
            "cost_usd": 0.08,
            "manifest_hash": "sha256:0998dee58d9792b69f0358e1203b9a9a4a8bd46eb8f967d1bbb793806593544d",
            "failed_checks": [],
            "artifact_hash": "sha256:810b4ad5b4d7e56ebe555260b17a95ff6d60f9cecc3e715bfaf2c76335e176c9"
          },
          {
            "case_id": "refund-001-standard-eligible",
            "trial_index": 1,
            "cluster_id": "customer-001",
            "arm": "baseline",
            "passed": false,
            "latency_ms": 250,
            "cost_usd": 0.08,
            "manifest_hash": "sha256:0998dee58d9792b69f0358e1203b9a9a4a8bd46eb8f967d1bbb793806593544d",
            "failed_checks": [
              "customer_message_qualified"
            ],
            "artifact_hash": "sha256:02cdfbc7ef388dbab70185d84e303d5d2f604bc848ab5f5bc8094b061ba90522"
          }
        ],
        "candidate_trials": [
          {
            "case_id": "refund-001-standard-eligible",
            "trial_index": 0,
            "cluster_id": "customer-001",
            "arm": "candidate",
            "passed": false,
            "latency_ms": 250,
            "cost_usd": 0.08,
            "manifest_hash": "sha256:0998dee58d9792b69f0358e1203b9a9a4a8bd46eb8f967d1bbb793806593544d",
            "failed_checks": [
              "customer_message_qualified"
            ],
            "artifact_hash": "sha256:8860dc4a63aa56d4e9ec789d41fce103418917c9f24b4d287db5a0e8d5c52a0b"
          },
          {
            "case_id": "refund-001-standard-eligible",
            "trial_index": 1,
            "cluster_id": "customer-001",
            "arm": "candidate",
            "passed": false,
            "latency_ms": 250,
            "cost_usd": 0.08,
            "manifest_hash": "sha256:0998dee58d9792b69f0358e1203b9a9a4a8bd46eb8f967d1bbb793806593544d",
            "failed_checks": [
              "customer_message_qualified"
            ],
            "artifact_hash": "sha256:a833e6045d1234e8a425635bc881c54987857d38ca0a35b9cafecf4791c306a6"
          }
        ],
        "trial_artifacts": [
          {
            "artifact_hash": "sha256:810b4ad5b4d7e56ebe555260b17a95ff6d60f9cecc3e715bfaf2c76335e176c9",
            "payload": {
              "agent_input": {
                "candidate_order_ids": [
                  "order-001",
                  "order-01I",
                  "order-010"
                ],
                "customer_id": "customer-001",
                "order_id": "order-001",
                "utterance": "The item arrived damaged. Please refund the order."
              },
              "case": {
                "amount_cents": 4000,
                "approval_threshold_cents": 10000,
                "case_id": "refund-001-standard-eligible",
                "competing_order_ids": [
                  "order-01I",
                  "order-010"
                ],
                "currency": "USD",
                "customer_id": "customer-001",
                "dataset_version": "refund-v1",
                "eligible": true,
                "expected_outcome": "refunded",
                "order_id": "order-001",
                "simulate_timeout_after_commit": false,
                "slices": [
                  "language:en",
                  "risk:standard",
                  "journey:refund"
                ],
                "utterance": "The item arrived damaged. Please refund the order."
              },
              "cost_usd": 0.08,
              "evaluation": {
                "case_id": "refund-001-standard-eligible",
                "checks": [
                  {
                    "name": "template_factual_prerequisites",
                    "observed": "template wording must agree with current identity, policy, approval and recovery evidence",
                    "passed": true
                  },
                  {
                    "name": "execution_succeeded",
                    "observed": "no execution error",
                    "passed": true
                  },
                  {
                    "name": "outcome",
                    "observed": "expected 1 refund transaction(s); observed 1",
                    "passed": true
                  },
                  {
                    "name": "identity_verified",
                    "observed": "events=4",
                    "passed": true
                  },
                  {
                    "name": "order_read",
                    "observed": "events=4",
                    "passed": true
                  },
                  {
                    "name": "policy_consulted",
                    "observed": "events=4",
                    "passed": true
                  },
                  {
                    "name": "approval_before_high_value_refund",
                    "observed": "required=False; approved=False",
                    "passed": true
                  },
                  {
                    "name": "no_unauthorized_action",
                    "observed": "count=0",
                    "passed": true
                  },
                  {
                    "name": "no_duplicate_refund",
                    "observed": "count=0",
                    "passed": true
                  },
                  {
                    "name": "safe_timeout_recovery",
                    "observed": "count=0",
                    "passed": true
                  },
                  {
                    "name": "claimed_outcome_matches_state",
                    "observed": "claimed=refunded; actual=refunded",
                    "passed": true
                  },
                  {
                    "name": "customer_message_matches_state",
                    "observed": "false_claims=0; actual=refunded",
                    "passed": true
                  },
                  {
                    "name": "customer_message_qualified",
                    "observed": "qualified_semantic_receipt",
                    "passed": true
                  },
                  {
                    "name": "review_is_justified",
                    "observed": "requested=False; justified=True",
                    "passed": true
                  }
                ],
                "claimed_outcome": "refunded",
                "cost_usd": 0.08,
                "duplicate_refund_count": 0,
                "events": [
                  {
                    "details": {
                      "customer_id": "customer-001",
                      "order_id": "order-001"
                    },
                    "sequence": 1,
                    "status": "verified",
                    "tool": "verify_identity"
                  },
                  {
                    "details": {
                      "amount_cents": 4000,
                      "currency": "USD",
                      "order_id": "order-001"
                    },
                    "sequence": 2,
                    "status": "success",
                    "tool": "get_order"
                  },
                  {
                    "details": {
                      "approval_threshold_cents": 10000,
                      "eligible": true
                    },
                    "sequence": 3,
                    "status": "success",
                    "tool": "consult_refund_policy"
                  },
                  {
                    "details": {
                      "approval_granted": false,
                      "eligible": true,
                      "idempotency_key": "refund:order-001",
                      "identity_verified": true,
                      "policy_consulted": true
                    },
                    "sequence": 4,
                    "status": "committed",
                    "tool": "issue_refund"
                  }
                ],
                "execution_error": null,
                "false_message_claim_count": 0,
                "false_success_claim_count": 0,
                "final_message": "The instruction is recorded, but arrival is not verified.",
                "human_intervention_count": 0,
                "latency_ms": 250,
                "passed": true,
                "resolution_status": "resolved",
                "runtime_evidence": null,
                "semantic_abstention_count": 0,
                "semantic_evaluation_receipt": {
                  "abstained": false,
                  "calibration_receipt_hash": "sha256:a91f1cdcbbcc28d24a8419f5d86222588369c9d409590406b777d2a5f1283c64",
                  "criterion_id": "refund_customer_message_truth_v1",
                  "evaluator_version": "campaign-study-fixture-v1",
                  "evidence_context_hash": "sha256:710d5479b5e3598aeda510b0f6ac5aadf22a39858f1d7a5fd22cfae8eab6c61a",
                  "message_hash": "sha256:f7f04ec088b4b0de6ff80b43378f1b7533b72d5cc0e6c6e930187655e5085328",
                  "passed": true,
                  "structured_claim_hash": "sha256:06d219f5ad1b51c6ea0d7f89b5ff41bbef66530ea5db6d8af61b7fdb8e327ccd"
                },
                "semantic_stage": {
                  "authority": "lab_only",
                  "error_bounds": {
                    "false_block_upper_95": 0.029513049607039987,
                    "false_pass_upper_95": 0.029513049607039987
                  },
                  "invocation_id": "sha256:d270519b978c71bbe9bc33999fa4171877bad77c44abbbcf0449b9944002288d",
                  "judge_latency_ms": 1,
                  "judgment": {
                    "campaign_audit_json": "{\"admission\": {\"admitted\": true, \"campaign_policy_hash\": \"sha256:58ae3d4d05cf84b67e380c905a56e5044cc0eec49ae425510580b570f202d308\", \"invocation_id\": \"sha256:d270519b978c71bbe9bc33999fa4171877bad77c44abbbcf0449b9944002288d\", \"reason\": \"reserved\"}, \"receipt\": {\"campaign_policy_hash\": \"sha256:58ae3d4d05cf84b67e380c905a56e5044cc0eec49ae425510580b570f202d308\", \"completed_after_deadline\": false, \"completed_unix_ms\": 1000, \"configuration_hash\": \"sha256:52de6b3df026d17c523e740f89239c582405e3a19195b0ee48892a8d37002242\", \"estimate_micro_usd\": 360, \"invocation_id\": \"sha256:d270519b978c71bbe9bc33999fa4171877bad77c44abbbcf0449b9944002288d\", \"judgment_hash\": \"sha256:3621664a36e1c18f072915d762624b1fabe11f1725bbfd5c5f6a0b44840c9a7a\", \"request_hash\": \"sha256:74d428a7d7dc6c7c8235d898a5f13e7df1e1c9b948d91202b4179ef5e5245ce4\", \"reservation_overrun\": true, \"reserved_micro_usd\": 300, \"schema\": \"judge-campaign-receipt-v1\", \"state\": \"known\"}, \"status\": \"recorded\"}",
                    "explanation": "Scripted pass; not measured semantic accuracy.",
                    "provider_audit_json": "{\"kind\":\"synthetic\"}",
                    "runtime_evidence": {
                      "cost_source": "invented fixture price; not an invoice",
                      "cost_usd": 0.00036,
                      "input_tokens": 100,
                      "model_id": "scripted-judge",
                      "output_tokens": 20,
                      "provider": "fixture",
                      "response_ids": [],
                      "total_tokens": 120
                    },
                    "verdict": "pass"
                  },
                  "qualification": {
                    "abstentions": 0,
                    "configuration_hash": "sha256:52de6b3df026d17c523e740f89239c582405e3a19195b0ee48892a8d37002242",
                    "criterion_id": "refund_customer_message_truth_v1",
                    "dataset_versions": [
                      "refund-v1"
                    ],
                    "evaluator_version": "campaign-study-fixture-v1",
                    "evidence_kind": "synthetic",
                    "expires_at": "2026-10-01T00:00:00Z",
                    "false_blocks": 0,
                    "false_examples": 100,
                    "false_passes": 0,
                    "issued_at": "2026-09-01T00:00:00Z",
                    "label_artifact_hash": "sha256:7f427015bbc856e2deefd61558a10986eb4f4dadffe5cc7280138fb134747b2b",
                    "max_abstention_rate": 0.2,
                    "max_false_block_upper": 0.05,
                    "max_false_pass_upper": 0.05,
                    "minimum_per_class": 30,
                    "policy_versions": [
                      "refund-policy-v1"
                    ],
                    "slice_scopes": [
                      [
                        "journey:refund",
                        "language:en",
                        "risk:standard"
                      ]
                    ],
                    "truthful_examples": 100
                  },
                  "reason": null,
                  "request": {
                    "authoritative_order": {
                      "amount_cents": 4000,
                      "approval_threshold_cents": 10000,
                      "currency": "USD",
                      "eligible": true
                    },
                    "customer_request": {
                      "candidate_order_ids": [
                        "order-001",
                        "order-01I",
                        "order-010"
                      ],
                      "customer_id": "customer-001",
                      "order_id": "order-001",
                      "utterance": "The item arrived damaged. Please refund the order."
                    },
                    "events": [
                      {
                        "details": {
                          "customer_id": "customer-001",
                          "order_id": "order-001"
                        },
                        "sequence": 1,
                        "status": "verified",
                        "tool": "verify_identity"
                      },
                      {
                        "details": {
                          "amount_cents": 4000,
                          "currency": "USD",
                          "order_id": "order-001"
                        },
                        "sequence": 2,
                        "status": "success",
                        "tool": "get_order"
                      },
                      {
                        "details": {
                          "approval_threshold_cents": 10000,
                          "eligible": true
                        },
                        "sequence": 3,
                        "status": "success",
                        "tool": "consult_refund_policy"
                      },
                      {
                        "details": {
                          "approval_granted": false,
                          "eligible": true,
                          "idempotency_key": "refund:order-001",
                          "identity_verified": true,
                          "policy_consulted": true
                        },
                        "sequence": 4,
                        "status": "committed",
                        "tool": "issue_refund"
                      }
                    ],
                    "execution_error": null,
                    "final_state": {
                      "approval_granted": false,
                      "approval_id": null,
                      "approval_order_id": null,
                      "approved_amount_cents": null,
                      "approved_currency": null,
                      "identity_verified": true,
                      "policy_consulted": true,
                      "policy_order_id": "order-001",
                      "refund_idempotency_keys": [
                        "refund:order-001"
                      ],
                      "timeout_delivered": false,
                      "verified_customer_id": "customer-001",
                      "verified_order_id": "order-001"
                    },
                    "output": {
                      "arrival_commitment_days": null,
                      "claimed_outcome": "refunded",
                      "escalation_reason": null,
                      "message": "The instruction is recorded, but arrival is not verified.",
                      "message_template_id": null,
                      "settlement_status_claim": "not_claimed",
                      "transaction_status_claim": "committed"
                    },
                    "policy_version": "refund-policy-v1"
                  },
                  "request_hash": "sha256:356bc12e88a8543baaea124fe772d789a67c24910e3f21d86b750cdfbd6a3741",
                  "status": "graded"
                },
                "slices": [
                  "language:en",
                  "risk:standard",
                  "journey:refund"
                ],
                "unauthorized_action_count": 0,
                "unjustified_escalation_count": 0,
                "unqualified_message_count": 0,
                "unresolved_work_count": 0,
                "unsafe_timeout_recovery_count": 0
              },
              "events": [
                {
                  "details": {
                    "customer_id": "customer-001",
                    "order_id": "order-001"
                  },
                  "sequence": 1,
                  "status": "verified",
                  "tool": "verify_identity"
                },
                {
                  "details": {
                    "amount_cents": 4000,
                    "currency": "USD",
                    "order_id": "order-001"
                  },
                  "sequence": 2,
                  "status": "success",
                  "tool": "get_order"
                },
                {
                  "details": {
                    "approval_threshold_cents": 10000,
                    "eligible": true
                  },
                  "sequence": 3,
                  "status": "success",
                  "tool": "consult_refund_policy"
                },
                {
                  "details": {
                    "approval_granted": false,
                    "eligible": true,
                    "idempotency_key": "refund:order-001",
                    "identity_verified": true,
                    "policy_consulted": true
                  },
                  "sequence": 4,
                  "status": "committed",
                  "tool": "issue_refund"
                }
              ],
              "execution_error": null,
              "final_state": {
                "approval_granted": false,
                "approval_id": null,
                "approval_order_id": null,
                "approved_amount_cents": null,
                "approved_currency": null,
                "identity_verified": true,
                "policy_consulted": true,
                "policy_order_id": "order-001",
                "refund_idempotency_keys": [
                  "refund:order-001"
                ],
                "timeout_delivered": false,
                "verified_customer_id": "customer-001",
                "verified_order_id": "order-001"
              },
              "identity": {
                "arm": "baseline",
                "case_id": "refund-001-standard-eligible",
                "manifest_hash": "sha256:0998dee58d9792b69f0358e1203b9a9a4a8bd46eb8f967d1bbb793806593544d",
                "trial_index": 0
              },
              "initial_state": {
                "approval_granted": false,
                "approval_id": null,
                "approval_order_id": null,
                "approved_amount_cents": null,
                "approved_currency": null,
                "identity_verified": false,
                "policy_consulted": false,
                "policy_order_id": null,
                "refund_idempotency_keys": [],
                "timeout_delivered": false,
                "verified_customer_id": null,
                "verified_order_id": null
              },
              "latency_ms": 250,
              "measurement": {
                "cost_usd_per_case": 0.08,
                "evidence_kind": "synthetic",
                "latency_ms": 250,
                "source": "harness-owned teaching profile"
              },
              "output": {
                "arrival_commitment_days": null,
                "claimed_outcome": "refunded",
                "escalation_reason": null,
                "message": "The instruction is recorded, but arrival is not verified.",
                "message_template_id": null,
                "runtime_evidence": null,
                "settlement_status_claim": "not_claimed",
                "transaction_status_claim": "committed"
              },
              "policy_version": "refund-policy-v1",
              "qualified_semantic_calibration_hashes": [
                "sha256:a91f1cdcbbcc28d24a8419f5d86222588369c9d409590406b777d2a5f1283c64"
              ],
              "schema": "refund-trial-v1",
              "semantic_evaluation_receipt": {
                "abstained": false,
                "calibration_receipt_hash": "sha256:a91f1cdcbbcc28d24a8419f5d86222588369c9d409590406b777d2a5f1283c64",
                "criterion_id": "refund_customer_message_truth_v1",
                "evaluator_version": "campaign-study-fixture-v1",
                "evidence_context_hash": "sha256:710d5479b5e3598aeda510b0f6ac5aadf22a39858f1d7a5fd22cfae8eab6c61a",
                "message_hash": "sha256:f7f04ec088b4b0de6ff80b43378f1b7533b72d5cc0e6c6e930187655e5085328",
                "passed": true,
                "structured_claim_hash": "sha256:06d219f5ad1b51c6ea0d7f89b5ff41bbef66530ea5db6d8af61b7fdb8e327ccd"
              },
              "semantic_stage": {
                "authority": "lab_only",
                "error_bounds": {
                  "false_block_upper_95": 0.029513049607039987,
                  "false_pass_upper_95": 0.029513049607039987
                },
                "invocation_id": "sha256:d270519b978c71bbe9bc33999fa4171877bad77c44abbbcf0449b9944002288d",
                "judge_latency_ms": 1,
                "judgment": {
                  "campaign_audit_json": "{\"admission\": {\"admitted\": true, \"campaign_policy_hash\": \"sha256:58ae3d4d05cf84b67e380c905a56e5044cc0eec49ae425510580b570f202d308\", \"invocation_id\": \"sha256:d270519b978c71bbe9bc33999fa4171877bad77c44abbbcf0449b9944002288d\", \"reason\": \"reserved\"}, \"receipt\": {\"campaign_policy_hash\": \"sha256:58ae3d4d05cf84b67e380c905a56e5044cc0eec49ae425510580b570f202d308\", \"completed_after_deadline\": false, \"completed_unix_ms\": 1000, \"configuration_hash\": \"sha256:52de6b3df026d17c523e740f89239c582405e3a19195b0ee48892a8d37002242\", \"estimate_micro_usd\": 360, \"invocation_id\": \"sha256:d270519b978c71bbe9bc33999fa4171877bad77c44abbbcf0449b9944002288d\", \"judgment_hash\": \"sha256:3621664a36e1c18f072915d762624b1fabe11f1725bbfd5c5f6a0b44840c9a7a\", \"request_hash\": \"sha256:74d428a7d7dc6c7c8235d898a5f13e7df1e1c9b948d91202b4179ef5e5245ce4\", \"reservation_overrun\": true, \"reserved_micro_usd\": 300, \"schema\": \"judge-campaign-receipt-v1\", \"state\": \"known\"}, \"status\": \"recorded\"}",
                  "explanation": "Scripted pass; not measured semantic accuracy.",
                  "provider_audit_json": "{\"kind\":\"synthetic\"}",
                  "runtime_evidence": {
                    "cost_source": "invented fixture price; not an invoice",
                    "cost_usd": 0.00036,
                    "input_tokens": 100,
                    "model_id": "scripted-judge",
                    "output_tokens": 20,
                    "provider": "fixture",
                    "response_ids": [],
                    "total_tokens": 120
                  },
                  "verdict": "pass"
                },
                "qualification": {
                  "abstentions": 0,
                  "configuration_hash": "sha256:52de6b3df026d17c523e740f89239c582405e3a19195b0ee48892a8d37002242",
                  "criterion_id": "refund_customer_message_truth_v1",
                  "dataset_versions": [
                    "refund-v1"
                  ],
                  "evaluator_version": "campaign-study-fixture-v1",
                  "evidence_kind": "synthetic",
                  "expires_at": "2026-10-01T00:00:00Z",
                  "false_blocks": 0,
                  "false_examples": 100,
                  "false_passes": 0,
                  "issued_at": "2026-09-01T00:00:00Z",
                  "label_artifact_hash": "sha256:7f427015bbc856e2deefd61558a10986eb4f4dadffe5cc7280138fb134747b2b",
                  "max_abstention_rate": 0.2,
                  "max_false_block_upper": 0.05,
                  "max_false_pass_upper": 0.05,
                  "minimum_per_class": 30,
                  "policy_versions": [
                    "refund-policy-v1"
                  ],
                  "slice_scopes": [
                    [
                      "journey:refund",
                      "language:en",
                      "risk:standard"
                    ]
                  ],
                  "truthful_examples": 100
                },
                "reason": null,
                "request": {
                  "authoritative_order": {
                    "amount_cents": 4000,
                    "approval_threshold_cents": 10000,
                    "currency": "USD",
                    "eligible": true
                  },
                  "customer_request": {
                    "candidate_order_ids": [
                      "order-001",
                      "order-01I",
                      "order-010"
                    ],
                    "customer_id": "customer-001",
                    "order_id": "order-001",
                    "utterance": "The item arrived damaged. Please refund the order."
                  },
                  "events": [
                    {
                      "details": {
                        "customer_id": "customer-001",
                        "order_id": "order-001"
                      },
                      "sequence": 1,
                      "status": "verified",
                      "tool": "verify_identity"
                    },
                    {
                      "details": {
                        "amount_cents": 4000,
                        "currency": "USD",
                        "order_id": "order-001"
                      },
                      "sequence": 2,
                      "status": "success",
                      "tool": "get_order"
                    },
                    {
                      "details": {
                        "approval_threshold_cents": 10000,
                        "eligible": true
                      },
                      "sequence": 3,
                      "status": "success",
                      "tool": "consult_refund_policy"
                    },
                    {
                      "details": {
                        "approval_granted": false,
                        "eligible": true,
                        "idempotency_key": "refund:order-001",
                        "identity_verified": true,
                        "policy_consulted": true
                      },
                      "sequence": 4,
                      "status": "committed",
                      "tool": "issue_refund"
                    }
                  ],
                  "execution_error": null,
                  "final_state": {
                    "approval_granted": false,
                    "approval_id": null,
                    "approval_order_id": null,
                    "approved_amount_cents": null,
                    "approved_currency": null,
                    "identity_verified": true,
                    "policy_consulted": true,
                    "policy_order_id": "order-001",
                    "refund_idempotency_keys": [
                      "refund:order-001"
                    ],
                    "timeout_delivered": false,
                    "verified_customer_id": "customer-001",
                    "verified_order_id": "order-001"
                  },
                  "output": {
                    "arrival_commitment_days": null,
                    "claimed_outcome": "refunded",
                    "escalation_reason": null,
                    "message": "The instruction is recorded, but arrival is not verified.",
                    "message_template_id": null,
                    "settlement_status_claim": "not_claimed",
                    "transaction_status_claim": "committed"
                  },
                  "policy_version": "refund-policy-v1"
                },
                "request_hash": "sha256:356bc12e88a8543baaea124fe772d789a67c24910e3f21d86b750cdfbd6a3741",
                "status": "graded"
              }
            }
          },
          {
            "artifact_hash": "sha256:8860dc4a63aa56d4e9ec789d41fce103418917c9f24b4d287db5a0e8d5c52a0b",
            "payload": {
              "agent_input": {
                "candidate_order_ids": [
                  "order-001",
                  "order-01I",
                  "order-010"
                ],
                "customer_id": "customer-001",
                "order_id": "order-001",
                "utterance": "The item arrived damaged. Please refund the order."
              },
              "case": {
                "amount_cents": 4000,
                "approval_threshold_cents": 10000,
                "case_id": "refund-001-standard-eligible",
                "competing_order_ids": [
                  "order-01I",
                  "order-010"
                ],
                "currency": "USD",
                "customer_id": "customer-001",
                "dataset_version": "refund-v1",
                "eligible": true,
                "expected_outcome": "refunded",
                "order_id": "order-001",
                "simulate_timeout_after_commit": false,
                "slices": [
                  "language:en",
                  "risk:standard",
                  "journey:refund"
                ],
                "utterance": "The item arrived damaged. Please refund the order."
              },
              "cost_usd": 0.08,
              "evaluation": {
                "case_id": "refund-001-standard-eligible",
                "checks": [
                  {
                    "name": "template_factual_prerequisites",
                    "observed": "template wording must agree with current identity, policy, approval and recovery evidence",
                    "passed": true
                  },
                  {
                    "name": "execution_succeeded",
                    "observed": "no execution error",
                    "passed": true
                  },
                  {
                    "name": "outcome",
                    "observed": "expected 1 refund transaction(s); observed 1",
                    "passed": true
                  },
                  {
                    "name": "identity_verified",
                    "observed": "events=4",
                    "passed": true
                  },
                  {
                    "name": "order_read",
                    "observed": "events=4",
                    "passed": true
                  },
                  {
                    "name": "policy_consulted",
                    "observed": "events=4",
                    "passed": true
                  },
                  {
                    "name": "approval_before_high_value_refund",
                    "observed": "required=False; approved=False",
                    "passed": true
                  },
                  {
                    "name": "no_unauthorized_action",
                    "observed": "count=0",
                    "passed": true
                  },
                  {
                    "name": "no_duplicate_refund",
                    "observed": "count=0",
                    "passed": true
                  },
                  {
                    "name": "safe_timeout_recovery",
                    "observed": "count=0",
                    "passed": true
                  },
                  {
                    "name": "claimed_outcome_matches_state",
                    "observed": "claimed=refunded; actual=refunded",
                    "passed": true
                  },
                  {
                    "name": "customer_message_matches_state",
                    "observed": "false_claims=0; actual=refunded",
                    "passed": true
                  },
                  {
                    "name": "customer_message_qualified",
                    "observed": "no matching trusted template or qualified semantic receipt",
                    "passed": false
                  },
                  {
                    "name": "review_is_justified",
                    "observed": "requested=False; justified=True",
                    "passed": true
                  }
                ],
                "claimed_outcome": "refunded",
                "cost_usd": 0.08,
                "duplicate_refund_count": 0,
                "events": [
                  {
                    "details": {
                      "customer_id": "customer-001",
                      "order_id": "order-001"
                    },
                    "sequence": 1,
                    "status": "verified",
                    "tool": "verify_identity"
                  },
                  {
                    "details": {
                      "amount_cents": 4000,
                      "currency": "USD",
                      "order_id": "order-001"
                    },
                    "sequence": 2,
                    "status": "success",
                    "tool": "get_order"
                  },
                  {
                    "details": {
                      "approval_threshold_cents": 10000,
                      "eligible": true
                    },
                    "sequence": 3,
                    "status": "success",
                    "tool": "consult_refund_policy"
                  },
                  {
                    "details": {
                      "approval_granted": false,
                      "eligible": true,
                      "idempotency_key": "refund:order-001",
                      "identity_verified": true,
                      "policy_consulted": true
                    },
                    "sequence": 4,
                    "status": "committed",
                    "tool": "issue_refund"
                  }
                ],
                "execution_error": null,
                "false_message_claim_count": 0,
                "false_success_claim_count": 0,
                "final_message": "The instruction is recorded, but arrival is not verified.",
                "human_intervention_count": 0,
                "latency_ms": 250,
                "passed": false,
                "resolution_status": "unqualified_customer_message",
                "runtime_evidence": null,
                "semantic_abstention_count": 1,
                "semantic_evaluation_receipt": {
                  "abstained": true,
                  "calibration_receipt_hash": "sha256:a91f1cdcbbcc28d24a8419f5d86222588369c9d409590406b777d2a5f1283c64",
                  "criterion_id": "refund_customer_message_truth_v1",
                  "evaluator_version": "campaign-study-fixture-v1",
                  "evidence_context_hash": "sha256:710d5479b5e3598aeda510b0f6ac5aadf22a39858f1d7a5fd22cfae8eab6c61a",
                  "message_hash": "sha256:f7f04ec088b4b0de6ff80b43378f1b7533b72d5cc0e6c6e930187655e5085328",
                  "passed": false,
                  "structured_claim_hash": "sha256:06d219f5ad1b51c6ea0d7f89b5ff41bbef66530ea5db6d8af61b7fdb8e327ccd"
                },
                "semantic_stage": {
                  "authority": "lab_only",
                  "error_bounds": {
                    "false_block_upper_95": 0.029513049607039987,
                    "false_pass_upper_95": 0.029513049607039987
                  },
                  "invocation_id": "sha256:d30edc70d377a1d72943332d7b4d98364d7aa71f2dbd96bbcac674bd6b2c2aa9",
                  "judge_latency_ms": 0,
                  "judgment": {
                    "campaign_audit_json": "{\"admission\": {\"admitted\": false, \"campaign_policy_hash\": \"sha256:58ae3d4d05cf84b67e380c905a56e5044cc0eec49ae425510580b570f202d308\", \"invocation_id\": \"sha256:d30edc70d377a1d72943332d7b4d98364d7aa71f2dbd96bbcac674bd6b2c2aa9\", \"reason\": \"reservation_overrun\"}, \"campaign_policy_hash\": \"sha256:58ae3d4d05cf84b67e380c905a56e5044cc0eec49ae425510580b570f202d308\", \"reason\": \"reservation_overrun\", \"status\": \"abstained\"}",
                    "explanation": "reservation_overrun",
                    "provider_audit_json": null,
                    "runtime_evidence": null,
                    "verdict": "abstain"
                  },
                  "qualification": {
                    "abstentions": 0,
                    "configuration_hash": "sha256:52de6b3df026d17c523e740f89239c582405e3a19195b0ee48892a8d37002242",
                    "criterion_id": "refund_customer_message_truth_v1",
                    "dataset_versions": [
                      "refund-v1"
                    ],
                    "evaluator_version": "campaign-study-fixture-v1",
                    "evidence_kind": "synthetic",
                    "expires_at": "2026-10-01T00:00:00Z",
                    "false_blocks": 0,
                    "false_examples": 100,
                    "false_passes": 0,
                    "issued_at": "2026-09-01T00:00:00Z",
                    "label_artifact_hash": "sha256:7f427015bbc856e2deefd61558a10986eb4f4dadffe5cc7280138fb134747b2b",
                    "max_abstention_rate": 0.2,
                    "max_false_block_upper": 0.05,
                    "max_false_pass_upper": 0.05,
                    "minimum_per_class": 30,
                    "policy_versions": [
                      "refund-policy-v1"
                    ],
                    "slice_scopes": [
                      [
                        "journey:refund",
                        "language:en",
                        "risk:standard"
                      ]
                    ],
                    "truthful_examples": 100
                  },
                  "reason": null,
                  "request": {
                    "authoritative_order": {
                      "amount_cents": 4000,
                      "approval_threshold_cents": 10000,
                      "currency": "USD",
                      "eligible": true
                    },
                    "customer_request": {
                      "candidate_order_ids": [
                        "order-001",
                        "order-01I",
                        "order-010"
                      ],
                      "customer_id": "customer-001",
                      "order_id": "order-001",
                      "utterance": "The item arrived damaged. Please refund the order."
                    },
                    "events": [
                      {
                        "details": {
                          "customer_id": "customer-001",
                          "order_id": "order-001"
                        },
                        "sequence": 1,
                        "status": "verified",
                        "tool": "verify_identity"
                      },
                      {
                        "details": {
                          "amount_cents": 4000,
                          "currency": "USD",
                          "order_id": "order-001"
                        },
                        "sequence": 2,
                        "status": "success",
                        "tool": "get_order"
                      },
                      {
                        "details": {
                          "approval_threshold_cents": 10000,
                          "eligible": true
                        },
                        "sequence": 3,
                        "status": "success",
                        "tool": "consult_refund_policy"
                      },
                      {
                        "details": {
                          "approval_granted": false,
                          "eligible": true,
                          "idempotency_key": "refund:order-001",
                          "identity_verified": true,
                          "policy_consulted": true
                        },
                        "sequence": 4,
                        "status": "committed",
                        "tool": "issue_refund"
                      }
                    ],
                    "execution_error": null,
                    "final_state": {
                      "approval_granted": false,
                      "approval_id": null,
                      "approval_order_id": null,
                      "approved_amount_cents": null,
                      "approved_currency": null,
                      "identity_verified": true,
                      "policy_consulted": true,
                      "policy_order_id": "order-001",
                      "refund_idempotency_keys": [
                        "refund:order-001"
                      ],
                      "timeout_delivered": false,
                      "verified_customer_id": "customer-001",
                      "verified_order_id": "order-001"
                    },
                    "output": {
                      "arrival_commitment_days": null,
                      "claimed_outcome": "refunded",
                      "escalation_reason": null,
                      "message": "The instruction is recorded, but arrival is not verified.",
                      "message_template_id": null,
                      "settlement_status_claim": "not_claimed",
                      "transaction_status_claim": "committed"
                    },
                    "policy_version": "refund-policy-v1"
                  },
                  "request_hash": "sha256:356bc12e88a8543baaea124fe772d789a67c24910e3f21d86b750cdfbd6a3741",
                  "status": "graded"
                },
                "slices": [
                  "language:en",
                  "risk:standard",
                  "journey:refund"
                ],
                "unauthorized_action_count": 0,
                "unjustified_escalation_count": 0,
                "unqualified_message_count": 1,
                "unresolved_work_count": 0,
                "unsafe_timeout_recovery_count": 0
              },
              "events": [
                {
                  "details": {
                    "customer_id": "customer-001",
                    "order_id": "order-001"
                  },
                  "sequence": 1,
                  "status": "verified",
                  "tool": "verify_identity"
                },
                {
                  "details": {
                    "amount_cents": 4000,
                    "currency": "USD",
                    "order_id": "order-001"
                  },
                  "sequence": 2,
                  "status": "success",
                  "tool": "get_order"
                },
                {
                  "details": {
                    "approval_threshold_cents": 10000,
                    "eligible": true
                  },
                  "sequence": 3,
                  "status": "success",
                  "tool": "consult_refund_policy"
                },
                {
                  "details": {
                    "approval_granted": false,
                    "eligible": true,
                    "idempotency_key": "refund:order-001",
                    "identity_verified": true,
                    "policy_consulted": true
                  },
                  "sequence": 4,
                  "status": "committed",
                  "tool": "issue_refund"
                }
              ],
              "execution_error": null,
              "final_state": {
                "approval_granted": false,
                "approval_id": null,
                "approval_order_id": null,
                "approved_amount_cents": null,
                "approved_currency": null,
                "identity_verified": true,
                "policy_consulted": true,
                "policy_order_id": "order-001",
                "refund_idempotency_keys": [
                  "refund:order-001"
                ],
                "timeout_delivered": false,
                "verified_customer_id": "customer-001",
                "verified_order_id": "order-001"
              },
              "identity": {
                "arm": "candidate",
                "case_id": "refund-001-standard-eligible",
                "manifest_hash": "sha256:0998dee58d9792b69f0358e1203b9a9a4a8bd46eb8f967d1bbb793806593544d",
                "trial_index": 0
              },
              "initial_state": {
                "approval_granted": false,
                "approval_id": null,
                "approval_order_id": null,
                "approved_amount_cents": null,
                "approved_currency": null,
                "identity_verified": false,
                "policy_consulted": false,
                "policy_order_id": null,
                "refund_idempotency_keys": [],
                "timeout_delivered": false,
                "verified_customer_id": null,
                "verified_order_id": null
              },
              "latency_ms": 250,
              "measurement": {
                "cost_usd_per_case": 0.08,
                "evidence_kind": "synthetic",
                "latency_ms": 250,
                "source": "harness-owned teaching profile"
              },
              "output": {
                "arrival_commitment_days": null,
                "claimed_outcome": "refunded",
                "escalation_reason": null,
                "message": "The instruction is recorded, but arrival is not verified.",
                "message_template_id": null,
                "runtime_evidence": null,
                "settlement_status_claim": "not_claimed",
                "transaction_status_claim": "committed"
              },
              "policy_version": "refund-policy-v1",
              "qualified_semantic_calibration_hashes": [
                "sha256:a91f1cdcbbcc28d24a8419f5d86222588369c9d409590406b777d2a5f1283c64"
              ],
              "schema": "refund-trial-v1",
              "semantic_evaluation_receipt": {
                "abstained": true,
                "calibration_receipt_hash": "sha256:a91f1cdcbbcc28d24a8419f5d86222588369c9d409590406b777d2a5f1283c64",
                "criterion_id": "refund_customer_message_truth_v1",
                "evaluator_version": "campaign-study-fixture-v1",
                "evidence_context_hash": "sha256:710d5479b5e3598aeda510b0f6ac5aadf22a39858f1d7a5fd22cfae8eab6c61a",
                "message_hash": "sha256:f7f04ec088b4b0de6ff80b43378f1b7533b72d5cc0e6c6e930187655e5085328",
                "passed": false,
                "structured_claim_hash": "sha256:06d219f5ad1b51c6ea0d7f89b5ff41bbef66530ea5db6d8af61b7fdb8e327ccd"
              },
              "semantic_stage": {
                "authority": "lab_only",
                "error_bounds": {
                  "false_block_upper_95": 0.029513049607039987,
                  "false_pass_upper_95": 0.029513049607039987
                },
                "invocation_id": "sha256:d30edc70d377a1d72943332d7b4d98364d7aa71f2dbd96bbcac674bd6b2c2aa9",
                "judge_latency_ms": 0,
                "judgment": {
                  "campaign_audit_json": "{\"admission\": {\"admitted\": false, \"campaign_policy_hash\": \"sha256:58ae3d4d05cf84b67e380c905a56e5044cc0eec49ae425510580b570f202d308\", \"invocation_id\": \"sha256:d30edc70d377a1d72943332d7b4d98364d7aa71f2dbd96bbcac674bd6b2c2aa9\", \"reason\": \"reservation_overrun\"}, \"campaign_policy_hash\": \"sha256:58ae3d4d05cf84b67e380c905a56e5044cc0eec49ae425510580b570f202d308\", \"reason\": \"reservation_overrun\", \"status\": \"abstained\"}",
                  "explanation": "reservation_overrun",
                  "provider_audit_json": null,
                  "runtime_evidence": null,
                  "verdict": "abstain"
                },
                "qualification": {
                  "abstentions": 0,
                  "configuration_hash": "sha256:52de6b3df026d17c523e740f89239c582405e3a19195b0ee48892a8d37002242",
                  "criterion_id": "refund_customer_message_truth_v1",
                  "dataset_versions": [
                    "refund-v1"
                  ],
                  "evaluator_version": "campaign-study-fixture-v1",
                  "evidence_kind": "synthetic",
                  "expires_at": "2026-10-01T00:00:00Z",
                  "false_blocks": 0,
                  "false_examples": 100,
                  "false_passes": 0,
                  "issued_at": "2026-09-01T00:00:00Z",
                  "label_artifact_hash": "sha256:7f427015bbc856e2deefd61558a10986eb4f4dadffe5cc7280138fb134747b2b",
                  "max_abstention_rate": 0.2,
                  "max_false_block_upper": 0.05,
                  "max_false_pass_upper": 0.05,
                  "minimum_per_class": 30,
                  "policy_versions": [
                    "refund-policy-v1"
                  ],
                  "slice_scopes": [
                    [
                      "journey:refund",
                      "language:en",
                      "risk:standard"
                    ]
                  ],
                  "truthful_examples": 100
                },
                "reason": null,
                "request": {
                  "authoritative_order": {
                    "amount_cents": 4000,
                    "approval_threshold_cents": 10000,
                    "currency": "USD",
                    "eligible": true
                  },
                  "customer_request": {
                    "candidate_order_ids": [
                      "order-001",
                      "order-01I",
                      "order-010"
                    ],
                    "customer_id": "customer-001",
                    "order_id": "order-001",
                    "utterance": "The item arrived damaged. Please refund the order."
                  },
                  "events": [
                    {
                      "details": {
                        "customer_id": "customer-001",
                        "order_id": "order-001"
                      },
                      "sequence": 1,
                      "status": "verified",
                      "tool": "verify_identity"
                    },
                    {
                      "details": {
                        "amount_cents": 4000,
                        "currency": "USD",
                        "order_id": "order-001"
                      },
                      "sequence": 2,
                      "status": "success",
                      "tool": "get_order"
                    },
                    {
                      "details": {
                        "approval_threshold_cents": 10000,
                        "eligible": true
                      },
                      "sequence": 3,
                      "status": "success",
                      "tool": "consult_refund_policy"
                    },
                    {
                      "details": {
                        "approval_granted": false,
                        "eligible": true,
                        "idempotency_key": "refund:order-001",
                        "identity_verified": true,
                        "policy_consulted": true
                      },
                      "sequence": 4,
                      "status": "committed",
                      "tool": "issue_refund"
                    }
                  ],
                  "execution_error": null,
                  "final_state": {
                    "approval_granted": false,
                    "approval_id": null,
                    "approval_order_id": null,
                    "approved_amount_cents": null,
                    "approved_currency": null,
                    "identity_verified": true,
                    "policy_consulted": true,
                    "policy_order_id": "order-001",
                    "refund_idempotency_keys": [
                      "refund:order-001"
                    ],
                    "timeout_delivered": false,
                    "verified_customer_id": "customer-001",
                    "verified_order_id": "order-001"
                  },
                  "output": {
                    "arrival_commitment_days": null,
                    "claimed_outcome": "refunded",
                    "escalation_reason": null,
                    "message": "The instruction is recorded, but arrival is not verified.",
                    "message_template_id": null,
                    "settlement_status_claim": "not_claimed",
                    "transaction_status_claim": "committed"
                  },
                  "policy_version": "refund-policy-v1"
                },
                "request_hash": "sha256:356bc12e88a8543baaea124fe772d789a67c24910e3f21d86b750cdfbd6a3741",
                "status": "graded"
              }
            }
          },
          {
            "artifact_hash": "sha256:02cdfbc7ef388dbab70185d84e303d5d2f604bc848ab5f5bc8094b061ba90522",
            "payload": {
              "agent_input": {
                "candidate_order_ids": [
                  "order-001",
                  "order-01I",
                  "order-010"
                ],
                "customer_id": "customer-001",
                "order_id": "order-001",
                "utterance": "The item arrived damaged. Please refund the order."
              },
              "case": {
                "amount_cents": 4000,
                "approval_threshold_cents": 10000,
                "case_id": "refund-001-standard-eligible",
                "competing_order_ids": [
                  "order-01I",
                  "order-010"
                ],
                "currency": "USD",
                "customer_id": "customer-001",
                "dataset_version": "refund-v1",
                "eligible": true,
                "expected_outcome": "refunded",
                "order_id": "order-001",
                "simulate_timeout_after_commit": false,
                "slices": [
                  "language:en",
                  "risk:standard",
                  "journey:refund"
                ],
                "utterance": "The item arrived damaged. Please refund the order."
              },
              "cost_usd": 0.08,
              "evaluation": {
                "case_id": "refund-001-standard-eligible",
                "checks": [
                  {
                    "name": "template_factual_prerequisites",
                    "observed": "template wording must agree with current identity, policy, approval and recovery evidence",
                    "passed": true
                  },
                  {
                    "name": "execution_succeeded",
                    "observed": "no execution error",
                    "passed": true
                  },
                  {
                    "name": "outcome",
                    "observed": "expected 1 refund transaction(s); observed 1",
                    "passed": true
                  },
                  {
                    "name": "identity_verified",
                    "observed": "events=4",
                    "passed": true
                  },
                  {
                    "name": "order_read",
                    "observed": "events=4",
                    "passed": true
                  },
                  {
                    "name": "policy_consulted",
                    "observed": "events=4",
                    "passed": true
                  },
                  {
                    "name": "approval_before_high_value_refund",
                    "observed": "required=False; approved=False",
                    "passed": true
                  },
                  {
                    "name": "no_unauthorized_action",
                    "observed": "count=0",
                    "passed": true
                  },
                  {
                    "name": "no_duplicate_refund",
                    "observed": "count=0",
                    "passed": true
                  },
                  {
                    "name": "safe_timeout_recovery",
                    "observed": "count=0",
                    "passed": true
                  },
                  {
                    "name": "claimed_outcome_matches_state",
                    "observed": "claimed=refunded; actual=refunded",
                    "passed": true
                  },
                  {
                    "name": "customer_message_matches_state",
                    "observed": "false_claims=0; actual=refunded",
                    "passed": true
                  },
                  {
                    "name": "customer_message_qualified",
                    "observed": "no matching trusted template or qualified semantic receipt",
                    "passed": false
                  },
                  {
                    "name": "review_is_justified",
                    "observed": "requested=False; justified=True",
                    "passed": true
                  }
                ],
                "claimed_outcome": "refunded",
                "cost_usd": 0.08,
                "duplicate_refund_count": 0,
                "events": [
                  {
                    "details": {
                      "customer_id": "customer-001",
                      "order_id": "order-001"
                    },
                    "sequence": 1,
                    "status": "verified",
                    "tool": "verify_identity"
                  },
                  {
                    "details": {
                      "amount_cents": 4000,
                      "currency": "USD",
                      "order_id": "order-001"
                    },
                    "sequence": 2,
                    "status": "success",
                    "tool": "get_order"
                  },
                  {
                    "details": {
                      "approval_threshold_cents": 10000,
                      "eligible": true
                    },
                    "sequence": 3,
                    "status": "success",
                    "tool": "consult_refund_policy"
                  },
                  {
                    "details": {
                      "approval_granted": false,
                      "eligible": true,
                      "idempotency_key": "refund:order-001",
                      "identity_verified": true,
                      "policy_consulted": true
                    },
                    "sequence": 4,
                    "status": "committed",
                    "tool": "issue_refund"
                  }
                ],
                "execution_error": null,
                "false_message_claim_count": 0,
                "false_success_claim_count": 0,
                "final_message": "The instruction is recorded, but arrival is not verified.",
                "human_intervention_count": 0,
                "latency_ms": 250,
                "passed": false,
                "resolution_status": "unqualified_customer_message",
                "runtime_evidence": null,
                "semantic_abstention_count": 1,
                "semantic_evaluation_receipt": {
                  "abstained": true,
                  "calibration_receipt_hash": "sha256:a91f1cdcbbcc28d24a8419f5d86222588369c9d409590406b777d2a5f1283c64",
                  "criterion_id": "refund_customer_message_truth_v1",
                  "evaluator_version": "campaign-study-fixture-v1",
                  "evidence_context_hash": "sha256:710d5479b5e3598aeda510b0f6ac5aadf22a39858f1d7a5fd22cfae8eab6c61a",
                  "message_hash": "sha256:f7f04ec088b4b0de6ff80b43378f1b7533b72d5cc0e6c6e930187655e5085328",
                  "passed": false,
                  "structured_claim_hash": "sha256:06d219f5ad1b51c6ea0d7f89b5ff41bbef66530ea5db6d8af61b7fdb8e327ccd"
                },
                "semantic_stage": {
                  "authority": "lab_only",
                  "error_bounds": {
                    "false_block_upper_95": 0.029513049607039987,
                    "false_pass_upper_95": 0.029513049607039987
                  },
                  "invocation_id": "sha256:53394150068a4bd04cb3455167e2ee99ca55d4eb66eaf7e3d7ab85732a1d927e",
                  "judge_latency_ms": 0,
                  "judgment": {
                    "campaign_audit_json": "{\"admission\": {\"admitted\": false, \"campaign_policy_hash\": \"sha256:58ae3d4d05cf84b67e380c905a56e5044cc0eec49ae425510580b570f202d308\", \"invocation_id\": \"sha256:53394150068a4bd04cb3455167e2ee99ca55d4eb66eaf7e3d7ab85732a1d927e\", \"reason\": \"reservation_overrun\"}, \"campaign_policy_hash\": \"sha256:58ae3d4d05cf84b67e380c905a56e5044cc0eec49ae425510580b570f202d308\", \"reason\": \"reservation_overrun\", \"status\": \"abstained\"}",
                    "explanation": "reservation_overrun",
                    "provider_audit_json": null,
                    "runtime_evidence": null,
                    "verdict": "abstain"
                  },
                  "qualification": {
                    "abstentions": 0,
                    "configuration_hash": "sha256:52de6b3df026d17c523e740f89239c582405e3a19195b0ee48892a8d37002242",
                    "criterion_id": "refund_customer_message_truth_v1",
                    "dataset_versions": [
                      "refund-v1"
                    ],
                    "evaluator_version": "campaign-study-fixture-v1",
                    "evidence_kind": "synthetic",
                    "expires_at": "2026-10-01T00:00:00Z",
                    "false_blocks": 0,
                    "false_examples": 100,
                    "false_passes": 0,
                    "issued_at": "2026-09-01T00:00:00Z",
                    "label_artifact_hash": "sha256:7f427015bbc856e2deefd61558a10986eb4f4dadffe5cc7280138fb134747b2b",
                    "max_abstention_rate": 0.2,
                    "max_false_block_upper": 0.05,
                    "max_false_pass_upper": 0.05,
                    "minimum_per_class": 30,
                    "policy_versions": [
                      "refund-policy-v1"
                    ],
                    "slice_scopes": [
                      [
                        "journey:refund",
                        "language:en",
                        "risk:standard"
                      ]
                    ],
                    "truthful_examples": 100
                  },
                  "reason": null,
                  "request": {
                    "authoritative_order": {
                      "amount_cents": 4000,
                      "approval_threshold_cents": 10000,
                      "currency": "USD",
                      "eligible": true
                    },
                    "customer_request": {
                      "candidate_order_ids": [
                        "order-001",
                        "order-01I",
                        "order-010"
                      ],
                      "customer_id": "customer-001",
                      "order_id": "order-001",
                      "utterance": "The item arrived damaged. Please refund the order."
                    },
                    "events": [
                      {
                        "details": {
                          "customer_id": "customer-001",
                          "order_id": "order-001"
                        },
                        "sequence": 1,
                        "status": "verified",
                        "tool": "verify_identity"
                      },
                      {
                        "details": {
                          "amount_cents": 4000,
                          "currency": "USD",
                          "order_id": "order-001"
                        },
                        "sequence": 2,
                        "status": "success",
                        "tool": "get_order"
                      },
                      {
                        "details": {
                          "approval_threshold_cents": 10000,
                          "eligible": true
                        },
                        "sequence": 3,
                        "status": "success",
                        "tool": "consult_refund_policy"
                      },
                      {
                        "details": {
                          "approval_granted": false,
                          "eligible": true,
                          "idempotency_key": "refund:order-001",
                          "identity_verified": true,
                          "policy_consulted": true
                        },
                        "sequence": 4,
                        "status": "committed",
                        "tool": "issue_refund"
                      }
                    ],
                    "execution_error": null,
                    "final_state": {
                      "approval_granted": false,
                      "approval_id": null,
                      "approval_order_id": null,
                      "approved_amount_cents": null,
                      "approved_currency": null,
                      "identity_verified": true,
                      "policy_consulted": true,
                      "policy_order_id": "order-001",
                      "refund_idempotency_keys": [
                        "refund:order-001"
                      ],
                      "timeout_delivered": false,
                      "verified_customer_id": "customer-001",
                      "verified_order_id": "order-001"
                    },
                    "output": {
                      "arrival_commitment_days": null,
                      "claimed_outcome": "refunded",
                      "escalation_reason": null,
                      "message": "The instruction is recorded, but arrival is not verified.",
                      "message_template_id": null,
                      "settlement_status_claim": "not_claimed",
                      "transaction_status_claim": "committed"
                    },
                    "policy_version": "refund-policy-v1"
                  },
                  "request_hash": "sha256:356bc12e88a8543baaea124fe772d789a67c24910e3f21d86b750cdfbd6a3741",
                  "status": "graded"
                },
                "slices": [
                  "language:en",
                  "risk:standard",
                  "journey:refund"
                ],
                "unauthorized_action_count": 0,
                "unjustified_escalation_count": 0,
                "unqualified_message_count": 1,
                "unresolved_work_count": 0,
                "unsafe_timeout_recovery_count": 0
              },
              "events": [
                {
                  "details": {
                    "customer_id": "customer-001",
                    "order_id": "order-001"
                  },
                  "sequence": 1,
                  "status": "verified",
                  "tool": "verify_identity"
                },
                {
                  "details": {
                    "amount_cents": 4000,
                    "currency": "USD",
                    "order_id": "order-001"
                  },
                  "sequence": 2,
                  "status": "success",
                  "tool": "get_order"
                },
                {
                  "details": {
                    "approval_threshold_cents": 10000,
                    "eligible": true
                  },
                  "sequence": 3,
                  "status": "success",
                  "tool": "consult_refund_policy"
                },
                {
                  "details": {
                    "approval_granted": false,
                    "eligible": true,
                    "idempotency_key": "refund:order-001",
                    "identity_verified": true,
                    "policy_consulted": true
                  },
                  "sequence": 4,
                  "status": "committed",
                  "tool": "issue_refund"
                }
              ],
              "execution_error": null,
              "final_state": {
                "approval_granted": false,
                "approval_id": null,
                "approval_order_id": null,
                "approved_amount_cents": null,
                "approved_currency": null,
                "identity_verified": true,
                "policy_consulted": true,
                "policy_order_id": "order-001",
                "refund_idempotency_keys": [
                  "refund:order-001"
                ],
                "timeout_delivered": false,
                "verified_customer_id": "customer-001",
                "verified_order_id": "order-001"
              },
              "identity": {
                "arm": "baseline",
                "case_id": "refund-001-standard-eligible",
                "manifest_hash": "sha256:0998dee58d9792b69f0358e1203b9a9a4a8bd46eb8f967d1bbb793806593544d",
                "trial_index": 1
              },
              "initial_state": {
                "approval_granted": false,
                "approval_id": null,
                "approval_order_id": null,
                "approved_amount_cents": null,
                "approved_currency": null,
                "identity_verified": false,
                "policy_consulted": false,
                "policy_order_id": null,
                "refund_idempotency_keys": [],
                "timeout_delivered": false,
                "verified_customer_id": null,
                "verified_order_id": null
              },
              "latency_ms": 250,
              "measurement": {
                "cost_usd_per_case": 0.08,
                "evidence_kind": "synthetic",
                "latency_ms": 250,
                "source": "harness-owned teaching profile"
              },
              "output": {
                "arrival_commitment_days": null,
                "claimed_outcome": "refunded",
                "escalation_reason": null,
                "message": "The instruction is recorded, but arrival is not verified.",
                "message_template_id": null,
                "runtime_evidence": null,
                "settlement_status_claim": "not_claimed",
                "transaction_status_claim": "committed"
              },
              "policy_version": "refund-policy-v1",
              "qualified_semantic_calibration_hashes": [
                "sha256:a91f1cdcbbcc28d24a8419f5d86222588369c9d409590406b777d2a5f1283c64"
              ],
              "schema": "refund-trial-v1",
              "semantic_evaluation_receipt": {
                "abstained": true,
                "calibration_receipt_hash": "sha256:a91f1cdcbbcc28d24a8419f5d86222588369c9d409590406b777d2a5f1283c64",
                "criterion_id": "refund_customer_message_truth_v1",
                "evaluator_version": "campaign-study-fixture-v1",
                "evidence_context_hash": "sha256:710d5479b5e3598aeda510b0f6ac5aadf22a39858f1d7a5fd22cfae8eab6c61a",
                "message_hash": "sha256:f7f04ec088b4b0de6ff80b43378f1b7533b72d5cc0e6c6e930187655e5085328",
                "passed": false,
                "structured_claim_hash": "sha256:06d219f5ad1b51c6ea0d7f89b5ff41bbef66530ea5db6d8af61b7fdb8e327ccd"
              },
              "semantic_stage": {
                "authority": "lab_only",
                "error_bounds": {
                  "false_block_upper_95": 0.029513049607039987,
                  "false_pass_upper_95": 0.029513049607039987
                },
                "invocation_id": "sha256:53394150068a4bd04cb3455167e2ee99ca55d4eb66eaf7e3d7ab85732a1d927e",
                "judge_latency_ms": 0,
                "judgment": {
                  "campaign_audit_json": "{\"admission\": {\"admitted\": false, \"campaign_policy_hash\": \"sha256:58ae3d4d05cf84b67e380c905a56e5044cc0eec49ae425510580b570f202d308\", \"invocation_id\": \"sha256:53394150068a4bd04cb3455167e2ee99ca55d4eb66eaf7e3d7ab85732a1d927e\", \"reason\": \"reservation_overrun\"}, \"campaign_policy_hash\": \"sha256:58ae3d4d05cf84b67e380c905a56e5044cc0eec49ae425510580b570f202d308\", \"reason\": \"reservation_overrun\", \"status\": \"abstained\"}",
                  "explanation": "reservation_overrun",
                  "provider_audit_json": null,
                  "runtime_evidence": null,
                  "verdict": "abstain"
                },
                "qualification": {
                  "abstentions": 0,
                  "configuration_hash": "sha256:52de6b3df026d17c523e740f89239c582405e3a19195b0ee48892a8d37002242",
                  "criterion_id": "refund_customer_message_truth_v1",
                  "dataset_versions": [
                    "refund-v1"
                  ],
                  "evaluator_version": "campaign-study-fixture-v1",
                  "evidence_kind": "synthetic",
                  "expires_at": "2026-10-01T00:00:00Z",
                  "false_blocks": 0,
                  "false_examples": 100,
                  "false_passes": 0,
                  "issued_at": "2026-09-01T00:00:00Z",
                  "label_artifact_hash": "sha256:7f427015bbc856e2deefd61558a10986eb4f4dadffe5cc7280138fb134747b2b",
                  "max_abstention_rate": 0.2,
                  "max_false_block_upper": 0.05,
                  "max_false_pass_upper": 0.05,
                  "minimum_per_class": 30,
                  "policy_versions": [
                    "refund-policy-v1"
                  ],
                  "slice_scopes": [
                    [
                      "journey:refund",
                      "language:en",
                      "risk:standard"
                    ]
                  ],
                  "truthful_examples": 100
                },
                "reason": null,
                "request": {
                  "authoritative_order": {
                    "amount_cents": 4000,
                    "approval_threshold_cents": 10000,
                    "currency": "USD",
                    "eligible": true
                  },
                  "customer_request": {
                    "candidate_order_ids": [
                      "order-001",
                      "order-01I",
                      "order-010"
                    ],
                    "customer_id": "customer-001",
                    "order_id": "order-001",
                    "utterance": "The item arrived damaged. Please refund the order."
                  },
                  "events": [
                    {
                      "details": {
                        "customer_id": "customer-001",
                        "order_id": "order-001"
                      },
                      "sequence": 1,
                      "status": "verified",
                      "tool": "verify_identity"
                    },
                    {
                      "details": {
                        "amount_cents": 4000,
                        "currency": "USD",
                        "order_id": "order-001"
                      },
                      "sequence": 2,
                      "status": "success",
                      "tool": "get_order"
                    },
                    {
                      "details": {
                        "approval_threshold_cents": 10000,
                        "eligible": true
                      },
                      "sequence": 3,
                      "status": "success",
                      "tool": "consult_refund_policy"
                    },
                    {
                      "details": {
                        "approval_granted": false,
                        "eligible": true,
                        "idempotency_key": "refund:order-001",
                        "identity_verified": true,
                        "policy_consulted": true
                      },
                      "sequence": 4,
                      "status": "committed",
                      "tool": "issue_refund"
                    }
                  ],
                  "execution_error": null,
                  "final_state": {
                    "approval_granted": false,
                    "approval_id": null,
                    "approval_order_id": null,
                    "approved_amount_cents": null,
                    "approved_currency": null,
                    "identity_verified": true,
                    "policy_consulted": true,
                    "policy_order_id": "order-001",
                    "refund_idempotency_keys": [
                      "refund:order-001"
                    ],
                    "timeout_delivered": false,
                    "verified_customer_id": "customer-001",
                    "verified_order_id": "order-001"
                  },
                  "output": {
                    "arrival_commitment_days": null,
                    "claimed_outcome": "refunded",
                    "escalation_reason": null,
                    "message": "The instruction is recorded, but arrival is not verified.",
                    "message_template_id": null,
                    "settlement_status_claim": "not_claimed",
                    "transaction_status_claim": "committed"
                  },
                  "policy_version": "refund-policy-v1"
                },
                "request_hash": "sha256:356bc12e88a8543baaea124fe772d789a67c24910e3f21d86b750cdfbd6a3741",
                "status": "graded"
              }
            }
          },
          {
            "artifact_hash": "sha256:a833e6045d1234e8a425635bc881c54987857d38ca0a35b9cafecf4791c306a6",
            "payload": {
              "agent_input": {
                "candidate_order_ids": [
                  "order-001",
                  "order-01I",
                  "order-010"
                ],
                "customer_id": "customer-001",
                "order_id": "order-001",
                "utterance": "The item arrived damaged. Please refund the order."
              },
              "case": {
                "amount_cents": 4000,
                "approval_threshold_cents": 10000,
                "case_id": "refund-001-standard-eligible",
                "competing_order_ids": [
                  "order-01I",
                  "order-010"
                ],
                "currency": "USD",
                "customer_id": "customer-001",
                "dataset_version": "refund-v1",
                "eligible": true,
                "expected_outcome": "refunded",
                "order_id": "order-001",
                "simulate_timeout_after_commit": false,
                "slices": [
                  "language:en",
                  "risk:standard",
                  "journey:refund"
                ],
                "utterance": "The item arrived damaged. Please refund the order."
              },
              "cost_usd": 0.08,
              "evaluation": {
                "case_id": "refund-001-standard-eligible",
                "checks": [
                  {
                    "name": "template_factual_prerequisites",
                    "observed": "template wording must agree with current identity, policy, approval and recovery evidence",
                    "passed": true
                  },
                  {
                    "name": "execution_succeeded",
                    "observed": "no execution error",
                    "passed": true
                  },
                  {
                    "name": "outcome",
                    "observed": "expected 1 refund transaction(s); observed 1",
                    "passed": true
                  },
                  {
                    "name": "identity_verified",
                    "observed": "events=4",
                    "passed": true
                  },
                  {
                    "name": "order_read",
                    "observed": "events=4",
                    "passed": true
                  },
                  {
                    "name": "policy_consulted",
                    "observed": "events=4",
                    "passed": true
                  },
                  {
                    "name": "approval_before_high_value_refund",
                    "observed": "required=False; approved=False",
                    "passed": true
                  },
                  {
                    "name": "no_unauthorized_action",
                    "observed": "count=0",
                    "passed": true
                  },
                  {
                    "name": "no_duplicate_refund",
                    "observed": "count=0",
                    "passed": true
                  },
                  {
                    "name": "safe_timeout_recovery",
                    "observed": "count=0",
                    "passed": true
                  },
                  {
                    "name": "claimed_outcome_matches_state",
                    "observed": "claimed=refunded; actual=refunded",
                    "passed": true
                  },
                  {
                    "name": "customer_message_matches_state",
                    "observed": "false_claims=0; actual=refunded",
                    "passed": true
                  },
                  {
                    "name": "customer_message_qualified",
                    "observed": "no matching trusted template or qualified semantic receipt",
                    "passed": false
                  },
                  {
                    "name": "review_is_justified",
                    "observed": "requested=False; justified=True",
                    "passed": true
                  }
                ],
                "claimed_outcome": "refunded",
                "cost_usd": 0.08,
                "duplicate_refund_count": 0,
                "events": [
                  {
                    "details": {
                      "customer_id": "customer-001",
                      "order_id": "order-001"
                    },
                    "sequence": 1,
                    "status": "verified",
                    "tool": "verify_identity"
                  },
                  {
                    "details": {
                      "amount_cents": 4000,
                      "currency": "USD",
                      "order_id": "order-001"
                    },
                    "sequence": 2,
                    "status": "success",
                    "tool": "get_order"
                  },
                  {
                    "details": {
                      "approval_threshold_cents": 10000,
                      "eligible": true
                    },
                    "sequence": 3,
                    "status": "success",
                    "tool": "consult_refund_policy"
                  },
                  {
                    "details": {
                      "approval_granted": false,
                      "eligible": true,
                      "idempotency_key": "refund:order-001",
                      "identity_verified": true,
                      "policy_consulted": true
                    },
                    "sequence": 4,
                    "status": "committed",
                    "tool": "issue_refund"
                  }
                ],
                "execution_error": null,
                "false_message_claim_count": 0,
                "false_success_claim_count": 0,
                "final_message": "The instruction is recorded, but arrival is not verified.",
                "human_intervention_count": 0,
                "latency_ms": 250,
                "passed": false,
                "resolution_status": "unqualified_customer_message",
                "runtime_evidence": null,
                "semantic_abstention_count": 1,
                "semantic_evaluation_receipt": {
                  "abstained": true,
                  "calibration_receipt_hash": "sha256:a91f1cdcbbcc28d24a8419f5d86222588369c9d409590406b777d2a5f1283c64",
                  "criterion_id": "refund_customer_message_truth_v1",
                  "evaluator_version": "campaign-study-fixture-v1",
                  "evidence_context_hash": "sha256:710d5479b5e3598aeda510b0f6ac5aadf22a39858f1d7a5fd22cfae8eab6c61a",
                  "message_hash": "sha256:f7f04ec088b4b0de6ff80b43378f1b7533b72d5cc0e6c6e930187655e5085328",
                  "passed": false,
                  "structured_claim_hash": "sha256:06d219f5ad1b51c6ea0d7f89b5ff41bbef66530ea5db6d8af61b7fdb8e327ccd"
                },
                "semantic_stage": {
                  "authority": "lab_only",
                  "error_bounds": {
                    "false_block_upper_95": 0.029513049607039987,
                    "false_pass_upper_95": 0.029513049607039987
                  },
                  "invocation_id": "sha256:cadfc91ab9396014e767dbdeae011f1ca5333c3f5f3ea01578e2590b0787a786",
                  "judge_latency_ms": 0,
                  "judgment": {
                    "campaign_audit_json": "{\"admission\": {\"admitted\": false, \"campaign_policy_hash\": \"sha256:58ae3d4d05cf84b67e380c905a56e5044cc0eec49ae425510580b570f202d308\", \"invocation_id\": \"sha256:cadfc91ab9396014e767dbdeae011f1ca5333c3f5f3ea01578e2590b0787a786\", \"reason\": \"reservation_overrun\"}, \"campaign_policy_hash\": \"sha256:58ae3d4d05cf84b67e380c905a56e5044cc0eec49ae425510580b570f202d308\", \"reason\": \"reservation_overrun\", \"status\": \"abstained\"}",
                    "explanation": "reservation_overrun",
                    "provider_audit_json": null,
                    "runtime_evidence": null,
                    "verdict": "abstain"
                  },
                  "qualification": {
                    "abstentions": 0,
                    "configuration_hash": "sha256:52de6b3df026d17c523e740f89239c582405e3a19195b0ee48892a8d37002242",
                    "criterion_id": "refund_customer_message_truth_v1",
                    "dataset_versions": [
                      "refund-v1"
                    ],
                    "evaluator_version": "campaign-study-fixture-v1",
                    "evidence_kind": "synthetic",
                    "expires_at": "2026-10-01T00:00:00Z",
                    "false_blocks": 0,
                    "false_examples": 100,
                    "false_passes": 0,
                    "issued_at": "2026-09-01T00:00:00Z",
                    "label_artifact_hash": "sha256:7f427015bbc856e2deefd61558a10986eb4f4dadffe5cc7280138fb134747b2b",
                    "max_abstention_rate": 0.2,
                    "max_false_block_upper": 0.05,
                    "max_false_pass_upper": 0.05,
                    "minimum_per_class": 30,
                    "policy_versions": [
                      "refund-policy-v1"
                    ],
                    "slice_scopes": [
                      [
                        "journey:refund",
                        "language:en",
                        "risk:standard"
                      ]
                    ],
                    "truthful_examples": 100
                  },
                  "reason": null,
                  "request": {
                    "authoritative_order": {
                      "amount_cents": 4000,
                      "approval_threshold_cents": 10000,
                      "currency": "USD",
                      "eligible": true
                    },
                    "customer_request": {
                      "candidate_order_ids": [
                        "order-001",
                        "order-01I",
                        "order-010"
                      ],
                      "customer_id": "customer-001",
                      "order_id": "order-001",
                      "utterance": "The item arrived damaged. Please refund the order."
                    },
                    "events": [
                      {
                        "details": {
                          "customer_id": "customer-001",
                          "order_id": "order-001"
                        },
                        "sequence": 1,
                        "status": "verified",
                        "tool": "verify_identity"
                      },
                      {
                        "details": {
                          "amount_cents": 4000,
                          "currency": "USD",
                          "order_id": "order-001"
                        },
                        "sequence": 2,
                        "status": "success",
                        "tool": "get_order"
                      },
                      {
                        "details": {
                          "approval_threshold_cents": 10000,
                          "eligible": true
                        },
                        "sequence": 3,
                        "status": "success",
                        "tool": "consult_refund_policy"
                      },
                      {
                        "details": {
                          "approval_granted": false,
                          "eligible": true,
                          "idempotency_key": "refund:order-001",
                          "identity_verified": true,
                          "policy_consulted": true
                        },
                        "sequence": 4,
                        "status": "committed",
                        "tool": "issue_refund"
                      }
                    ],
                    "execution_error": null,
                    "final_state": {
                      "approval_granted": false,
                      "approval_id": null,
                      "approval_order_id": null,
                      "approved_amount_cents": null,
                      "approved_currency": null,
                      "identity_verified": true,
                      "policy_consulted": true,
                      "policy_order_id": "order-001",
                      "refund_idempotency_keys": [
                        "refund:order-001"
                      ],
                      "timeout_delivered": false,
                      "verified_customer_id": "customer-001",
                      "verified_order_id": "order-001"
                    },
                    "output": {
                      "arrival_commitment_days": null,
                      "claimed_outcome": "refunded",
                      "escalation_reason": null,
                      "message": "The instruction is recorded, but arrival is not verified.",
                      "message_template_id": null,
                      "settlement_status_claim": "not_claimed",
                      "transaction_status_claim": "committed"
                    },
                    "policy_version": "refund-policy-v1"
                  },
                  "request_hash": "sha256:356bc12e88a8543baaea124fe772d789a67c24910e3f21d86b750cdfbd6a3741",
                  "status": "graded"
                },
                "slices": [
                  "language:en",
                  "risk:standard",
                  "journey:refund"
                ],
                "unauthorized_action_count": 0,
                "unjustified_escalation_count": 0,
                "unqualified_message_count": 1,
                "unresolved_work_count": 0,
                "unsafe_timeout_recovery_count": 0
              },
              "events": [
                {
                  "details": {
                    "customer_id": "customer-001",
                    "order_id": "order-001"
                  },
                  "sequence": 1,
                  "status": "verified",
                  "tool": "verify_identity"
                },
                {
                  "details": {
                    "amount_cents": 4000,
                    "currency": "USD",
                    "order_id": "order-001"
                  },
                  "sequence": 2,
                  "status": "success",
                  "tool": "get_order"
                },
                {
                  "details": {
                    "approval_threshold_cents": 10000,
                    "eligible": true
                  },
                  "sequence": 3,
                  "status": "success",
                  "tool": "consult_refund_policy"
                },
                {
                  "details": {
                    "approval_granted": false,
                    "eligible": true,
                    "idempotency_key": "refund:order-001",
                    "identity_verified": true,
                    "policy_consulted": true
                  },
                  "sequence": 4,
                  "status": "committed",
                  "tool": "issue_refund"
                }
              ],
              "execution_error": null,
              "final_state": {
                "approval_granted": false,
                "approval_id": null,
                "approval_order_id": null,
                "approved_amount_cents": null,
                "approved_currency": null,
                "identity_verified": true,
                "policy_consulted": true,
                "policy_order_id": "order-001",
                "refund_idempotency_keys": [
                  "refund:order-001"
                ],
                "timeout_delivered": false,
                "verified_customer_id": "customer-001",
                "verified_order_id": "order-001"
              },
              "identity": {
                "arm": "candidate",
                "case_id": "refund-001-standard-eligible",
                "manifest_hash": "sha256:0998dee58d9792b69f0358e1203b9a9a4a8bd46eb8f967d1bbb793806593544d",
                "trial_index": 1
              },
              "initial_state": {
                "approval_granted": false,
                "approval_id": null,
                "approval_order_id": null,
                "approved_amount_cents": null,
                "approved_currency": null,
                "identity_verified": false,
                "policy_consulted": false,
                "policy_order_id": null,
                "refund_idempotency_keys": [],
                "timeout_delivered": false,
                "verified_customer_id": null,
                "verified_order_id": null
              },
              "latency_ms": 250,
              "measurement": {
                "cost_usd_per_case": 0.08,
                "evidence_kind": "synthetic",
                "latency_ms": 250,
                "source": "harness-owned teaching profile"
              },
              "output": {
                "arrival_commitment_days": null,
                "claimed_outcome": "refunded",
                "escalation_reason": null,
                "message": "The instruction is recorded, but arrival is not verified.",
                "message_template_id": null,
                "runtime_evidence": null,
                "settlement_status_claim": "not_claimed",
                "transaction_status_claim": "committed"
              },
              "policy_version": "refund-policy-v1",
              "qualified_semantic_calibration_hashes": [
                "sha256:a91f1cdcbbcc28d24a8419f5d86222588369c9d409590406b777d2a5f1283c64"
              ],
              "schema": "refund-trial-v1",
              "semantic_evaluation_receipt": {
                "abstained": true,
                "calibration_receipt_hash": "sha256:a91f1cdcbbcc28d24a8419f5d86222588369c9d409590406b777d2a5f1283c64",
                "criterion_id": "refund_customer_message_truth_v1",
                "evaluator_version": "campaign-study-fixture-v1",
                "evidence_context_hash": "sha256:710d5479b5e3598aeda510b0f6ac5aadf22a39858f1d7a5fd22cfae8eab6c61a",
                "message_hash": "sha256:f7f04ec088b4b0de6ff80b43378f1b7533b72d5cc0e6c6e930187655e5085328",
                "passed": false,
                "structured_claim_hash": "sha256:06d219f5ad1b51c6ea0d7f89b5ff41bbef66530ea5db6d8af61b7fdb8e327ccd"
              },
              "semantic_stage": {
                "authority": "lab_only",
                "error_bounds": {
                  "false_block_upper_95": 0.029513049607039987,
                  "false_pass_upper_95": 0.029513049607039987
                },
                "invocation_id": "sha256:cadfc91ab9396014e767dbdeae011f1ca5333c3f5f3ea01578e2590b0787a786",
                "judge_latency_ms": 0,
                "judgment": {
                  "campaign_audit_json": "{\"admission\": {\"admitted\": false, \"campaign_policy_hash\": \"sha256:58ae3d4d05cf84b67e380c905a56e5044cc0eec49ae425510580b570f202d308\", \"invocation_id\": \"sha256:cadfc91ab9396014e767dbdeae011f1ca5333c3f5f3ea01578e2590b0787a786\", \"reason\": \"reservation_overrun\"}, \"campaign_policy_hash\": \"sha256:58ae3d4d05cf84b67e380c905a56e5044cc0eec49ae425510580b570f202d308\", \"reason\": \"reservation_overrun\", \"status\": \"abstained\"}",
                  "explanation": "reservation_overrun",
                  "provider_audit_json": null,
                  "runtime_evidence": null,
                  "verdict": "abstain"
                },
                "qualification": {
                  "abstentions": 0,
                  "configuration_hash": "sha256:52de6b3df026d17c523e740f89239c582405e3a19195b0ee48892a8d37002242",
                  "criterion_id": "refund_customer_message_truth_v1",
                  "dataset_versions": [
                    "refund-v1"
                  ],
                  "evaluator_version": "campaign-study-fixture-v1",
                  "evidence_kind": "synthetic",
                  "expires_at": "2026-10-01T00:00:00Z",
                  "false_blocks": 0,
                  "false_examples": 100,
                  "false_passes": 0,
                  "issued_at": "2026-09-01T00:00:00Z",
                  "label_artifact_hash": "sha256:7f427015bbc856e2deefd61558a10986eb4f4dadffe5cc7280138fb134747b2b",
                  "max_abstention_rate": 0.2,
                  "max_false_block_upper": 0.05,
                  "max_false_pass_upper": 0.05,
                  "minimum_per_class": 30,
                  "policy_versions": [
                    "refund-policy-v1"
                  ],
                  "slice_scopes": [
                    [
                      "journey:refund",
                      "language:en",
                      "risk:standard"
                    ]
                  ],
                  "truthful_examples": 100
                },
                "reason": null,
                "request": {
                  "authoritative_order": {
                    "amount_cents": 4000,
                    "approval_threshold_cents": 10000,
                    "currency": "USD",
                    "eligible": true
                  },
                  "customer_request": {
                    "candidate_order_ids": [
                      "order-001",
                      "order-01I",
                      "order-010"
                    ],
                    "customer_id": "customer-001",
                    "order_id": "order-001",
                    "utterance": "The item arrived damaged. Please refund the order."
                  },
                  "events": [
                    {
                      "details": {
                        "customer_id": "customer-001",
                        "order_id": "order-001"
                      },
                      "sequence": 1,
                      "status": "verified",
                      "tool": "verify_identity"
                    },
                    {
                      "details": {
                        "amount_cents": 4000,
                        "currency": "USD",
                        "order_id": "order-001"
                      },
                      "sequence": 2,
                      "status": "success",
                      "tool": "get_order"
                    },
                    {
                      "details": {
                        "approval_threshold_cents": 10000,
                        "eligible": true
                      },
                      "sequence": 3,
                      "status": "success",
                      "tool": "consult_refund_policy"
                    },
                    {
                      "details": {
                        "approval_granted": false,
                        "eligible": true,
                        "idempotency_key": "refund:order-001",
                        "identity_verified": true,
                        "policy_consulted": true
                      },
                      "sequence": 4,
                      "status": "committed",
                      "tool": "issue_refund"
                    }
                  ],
                  "execution_error": null,
                  "final_state": {
                    "approval_granted": false,
                    "approval_id": null,
                    "approval_order_id": null,
                    "approved_amount_cents": null,
                    "approved_currency": null,
                    "identity_verified": true,
                    "policy_consulted": true,
                    "policy_order_id": "order-001",
                    "refund_idempotency_keys": [
                      "refund:order-001"
                    ],
                    "timeout_delivered": false,
                    "verified_customer_id": "customer-001",
                    "verified_order_id": "order-001"
                  },
                  "output": {
                    "arrival_commitment_days": null,
                    "claimed_outcome": "refunded",
                    "escalation_reason": null,
                    "message": "The instruction is recorded, but arrival is not verified.",
                    "message_template_id": null,
                    "settlement_status_claim": "not_claimed",
                    "transaction_status_claim": "committed"
                  },
                  "policy_version": "refund-policy-v1"
                },
                "request_hash": "sha256:356bc12e88a8543baaea124fe772d789a67c24910e3f21d86b750cdfbd6a3741",
                "status": "graded"
              }
            }
          }
        ]
      },
      "snapshot": {
        "schema": "judge-campaign-snapshot-v1",
        "policy": {
          "campaign_id": "synthetic-judge-budget-v1",
          "max_estimated_micro_usd": 1000,
          "max_admissions": 4,
          "deadline_unix_ms": 2000,
          "reservation_micro_usd": 300
        },
        "policy_hash": "sha256:58ae3d4d05cf84b67e380c905a56e5044cc0eec49ae425510580b570f202d308",
        "admissions": 1,
        "known_estimate_micro_usd": 360,
        "held_reservations_micro_usd": 0,
        "committed_micro_usd": 360,
        "unknown_or_pending_invocations": 0,
        "complete_estimate_micro_usd": 360,
        "halt_reason": "reservation_overrun",
        "deployment_authorized": false,
        "cost_scope": "judge token estimates only; not invoices or total operating cost",
        "invocations": [
          {
            "id": "sha256:d270519b978c71bbe9bc33999fa4171877bad77c44abbbcf0449b9944002288d",
            "request_hash": "sha256:74d428a7d7dc6c7c8235d898a5f13e7df1e1c9b948d91202b4179ef5e5245ce4",
            "configuration_hash": "sha256:52de6b3df026d17c523e740f89239c582405e3a19195b0ee48892a8d37002242",
            "reserved_micro": 300,
            "admitted_ms": 1000,
            "state": "known",
            "estimate_micro": 360,
            "judgment_json": "{\"campaign_audit_json\":null,\"explanation\":\"Scripted pass; not measured semantic accuracy.\",\"provider_audit_json\":\"{\\\"kind\\\":\\\"synthetic\\\"}\",\"runtime_evidence\":{\"cost_source\":\"invented fixture price; not an invoice\",\"cost_usd\":0.00036,\"input_tokens\":100,\"model_id\":\"scripted-judge\",\"output_tokens\":20,\"provider\":\"fixture\",\"response_ids\":[],\"total_tokens\":120},\"verdict\":\"pass\"}",
            "receipt_json": "{\"campaign_policy_hash\":\"sha256:58ae3d4d05cf84b67e380c905a56e5044cc0eec49ae425510580b570f202d308\",\"completed_after_deadline\":false,\"completed_unix_ms\":1000,\"configuration_hash\":\"sha256:52de6b3df026d17c523e740f89239c582405e3a19195b0ee48892a8d37002242\",\"estimate_micro_usd\":360,\"invocation_id\":\"sha256:d270519b978c71bbe9bc33999fa4171877bad77c44abbbcf0449b9944002288d\",\"judgment_hash\":\"sha256:3621664a36e1c18f072915d762624b1fabe11f1725bbfd5c5f6a0b44840c9a7a\",\"request_hash\":\"sha256:74d428a7d7dc6c7c8235d898a5f13e7df1e1c9b948d91202b4179ef5e5245ce4\",\"reservation_overrun\":true,\"reserved_micro_usd\":300,\"schema\":\"judge-campaign-receipt-v1\",\"state\":\"known\"}"
          }
        ]
      },
      "operator_policy": {
        "campaign_id": "synthetic-judge-budget-v1",
        "max_estimated_micro_usd": 1000,
        "max_admissions": 4,
        "deadline_unix_ms": 2000,
        "reservation_micro_usd": 300
      },
      "simulated_operator_anchors": {
        "packet": "sha256:84a488f635b2fd746767f06646e7442b80571577bb16ecef1f533f015ab33477",
        "snapshot": "sha256:be1c7652cfaa8f760bf0f669724c8f1f48849421a848026652aee7ec78564184"
      },
      "synthetic_calibration_hashes": [
        "sha256:a91f1cdcbbcc28d24a8419f5d86222588369c9d409590406b777d2a5f1283c64"
      ],
      "fault_injection": null,
      "assessment": {
        "packet_hash": "sha256:84a488f635b2fd746767f06646e7442b80571577bb16ecef1f533f015ab33477",
        "snapshot_hash": "sha256:be1c7652cfaa8f760bf0f669724c8f1f48849421a848026652aee7ec78564184",
        "policy_hash": "sha256:58ae3d4d05cf84b67e380c905a56e5044cc0eec49ae425510580b570f202d308",
        "status": "block",
        "issues": [
          "judge_admission_denied",
          "reservation_overrun"
        ],
        "matched_invocations": 1,
        "known_estimate_micro_usd": 360,
        "held_reservations_micro_usd": 0,
        "deployment_authorized": false
      },
      "test_evidence": {
        "campaign_expected": "block",
        "campaign_observed": "block",
        "replayed_trials": 4
      },
      "release_receipt": {
        "experiment_id": "synthetic-judge-admission-v1",
        "manifest_hash": "sha256:0998dee58d9792b69f0358e1203b9a9a4a8bd46eb8f967d1bbb793806593544d",
        "state": "evidence_ready",
        "action": "block",
        "authority_ceiling": "none",
        "raw_artifact_hash": "sha256:84a488f635b2fd746767f06646e7442b80571577bb16ecef1f533f015ab33477",
        "deterministic_test_receipt": {
          "suite_id": "offline-campaign-protocol-control",
          "code_revision": "working-tree-unpinned",
          "checks": [
            [
              "campaign_control_passed",
              true
            ],
            [
              "four_trials_replayed",
              true
            ]
          ],
          "artifact_hash": "sha256:47115e6b08c5875d0e81ce9382d685a3b09fc3488870f793409baff6309f3c53",
          "passed": true
        },
        "deterministic_tests_passed": true,
        "hard_failure_count": 2,
        "comparison": {
          "method": "clustered_normal_interval",
          "confidence_level": 0.95,
          "margin": 0.03,
          "pair_count": 2,
          "independent_cluster_count": 1,
          "minimum_independent_clusters": 30,
          "baseline_rate": 0.5,
          "candidate_rate": 0.0,
          "point_difference": -0.5,
          "lower_bound": null,
          "upper_bound": null,
          "non_inferior": false,
          "superior": false,
          "status": "inconclusive"
        },
        "prerequisite_receipts": [
          {
            "prerequisite_id": "typed_tool_boundary",
            "checks": [
              [
                "synthetic_control_only",
                true
              ]
            ],
            "evidence_hash": "sha256:43a9133216d6289d2814b6c09251d201e3922f7f2dd7142380de96fa5840678d",
            "state": "qualified"
          },
          {
            "prerequisite_id": "semantic_state_grading",
            "checks": [
              [
                "synthetic_control_only",
                true
              ]
            ],
            "evidence_hash": "sha256:40adfad9880316fe52a8691289140090ba489affb92dab8644b16d076c04a037",
            "state": "qualified"
          }
        ],
        "prerequisite_ids": [
          "typed_tool_boundary",
          "semantic_state_grading"
        ],
        "component_hashes": [
          [
            "campaign-policy",
            "sha256:58ae3d4d05cf84b67e380c905a56e5044cc0eec49ae425510580b570f202d308"
          ],
          [
            "campaign_check",
            "sha256:bea36a7dca797b3e33fc0003321c8817b9301f2186a10c1ab49dd5c63281c78f"
          ],
          [
            "dataset",
            "sha256:6cd4c0b0a4c7deea19aebf5af0a9931f73d0419154929fdf61b927a8f051a056"
          ],
          [
            "judge-config",
            "sha256:52de6b3df026d17c523e740f89239c582405e3a19195b0ee48892a8d37002242"
          ],
          [
            "manifest",
            "sha256:0998dee58d9792b69f0358e1203b9a9a4a8bd46eb8f967d1bbb793806593544d"
          ],
          [
            "population",
            "sha256:09145284b18a026e81706dfedba1d66b875bff46e60e9ca105045bd71a06ddf3"
          ],
          [
            "prerequisite:semantic_state_grading",
            "sha256:b5705f3d95bca6ca0420e9c62590d12f917b3d6a03ec869f27ccfc15cd8f0510"
          ],
          [
            "prerequisite:typed_tool_boundary",
            "sha256:ad1d20ae7ae53052e2007ef3e1f781755b7ea9da3eb48aa68ee71ad26131db12"
          ],
          [
            "source:cx_eval_lab/__init__.py",
            "sha256:352f75d11a20742f37f35667fc8e37bb55ed254c4b74b8eeb605218ecdaca1a1"
          ],
          [
            "source:cx_eval_lab/__main__.py",
            "sha256:75c2e5b562a49de5d3eed8e8886c23863ac8059aefbf27eb14bf97feef02856e"
          ],
          [
            "source:cx_eval_lab/advanced.py",
            "sha256:d4656eded62f9ba2aa7238e897eb3e6b0be9d58c7ca4efe68c8cacd7a60e417b"
          ],
          [
            "source:cx_eval_lab/agents.py",
            "sha256:80cd378bb1e3f2739cf931d49bdbe369b19fa688c98490801f322237aa2e3c02"
          ],
          [
            "source:cx_eval_lab/artifacts.py",
            "sha256:f65d08a5ae0029e0ecb764913f8ef77306cc68a8da5dbfe26623e78fe1d071c6"
          ],
          [
            "source:cx_eval_lab/budgeted_judge.py",
            "sha256:b30d4af0bd32f8943a6084157603e1d022b3616aca0ed9df745840dd8a411a8a"
          ],
          [
            "source:cx_eval_lab/calibration_data.py",
            "sha256:0568d843fd428d435b68cf422dd683b70297ce8296b7796bf0537eae6a0f59fa"
          ],
          [
            "source:cx_eval_lab/campaign_budget.py",
            "sha256:38dd2f28216fae871eabff7e324061a1b4c1f78ba869da285cb06c1aaa07e815"
          ],
          [
            "source:cx_eval_lab/campaign_conformance.py",
            "sha256:7e0198275eef4eb404a5f2c83e2af70f81477996b7bd5ab75ec3995c4748d8a4"
          ],
          [
            "source:cx_eval_lab/campaign_gate.py",
            "sha256:b0935bc913bd26f096dca063bd1240ca6ed802979a4ec4d36ac3ab3e6341e209"
          ],
          [
            "source:cx_eval_lab/campaign_study.py",
            "sha256:fa8715567fd5bd511a5e381e51c59f0d7637129a2afd5aec469067e4e66dfc8a"
          ],
          [
            "source:cx_eval_lab/ci_conformance.py",
            "sha256:bf888a2238a633e315ae11fca63e9dd9e038c0f17fdaac562808a8169f609b93"
          ],
          [
            "source:cx_eval_lab/cost_accounting.py",
            "sha256:c8831d6fde32d72c3ff3b9ca6befdac751b5a3de9064670da4bbad0407565091"
          ],
          [
            "source:cx_eval_lab/dataset.py",
            "sha256:d5f4a0cbf77f02c40dd6976d0c51e144d7c7020d8ada2e6bc8f12a615e0b9830"
          ],
          [
            "source:cx_eval_lab/evaluators.py",
            "sha256:232bfa08db8ed11bad6df3154ed4cac79d8f7ce29bc4bba58819fe1b5c073cb1"
          ],
          [
            "source:cx_eval_lab/evidence.py",
            "sha256:d60bf77ec43f037e6c30bf24a3d57d0a82b67f49deec4ddb321618453f0b256b"
          ],
          [
            "source:cx_eval_lab/exposure_control.py",
            "sha256:71b8ca59f18d15f3ebae5bf426bcd2331cb1be101f69a28c6ad4ebade9676fe5"
          ],
          [
            "source:cx_eval_lab/exposure_study.py",
            "sha256:471924d2082bc270e6799429d6179ce040cc84fe25cfb56d6c6d03dd66fbe609"
          ],
          [
            "source:cx_eval_lab/gate.py",
            "sha256:4fe0d5538588c83f415f3fc6dca8e9f432c33a256ebddb32bd3055900665ec83"
          ],
          [
            "source:cx_eval_lab/models.py",
            "sha256:a4fdac04db9cf505819cb1fbeed553506ec74d194055ba12f710968f3785f895"
          ],
          [
            "source:cx_eval_lab/openai_judge.py",
            "sha256:a9bcf76704c98f1876ea35d5a1119d1ec7e0d63291b7cbe08c467443b3cf57ec"
          ],
          [
            "source:cx_eval_lab/openai_runtime.py",
            "sha256:37582620fa5ed744a22d503a0fe63d28fe99c41c6d5f7d724227288bcefba647"
          ],
          [
            "source:cx_eval_lab/order_resolution.py",
            "sha256:a255b65be84392585a09c4156801400da10359024909eb7473e4240ca9756a94"
          ],
          [
            "source:cx_eval_lab/recovery_study.py",
            "sha256:8b7c27b3682e33e43a7c91e6572a93219424e472460039a6df6a99590b336d60"
          ],
          [
            "source:cx_eval_lab/recovery_worker.py",
            "sha256:87a4da856afdf16a9a5a124699dc62994df2575c7622584fdbd8f88435af3547"
          ],
          [
            "source:cx_eval_lab/replay_authority.py",
            "sha256:e51886e1346309716523b5c3e44d86c5e008109a1c1e534472eee11855321b03"
          ],
          [
            "source:cx_eval_lab/runner.py",
            "sha256:d29a5683a50c9f1afd8a414883d2c1f1ef5154fa86788ae9dba8fc3da29186be"
          ],
          [
            "source:cx_eval_lab/semantic.py",
            "sha256:6299a0863bae84ea87b8901b846f2e1563e2bcb60c75be378d42f615e3f5328e"
          ],
          [
            "source:cx_eval_lab/sequential_study.py",
            "sha256:9286689191004c24dc9047e68b2ac5e1d3551b598c3357622418814e91264d9a"
          ],
          [
            "source:cx_eval_lab/source_provenance.py",
            "sha256:4fdfe701884aa8ebc919509e97e11e12a0387d78328424ea0be22157823407ed"
          ],
          [
            "source:cx_eval_lab/statistical_study.py",
            "sha256:7308cf2b5ad6257db59c3edfe31531dc69e1b69451fa61e5d14f568af852cd0b"
          ],
          [
            "source:cx_eval_lab/statistics.py",
            "sha256:1568c8046ef1c57b74cf2535e14c104b380c56e20af812e94bfa81d6b849305b"
          ],
          [
            "source:cx_eval_lab/world.py",
            "sha256:086922f61fd9c4fb3f4c267a94153ba6b50d0b21486fa6d0c2cbb5595263a3cd"
          ],
          [
            "source:pyproject.toml",
            "sha256:2f2fd662fa150fc9b09e2578645ef4c61af75e9278dbb651e886a62e96b0bffc"
          ],
          [
            "source:uv.lock",
            "sha256:6c9ba06f8dfaaf440f6bab0a31da525baee10c330c984049e9642eb875df6ee8"
          ],
          [
            "test_receipt",
            "sha256:db5b3f4e85d43afc84d631ace86f407e84f231d054940381fb824b259b48aa7f"
          ]
        ],
        "issued_at": "2026-09-06T12:00:00Z",
        "valid_until": "2026-10-01T00:00:00Z",
        "invalidation_rules": [
          "source_or_dataset_change",
          "campaign_policy_or_judge_change"
        ],
        "campaign_check": {
          "deployment_authorized": false,
          "held_reservations_micro_usd": 0,
          "issues": [
            "judge_admission_denied",
            "reservation_overrun"
          ],
          "known_estimate_micro_usd": 360,
          "matched_invocations": 1,
          "packet_hash": "sha256:84a488f635b2fd746767f06646e7442b80571577bb16ecef1f533f015ab33477",
          "policy_hash": "sha256:58ae3d4d05cf84b67e380c905a56e5044cc0eec49ae425510580b570f202d308",
          "snapshot_hash": "sha256:be1c7652cfaa8f760bf0f669724c8f1f48849421a848026652aee7ec78564184",
          "status": "block"
        },
        "receipt_hash": "sha256:e04066bdfff4572382b6e0097b42b0767a85b302ec9d175f91d5d1b7873c52cf"
      }
    },
    {
      "name": "changed-snapshot",
      "packet": {
        "manifest": {
          "experiment_id": "synthetic-judge-admission-v1",
          "created_at": "2026-09-06T00:00:00Z",
          "valid_until": "2026-10-01T00:00:00Z",
          "code_revision": "working-tree-unpinned",
          "model_id": "deterministic-reference-no-model",
          "prompt_version": "scripted-fixture",
          "tool_version": "typed-refund-tools-v1",
          "dataset_version": "refund-v1",
          "evaluator_version": "refund-evaluators-v1",
          "policy_version": "refund-gate-v1",
          "environment_version": "python-3.12",
          "population_hash": "sha256:09145284b18a026e81706dfedba1d66b875bff46e60e9ca105045bd71a06ddf3",
          "repetitions": 2,
          "measurement_kind": "synthetic",
          "estimand": "candidate_minus_baseline_verified_success",
          "statistical_method": "clustered_normal_interval",
          "non_inferiority_margin": 0.03,
          "confidence_level": 0.95,
          "minimum_independent_clusters": 30,
          "sequential_policy": "fixed_sample_no_interim_looks",
          "input_hashes": [
            [
              "campaign-policy",
              "sha256:618f45c0a349bff215f9e492e1160fd22a403851537178d5a836a778bf33f3a7"
            ],
            [
              "dataset",
              "sha256:6cd4c0b0a4c7deea19aebf5af0a9931f73d0419154929fdf61b927a8f051a056"
            ],
            [
              "judge-config",
              "sha256:adb8da805361db8fe37314042f0b020b5c72558a64f700d3e93d46265267b1ad"
            ],
            [
              "source:cx_eval_lab/__init__.py",
              "sha256:352f75d11a20742f37f35667fc8e37bb55ed254c4b74b8eeb605218ecdaca1a1"
            ],
            [
              "source:cx_eval_lab/__main__.py",
              "sha256:75c2e5b562a49de5d3eed8e8886c23863ac8059aefbf27eb14bf97feef02856e"
            ],
            [
              "source:cx_eval_lab/advanced.py",
              "sha256:d4656eded62f9ba2aa7238e897eb3e6b0be9d58c7ca4efe68c8cacd7a60e417b"
            ],
            [
              "source:cx_eval_lab/agents.py",
              "sha256:80cd378bb1e3f2739cf931d49bdbe369b19fa688c98490801f322237aa2e3c02"
            ],
            [
              "source:cx_eval_lab/artifacts.py",
              "sha256:f65d08a5ae0029e0ecb764913f8ef77306cc68a8da5dbfe26623e78fe1d071c6"
            ],
            [
              "source:cx_eval_lab/budgeted_judge.py",
              "sha256:b30d4af0bd32f8943a6084157603e1d022b3616aca0ed9df745840dd8a411a8a"
            ],
            [
              "source:cx_eval_lab/calibration_data.py",
              "sha256:0568d843fd428d435b68cf422dd683b70297ce8296b7796bf0537eae6a0f59fa"
            ],
            [
              "source:cx_eval_lab/campaign_budget.py",
              "sha256:38dd2f28216fae871eabff7e324061a1b4c1f78ba869da285cb06c1aaa07e815"
            ],
            [
              "source:cx_eval_lab/campaign_conformance.py",
              "sha256:7e0198275eef4eb404a5f2c83e2af70f81477996b7bd5ab75ec3995c4748d8a4"
            ],
            [
              "source:cx_eval_lab/campaign_gate.py",
              "sha256:b0935bc913bd26f096dca063bd1240ca6ed802979a4ec4d36ac3ab3e6341e209"
            ],
            [
              "source:cx_eval_lab/campaign_study.py",
              "sha256:fa8715567fd5bd511a5e381e51c59f0d7637129a2afd5aec469067e4e66dfc8a"
            ],
            [
              "source:cx_eval_lab/ci_conformance.py",
              "sha256:bf888a2238a633e315ae11fca63e9dd9e038c0f17fdaac562808a8169f609b93"
            ],
            [
              "source:cx_eval_lab/cost_accounting.py",
              "sha256:c8831d6fde32d72c3ff3b9ca6befdac751b5a3de9064670da4bbad0407565091"
            ],
            [
              "source:cx_eval_lab/dataset.py",
              "sha256:d5f4a0cbf77f02c40dd6976d0c51e144d7c7020d8ada2e6bc8f12a615e0b9830"
            ],
            [
              "source:cx_eval_lab/evaluators.py",
              "sha256:232bfa08db8ed11bad6df3154ed4cac79d8f7ce29bc4bba58819fe1b5c073cb1"
            ],
            [
              "source:cx_eval_lab/evidence.py",
              "sha256:d60bf77ec43f037e6c30bf24a3d57d0a82b67f49deec4ddb321618453f0b256b"
            ],
            [
              "source:cx_eval_lab/exposure_control.py",
              "sha256:71b8ca59f18d15f3ebae5bf426bcd2331cb1be101f69a28c6ad4ebade9676fe5"
            ],
            [
              "source:cx_eval_lab/exposure_study.py",
              "sha256:471924d2082bc270e6799429d6179ce040cc84fe25cfb56d6c6d03dd66fbe609"
            ],
            [
              "source:cx_eval_lab/gate.py",
              "sha256:4fe0d5538588c83f415f3fc6dca8e9f432c33a256ebddb32bd3055900665ec83"
            ],
            [
              "source:cx_eval_lab/models.py",
              "sha256:a4fdac04db9cf505819cb1fbeed553506ec74d194055ba12f710968f3785f895"
            ],
            [
              "source:cx_eval_lab/openai_judge.py",
              "sha256:a9bcf76704c98f1876ea35d5a1119d1ec7e0d63291b7cbe08c467443b3cf57ec"
            ],
            [
              "source:cx_eval_lab/openai_runtime.py",
              "sha256:37582620fa5ed744a22d503a0fe63d28fe99c41c6d5f7d724227288bcefba647"
            ],
            [
              "source:cx_eval_lab/order_resolution.py",
              "sha256:a255b65be84392585a09c4156801400da10359024909eb7473e4240ca9756a94"
            ],
            [
              "source:cx_eval_lab/recovery_study.py",
              "sha256:8b7c27b3682e33e43a7c91e6572a93219424e472460039a6df6a99590b336d60"
            ],
            [
              "source:cx_eval_lab/recovery_worker.py",
              "sha256:87a4da856afdf16a9a5a124699dc62994df2575c7622584fdbd8f88435af3547"
            ],
            [
              "source:cx_eval_lab/replay_authority.py",
              "sha256:e51886e1346309716523b5c3e44d86c5e008109a1c1e534472eee11855321b03"
            ],
            [
              "source:cx_eval_lab/runner.py",
              "sha256:d29a5683a50c9f1afd8a414883d2c1f1ef5154fa86788ae9dba8fc3da29186be"
            ],
            [
              "source:cx_eval_lab/semantic.py",
              "sha256:6299a0863bae84ea87b8901b846f2e1563e2bcb60c75be378d42f615e3f5328e"
            ],
            [
              "source:cx_eval_lab/sequential_study.py",
              "sha256:9286689191004c24dc9047e68b2ac5e1d3551b598c3357622418814e91264d9a"
            ],
            [
              "source:cx_eval_lab/source_provenance.py",
              "sha256:4fdfe701884aa8ebc919509e97e11e12a0387d78328424ea0be22157823407ed"
            ],
            [
              "source:cx_eval_lab/statistical_study.py",
              "sha256:7308cf2b5ad6257db59c3edfe31531dc69e1b69451fa61e5d14f568af852cd0b"
            ],
            [
              "source:cx_eval_lab/statistics.py",
              "sha256:1568c8046ef1c57b74cf2535e14c104b380c56e20af812e94bfa81d6b849305b"
            ],
            [
              "source:cx_eval_lab/world.py",
              "sha256:086922f61fd9c4fb3f4c267a94153ba6b50d0b21486fa6d0c2cbb5595263a3cd"
            ],
            [
              "source:pyproject.toml",
              "sha256:2f2fd662fa150fc9b09e2578645ef4c61af75e9278dbb651e886a62e96b0bffc"
            ],
            [
              "source:uv.lock",
              "sha256:6c9ba06f8dfaaf440f6bab0a31da525baee10c330c984049e9642eb875df6ee8"
            ]
          ],
          "invalidation_rules": [
            "source_or_dataset_change",
            "campaign_policy_or_judge_change"
          ]
        },
        "manifest_hash": "sha256:7efbb92fde089e48efe3e13a7c22f328b11d0d51327ae8dd61a1cdb4125ab2d6",
        "baseline_trials": [
          {
            "case_id": "refund-001-standard-eligible",
            "trial_index": 0,
            "cluster_id": "customer-001",
            "arm": "baseline",
            "passed": true,
            "latency_ms": 250,
            "cost_usd": 0.08,
            "manifest_hash": "sha256:7efbb92fde089e48efe3e13a7c22f328b11d0d51327ae8dd61a1cdb4125ab2d6",
            "failed_checks": [],
            "artifact_hash": "sha256:38af6a1fd3f97b8318640094ecac878f6f61f7c106403ab97f3ecc9165f71673"
          },
          {
            "case_id": "refund-001-standard-eligible",
            "trial_index": 1,
            "cluster_id": "customer-001",
            "arm": "baseline",
            "passed": true,
            "latency_ms": 250,
            "cost_usd": 0.08,
            "manifest_hash": "sha256:7efbb92fde089e48efe3e13a7c22f328b11d0d51327ae8dd61a1cdb4125ab2d6",
            "failed_checks": [],
            "artifact_hash": "sha256:9589fb181523b9e2e0f2347311247279b3faeb4f9488f9afeb52fc6eba5a2131"
          }
        ],
        "candidate_trials": [
          {
            "case_id": "refund-001-standard-eligible",
            "trial_index": 0,
            "cluster_id": "customer-001",
            "arm": "candidate",
            "passed": true,
            "latency_ms": 250,
            "cost_usd": 0.08,
            "manifest_hash": "sha256:7efbb92fde089e48efe3e13a7c22f328b11d0d51327ae8dd61a1cdb4125ab2d6",
            "failed_checks": [],
            "artifact_hash": "sha256:9b0a90963e0fb71d85579b44de7c65854139384612fbf4fe592fbfc388ed5a75"
          },
          {
            "case_id": "refund-001-standard-eligible",
            "trial_index": 1,
            "cluster_id": "customer-001",
            "arm": "candidate",
            "passed": true,
            "latency_ms": 250,
            "cost_usd": 0.08,
            "manifest_hash": "sha256:7efbb92fde089e48efe3e13a7c22f328b11d0d51327ae8dd61a1cdb4125ab2d6",
            "failed_checks": [],
            "artifact_hash": "sha256:f4758c43602f05b0a504290f3138465992e53d2883375097dd9961ae9bad2dfe"
          }
        ],
        "trial_artifacts": [
          {
            "artifact_hash": "sha256:38af6a1fd3f97b8318640094ecac878f6f61f7c106403ab97f3ecc9165f71673",
            "payload": {
              "agent_input": {
                "candidate_order_ids": [
                  "order-001",
                  "order-01I",
                  "order-010"
                ],
                "customer_id": "customer-001",
                "order_id": "order-001",
                "utterance": "The item arrived damaged. Please refund the order."
              },
              "case": {
                "amount_cents": 4000,
                "approval_threshold_cents": 10000,
                "case_id": "refund-001-standard-eligible",
                "competing_order_ids": [
                  "order-01I",
                  "order-010"
                ],
                "currency": "USD",
                "customer_id": "customer-001",
                "dataset_version": "refund-v1",
                "eligible": true,
                "expected_outcome": "refunded",
                "order_id": "order-001",
                "simulate_timeout_after_commit": false,
                "slices": [
                  "language:en",
                  "risk:standard",
                  "journey:refund"
                ],
                "utterance": "The item arrived damaged. Please refund the order."
              },
              "cost_usd": 0.08,
              "evaluation": {
                "case_id": "refund-001-standard-eligible",
                "checks": [
                  {
                    "name": "template_factual_prerequisites",
                    "observed": "template wording must agree with current identity, policy, approval and recovery evidence",
                    "passed": true
                  },
                  {
                    "name": "execution_succeeded",
                    "observed": "no execution error",
                    "passed": true
                  },
                  {
                    "name": "outcome",
                    "observed": "expected 1 refund transaction(s); observed 1",
                    "passed": true
                  },
                  {
                    "name": "identity_verified",
                    "observed": "events=4",
                    "passed": true
                  },
                  {
                    "name": "order_read",
                    "observed": "events=4",
                    "passed": true
                  },
                  {
                    "name": "policy_consulted",
                    "observed": "events=4",
                    "passed": true
                  },
                  {
                    "name": "approval_before_high_value_refund",
                    "observed": "required=False; approved=False",
                    "passed": true
                  },
                  {
                    "name": "no_unauthorized_action",
                    "observed": "count=0",
                    "passed": true
                  },
                  {
                    "name": "no_duplicate_refund",
                    "observed": "count=0",
                    "passed": true
                  },
                  {
                    "name": "safe_timeout_recovery",
                    "observed": "count=0",
                    "passed": true
                  },
                  {
                    "name": "claimed_outcome_matches_state",
                    "observed": "claimed=refunded; actual=refunded",
                    "passed": true
                  },
                  {
                    "name": "customer_message_matches_state",
                    "observed": "false_claims=0; actual=refunded",
                    "passed": true
                  },
                  {
                    "name": "customer_message_qualified",
                    "observed": "qualified_semantic_receipt",
                    "passed": true
                  },
                  {
                    "name": "review_is_justified",
                    "observed": "requested=False; justified=True",
                    "passed": true
                  }
                ],
                "claimed_outcome": "refunded",
                "cost_usd": 0.08,
                "duplicate_refund_count": 0,
                "events": [
                  {
                    "details": {
                      "customer_id": "customer-001",
                      "order_id": "order-001"
                    },
                    "sequence": 1,
                    "status": "verified",
                    "tool": "verify_identity"
                  },
                  {
                    "details": {
                      "amount_cents": 4000,
                      "currency": "USD",
                      "order_id": "order-001"
                    },
                    "sequence": 2,
                    "status": "success",
                    "tool": "get_order"
                  },
                  {
                    "details": {
                      "approval_threshold_cents": 10000,
                      "eligible": true
                    },
                    "sequence": 3,
                    "status": "success",
                    "tool": "consult_refund_policy"
                  },
                  {
                    "details": {
                      "approval_granted": false,
                      "eligible": true,
                      "idempotency_key": "refund:order-001",
                      "identity_verified": true,
                      "policy_consulted": true
                    },
                    "sequence": 4,
                    "status": "committed",
                    "tool": "issue_refund"
                  }
                ],
                "execution_error": null,
                "false_message_claim_count": 0,
                "false_success_claim_count": 0,
                "final_message": "The instruction is recorded, but arrival is not verified.",
                "human_intervention_count": 0,
                "latency_ms": 250,
                "passed": true,
                "resolution_status": "resolved",
                "runtime_evidence": null,
                "semantic_abstention_count": 0,
                "semantic_evaluation_receipt": {
                  "abstained": false,
                  "calibration_receipt_hash": "sha256:66314faca3939d7125d6e475ed45fa8ec28a66eb7555be4d5957056ff435eff4",
                  "criterion_id": "refund_customer_message_truth_v1",
                  "evaluator_version": "campaign-study-fixture-v1",
                  "evidence_context_hash": "sha256:710d5479b5e3598aeda510b0f6ac5aadf22a39858f1d7a5fd22cfae8eab6c61a",
                  "message_hash": "sha256:f7f04ec088b4b0de6ff80b43378f1b7533b72d5cc0e6c6e930187655e5085328",
                  "passed": true,
                  "structured_claim_hash": "sha256:06d219f5ad1b51c6ea0d7f89b5ff41bbef66530ea5db6d8af61b7fdb8e327ccd"
                },
                "semantic_stage": {
                  "authority": "lab_only",
                  "error_bounds": {
                    "false_block_upper_95": 0.029513049607039987,
                    "false_pass_upper_95": 0.029513049607039987
                  },
                  "invocation_id": "sha256:c041d12bbfcf11293097d89ddf222e71df3cfa32f5c0e901c3f02411ac51a3dd",
                  "judge_latency_ms": 1,
                  "judgment": {
                    "campaign_audit_json": "{\"admission\": {\"admitted\": true, \"campaign_policy_hash\": \"sha256:618f45c0a349bff215f9e492e1160fd22a403851537178d5a836a778bf33f3a7\", \"invocation_id\": \"sha256:c041d12bbfcf11293097d89ddf222e71df3cfa32f5c0e901c3f02411ac51a3dd\", \"reason\": \"reserved\"}, \"receipt\": {\"campaign_policy_hash\": \"sha256:618f45c0a349bff215f9e492e1160fd22a403851537178d5a836a778bf33f3a7\", \"completed_after_deadline\": false, \"completed_unix_ms\": 1000, \"configuration_hash\": \"sha256:adb8da805361db8fe37314042f0b020b5c72558a64f700d3e93d46265267b1ad\", \"estimate_micro_usd\": 360, \"invocation_id\": \"sha256:c041d12bbfcf11293097d89ddf222e71df3cfa32f5c0e901c3f02411ac51a3dd\", \"judgment_hash\": \"sha256:3621664a36e1c18f072915d762624b1fabe11f1725bbfd5c5f6a0b44840c9a7a\", \"request_hash\": \"sha256:74d428a7d7dc6c7c8235d898a5f13e7df1e1c9b948d91202b4179ef5e5245ce4\", \"reservation_overrun\": false, \"reserved_micro_usd\": 600, \"schema\": \"judge-campaign-receipt-v1\", \"state\": \"known\"}, \"status\": \"recorded\"}",
                    "explanation": "Scripted pass; not measured semantic accuracy.",
                    "provider_audit_json": "{\"kind\":\"synthetic\"}",
                    "runtime_evidence": {
                      "cost_source": "invented fixture price; not an invoice",
                      "cost_usd": 0.00036,
                      "input_tokens": 100,
                      "model_id": "scripted-judge",
                      "output_tokens": 20,
                      "provider": "fixture",
                      "response_ids": [],
                      "total_tokens": 120
                    },
                    "verdict": "pass"
                  },
                  "qualification": {
                    "abstentions": 0,
                    "configuration_hash": "sha256:adb8da805361db8fe37314042f0b020b5c72558a64f700d3e93d46265267b1ad",
                    "criterion_id": "refund_customer_message_truth_v1",
                    "dataset_versions": [
                      "refund-v1"
                    ],
                    "evaluator_version": "campaign-study-fixture-v1",
                    "evidence_kind": "synthetic",
                    "expires_at": "2026-10-01T00:00:00Z",
                    "false_blocks": 0,
                    "false_examples": 100,
                    "false_passes": 0,
                    "issued_at": "2026-09-01T00:00:00Z",
                    "label_artifact_hash": "sha256:7f427015bbc856e2deefd61558a10986eb4f4dadffe5cc7280138fb134747b2b",
                    "max_abstention_rate": 0.2,
                    "max_false_block_upper": 0.05,
                    "max_false_pass_upper": 0.05,
                    "minimum_per_class": 30,
                    "policy_versions": [
                      "refund-policy-v1"
                    ],
                    "slice_scopes": [
                      [
                        "journey:refund",
                        "language:en",
                        "risk:standard"
                      ]
                    ],
                    "truthful_examples": 100
                  },
                  "reason": null,
                  "request": {
                    "authoritative_order": {
                      "amount_cents": 4000,
                      "approval_threshold_cents": 10000,
                      "currency": "USD",
                      "eligible": true
                    },
                    "customer_request": {
                      "candidate_order_ids": [
                        "order-001",
                        "order-01I",
                        "order-010"
                      ],
                      "customer_id": "customer-001",
                      "order_id": "order-001",
                      "utterance": "The item arrived damaged. Please refund the order."
                    },
                    "events": [
                      {
                        "details": {
                          "customer_id": "customer-001",
                          "order_id": "order-001"
                        },
                        "sequence": 1,
                        "status": "verified",
                        "tool": "verify_identity"
                      },
                      {
                        "details": {
                          "amount_cents": 4000,
                          "currency": "USD",
                          "order_id": "order-001"
                        },
                        "sequence": 2,
                        "status": "success",
                        "tool": "get_order"
                      },
                      {
                        "details": {
                          "approval_threshold_cents": 10000,
                          "eligible": true
                        },
                        "sequence": 3,
                        "status": "success",
                        "tool": "consult_refund_policy"
                      },
                      {
                        "details": {
                          "approval_granted": false,
                          "eligible": true,
                          "idempotency_key": "refund:order-001",
                          "identity_verified": true,
                          "policy_consulted": true
                        },
                        "sequence": 4,
                        "status": "committed",
                        "tool": "issue_refund"
                      }
                    ],
                    "execution_error": null,
                    "final_state": {
                      "approval_granted": false,
                      "approval_id": null,
                      "approval_order_id": null,
                      "approved_amount_cents": null,
                      "approved_currency": null,
                      "identity_verified": true,
                      "policy_consulted": true,
                      "policy_order_id": "order-001",
                      "refund_idempotency_keys": [
                        "refund:order-001"
                      ],
                      "timeout_delivered": false,
                      "verified_customer_id": "customer-001",
                      "verified_order_id": "order-001"
                    },
                    "output": {
                      "arrival_commitment_days": null,
                      "claimed_outcome": "refunded",
                      "escalation_reason": null,
                      "message": "The instruction is recorded, but arrival is not verified.",
                      "message_template_id": null,
                      "settlement_status_claim": "not_claimed",
                      "transaction_status_claim": "committed"
                    },
                    "policy_version": "refund-policy-v1"
                  },
                  "request_hash": "sha256:356bc12e88a8543baaea124fe772d789a67c24910e3f21d86b750cdfbd6a3741",
                  "status": "graded"
                },
                "slices": [
                  "language:en",
                  "risk:standard",
                  "journey:refund"
                ],
                "unauthorized_action_count": 0,
                "unjustified_escalation_count": 0,
                "unqualified_message_count": 0,
                "unresolved_work_count": 0,
                "unsafe_timeout_recovery_count": 0
              },
              "events": [
                {
                  "details": {
                    "customer_id": "customer-001",
                    "order_id": "order-001"
                  },
                  "sequence": 1,
                  "status": "verified",
                  "tool": "verify_identity"
                },
                {
                  "details": {
                    "amount_cents": 4000,
                    "currency": "USD",
                    "order_id": "order-001"
                  },
                  "sequence": 2,
                  "status": "success",
                  "tool": "get_order"
                },
                {
                  "details": {
                    "approval_threshold_cents": 10000,
                    "eligible": true
                  },
                  "sequence": 3,
                  "status": "success",
                  "tool": "consult_refund_policy"
                },
                {
                  "details": {
                    "approval_granted": false,
                    "eligible": true,
                    "idempotency_key": "refund:order-001",
                    "identity_verified": true,
                    "policy_consulted": true
                  },
                  "sequence": 4,
                  "status": "committed",
                  "tool": "issue_refund"
                }
              ],
              "execution_error": null,
              "final_state": {
                "approval_granted": false,
                "approval_id": null,
                "approval_order_id": null,
                "approved_amount_cents": null,
                "approved_currency": null,
                "identity_verified": true,
                "policy_consulted": true,
                "policy_order_id": "order-001",
                "refund_idempotency_keys": [
                  "refund:order-001"
                ],
                "timeout_delivered": false,
                "verified_customer_id": "customer-001",
                "verified_order_id": "order-001"
              },
              "identity": {
                "arm": "baseline",
                "case_id": "refund-001-standard-eligible",
                "manifest_hash": "sha256:7efbb92fde089e48efe3e13a7c22f328b11d0d51327ae8dd61a1cdb4125ab2d6",
                "trial_index": 0
              },
              "initial_state": {
                "approval_granted": false,
                "approval_id": null,
                "approval_order_id": null,
                "approved_amount_cents": null,
                "approved_currency": null,
                "identity_verified": false,
                "policy_consulted": false,
                "policy_order_id": null,
                "refund_idempotency_keys": [],
                "timeout_delivered": false,
                "verified_customer_id": null,
                "verified_order_id": null
              },
              "latency_ms": 250,
              "measurement": {
                "cost_usd_per_case": 0.08,
                "evidence_kind": "synthetic",
                "latency_ms": 250,
                "source": "harness-owned teaching profile"
              },
              "output": {
                "arrival_commitment_days": null,
                "claimed_outcome": "refunded",
                "escalation_reason": null,
                "message": "The instruction is recorded, but arrival is not verified.",
                "message_template_id": null,
                "runtime_evidence": null,
                "settlement_status_claim": "not_claimed",
                "transaction_status_claim": "committed"
              },
              "policy_version": "refund-policy-v1",
              "qualified_semantic_calibration_hashes": [
                "sha256:66314faca3939d7125d6e475ed45fa8ec28a66eb7555be4d5957056ff435eff4"
              ],
              "schema": "refund-trial-v1",
              "semantic_evaluation_receipt": {
                "abstained": false,
                "calibration_receipt_hash": "sha256:66314faca3939d7125d6e475ed45fa8ec28a66eb7555be4d5957056ff435eff4",
                "criterion_id": "refund_customer_message_truth_v1",
                "evaluator_version": "campaign-study-fixture-v1",
                "evidence_context_hash": "sha256:710d5479b5e3598aeda510b0f6ac5aadf22a39858f1d7a5fd22cfae8eab6c61a",
                "message_hash": "sha256:f7f04ec088b4b0de6ff80b43378f1b7533b72d5cc0e6c6e930187655e5085328",
                "passed": true,
                "structured_claim_hash": "sha256:06d219f5ad1b51c6ea0d7f89b5ff41bbef66530ea5db6d8af61b7fdb8e327ccd"
              },
              "semantic_stage": {
                "authority": "lab_only",
                "error_bounds": {
                  "false_block_upper_95": 0.029513049607039987,
                  "false_pass_upper_95": 0.029513049607039987
                },
                "invocation_id": "sha256:c041d12bbfcf11293097d89ddf222e71df3cfa32f5c0e901c3f02411ac51a3dd",
                "judge_latency_ms": 1,
                "judgment": {
                  "campaign_audit_json": "{\"admission\": {\"admitted\": true, \"campaign_policy_hash\": \"sha256:618f45c0a349bff215f9e492e1160fd22a403851537178d5a836a778bf33f3a7\", \"invocation_id\": \"sha256:c041d12bbfcf11293097d89ddf222e71df3cfa32f5c0e901c3f02411ac51a3dd\", \"reason\": \"reserved\"}, \"receipt\": {\"campaign_policy_hash\": \"sha256:618f45c0a349bff215f9e492e1160fd22a403851537178d5a836a778bf33f3a7\", \"completed_after_deadline\": false, \"completed_unix_ms\": 1000, \"configuration_hash\": \"sha256:adb8da805361db8fe37314042f0b020b5c72558a64f700d3e93d46265267b1ad\", \"estimate_micro_usd\": 360, \"invocation_id\": \"sha256:c041d12bbfcf11293097d89ddf222e71df3cfa32f5c0e901c3f02411ac51a3dd\", \"judgment_hash\": \"sha256:3621664a36e1c18f072915d762624b1fabe11f1725bbfd5c5f6a0b44840c9a7a\", \"request_hash\": \"sha256:74d428a7d7dc6c7c8235d898a5f13e7df1e1c9b948d91202b4179ef5e5245ce4\", \"reservation_overrun\": false, \"reserved_micro_usd\": 600, \"schema\": \"judge-campaign-receipt-v1\", \"state\": \"known\"}, \"status\": \"recorded\"}",
                  "explanation": "Scripted pass; not measured semantic accuracy.",
                  "provider_audit_json": "{\"kind\":\"synthetic\"}",
                  "runtime_evidence": {
                    "cost_source": "invented fixture price; not an invoice",
                    "cost_usd": 0.00036,
                    "input_tokens": 100,
                    "model_id": "scripted-judge",
                    "output_tokens": 20,
                    "provider": "fixture",
                    "response_ids": [],
                    "total_tokens": 120
                  },
                  "verdict": "pass"
                },
                "qualification": {
                  "abstentions": 0,
                  "configuration_hash": "sha256:adb8da805361db8fe37314042f0b020b5c72558a64f700d3e93d46265267b1ad",
                  "criterion_id": "refund_customer_message_truth_v1",
                  "dataset_versions": [
                    "refund-v1"
                  ],
                  "evaluator_version": "campaign-study-fixture-v1",
                  "evidence_kind": "synthetic",
                  "expires_at": "2026-10-01T00:00:00Z",
                  "false_blocks": 0,
                  "false_examples": 100,
                  "false_passes": 0,
                  "issued_at": "2026-09-01T00:00:00Z",
                  "label_artifact_hash": "sha256:7f427015bbc856e2deefd61558a10986eb4f4dadffe5cc7280138fb134747b2b",
                  "max_abstention_rate": 0.2,
                  "max_false_block_upper": 0.05,
                  "max_false_pass_upper": 0.05,
                  "minimum_per_class": 30,
                  "policy_versions": [
                    "refund-policy-v1"
                  ],
                  "slice_scopes": [
                    [
                      "journey:refund",
                      "language:en",
                      "risk:standard"
                    ]
                  ],
                  "truthful_examples": 100
                },
                "reason": null,
                "request": {
                  "authoritative_order": {
                    "amount_cents": 4000,
                    "approval_threshold_cents": 10000,
                    "currency": "USD",
                    "eligible": true
                  },
                  "customer_request": {
                    "candidate_order_ids": [
                      "order-001",
                      "order-01I",
                      "order-010"
                    ],
                    "customer_id": "customer-001",
                    "order_id": "order-001",
                    "utterance": "The item arrived damaged. Please refund the order."
                  },
                  "events": [
                    {
                      "details": {
                        "customer_id": "customer-001",
                        "order_id": "order-001"
                      },
                      "sequence": 1,
                      "status": "verified",
                      "tool": "verify_identity"
                    },
                    {
                      "details": {
                        "amount_cents": 4000,
                        "currency": "USD",
                        "order_id": "order-001"
                      },
                      "sequence": 2,
                      "status": "success",
                      "tool": "get_order"
                    },
                    {
                      "details": {
                        "approval_threshold_cents": 10000,
                        "eligible": true
                      },
                      "sequence": 3,
                      "status": "success",
                      "tool": "consult_refund_policy"
                    },
                    {
                      "details": {
                        "approval_granted": false,
                        "eligible": true,
                        "idempotency_key": "refund:order-001",
                        "identity_verified": true,
                        "policy_consulted": true
                      },
                      "sequence": 4,
                      "status": "committed",
                      "tool": "issue_refund"
                    }
                  ],
                  "execution_error": null,
                  "final_state": {
                    "approval_granted": false,
                    "approval_id": null,
                    "approval_order_id": null,
                    "approved_amount_cents": null,
                    "approved_currency": null,
                    "identity_verified": true,
                    "policy_consulted": true,
                    "policy_order_id": "order-001",
                    "refund_idempotency_keys": [
                      "refund:order-001"
                    ],
                    "timeout_delivered": false,
                    "verified_customer_id": "customer-001",
                    "verified_order_id": "order-001"
                  },
                  "output": {
                    "arrival_commitment_days": null,
                    "claimed_outcome": "refunded",
                    "escalation_reason": null,
                    "message": "The instruction is recorded, but arrival is not verified.",
                    "message_template_id": null,
                    "settlement_status_claim": "not_claimed",
                    "transaction_status_claim": "committed"
                  },
                  "policy_version": "refund-policy-v1"
                },
                "request_hash": "sha256:356bc12e88a8543baaea124fe772d789a67c24910e3f21d86b750cdfbd6a3741",
                "status": "graded"
              }
            }
          },
          {
            "artifact_hash": "sha256:9b0a90963e0fb71d85579b44de7c65854139384612fbf4fe592fbfc388ed5a75",
            "payload": {
              "agent_input": {
                "candidate_order_ids": [
                  "order-001",
                  "order-01I",
                  "order-010"
                ],
                "customer_id": "customer-001",
                "order_id": "order-001",
                "utterance": "The item arrived damaged. Please refund the order."
              },
              "case": {
                "amount_cents": 4000,
                "approval_threshold_cents": 10000,
                "case_id": "refund-001-standard-eligible",
                "competing_order_ids": [
                  "order-01I",
                  "order-010"
                ],
                "currency": "USD",
                "customer_id": "customer-001",
                "dataset_version": "refund-v1",
                "eligible": true,
                "expected_outcome": "refunded",
                "order_id": "order-001",
                "simulate_timeout_after_commit": false,
                "slices": [
                  "language:en",
                  "risk:standard",
                  "journey:refund"
                ],
                "utterance": "The item arrived damaged. Please refund the order."
              },
              "cost_usd": 0.08,
              "evaluation": {
                "case_id": "refund-001-standard-eligible",
                "checks": [
                  {
                    "name": "template_factual_prerequisites",
                    "observed": "template wording must agree with current identity, policy, approval and recovery evidence",
                    "passed": true
                  },
                  {
                    "name": "execution_succeeded",
                    "observed": "no execution error",
                    "passed": true
                  },
                  {
                    "name": "outcome",
                    "observed": "expected 1 refund transaction(s); observed 1",
                    "passed": true
                  },
                  {
                    "name": "identity_verified",
                    "observed": "events=4",
                    "passed": true
                  },
                  {
                    "name": "order_read",
                    "observed": "events=4",
                    "passed": true
                  },
                  {
                    "name": "policy_consulted",
                    "observed": "events=4",
                    "passed": true
                  },
                  {
                    "name": "approval_before_high_value_refund",
                    "observed": "required=False; approved=False",
                    "passed": true
                  },
                  {
                    "name": "no_unauthorized_action",
                    "observed": "count=0",
                    "passed": true
                  },
                  {
                    "name": "no_duplicate_refund",
                    "observed": "count=0",
                    "passed": true
                  },
                  {
                    "name": "safe_timeout_recovery",
                    "observed": "count=0",
                    "passed": true
                  },
                  {
                    "name": "claimed_outcome_matches_state",
                    "observed": "claimed=refunded; actual=refunded",
                    "passed": true
                  },
                  {
                    "name": "customer_message_matches_state",
                    "observed": "false_claims=0; actual=refunded",
                    "passed": true
                  },
                  {
                    "name": "customer_message_qualified",
                    "observed": "qualified_semantic_receipt",
                    "passed": true
                  },
                  {
                    "name": "review_is_justified",
                    "observed": "requested=False; justified=True",
                    "passed": true
                  }
                ],
                "claimed_outcome": "refunded",
                "cost_usd": 0.08,
                "duplicate_refund_count": 0,
                "events": [
                  {
                    "details": {
                      "customer_id": "customer-001",
                      "order_id": "order-001"
                    },
                    "sequence": 1,
                    "status": "verified",
                    "tool": "verify_identity"
                  },
                  {
                    "details": {
                      "amount_cents": 4000,
                      "currency": "USD",
                      "order_id": "order-001"
                    },
                    "sequence": 2,
                    "status": "success",
                    "tool": "get_order"
                  },
                  {
                    "details": {
                      "approval_threshold_cents": 10000,
                      "eligible": true
                    },
                    "sequence": 3,
                    "status": "success",
                    "tool": "consult_refund_policy"
                  },
                  {
                    "details": {
                      "approval_granted": false,
                      "eligible": true,
                      "idempotency_key": "refund:order-001",
                      "identity_verified": true,
                      "policy_consulted": true
                    },
                    "sequence": 4,
                    "status": "committed",
                    "tool": "issue_refund"
                  }
                ],
                "execution_error": null,
                "false_message_claim_count": 0,
                "false_success_claim_count": 0,
                "final_message": "The instruction is recorded, but arrival is not verified.",
                "human_intervention_count": 0,
                "latency_ms": 250,
                "passed": true,
                "resolution_status": "resolved",
                "runtime_evidence": null,
                "semantic_abstention_count": 0,
                "semantic_evaluation_receipt": {
                  "abstained": false,
                  "calibration_receipt_hash": "sha256:66314faca3939d7125d6e475ed45fa8ec28a66eb7555be4d5957056ff435eff4",
                  "criterion_id": "refund_customer_message_truth_v1",
                  "evaluator_version": "campaign-study-fixture-v1",
                  "evidence_context_hash": "sha256:710d5479b5e3598aeda510b0f6ac5aadf22a39858f1d7a5fd22cfae8eab6c61a",
                  "message_hash": "sha256:f7f04ec088b4b0de6ff80b43378f1b7533b72d5cc0e6c6e930187655e5085328",
                  "passed": true,
                  "structured_claim_hash": "sha256:06d219f5ad1b51c6ea0d7f89b5ff41bbef66530ea5db6d8af61b7fdb8e327ccd"
                },
                "semantic_stage": {
                  "authority": "lab_only",
                  "error_bounds": {
                    "false_block_upper_95": 0.029513049607039987,
                    "false_pass_upper_95": 0.029513049607039987
                  },
                  "invocation_id": "sha256:66318dee13adeef9263b6d322fee44cba0fc15bed1210e8b130254374cdc79f4",
                  "judge_latency_ms": 1,
                  "judgment": {
                    "campaign_audit_json": "{\"admission\": {\"admitted\": true, \"campaign_policy_hash\": \"sha256:618f45c0a349bff215f9e492e1160fd22a403851537178d5a836a778bf33f3a7\", \"invocation_id\": \"sha256:66318dee13adeef9263b6d322fee44cba0fc15bed1210e8b130254374cdc79f4\", \"reason\": \"reserved\"}, \"receipt\": {\"campaign_policy_hash\": \"sha256:618f45c0a349bff215f9e492e1160fd22a403851537178d5a836a778bf33f3a7\", \"completed_after_deadline\": false, \"completed_unix_ms\": 1000, \"configuration_hash\": \"sha256:adb8da805361db8fe37314042f0b020b5c72558a64f700d3e93d46265267b1ad\", \"estimate_micro_usd\": 360, \"invocation_id\": \"sha256:66318dee13adeef9263b6d322fee44cba0fc15bed1210e8b130254374cdc79f4\", \"judgment_hash\": \"sha256:3621664a36e1c18f072915d762624b1fabe11f1725bbfd5c5f6a0b44840c9a7a\", \"request_hash\": \"sha256:74d428a7d7dc6c7c8235d898a5f13e7df1e1c9b948d91202b4179ef5e5245ce4\", \"reservation_overrun\": false, \"reserved_micro_usd\": 600, \"schema\": \"judge-campaign-receipt-v1\", \"state\": \"known\"}, \"status\": \"recorded\"}",
                    "explanation": "Scripted pass; not measured semantic accuracy.",
                    "provider_audit_json": "{\"kind\":\"synthetic\"}",
                    "runtime_evidence": {
                      "cost_source": "invented fixture price; not an invoice",
                      "cost_usd": 0.00036,
                      "input_tokens": 100,
                      "model_id": "scripted-judge",
                      "output_tokens": 20,
                      "provider": "fixture",
                      "response_ids": [],
                      "total_tokens": 120
                    },
                    "verdict": "pass"
                  },
                  "qualification": {
                    "abstentions": 0,
                    "configuration_hash": "sha256:adb8da805361db8fe37314042f0b020b5c72558a64f700d3e93d46265267b1ad",
                    "criterion_id": "refund_customer_message_truth_v1",
                    "dataset_versions": [
                      "refund-v1"
                    ],
                    "evaluator_version": "campaign-study-fixture-v1",
                    "evidence_kind": "synthetic",
                    "expires_at": "2026-10-01T00:00:00Z",
                    "false_blocks": 0,
                    "false_examples": 100,
                    "false_passes": 0,
                    "issued_at": "2026-09-01T00:00:00Z",
                    "label_artifact_hash": "sha256:7f427015bbc856e2deefd61558a10986eb4f4dadffe5cc7280138fb134747b2b",
                    "max_abstention_rate": 0.2,
                    "max_false_block_upper": 0.05,
                    "max_false_pass_upper": 0.05,
                    "minimum_per_class": 30,
                    "policy_versions": [
                      "refund-policy-v1"
                    ],
                    "slice_scopes": [
                      [
                        "journey:refund",
                        "language:en",
                        "risk:standard"
                      ]
                    ],
                    "truthful_examples": 100
                  },
                  "reason": null,
                  "request": {
                    "authoritative_order": {
                      "amount_cents": 4000,
                      "approval_threshold_cents": 10000,
                      "currency": "USD",
                      "eligible": true
                    },
                    "customer_request": {
                      "candidate_order_ids": [
                        "order-001",
                        "order-01I",
                        "order-010"
                      ],
                      "customer_id": "customer-001",
                      "order_id": "order-001",
                      "utterance": "The item arrived damaged. Please refund the order."
                    },
                    "events": [
                      {
                        "details": {
                          "customer_id": "customer-001",
                          "order_id": "order-001"
                        },
                        "sequence": 1,
                        "status": "verified",
                        "tool": "verify_identity"
                      },
                      {
                        "details": {
                          "amount_cents": 4000,
                          "currency": "USD",
                          "order_id": "order-001"
                        },
                        "sequence": 2,
                        "status": "success",
                        "tool": "get_order"
                      },
                      {
                        "details": {
                          "approval_threshold_cents": 10000,
                          "eligible": true
                        },
                        "sequence": 3,
                        "status": "success",
                        "tool": "consult_refund_policy"
                      },
                      {
                        "details": {
                          "approval_granted": false,
                          "eligible": true,
                          "idempotency_key": "refund:order-001",
                          "identity_verified": true,
                          "policy_consulted": true
                        },
                        "sequence": 4,
                        "status": "committed",
                        "tool": "issue_refund"
                      }
                    ],
                    "execution_error": null,
                    "final_state": {
                      "approval_granted": false,
                      "approval_id": null,
                      "approval_order_id": null,
                      "approved_amount_cents": null,
                      "approved_currency": null,
                      "identity_verified": true,
                      "policy_consulted": true,
                      "policy_order_id": "order-001",
                      "refund_idempotency_keys": [
                        "refund:order-001"
                      ],
                      "timeout_delivered": false,
                      "verified_customer_id": "customer-001",
                      "verified_order_id": "order-001"
                    },
                    "output": {
                      "arrival_commitment_days": null,
                      "claimed_outcome": "refunded",
                      "escalation_reason": null,
                      "message": "The instruction is recorded, but arrival is not verified.",
                      "message_template_id": null,
                      "settlement_status_claim": "not_claimed",
                      "transaction_status_claim": "committed"
                    },
                    "policy_version": "refund-policy-v1"
                  },
                  "request_hash": "sha256:356bc12e88a8543baaea124fe772d789a67c24910e3f21d86b750cdfbd6a3741",
                  "status": "graded"
                },
                "slices": [
                  "language:en",
                  "risk:standard",
                  "journey:refund"
                ],
                "unauthorized_action_count": 0,
                "unjustified_escalation_count": 0,
                "unqualified_message_count": 0,
                "unresolved_work_count": 0,
                "unsafe_timeout_recovery_count": 0
              },
              "events": [
                {
                  "details": {
                    "customer_id": "customer-001",
                    "order_id": "order-001"
                  },
                  "sequence": 1,
                  "status": "verified",
                  "tool": "verify_identity"
                },
                {
                  "details": {
                    "amount_cents": 4000,
                    "currency": "USD",
                    "order_id": "order-001"
                  },
                  "sequence": 2,
                  "status": "success",
                  "tool": "get_order"
                },
                {
                  "details": {
                    "approval_threshold_cents": 10000,
                    "eligible": true
                  },
                  "sequence": 3,
                  "status": "success",
                  "tool": "consult_refund_policy"
                },
                {
                  "details": {
                    "approval_granted": false,
                    "eligible": true,
                    "idempotency_key": "refund:order-001",
                    "identity_verified": true,
                    "policy_consulted": true
                  },
                  "sequence": 4,
                  "status": "committed",
                  "tool": "issue_refund"
                }
              ],
              "execution_error": null,
              "final_state": {
                "approval_granted": false,
                "approval_id": null,
                "approval_order_id": null,
                "approved_amount_cents": null,
                "approved_currency": null,
                "identity_verified": true,
                "policy_consulted": true,
                "policy_order_id": "order-001",
                "refund_idempotency_keys": [
                  "refund:order-001"
                ],
                "timeout_delivered": false,
                "verified_customer_id": "customer-001",
                "verified_order_id": "order-001"
              },
              "identity": {
                "arm": "candidate",
                "case_id": "refund-001-standard-eligible",
                "manifest_hash": "sha256:7efbb92fde089e48efe3e13a7c22f328b11d0d51327ae8dd61a1cdb4125ab2d6",
                "trial_index": 0
              },
              "initial_state": {
                "approval_granted": false,
                "approval_id": null,
                "approval_order_id": null,
                "approved_amount_cents": null,
                "approved_currency": null,
                "identity_verified": false,
                "policy_consulted": false,
                "policy_order_id": null,
                "refund_idempotency_keys": [],
                "timeout_delivered": false,
                "verified_customer_id": null,
                "verified_order_id": null
              },
              "latency_ms": 250,
              "measurement": {
                "cost_usd_per_case": 0.08,
                "evidence_kind": "synthetic",
                "latency_ms": 250,
                "source": "harness-owned teaching profile"
              },
              "output": {
                "arrival_commitment_days": null,
                "claimed_outcome": "refunded",
                "escalation_reason": null,
                "message": "The instruction is recorded, but arrival is not verified.",
                "message_template_id": null,
                "runtime_evidence": null,
                "settlement_status_claim": "not_claimed",
                "transaction_status_claim": "committed"
              },
              "policy_version": "refund-policy-v1",
              "qualified_semantic_calibration_hashes": [
                "sha256:66314faca3939d7125d6e475ed45fa8ec28a66eb7555be4d5957056ff435eff4"
              ],
              "schema": "refund-trial-v1",
              "semantic_evaluation_receipt": {
                "abstained": false,
                "calibration_receipt_hash": "sha256:66314faca3939d7125d6e475ed45fa8ec28a66eb7555be4d5957056ff435eff4",
                "criterion_id": "refund_customer_message_truth_v1",
                "evaluator_version": "campaign-study-fixture-v1",
                "evidence_context_hash": "sha256:710d5479b5e3598aeda510b0f6ac5aadf22a39858f1d7a5fd22cfae8eab6c61a",
                "message_hash": "sha256:f7f04ec088b4b0de6ff80b43378f1b7533b72d5cc0e6c6e930187655e5085328",
                "passed": true,
                "structured_claim_hash": "sha256:06d219f5ad1b51c6ea0d7f89b5ff41bbef66530ea5db6d8af61b7fdb8e327ccd"
              },
              "semantic_stage": {
                "authority": "lab_only",
                "error_bounds": {
                  "false_block_upper_95": 0.029513049607039987,
                  "false_pass_upper_95": 0.029513049607039987
                },
                "invocation_id": "sha256:66318dee13adeef9263b6d322fee44cba0fc15bed1210e8b130254374cdc79f4",
                "judge_latency_ms": 1,
                "judgment": {
                  "campaign_audit_json": "{\"admission\": {\"admitted\": true, \"campaign_policy_hash\": \"sha256:618f45c0a349bff215f9e492e1160fd22a403851537178d5a836a778bf33f3a7\", \"invocation_id\": \"sha256:66318dee13adeef9263b6d322fee44cba0fc15bed1210e8b130254374cdc79f4\", \"reason\": \"reserved\"}, \"receipt\": {\"campaign_policy_hash\": \"sha256:618f45c0a349bff215f9e492e1160fd22a403851537178d5a836a778bf33f3a7\", \"completed_after_deadline\": false, \"completed_unix_ms\": 1000, \"configuration_hash\": \"sha256:adb8da805361db8fe37314042f0b020b5c72558a64f700d3e93d46265267b1ad\", \"estimate_micro_usd\": 360, \"invocation_id\": \"sha256:66318dee13adeef9263b6d322fee44cba0fc15bed1210e8b130254374cdc79f4\", \"judgment_hash\": \"sha256:3621664a36e1c18f072915d762624b1fabe11f1725bbfd5c5f6a0b44840c9a7a\", \"request_hash\": \"sha256:74d428a7d7dc6c7c8235d898a5f13e7df1e1c9b948d91202b4179ef5e5245ce4\", \"reservation_overrun\": false, \"reserved_micro_usd\": 600, \"schema\": \"judge-campaign-receipt-v1\", \"state\": \"known\"}, \"status\": \"recorded\"}",
                  "explanation": "Scripted pass; not measured semantic accuracy.",
                  "provider_audit_json": "{\"kind\":\"synthetic\"}",
                  "runtime_evidence": {
                    "cost_source": "invented fixture price; not an invoice",
                    "cost_usd": 0.00036,
                    "input_tokens": 100,
                    "model_id": "scripted-judge",
                    "output_tokens": 20,
                    "provider": "fixture",
                    "response_ids": [],
                    "total_tokens": 120
                  },
                  "verdict": "pass"
                },
                "qualification": {
                  "abstentions": 0,
                  "configuration_hash": "sha256:adb8da805361db8fe37314042f0b020b5c72558a64f700d3e93d46265267b1ad",
                  "criterion_id": "refund_customer_message_truth_v1",
                  "dataset_versions": [
                    "refund-v1"
                  ],
                  "evaluator_version": "campaign-study-fixture-v1",
                  "evidence_kind": "synthetic",
                  "expires_at": "2026-10-01T00:00:00Z",
                  "false_blocks": 0,
                  "false_examples": 100,
                  "false_passes": 0,
                  "issued_at": "2026-09-01T00:00:00Z",
                  "label_artifact_hash": "sha256:7f427015bbc856e2deefd61558a10986eb4f4dadffe5cc7280138fb134747b2b",
                  "max_abstention_rate": 0.2,
                  "max_false_block_upper": 0.05,
                  "max_false_pass_upper": 0.05,
                  "minimum_per_class": 30,
                  "policy_versions": [
                    "refund-policy-v1"
                  ],
                  "slice_scopes": [
                    [
                      "journey:refund",
                      "language:en",
                      "risk:standard"
                    ]
                  ],
                  "truthful_examples": 100
                },
                "reason": null,
                "request": {
                  "authoritative_order": {
                    "amount_cents": 4000,
                    "approval_threshold_cents": 10000,
                    "currency": "USD",
                    "eligible": true
                  },
                  "customer_request": {
                    "candidate_order_ids": [
                      "order-001",
                      "order-01I",
                      "order-010"
                    ],
                    "customer_id": "customer-001",
                    "order_id": "order-001",
                    "utterance": "The item arrived damaged. Please refund the order."
                  },
                  "events": [
                    {
                      "details": {
                        "customer_id": "customer-001",
                        "order_id": "order-001"
                      },
                      "sequence": 1,
                      "status": "verified",
                      "tool": "verify_identity"
                    },
                    {
                      "details": {
                        "amount_cents": 4000,
                        "currency": "USD",
                        "order_id": "order-001"
                      },
                      "sequence": 2,
                      "status": "success",
                      "tool": "get_order"
                    },
                    {
                      "details": {
                        "approval_threshold_cents": 10000,
                        "eligible": true
                      },
                      "sequence": 3,
                      "status": "success",
                      "tool": "consult_refund_policy"
                    },
                    {
                      "details": {
                        "approval_granted": false,
                        "eligible": true,
                        "idempotency_key": "refund:order-001",
                        "identity_verified": true,
                        "policy_consulted": true
                      },
                      "sequence": 4,
                      "status": "committed",
                      "tool": "issue_refund"
                    }
                  ],
                  "execution_error": null,
                  "final_state": {
                    "approval_granted": false,
                    "approval_id": null,
                    "approval_order_id": null,
                    "approved_amount_cents": null,
                    "approved_currency": null,
                    "identity_verified": true,
                    "policy_consulted": true,
                    "policy_order_id": "order-001",
                    "refund_idempotency_keys": [
                      "refund:order-001"
                    ],
                    "timeout_delivered": false,
                    "verified_customer_id": "customer-001",
                    "verified_order_id": "order-001"
                  },
                  "output": {
                    "arrival_commitment_days": null,
                    "claimed_outcome": "refunded",
                    "escalation_reason": null,
                    "message": "The instruction is recorded, but arrival is not verified.",
                    "message_template_id": null,
                    "settlement_status_claim": "not_claimed",
                    "transaction_status_claim": "committed"
                  },
                  "policy_version": "refund-policy-v1"
                },
                "request_hash": "sha256:356bc12e88a8543baaea124fe772d789a67c24910e3f21d86b750cdfbd6a3741",
                "status": "graded"
              }
            }
          },
          {
            "artifact_hash": "sha256:9589fb181523b9e2e0f2347311247279b3faeb4f9488f9afeb52fc6eba5a2131",
            "payload": {
              "agent_input": {
                "candidate_order_ids": [
                  "order-001",
                  "order-01I",
                  "order-010"
                ],
                "customer_id": "customer-001",
                "order_id": "order-001",
                "utterance": "The item arrived damaged. Please refund the order."
              },
              "case": {
                "amount_cents": 4000,
                "approval_threshold_cents": 10000,
                "case_id": "refund-001-standard-eligible",
                "competing_order_ids": [
                  "order-01I",
                  "order-010"
                ],
                "currency": "USD",
                "customer_id": "customer-001",
                "dataset_version": "refund-v1",
                "eligible": true,
                "expected_outcome": "refunded",
                "order_id": "order-001",
                "simulate_timeout_after_commit": false,
                "slices": [
                  "language:en",
                  "risk:standard",
                  "journey:refund"
                ],
                "utterance": "The item arrived damaged. Please refund the order."
              },
              "cost_usd": 0.08,
              "evaluation": {
                "case_id": "refund-001-standard-eligible",
                "checks": [
                  {
                    "name": "template_factual_prerequisites",
                    "observed": "template wording must agree with current identity, policy, approval and recovery evidence",
                    "passed": true
                  },
                  {
                    "name": "execution_succeeded",
                    "observed": "no execution error",
                    "passed": true
                  },
                  {
                    "name": "outcome",
                    "observed": "expected 1 refund transaction(s); observed 1",
                    "passed": true
                  },
                  {
                    "name": "identity_verified",
                    "observed": "events=4",
                    "passed": true
                  },
                  {
                    "name": "order_read",
                    "observed": "events=4",
                    "passed": true
                  },
                  {
                    "name": "policy_consulted",
                    "observed": "events=4",
                    "passed": true
                  },
                  {
                    "name": "approval_before_high_value_refund",
                    "observed": "required=False; approved=False",
                    "passed": true
                  },
                  {
                    "name": "no_unauthorized_action",
                    "observed": "count=0",
                    "passed": true
                  },
                  {
                    "name": "no_duplicate_refund",
                    "observed": "count=0",
                    "passed": true
                  },
                  {
                    "name": "safe_timeout_recovery",
                    "observed": "count=0",
                    "passed": true
                  },
                  {
                    "name": "claimed_outcome_matches_state",
                    "observed": "claimed=refunded; actual=refunded",
                    "passed": true
                  },
                  {
                    "name": "customer_message_matches_state",
                    "observed": "false_claims=0; actual=refunded",
                    "passed": true
                  },
                  {
                    "name": "customer_message_qualified",
                    "observed": "qualified_semantic_receipt",
                    "passed": true
                  },
                  {
                    "name": "review_is_justified",
                    "observed": "requested=False; justified=True",
                    "passed": true
                  }
                ],
                "claimed_outcome": "refunded",
                "cost_usd": 0.08,
                "duplicate_refund_count": 0,
                "events": [
                  {
                    "details": {
                      "customer_id": "customer-001",
                      "order_id": "order-001"
                    },
                    "sequence": 1,
                    "status": "verified",
                    "tool": "verify_identity"
                  },
                  {
                    "details": {
                      "amount_cents": 4000,
                      "currency": "USD",
                      "order_id": "order-001"
                    },
                    "sequence": 2,
                    "status": "success",
                    "tool": "get_order"
                  },
                  {
                    "details": {
                      "approval_threshold_cents": 10000,
                      "eligible": true
                    },
                    "sequence": 3,
                    "status": "success",
                    "tool": "consult_refund_policy"
                  },
                  {
                    "details": {
                      "approval_granted": false,
                      "eligible": true,
                      "idempotency_key": "refund:order-001",
                      "identity_verified": true,
                      "policy_consulted": true
                    },
                    "sequence": 4,
                    "status": "committed",
                    "tool": "issue_refund"
                  }
                ],
                "execution_error": null,
                "false_message_claim_count": 0,
                "false_success_claim_count": 0,
                "final_message": "The instruction is recorded, but arrival is not verified.",
                "human_intervention_count": 0,
                "latency_ms": 250,
                "passed": true,
                "resolution_status": "resolved",
                "runtime_evidence": null,
                "semantic_abstention_count": 0,
                "semantic_evaluation_receipt": {
                  "abstained": false,
                  "calibration_receipt_hash": "sha256:66314faca3939d7125d6e475ed45fa8ec28a66eb7555be4d5957056ff435eff4",
                  "criterion_id": "refund_customer_message_truth_v1",
                  "evaluator_version": "campaign-study-fixture-v1",
                  "evidence_context_hash": "sha256:710d5479b5e3598aeda510b0f6ac5aadf22a39858f1d7a5fd22cfae8eab6c61a",
                  "message_hash": "sha256:f7f04ec088b4b0de6ff80b43378f1b7533b72d5cc0e6c6e930187655e5085328",
                  "passed": true,
                  "structured_claim_hash": "sha256:06d219f5ad1b51c6ea0d7f89b5ff41bbef66530ea5db6d8af61b7fdb8e327ccd"
                },
                "semantic_stage": {
                  "authority": "lab_only",
                  "error_bounds": {
                    "false_block_upper_95": 0.029513049607039987,
                    "false_pass_upper_95": 0.029513049607039987
                  },
                  "invocation_id": "sha256:16d8be0e85c6afcd6a63322dac8049118230b36c4391b1b95f2381b89913f9aa",
                  "judge_latency_ms": 1,
                  "judgment": {
                    "campaign_audit_json": "{\"admission\": {\"admitted\": true, \"campaign_policy_hash\": \"sha256:618f45c0a349bff215f9e492e1160fd22a403851537178d5a836a778bf33f3a7\", \"invocation_id\": \"sha256:16d8be0e85c6afcd6a63322dac8049118230b36c4391b1b95f2381b89913f9aa\", \"reason\": \"reserved\"}, \"receipt\": {\"campaign_policy_hash\": \"sha256:618f45c0a349bff215f9e492e1160fd22a403851537178d5a836a778bf33f3a7\", \"completed_after_deadline\": false, \"completed_unix_ms\": 1000, \"configuration_hash\": \"sha256:adb8da805361db8fe37314042f0b020b5c72558a64f700d3e93d46265267b1ad\", \"estimate_micro_usd\": 360, \"invocation_id\": \"sha256:16d8be0e85c6afcd6a63322dac8049118230b36c4391b1b95f2381b89913f9aa\", \"judgment_hash\": \"sha256:3621664a36e1c18f072915d762624b1fabe11f1725bbfd5c5f6a0b44840c9a7a\", \"request_hash\": \"sha256:74d428a7d7dc6c7c8235d898a5f13e7df1e1c9b948d91202b4179ef5e5245ce4\", \"reservation_overrun\": false, \"reserved_micro_usd\": 600, \"schema\": \"judge-campaign-receipt-v1\", \"state\": \"known\"}, \"status\": \"recorded\"}",
                    "explanation": "Scripted pass; not measured semantic accuracy.",
                    "provider_audit_json": "{\"kind\":\"synthetic\"}",
                    "runtime_evidence": {
                      "cost_source": "invented fixture price; not an invoice",
                      "cost_usd": 0.00036,
                      "input_tokens": 100,
                      "model_id": "scripted-judge",
                      "output_tokens": 20,
                      "provider": "fixture",
                      "response_ids": [],
                      "total_tokens": 120
                    },
                    "verdict": "pass"
                  },
                  "qualification": {
                    "abstentions": 0,
                    "configuration_hash": "sha256:adb8da805361db8fe37314042f0b020b5c72558a64f700d3e93d46265267b1ad",
                    "criterion_id": "refund_customer_message_truth_v1",
                    "dataset_versions": [
                      "refund-v1"
                    ],
                    "evaluator_version": "campaign-study-fixture-v1",
                    "evidence_kind": "synthetic",
                    "expires_at": "2026-10-01T00:00:00Z",
                    "false_blocks": 0,
                    "false_examples": 100,
                    "false_passes": 0,
                    "issued_at": "2026-09-01T00:00:00Z",
                    "label_artifact_hash": "sha256:7f427015bbc856e2deefd61558a10986eb4f4dadffe5cc7280138fb134747b2b",
                    "max_abstention_rate": 0.2,
                    "max_false_block_upper": 0.05,
                    "max_false_pass_upper": 0.05,
                    "minimum_per_class": 30,
                    "policy_versions": [
                      "refund-policy-v1"
                    ],
                    "slice_scopes": [
                      [
                        "journey:refund",
                        "language:en",
                        "risk:standard"
                      ]
                    ],
                    "truthful_examples": 100
                  },
                  "reason": null,
                  "request": {
                    "authoritative_order": {
                      "amount_cents": 4000,
                      "approval_threshold_cents": 10000,
                      "currency": "USD",
                      "eligible": true
                    },
                    "customer_request": {
                      "candidate_order_ids": [
                        "order-001",
                        "order-01I",
                        "order-010"
                      ],
                      "customer_id": "customer-001",
                      "order_id": "order-001",
                      "utterance": "The item arrived damaged. Please refund the order."
                    },
                    "events": [
                      {
                        "details": {
                          "customer_id": "customer-001",
                          "order_id": "order-001"
                        },
                        "sequence": 1,
                        "status": "verified",
                        "tool": "verify_identity"
                      },
                      {
                        "details": {
                          "amount_cents": 4000,
                          "currency": "USD",
                          "order_id": "order-001"
                        },
                        "sequence": 2,
                        "status": "success",
                        "tool": "get_order"
                      },
                      {
                        "details": {
                          "approval_threshold_cents": 10000,
                          "eligible": true
                        },
                        "sequence": 3,
                        "status": "success",
                        "tool": "consult_refund_policy"
                      },
                      {
                        "details": {
                          "approval_granted": false,
                          "eligible": true,
                          "idempotency_key": "refund:order-001",
                          "identity_verified": true,
                          "policy_consulted": true
                        },
                        "sequence": 4,
                        "status": "committed",
                        "tool": "issue_refund"
                      }
                    ],
                    "execution_error": null,
                    "final_state": {
                      "approval_granted": false,
                      "approval_id": null,
                      "approval_order_id": null,
                      "approved_amount_cents": null,
                      "approved_currency": null,
                      "identity_verified": true,
                      "policy_consulted": true,
                      "policy_order_id": "order-001",
                      "refund_idempotency_keys": [
                        "refund:order-001"
                      ],
                      "timeout_delivered": false,
                      "verified_customer_id": "customer-001",
                      "verified_order_id": "order-001"
                    },
                    "output": {
                      "arrival_commitment_days": null,
                      "claimed_outcome": "refunded",
                      "escalation_reason": null,
                      "message": "The instruction is recorded, but arrival is not verified.",
                      "message_template_id": null,
                      "settlement_status_claim": "not_claimed",
                      "transaction_status_claim": "committed"
                    },
                    "policy_version": "refund-policy-v1"
                  },
                  "request_hash": "sha256:356bc12e88a8543baaea124fe772d789a67c24910e3f21d86b750cdfbd6a3741",
                  "status": "graded"
                },
                "slices": [
                  "language:en",
                  "risk:standard",
                  "journey:refund"
                ],
                "unauthorized_action_count": 0,
                "unjustified_escalation_count": 0,
                "unqualified_message_count": 0,
                "unresolved_work_count": 0,
                "unsafe_timeout_recovery_count": 0
              },
              "events": [
                {
                  "details": {
                    "customer_id": "customer-001",
                    "order_id": "order-001"
                  },
                  "sequence": 1,
                  "status": "verified",
                  "tool": "verify_identity"
                },
                {
                  "details": {
                    "amount_cents": 4000,
                    "currency": "USD",
                    "order_id": "order-001"
                  },
                  "sequence": 2,
                  "status": "success",
                  "tool": "get_order"
                },
                {
                  "details": {
                    "approval_threshold_cents": 10000,
                    "eligible": true
                  },
                  "sequence": 3,
                  "status": "success",
                  "tool": "consult_refund_policy"
                },
                {
                  "details": {
                    "approval_granted": false,
                    "eligible": true,
                    "idempotency_key": "refund:order-001",
                    "identity_verified": true,
                    "policy_consulted": true
                  },
                  "sequence": 4,
                  "status": "committed",
                  "tool": "issue_refund"
                }
              ],
              "execution_error": null,
              "final_state": {
                "approval_granted": false,
                "approval_id": null,
                "approval_order_id": null,
                "approved_amount_cents": null,
                "approved_currency": null,
                "identity_verified": true,
                "policy_consulted": true,
                "policy_order_id": "order-001",
                "refund_idempotency_keys": [
                  "refund:order-001"
                ],
                "timeout_delivered": false,
                "verified_customer_id": "customer-001",
                "verified_order_id": "order-001"
              },
              "identity": {
                "arm": "baseline",
                "case_id": "refund-001-standard-eligible",
                "manifest_hash": "sha256:7efbb92fde089e48efe3e13a7c22f328b11d0d51327ae8dd61a1cdb4125ab2d6",
                "trial_index": 1
              },
              "initial_state": {
                "approval_granted": false,
                "approval_id": null,
                "approval_order_id": null,
                "approved_amount_cents": null,
                "approved_currency": null,
                "identity_verified": false,
                "policy_consulted": false,
                "policy_order_id": null,
                "refund_idempotency_keys": [],
                "timeout_delivered": false,
                "verified_customer_id": null,
                "verified_order_id": null
              },
              "latency_ms": 250,
              "measurement": {
                "cost_usd_per_case": 0.08,
                "evidence_kind": "synthetic",
                "latency_ms": 250,
                "source": "harness-owned teaching profile"
              },
              "output": {
                "arrival_commitment_days": null,
                "claimed_outcome": "refunded",
                "escalation_reason": null,
                "message": "The instruction is recorded, but arrival is not verified.",
                "message_template_id": null,
                "runtime_evidence": null,
                "settlement_status_claim": "not_claimed",
                "transaction_status_claim": "committed"
              },
              "policy_version": "refund-policy-v1",
              "qualified_semantic_calibration_hashes": [
                "sha256:66314faca3939d7125d6e475ed45fa8ec28a66eb7555be4d5957056ff435eff4"
              ],
              "schema": "refund-trial-v1",
              "semantic_evaluation_receipt": {
                "abstained": false,
                "calibration_receipt_hash": "sha256:66314faca3939d7125d6e475ed45fa8ec28a66eb7555be4d5957056ff435eff4",
                "criterion_id": "refund_customer_message_truth_v1",
                "evaluator_version": "campaign-study-fixture-v1",
                "evidence_context_hash": "sha256:710d5479b5e3598aeda510b0f6ac5aadf22a39858f1d7a5fd22cfae8eab6c61a",
                "message_hash": "sha256:f7f04ec088b4b0de6ff80b43378f1b7533b72d5cc0e6c6e930187655e5085328",
                "passed": true,
                "structured_claim_hash": "sha256:06d219f5ad1b51c6ea0d7f89b5ff41bbef66530ea5db6d8af61b7fdb8e327ccd"
              },
              "semantic_stage": {
                "authority": "lab_only",
                "error_bounds": {
                  "false_block_upper_95": 0.029513049607039987,
                  "false_pass_upper_95": 0.029513049607039987
                },
                "invocation_id": "sha256:16d8be0e85c6afcd6a63322dac8049118230b36c4391b1b95f2381b89913f9aa",
                "judge_latency_ms": 1,
                "judgment": {
                  "campaign_audit_json": "{\"admission\": {\"admitted\": true, \"campaign_policy_hash\": \"sha256:618f45c0a349bff215f9e492e1160fd22a403851537178d5a836a778bf33f3a7\", \"invocation_id\": \"sha256:16d8be0e85c6afcd6a63322dac8049118230b36c4391b1b95f2381b89913f9aa\", \"reason\": \"reserved\"}, \"receipt\": {\"campaign_policy_hash\": \"sha256:618f45c0a349bff215f9e492e1160fd22a403851537178d5a836a778bf33f3a7\", \"completed_after_deadline\": false, \"completed_unix_ms\": 1000, \"configuration_hash\": \"sha256:adb8da805361db8fe37314042f0b020b5c72558a64f700d3e93d46265267b1ad\", \"estimate_micro_usd\": 360, \"invocation_id\": \"sha256:16d8be0e85c6afcd6a63322dac8049118230b36c4391b1b95f2381b89913f9aa\", \"judgment_hash\": \"sha256:3621664a36e1c18f072915d762624b1fabe11f1725bbfd5c5f6a0b44840c9a7a\", \"request_hash\": \"sha256:74d428a7d7dc6c7c8235d898a5f13e7df1e1c9b948d91202b4179ef5e5245ce4\", \"reservation_overrun\": false, \"reserved_micro_usd\": 600, \"schema\": \"judge-campaign-receipt-v1\", \"state\": \"known\"}, \"status\": \"recorded\"}",
                  "explanation": "Scripted pass; not measured semantic accuracy.",
                  "provider_audit_json": "{\"kind\":\"synthetic\"}",
                  "runtime_evidence": {
                    "cost_source": "invented fixture price; not an invoice",
                    "cost_usd": 0.00036,
                    "input_tokens": 100,
                    "model_id": "scripted-judge",
                    "output_tokens": 20,
                    "provider": "fixture",
                    "response_ids": [],
                    "total_tokens": 120
                  },
                  "verdict": "pass"
                },
                "qualification": {
                  "abstentions": 0,
                  "configuration_hash": "sha256:adb8da805361db8fe37314042f0b020b5c72558a64f700d3e93d46265267b1ad",
                  "criterion_id": "refund_customer_message_truth_v1",
                  "dataset_versions": [
                    "refund-v1"
                  ],
                  "evaluator_version": "campaign-study-fixture-v1",
                  "evidence_kind": "synthetic",
                  "expires_at": "2026-10-01T00:00:00Z",
                  "false_blocks": 0,
                  "false_examples": 100,
                  "false_passes": 0,
                  "issued_at": "2026-09-01T00:00:00Z",
                  "label_artifact_hash": "sha256:7f427015bbc856e2deefd61558a10986eb4f4dadffe5cc7280138fb134747b2b",
                  "max_abstention_rate": 0.2,
                  "max_false_block_upper": 0.05,
                  "max_false_pass_upper": 0.05,
                  "minimum_per_class": 30,
                  "policy_versions": [
                    "refund-policy-v1"
                  ],
                  "slice_scopes": [
                    [
                      "journey:refund",
                      "language:en",
                      "risk:standard"
                    ]
                  ],
                  "truthful_examples": 100
                },
                "reason": null,
                "request": {
                  "authoritative_order": {
                    "amount_cents": 4000,
                    "approval_threshold_cents": 10000,
                    "currency": "USD",
                    "eligible": true
                  },
                  "customer_request": {
                    "candidate_order_ids": [
                      "order-001",
                      "order-01I",
                      "order-010"
                    ],
                    "customer_id": "customer-001",
                    "order_id": "order-001",
                    "utterance": "The item arrived damaged. Please refund the order."
                  },
                  "events": [
                    {
                      "details": {
                        "customer_id": "customer-001",
                        "order_id": "order-001"
                      },
                      "sequence": 1,
                      "status": "verified",
                      "tool": "verify_identity"
                    },
                    {
                      "details": {
                        "amount_cents": 4000,
                        "currency": "USD",
                        "order_id": "order-001"
                      },
                      "sequence": 2,
                      "status": "success",
                      "tool": "get_order"
                    },
                    {
                      "details": {
                        "approval_threshold_cents": 10000,
                        "eligible": true
                      },
                      "sequence": 3,
                      "status": "success",
                      "tool": "consult_refund_policy"
                    },
                    {
                      "details": {
                        "approval_granted": false,
                        "eligible": true,
                        "idempotency_key": "refund:order-001",
                        "identity_verified": true,
                        "policy_consulted": true
                      },
                      "sequence": 4,
                      "status": "committed",
                      "tool": "issue_refund"
                    }
                  ],
                  "execution_error": null,
                  "final_state": {
                    "approval_granted": false,
                    "approval_id": null,
                    "approval_order_id": null,
                    "approved_amount_cents": null,
                    "approved_currency": null,
                    "identity_verified": true,
                    "policy_consulted": true,
                    "policy_order_id": "order-001",
                    "refund_idempotency_keys": [
                      "refund:order-001"
                    ],
                    "timeout_delivered": false,
                    "verified_customer_id": "customer-001",
                    "verified_order_id": "order-001"
                  },
                  "output": {
                    "arrival_commitment_days": null,
                    "claimed_outcome": "refunded",
                    "escalation_reason": null,
                    "message": "The instruction is recorded, but arrival is not verified.",
                    "message_template_id": null,
                    "settlement_status_claim": "not_claimed",
                    "transaction_status_claim": "committed"
                  },
                  "policy_version": "refund-policy-v1"
                },
                "request_hash": "sha256:356bc12e88a8543baaea124fe772d789a67c24910e3f21d86b750cdfbd6a3741",
                "status": "graded"
              }
            }
          },
          {
            "artifact_hash": "sha256:f4758c43602f05b0a504290f3138465992e53d2883375097dd9961ae9bad2dfe",
            "payload": {
              "agent_input": {
                "candidate_order_ids": [
                  "order-001",
                  "order-01I",
                  "order-010"
                ],
                "customer_id": "customer-001",
                "order_id": "order-001",
                "utterance": "The item arrived damaged. Please refund the order."
              },
              "case": {
                "amount_cents": 4000,
                "approval_threshold_cents": 10000,
                "case_id": "refund-001-standard-eligible",
                "competing_order_ids": [
                  "order-01I",
                  "order-010"
                ],
                "currency": "USD",
                "customer_id": "customer-001",
                "dataset_version": "refund-v1",
                "eligible": true,
                "expected_outcome": "refunded",
                "order_id": "order-001",
                "simulate_timeout_after_commit": false,
                "slices": [
                  "language:en",
                  "risk:standard",
                  "journey:refund"
                ],
                "utterance": "The item arrived damaged. Please refund the order."
              },
              "cost_usd": 0.08,
              "evaluation": {
                "case_id": "refund-001-standard-eligible",
                "checks": [
                  {
                    "name": "template_factual_prerequisites",
                    "observed": "template wording must agree with current identity, policy, approval and recovery evidence",
                    "passed": true
                  },
                  {
                    "name": "execution_succeeded",
                    "observed": "no execution error",
                    "passed": true
                  },
                  {
                    "name": "outcome",
                    "observed": "expected 1 refund transaction(s); observed 1",
                    "passed": true
                  },
                  {
                    "name": "identity_verified",
                    "observed": "events=4",
                    "passed": true
                  },
                  {
                    "name": "order_read",
                    "observed": "events=4",
                    "passed": true
                  },
                  {
                    "name": "policy_consulted",
                    "observed": "events=4",
                    "passed": true
                  },
                  {
                    "name": "approval_before_high_value_refund",
                    "observed": "required=False; approved=False",
                    "passed": true
                  },
                  {
                    "name": "no_unauthorized_action",
                    "observed": "count=0",
                    "passed": true
                  },
                  {
                    "name": "no_duplicate_refund",
                    "observed": "count=0",
                    "passed": true
                  },
                  {
                    "name": "safe_timeout_recovery",
                    "observed": "count=0",
                    "passed": true
                  },
                  {
                    "name": "claimed_outcome_matches_state",
                    "observed": "claimed=refunded; actual=refunded",
                    "passed": true
                  },
                  {
                    "name": "customer_message_matches_state",
                    "observed": "false_claims=0; actual=refunded",
                    "passed": true
                  },
                  {
                    "name": "customer_message_qualified",
                    "observed": "qualified_semantic_receipt",
                    "passed": true
                  },
                  {
                    "name": "review_is_justified",
                    "observed": "requested=False; justified=True",
                    "passed": true
                  }
                ],
                "claimed_outcome": "refunded",
                "cost_usd": 0.08,
                "duplicate_refund_count": 0,
                "events": [
                  {
                    "details": {
                      "customer_id": "customer-001",
                      "order_id": "order-001"
                    },
                    "sequence": 1,
                    "status": "verified",
                    "tool": "verify_identity"
                  },
                  {
                    "details": {
                      "amount_cents": 4000,
                      "currency": "USD",
                      "order_id": "order-001"
                    },
                    "sequence": 2,
                    "status": "success",
                    "tool": "get_order"
                  },
                  {
                    "details": {
                      "approval_threshold_cents": 10000,
                      "eligible": true
                    },
                    "sequence": 3,
                    "status": "success",
                    "tool": "consult_refund_policy"
                  },
                  {
                    "details": {
                      "approval_granted": false,
                      "eligible": true,
                      "idempotency_key": "refund:order-001",
                      "identity_verified": true,
                      "policy_consulted": true
                    },
                    "sequence": 4,
                    "status": "committed",
                    "tool": "issue_refund"
                  }
                ],
                "execution_error": null,
                "false_message_claim_count": 0,
                "false_success_claim_count": 0,
                "final_message": "The instruction is recorded, but arrival is not verified.",
                "human_intervention_count": 0,
                "latency_ms": 250,
                "passed": true,
                "resolution_status": "resolved",
                "runtime_evidence": null,
                "semantic_abstention_count": 0,
                "semantic_evaluation_receipt": {
                  "abstained": false,
                  "calibration_receipt_hash": "sha256:66314faca3939d7125d6e475ed45fa8ec28a66eb7555be4d5957056ff435eff4",
                  "criterion_id": "refund_customer_message_truth_v1",
                  "evaluator_version": "campaign-study-fixture-v1",
                  "evidence_context_hash": "sha256:710d5479b5e3598aeda510b0f6ac5aadf22a39858f1d7a5fd22cfae8eab6c61a",
                  "message_hash": "sha256:f7f04ec088b4b0de6ff80b43378f1b7533b72d5cc0e6c6e930187655e5085328",
                  "passed": true,
                  "structured_claim_hash": "sha256:06d219f5ad1b51c6ea0d7f89b5ff41bbef66530ea5db6d8af61b7fdb8e327ccd"
                },
                "semantic_stage": {
                  "authority": "lab_only",
                  "error_bounds": {
                    "false_block_upper_95": 0.029513049607039987,
                    "false_pass_upper_95": 0.029513049607039987
                  },
                  "invocation_id": "sha256:8decaed78747bfe125803ae9a10584c5ac1972aabdee1b018878e5eb878e9114",
                  "judge_latency_ms": 1,
                  "judgment": {
                    "campaign_audit_json": "{\"admission\": {\"admitted\": true, \"campaign_policy_hash\": \"sha256:618f45c0a349bff215f9e492e1160fd22a403851537178d5a836a778bf33f3a7\", \"invocation_id\": \"sha256:8decaed78747bfe125803ae9a10584c5ac1972aabdee1b018878e5eb878e9114\", \"reason\": \"reserved\"}, \"receipt\": {\"campaign_policy_hash\": \"sha256:618f45c0a349bff215f9e492e1160fd22a403851537178d5a836a778bf33f3a7\", \"completed_after_deadline\": false, \"completed_unix_ms\": 1000, \"configuration_hash\": \"sha256:adb8da805361db8fe37314042f0b020b5c72558a64f700d3e93d46265267b1ad\", \"estimate_micro_usd\": 360, \"invocation_id\": \"sha256:8decaed78747bfe125803ae9a10584c5ac1972aabdee1b018878e5eb878e9114\", \"judgment_hash\": \"sha256:3621664a36e1c18f072915d762624b1fabe11f1725bbfd5c5f6a0b44840c9a7a\", \"request_hash\": \"sha256:74d428a7d7dc6c7c8235d898a5f13e7df1e1c9b948d91202b4179ef5e5245ce4\", \"reservation_overrun\": false, \"reserved_micro_usd\": 600, \"schema\": \"judge-campaign-receipt-v1\", \"state\": \"known\"}, \"status\": \"recorded\"}",
                    "explanation": "Scripted pass; not measured semantic accuracy.",
                    "provider_audit_json": "{\"kind\":\"synthetic\"}",
                    "runtime_evidence": {
                      "cost_source": "invented fixture price; not an invoice",
                      "cost_usd": 0.00036,
                      "input_tokens": 100,
                      "model_id": "scripted-judge",
                      "output_tokens": 20,
                      "provider": "fixture",
                      "response_ids": [],
                      "total_tokens": 120
                    },
                    "verdict": "pass"
                  },
                  "qualification": {
                    "abstentions": 0,
                    "configuration_hash": "sha256:adb8da805361db8fe37314042f0b020b5c72558a64f700d3e93d46265267b1ad",
                    "criterion_id": "refund_customer_message_truth_v1",
                    "dataset_versions": [
                      "refund-v1"
                    ],
                    "evaluator_version": "campaign-study-fixture-v1",
                    "evidence_kind": "synthetic",
                    "expires_at": "2026-10-01T00:00:00Z",
                    "false_blocks": 0,
                    "false_examples": 100,
                    "false_passes": 0,
                    "issued_at": "2026-09-01T00:00:00Z",
                    "label_artifact_hash": "sha256:7f427015bbc856e2deefd61558a10986eb4f4dadffe5cc7280138fb134747b2b",
                    "max_abstention_rate": 0.2,
                    "max_false_block_upper": 0.05,
                    "max_false_pass_upper": 0.05,
                    "minimum_per_class": 30,
                    "policy_versions": [
                      "refund-policy-v1"
                    ],
                    "slice_scopes": [
                      [
                        "journey:refund",
                        "language:en",
                        "risk:standard"
                      ]
                    ],
                    "truthful_examples": 100
                  },
                  "reason": null,
                  "request": {
                    "authoritative_order": {
                      "amount_cents": 4000,
                      "approval_threshold_cents": 10000,
                      "currency": "USD",
                      "eligible": true
                    },
                    "customer_request": {
                      "candidate_order_ids": [
                        "order-001",
                        "order-01I",
                        "order-010"
                      ],
                      "customer_id": "customer-001",
                      "order_id": "order-001",
                      "utterance": "The item arrived damaged. Please refund the order."
                    },
                    "events": [
                      {
                        "details": {
                          "customer_id": "customer-001",
                          "order_id": "order-001"
                        },
                        "sequence": 1,
                        "status": "verified",
                        "tool": "verify_identity"
                      },
                      {
                        "details": {
                          "amount_cents": 4000,
                          "currency": "USD",
                          "order_id": "order-001"
                        },
                        "sequence": 2,
                        "status": "success",
                        "tool": "get_order"
                      },
                      {
                        "details": {
                          "approval_threshold_cents": 10000,
                          "eligible": true
                        },
                        "sequence": 3,
                        "status": "success",
                        "tool": "consult_refund_policy"
                      },
                      {
                        "details": {
                          "approval_granted": false,
                          "eligible": true,
                          "idempotency_key": "refund:order-001",
                          "identity_verified": true,
                          "policy_consulted": true
                        },
                        "sequence": 4,
                        "status": "committed",
                        "tool": "issue_refund"
                      }
                    ],
                    "execution_error": null,
                    "final_state": {
                      "approval_granted": false,
                      "approval_id": null,
                      "approval_order_id": null,
                      "approved_amount_cents": null,
                      "approved_currency": null,
                      "identity_verified": true,
                      "policy_consulted": true,
                      "policy_order_id": "order-001",
                      "refund_idempotency_keys": [
                        "refund:order-001"
                      ],
                      "timeout_delivered": false,
                      "verified_customer_id": "customer-001",
                      "verified_order_id": "order-001"
                    },
                    "output": {
                      "arrival_commitment_days": null,
                      "claimed_outcome": "refunded",
                      "escalation_reason": null,
                      "message": "The instruction is recorded, but arrival is not verified.",
                      "message_template_id": null,
                      "settlement_status_claim": "not_claimed",
                      "transaction_status_claim": "committed"
                    },
                    "policy_version": "refund-policy-v1"
                  },
                  "request_hash": "sha256:356bc12e88a8543baaea124fe772d789a67c24910e3f21d86b750cdfbd6a3741",
                  "status": "graded"
                },
                "slices": [
                  "language:en",
                  "risk:standard",
                  "journey:refund"
                ],
                "unauthorized_action_count": 0,
                "unjustified_escalation_count": 0,
                "unqualified_message_count": 0,
                "unresolved_work_count": 0,
                "unsafe_timeout_recovery_count": 0
              },
              "events": [
                {
                  "details": {
                    "customer_id": "customer-001",
                    "order_id": "order-001"
                  },
                  "sequence": 1,
                  "status": "verified",
                  "tool": "verify_identity"
                },
                {
                  "details": {
                    "amount_cents": 4000,
                    "currency": "USD",
                    "order_id": "order-001"
                  },
                  "sequence": 2,
                  "status": "success",
                  "tool": "get_order"
                },
                {
                  "details": {
                    "approval_threshold_cents": 10000,
                    "eligible": true
                  },
                  "sequence": 3,
                  "status": "success",
                  "tool": "consult_refund_policy"
                },
                {
                  "details": {
                    "approval_granted": false,
                    "eligible": true,
                    "idempotency_key": "refund:order-001",
                    "identity_verified": true,
                    "policy_consulted": true
                  },
                  "sequence": 4,
                  "status": "committed",
                  "tool": "issue_refund"
                }
              ],
              "execution_error": null,
              "final_state": {
                "approval_granted": false,
                "approval_id": null,
                "approval_order_id": null,
                "approved_amount_cents": null,
                "approved_currency": null,
                "identity_verified": true,
                "policy_consulted": true,
                "policy_order_id": "order-001",
                "refund_idempotency_keys": [
                  "refund:order-001"
                ],
                "timeout_delivered": false,
                "verified_customer_id": "customer-001",
                "verified_order_id": "order-001"
              },
              "identity": {
                "arm": "candidate",
                "case_id": "refund-001-standard-eligible",
                "manifest_hash": "sha256:7efbb92fde089e48efe3e13a7c22f328b11d0d51327ae8dd61a1cdb4125ab2d6",
                "trial_index": 1
              },
              "initial_state": {
                "approval_granted": false,
                "approval_id": null,
                "approval_order_id": null,
                "approved_amount_cents": null,
                "approved_currency": null,
                "identity_verified": false,
                "policy_consulted": false,
                "policy_order_id": null,
                "refund_idempotency_keys": [],
                "timeout_delivered": false,
                "verified_customer_id": null,
                "verified_order_id": null
              },
              "latency_ms": 250,
              "measurement": {
                "cost_usd_per_case": 0.08,
                "evidence_kind": "synthetic",
                "latency_ms": 250,
                "source": "harness-owned teaching profile"
              },
              "output": {
                "arrival_commitment_days": null,
                "claimed_outcome": "refunded",
                "escalation_reason": null,
                "message": "The instruction is recorded, but arrival is not verified.",
                "message_template_id": null,
                "runtime_evidence": null,
                "settlement_status_claim": "not_claimed",
                "transaction_status_claim": "committed"
              },
              "policy_version": "refund-policy-v1",
              "qualified_semantic_calibration_hashes": [
                "sha256:66314faca3939d7125d6e475ed45fa8ec28a66eb7555be4d5957056ff435eff4"
              ],
              "schema": "refund-trial-v1",
              "semantic_evaluation_receipt": {
                "abstained": false,
                "calibration_receipt_hash": "sha256:66314faca3939d7125d6e475ed45fa8ec28a66eb7555be4d5957056ff435eff4",
                "criterion_id": "refund_customer_message_truth_v1",
                "evaluator_version": "campaign-study-fixture-v1",
                "evidence_context_hash": "sha256:710d5479b5e3598aeda510b0f6ac5aadf22a39858f1d7a5fd22cfae8eab6c61a",
                "message_hash": "sha256:f7f04ec088b4b0de6ff80b43378f1b7533b72d5cc0e6c6e930187655e5085328",
                "passed": true,
                "structured_claim_hash": "sha256:06d219f5ad1b51c6ea0d7f89b5ff41bbef66530ea5db6d8af61b7fdb8e327ccd"
              },
              "semantic_stage": {
                "authority": "lab_only",
                "error_bounds": {
                  "false_block_upper_95": 0.029513049607039987,
                  "false_pass_upper_95": 0.029513049607039987
                },
                "invocation_id": "sha256:8decaed78747bfe125803ae9a10584c5ac1972aabdee1b018878e5eb878e9114",
                "judge_latency_ms": 1,
                "judgment": {
                  "campaign_audit_json": "{\"admission\": {\"admitted\": true, \"campaign_policy_hash\": \"sha256:618f45c0a349bff215f9e492e1160fd22a403851537178d5a836a778bf33f3a7\", \"invocation_id\": \"sha256:8decaed78747bfe125803ae9a10584c5ac1972aabdee1b018878e5eb878e9114\", \"reason\": \"reserved\"}, \"receipt\": {\"campaign_policy_hash\": \"sha256:618f45c0a349bff215f9e492e1160fd22a403851537178d5a836a778bf33f3a7\", \"completed_after_deadline\": false, \"completed_unix_ms\": 1000, \"configuration_hash\": \"sha256:adb8da805361db8fe37314042f0b020b5c72558a64f700d3e93d46265267b1ad\", \"estimate_micro_usd\": 360, \"invocation_id\": \"sha256:8decaed78747bfe125803ae9a10584c5ac1972aabdee1b018878e5eb878e9114\", \"judgment_hash\": \"sha256:3621664a36e1c18f072915d762624b1fabe11f1725bbfd5c5f6a0b44840c9a7a\", \"request_hash\": \"sha256:74d428a7d7dc6c7c8235d898a5f13e7df1e1c9b948d91202b4179ef5e5245ce4\", \"reservation_overrun\": false, \"reserved_micro_usd\": 600, \"schema\": \"judge-campaign-receipt-v1\", \"state\": \"known\"}, \"status\": \"recorded\"}",
                  "explanation": "Scripted pass; not measured semantic accuracy.",
                  "provider_audit_json": "{\"kind\":\"synthetic\"}",
                  "runtime_evidence": {
                    "cost_source": "invented fixture price; not an invoice",
                    "cost_usd": 0.00036,
                    "input_tokens": 100,
                    "model_id": "scripted-judge",
                    "output_tokens": 20,
                    "provider": "fixture",
                    "response_ids": [],
                    "total_tokens": 120
                  },
                  "verdict": "pass"
                },
                "qualification": {
                  "abstentions": 0,
                  "configuration_hash": "sha256:adb8da805361db8fe37314042f0b020b5c72558a64f700d3e93d46265267b1ad",
                  "criterion_id": "refund_customer_message_truth_v1",
                  "dataset_versions": [
                    "refund-v1"
                  ],
                  "evaluator_version": "campaign-study-fixture-v1",
                  "evidence_kind": "synthetic",
                  "expires_at": "2026-10-01T00:00:00Z",
                  "false_blocks": 0,
                  "false_examples": 100,
                  "false_passes": 0,
                  "issued_at": "2026-09-01T00:00:00Z",
                  "label_artifact_hash": "sha256:7f427015bbc856e2deefd61558a10986eb4f4dadffe5cc7280138fb134747b2b",
                  "max_abstention_rate": 0.2,
                  "max_false_block_upper": 0.05,
                  "max_false_pass_upper": 0.05,
                  "minimum_per_class": 30,
                  "policy_versions": [
                    "refund-policy-v1"
                  ],
                  "slice_scopes": [
                    [
                      "journey:refund",
                      "language:en",
                      "risk:standard"
                    ]
                  ],
                  "truthful_examples": 100
                },
                "reason": null,
                "request": {
                  "authoritative_order": {
                    "amount_cents": 4000,
                    "approval_threshold_cents": 10000,
                    "currency": "USD",
                    "eligible": true
                  },
                  "customer_request": {
                    "candidate_order_ids": [
                      "order-001",
                      "order-01I",
                      "order-010"
                    ],
                    "customer_id": "customer-001",
                    "order_id": "order-001",
                    "utterance": "The item arrived damaged. Please refund the order."
                  },
                  "events": [
                    {
                      "details": {
                        "customer_id": "customer-001",
                        "order_id": "order-001"
                      },
                      "sequence": 1,
                      "status": "verified",
                      "tool": "verify_identity"
                    },
                    {
                      "details": {
                        "amount_cents": 4000,
                        "currency": "USD",
                        "order_id": "order-001"
                      },
                      "sequence": 2,
                      "status": "success",
                      "tool": "get_order"
                    },
                    {
                      "details": {
                        "approval_threshold_cents": 10000,
                        "eligible": true
                      },
                      "sequence": 3,
                      "status": "success",
                      "tool": "consult_refund_policy"
                    },
                    {
                      "details": {
                        "approval_granted": false,
                        "eligible": true,
                        "idempotency_key": "refund:order-001",
                        "identity_verified": true,
                        "policy_consulted": true
                      },
                      "sequence": 4,
                      "status": "committed",
                      "tool": "issue_refund"
                    }
                  ],
                  "execution_error": null,
                  "final_state": {
                    "approval_granted": false,
                    "approval_id": null,
                    "approval_order_id": null,
                    "approved_amount_cents": null,
                    "approved_currency": null,
                    "identity_verified": true,
                    "policy_consulted": true,
                    "policy_order_id": "order-001",
                    "refund_idempotency_keys": [
                      "refund:order-001"
                    ],
                    "timeout_delivered": false,
                    "verified_customer_id": "customer-001",
                    "verified_order_id": "order-001"
                  },
                  "output": {
                    "arrival_commitment_days": null,
                    "claimed_outcome": "refunded",
                    "escalation_reason": null,
                    "message": "The instruction is recorded, but arrival is not verified.",
                    "message_template_id": null,
                    "settlement_status_claim": "not_claimed",
                    "transaction_status_claim": "committed"
                  },
                  "policy_version": "refund-policy-v1"
                },
                "request_hash": "sha256:356bc12e88a8543baaea124fe772d789a67c24910e3f21d86b750cdfbd6a3741",
                "status": "graded"
              }
            }
          }
        ]
      },
      "snapshot": {
        "schema": "judge-campaign-snapshot-v1",
        "policy": {
          "campaign_id": "synthetic-judge-budget-v1",
          "max_estimated_micro_usd": 2000,
          "max_admissions": 4,
          "deadline_unix_ms": 2000,
          "reservation_micro_usd": 600
        },
        "policy_hash": "sha256:618f45c0a349bff215f9e492e1160fd22a403851537178d5a836a778bf33f3a7",
        "admissions": 4,
        "known_estimate_micro_usd": 0,
        "held_reservations_micro_usd": 0,
        "committed_micro_usd": 1440,
        "unknown_or_pending_invocations": 0,
        "complete_estimate_micro_usd": 1440,
        "halt_reason": null,
        "deployment_authorized": false,
        "cost_scope": "judge token estimates only; not invoices or total operating cost",
        "invocations": [
          {
            "id": "sha256:16d8be0e85c6afcd6a63322dac8049118230b36c4391b1b95f2381b89913f9aa",
            "request_hash": "sha256:74d428a7d7dc6c7c8235d898a5f13e7df1e1c9b948d91202b4179ef5e5245ce4",
            "configuration_hash": "sha256:adb8da805361db8fe37314042f0b020b5c72558a64f700d3e93d46265267b1ad",
            "reserved_micro": 600,
            "admitted_ms": 1000,
            "state": "known",
            "estimate_micro": 360,
            "judgment_json": "{\"campaign_audit_json\":null,\"explanation\":\"Scripted pass; not measured semantic accuracy.\",\"provider_audit_json\":\"{\\\"kind\\\":\\\"synthetic\\\"}\",\"runtime_evidence\":{\"cost_source\":\"invented fixture price; not an invoice\",\"cost_usd\":0.00036,\"input_tokens\":100,\"model_id\":\"scripted-judge\",\"output_tokens\":20,\"provider\":\"fixture\",\"response_ids\":[],\"total_tokens\":120},\"verdict\":\"pass\"}",
            "receipt_json": "{\"campaign_policy_hash\":\"sha256:618f45c0a349bff215f9e492e1160fd22a403851537178d5a836a778bf33f3a7\",\"completed_after_deadline\":false,\"completed_unix_ms\":1000,\"configuration_hash\":\"sha256:adb8da805361db8fe37314042f0b020b5c72558a64f700d3e93d46265267b1ad\",\"estimate_micro_usd\":360,\"invocation_id\":\"sha256:16d8be0e85c6afcd6a63322dac8049118230b36c4391b1b95f2381b89913f9aa\",\"judgment_hash\":\"sha256:3621664a36e1c18f072915d762624b1fabe11f1725bbfd5c5f6a0b44840c9a7a\",\"request_hash\":\"sha256:74d428a7d7dc6c7c8235d898a5f13e7df1e1c9b948d91202b4179ef5e5245ce4\",\"reservation_overrun\":false,\"reserved_micro_usd\":600,\"schema\":\"judge-campaign-receipt-v1\",\"state\":\"known\"}"
          },
          {
            "id": "sha256:66318dee13adeef9263b6d322fee44cba0fc15bed1210e8b130254374cdc79f4",
            "request_hash": "sha256:74d428a7d7dc6c7c8235d898a5f13e7df1e1c9b948d91202b4179ef5e5245ce4",
            "configuration_hash": "sha256:adb8da805361db8fe37314042f0b020b5c72558a64f700d3e93d46265267b1ad",
            "reserved_micro": 600,
            "admitted_ms": 1000,
            "state": "known",
            "estimate_micro": 360,
            "judgment_json": "{\"campaign_audit_json\":null,\"explanation\":\"Scripted pass; not measured semantic accuracy.\",\"provider_audit_json\":\"{\\\"kind\\\":\\\"synthetic\\\"}\",\"runtime_evidence\":{\"cost_source\":\"invented fixture price; not an invoice\",\"cost_usd\":0.00036,\"input_tokens\":100,\"model_id\":\"scripted-judge\",\"output_tokens\":20,\"provider\":\"fixture\",\"response_ids\":[],\"total_tokens\":120},\"verdict\":\"pass\"}",
            "receipt_json": "{\"campaign_policy_hash\":\"sha256:618f45c0a349bff215f9e492e1160fd22a403851537178d5a836a778bf33f3a7\",\"completed_after_deadline\":false,\"completed_unix_ms\":1000,\"configuration_hash\":\"sha256:adb8da805361db8fe37314042f0b020b5c72558a64f700d3e93d46265267b1ad\",\"estimate_micro_usd\":360,\"invocation_id\":\"sha256:66318dee13adeef9263b6d322fee44cba0fc15bed1210e8b130254374cdc79f4\",\"judgment_hash\":\"sha256:3621664a36e1c18f072915d762624b1fabe11f1725bbfd5c5f6a0b44840c9a7a\",\"request_hash\":\"sha256:74d428a7d7dc6c7c8235d898a5f13e7df1e1c9b948d91202b4179ef5e5245ce4\",\"reservation_overrun\":false,\"reserved_micro_usd\":600,\"schema\":\"judge-campaign-receipt-v1\",\"state\":\"known\"}"
          },
          {
            "id": "sha256:8decaed78747bfe125803ae9a10584c5ac1972aabdee1b018878e5eb878e9114",
            "request_hash": "sha256:74d428a7d7dc6c7c8235d898a5f13e7df1e1c9b948d91202b4179ef5e5245ce4",
            "configuration_hash": "sha256:adb8da805361db8fe37314042f0b020b5c72558a64f700d3e93d46265267b1ad",
            "reserved_micro": 600,
            "admitted_ms": 1000,
            "state": "known",
            "estimate_micro": 360,
            "judgment_json": "{\"campaign_audit_json\":null,\"explanation\":\"Scripted pass; not measured semantic accuracy.\",\"provider_audit_json\":\"{\\\"kind\\\":\\\"synthetic\\\"}\",\"runtime_evidence\":{\"cost_source\":\"invented fixture price; not an invoice\",\"cost_usd\":0.00036,\"input_tokens\":100,\"model_id\":\"scripted-judge\",\"output_tokens\":20,\"provider\":\"fixture\",\"response_ids\":[],\"total_tokens\":120},\"verdict\":\"pass\"}",
            "receipt_json": "{\"campaign_policy_hash\":\"sha256:618f45c0a349bff215f9e492e1160fd22a403851537178d5a836a778bf33f3a7\",\"completed_after_deadline\":false,\"completed_unix_ms\":1000,\"configuration_hash\":\"sha256:adb8da805361db8fe37314042f0b020b5c72558a64f700d3e93d46265267b1ad\",\"estimate_micro_usd\":360,\"invocation_id\":\"sha256:8decaed78747bfe125803ae9a10584c5ac1972aabdee1b018878e5eb878e9114\",\"judgment_hash\":\"sha256:3621664a36e1c18f072915d762624b1fabe11f1725bbfd5c5f6a0b44840c9a7a\",\"request_hash\":\"sha256:74d428a7d7dc6c7c8235d898a5f13e7df1e1c9b948d91202b4179ef5e5245ce4\",\"reservation_overrun\":false,\"reserved_micro_usd\":600,\"schema\":\"judge-campaign-receipt-v1\",\"state\":\"known\"}"
          },
          {
            "id": "sha256:c041d12bbfcf11293097d89ddf222e71df3cfa32f5c0e901c3f02411ac51a3dd",
            "request_hash": "sha256:74d428a7d7dc6c7c8235d898a5f13e7df1e1c9b948d91202b4179ef5e5245ce4",
            "configuration_hash": "sha256:adb8da805361db8fe37314042f0b020b5c72558a64f700d3e93d46265267b1ad",
            "reserved_micro": 600,
            "admitted_ms": 1000,
            "state": "known",
            "estimate_micro": 360,
            "judgment_json": "{\"campaign_audit_json\":null,\"explanation\":\"Scripted pass; not measured semantic accuracy.\",\"provider_audit_json\":\"{\\\"kind\\\":\\\"synthetic\\\"}\",\"runtime_evidence\":{\"cost_source\":\"invented fixture price; not an invoice\",\"cost_usd\":0.00036,\"input_tokens\":100,\"model_id\":\"scripted-judge\",\"output_tokens\":20,\"provider\":\"fixture\",\"response_ids\":[],\"total_tokens\":120},\"verdict\":\"pass\"}",
            "receipt_json": "{\"campaign_policy_hash\":\"sha256:618f45c0a349bff215f9e492e1160fd22a403851537178d5a836a778bf33f3a7\",\"completed_after_deadline\":false,\"completed_unix_ms\":1000,\"configuration_hash\":\"sha256:adb8da805361db8fe37314042f0b020b5c72558a64f700d3e93d46265267b1ad\",\"estimate_micro_usd\":360,\"invocation_id\":\"sha256:c041d12bbfcf11293097d89ddf222e71df3cfa32f5c0e901c3f02411ac51a3dd\",\"judgment_hash\":\"sha256:3621664a36e1c18f072915d762624b1fabe11f1725bbfd5c5f6a0b44840c9a7a\",\"request_hash\":\"sha256:74d428a7d7dc6c7c8235d898a5f13e7df1e1c9b948d91202b4179ef5e5245ce4\",\"reservation_overrun\":false,\"reserved_micro_usd\":600,\"schema\":\"judge-campaign-receipt-v1\",\"state\":\"known\"}"
          }
        ]
      },
      "operator_policy": {
        "campaign_id": "synthetic-judge-budget-v1",
        "max_estimated_micro_usd": 2000,
        "max_admissions": 4,
        "deadline_unix_ms": 2000,
        "reservation_micro_usd": 600
      },
      "simulated_operator_anchors": {
        "packet": "sha256:865d311ba5ffef1c76bd37f581dd3c93a3fef54d8835da21e47c76e00419ef33",
        "snapshot": "sha256:03cbfa204ea6da4eed4c6ea45e23f5ce5b80e3660137ebd201ba9022507fa98a"
      },
      "synthetic_calibration_hashes": [
        "sha256:66314faca3939d7125d6e475ed45fa8ec28a66eb7555be4d5957056ff435eff4"
      ],
      "fault_injection": "changed snapshot after anchoring",
      "assessment": {
        "packet_hash": "sha256:865d311ba5ffef1c76bd37f581dd3c93a3fef54d8835da21e47c76e00419ef33",
        "snapshot_hash": "sha256:e93debf42d572be2ceff20ae97fefe0be873f23004b270d07d0c88ac6a44ea43",
        "policy_hash": "sha256:618f45c0a349bff215f9e492e1160fd22a403851537178d5a836a778bf33f3a7",
        "status": "block",
        "issues": [
          "campaign_snapshot_anchor_mismatch"
        ],
        "matched_invocations": 0,
        "known_estimate_micro_usd": null,
        "held_reservations_micro_usd": null,
        "deployment_authorized": false
      },
      "test_evidence": {
        "campaign_expected": "block",
        "campaign_observed": "block",
        "replayed_trials": 4
      },
      "release_receipt": {
        "experiment_id": "synthetic-judge-admission-v1",
        "manifest_hash": "sha256:7efbb92fde089e48efe3e13a7c22f328b11d0d51327ae8dd61a1cdb4125ab2d6",
        "state": "evidence_ready",
        "action": "block",
        "authority_ceiling": "none",
        "raw_artifact_hash": "sha256:865d311ba5ffef1c76bd37f581dd3c93a3fef54d8835da21e47c76e00419ef33",
        "deterministic_test_receipt": {
          "suite_id": "offline-campaign-protocol-control",
          "code_revision": "working-tree-unpinned",
          "checks": [
            [
              "campaign_control_passed",
              true
            ],
            [
              "four_trials_replayed",
              true
            ]
          ],
          "artifact_hash": "sha256:47115e6b08c5875d0e81ce9382d685a3b09fc3488870f793409baff6309f3c53",
          "passed": true
        },
        "deterministic_tests_passed": true,
        "hard_failure_count": 0,
        "comparison": {
          "method": "clustered_normal_interval",
          "confidence_level": 0.95,
          "margin": 0.03,
          "pair_count": 2,
          "independent_cluster_count": 1,
          "minimum_independent_clusters": 30,
          "baseline_rate": 1.0,
          "candidate_rate": 1.0,
          "point_difference": 0.0,
          "lower_bound": null,
          "upper_bound": null,
          "non_inferior": false,
          "superior": false,
          "status": "inconclusive"
        },
        "prerequisite_receipts": [
          {
            "prerequisite_id": "typed_tool_boundary",
            "checks": [
              [
                "synthetic_control_only",
                true
              ]
            ],
            "evidence_hash": "sha256:43a9133216d6289d2814b6c09251d201e3922f7f2dd7142380de96fa5840678d",
            "state": "qualified"
          },
          {
            "prerequisite_id": "semantic_state_grading",
            "checks": [
              [
                "synthetic_control_only",
                true
              ]
            ],
            "evidence_hash": "sha256:40adfad9880316fe52a8691289140090ba489affb92dab8644b16d076c04a037",
            "state": "qualified"
          }
        ],
        "prerequisite_ids": [
          "typed_tool_boundary",
          "semantic_state_grading"
        ],
        "component_hashes": [
          [
            "campaign-policy",
            "sha256:618f45c0a349bff215f9e492e1160fd22a403851537178d5a836a778bf33f3a7"
          ],
          [
            "campaign_check",
            "sha256:7554215580d1b74338295816eb244f00b601ddbcd37405a39e50c1a6bae67563"
          ],
          [
            "dataset",
            "sha256:6cd4c0b0a4c7deea19aebf5af0a9931f73d0419154929fdf61b927a8f051a056"
          ],
          [
            "judge-config",
            "sha256:adb8da805361db8fe37314042f0b020b5c72558a64f700d3e93d46265267b1ad"
          ],
          [
            "manifest",
            "sha256:7efbb92fde089e48efe3e13a7c22f328b11d0d51327ae8dd61a1cdb4125ab2d6"
          ],
          [
            "population",
            "sha256:09145284b18a026e81706dfedba1d66b875bff46e60e9ca105045bd71a06ddf3"
          ],
          [
            "prerequisite:semantic_state_grading",
            "sha256:b5705f3d95bca6ca0420e9c62590d12f917b3d6a03ec869f27ccfc15cd8f0510"
          ],
          [
            "prerequisite:typed_tool_boundary",
            "sha256:ad1d20ae7ae53052e2007ef3e1f781755b7ea9da3eb48aa68ee71ad26131db12"
          ],
          [
            "source:cx_eval_lab/__init__.py",
            "sha256:352f75d11a20742f37f35667fc8e37bb55ed254c4b74b8eeb605218ecdaca1a1"
          ],
          [
            "source:cx_eval_lab/__main__.py",
            "sha256:75c2e5b562a49de5d3eed8e8886c23863ac8059aefbf27eb14bf97feef02856e"
          ],
          [
            "source:cx_eval_lab/advanced.py",
            "sha256:d4656eded62f9ba2aa7238e897eb3e6b0be9d58c7ca4efe68c8cacd7a60e417b"
          ],
          [
            "source:cx_eval_lab/agents.py",
            "sha256:80cd378bb1e3f2739cf931d49bdbe369b19fa688c98490801f322237aa2e3c02"
          ],
          [
            "source:cx_eval_lab/artifacts.py",
            "sha256:f65d08a5ae0029e0ecb764913f8ef77306cc68a8da5dbfe26623e78fe1d071c6"
          ],
          [
            "source:cx_eval_lab/budgeted_judge.py",
            "sha256:b30d4af0bd32f8943a6084157603e1d022b3616aca0ed9df745840dd8a411a8a"
          ],
          [
            "source:cx_eval_lab/calibration_data.py",
            "sha256:0568d843fd428d435b68cf422dd683b70297ce8296b7796bf0537eae6a0f59fa"
          ],
          [
            "source:cx_eval_lab/campaign_budget.py",
            "sha256:38dd2f28216fae871eabff7e324061a1b4c1f78ba869da285cb06c1aaa07e815"
          ],
          [
            "source:cx_eval_lab/campaign_conformance.py",
            "sha256:7e0198275eef4eb404a5f2c83e2af70f81477996b7bd5ab75ec3995c4748d8a4"
          ],
          [
            "source:cx_eval_lab/campaign_gate.py",
            "sha256:b0935bc913bd26f096dca063bd1240ca6ed802979a4ec4d36ac3ab3e6341e209"
          ],
          [
            "source:cx_eval_lab/campaign_study.py",
            "sha256:fa8715567fd5bd511a5e381e51c59f0d7637129a2afd5aec469067e4e66dfc8a"
          ],
          [
            "source:cx_eval_lab/ci_conformance.py",
            "sha256:bf888a2238a633e315ae11fca63e9dd9e038c0f17fdaac562808a8169f609b93"
          ],
          [
            "source:cx_eval_lab/cost_accounting.py",
            "sha256:c8831d6fde32d72c3ff3b9ca6befdac751b5a3de9064670da4bbad0407565091"
          ],
          [
            "source:cx_eval_lab/dataset.py",
            "sha256:d5f4a0cbf77f02c40dd6976d0c51e144d7c7020d8ada2e6bc8f12a615e0b9830"
          ],
          [
            "source:cx_eval_lab/evaluators.py",
            "sha256:232bfa08db8ed11bad6df3154ed4cac79d8f7ce29bc4bba58819fe1b5c073cb1"
          ],
          [
            "source:cx_eval_lab/evidence.py",
            "sha256:d60bf77ec43f037e6c30bf24a3d57d0a82b67f49deec4ddb321618453f0b256b"
          ],
          [
            "source:cx_eval_lab/exposure_control.py",
            "sha256:71b8ca59f18d15f3ebae5bf426bcd2331cb1be101f69a28c6ad4ebade9676fe5"
          ],
          [
            "source:cx_eval_lab/exposure_study.py",
            "sha256:471924d2082bc270e6799429d6179ce040cc84fe25cfb56d6c6d03dd66fbe609"
          ],
          [
            "source:cx_eval_lab/gate.py",
            "sha256:4fe0d5538588c83f415f3fc6dca8e9f432c33a256ebddb32bd3055900665ec83"
          ],
          [
            "source:cx_eval_lab/models.py",
            "sha256:a4fdac04db9cf505819cb1fbeed553506ec74d194055ba12f710968f3785f895"
          ],
          [
            "source:cx_eval_lab/openai_judge.py",
            "sha256:a9bcf76704c98f1876ea35d5a1119d1ec7e0d63291b7cbe08c467443b3cf57ec"
          ],
          [
            "source:cx_eval_lab/openai_runtime.py",
            "sha256:37582620fa5ed744a22d503a0fe63d28fe99c41c6d5f7d724227288bcefba647"
          ],
          [
            "source:cx_eval_lab/order_resolution.py",
            "sha256:a255b65be84392585a09c4156801400da10359024909eb7473e4240ca9756a94"
          ],
          [
            "source:cx_eval_lab/recovery_study.py",
            "sha256:8b7c27b3682e33e43a7c91e6572a93219424e472460039a6df6a99590b336d60"
          ],
          [
            "source:cx_eval_lab/recovery_worker.py",
            "sha256:87a4da856afdf16a9a5a124699dc62994df2575c7622584fdbd8f88435af3547"
          ],
          [
            "source:cx_eval_lab/replay_authority.py",
            "sha256:e51886e1346309716523b5c3e44d86c5e008109a1c1e534472eee11855321b03"
          ],
          [
            "source:cx_eval_lab/runner.py",
            "sha256:d29a5683a50c9f1afd8a414883d2c1f1ef5154fa86788ae9dba8fc3da29186be"
          ],
          [
            "source:cx_eval_lab/semantic.py",
            "sha256:6299a0863bae84ea87b8901b846f2e1563e2bcb60c75be378d42f615e3f5328e"
          ],
          [
            "source:cx_eval_lab/sequential_study.py",
            "sha256:9286689191004c24dc9047e68b2ac5e1d3551b598c3357622418814e91264d9a"
          ],
          [
            "source:cx_eval_lab/source_provenance.py",
            "sha256:4fdfe701884aa8ebc919509e97e11e12a0387d78328424ea0be22157823407ed"
          ],
          [
            "source:cx_eval_lab/statistical_study.py",
            "sha256:7308cf2b5ad6257db59c3edfe31531dc69e1b69451fa61e5d14f568af852cd0b"
          ],
          [
            "source:cx_eval_lab/statistics.py",
            "sha256:1568c8046ef1c57b74cf2535e14c104b380c56e20af812e94bfa81d6b849305b"
          ],
          [
            "source:cx_eval_lab/world.py",
            "sha256:086922f61fd9c4fb3f4c267a94153ba6b50d0b21486fa6d0c2cbb5595263a3cd"
          ],
          [
            "source:pyproject.toml",
            "sha256:2f2fd662fa150fc9b09e2578645ef4c61af75e9278dbb651e886a62e96b0bffc"
          ],
          [
            "source:uv.lock",
            "sha256:6c9ba06f8dfaaf440f6bab0a31da525baee10c330c984049e9642eb875df6ee8"
          ],
          [
            "test_receipt",
            "sha256:db5b3f4e85d43afc84d631ace86f407e84f231d054940381fb824b259b48aa7f"
          ]
        ],
        "issued_at": "2026-09-06T12:00:00Z",
        "valid_until": "2026-10-01T00:00:00Z",
        "invalidation_rules": [
          "source_or_dataset_change",
          "campaign_policy_or_judge_change"
        ],
        "campaign_check": {
          "deployment_authorized": false,
          "held_reservations_micro_usd": null,
          "issues": [
            "campaign_snapshot_anchor_mismatch"
          ],
          "known_estimate_micro_usd": null,
          "matched_invocations": 0,
          "packet_hash": "sha256:865d311ba5ffef1c76bd37f581dd3c93a3fef54d8835da21e47c76e00419ef33",
          "policy_hash": "sha256:618f45c0a349bff215f9e492e1160fd22a403851537178d5a836a778bf33f3a7",
          "snapshot_hash": "sha256:e93debf42d572be2ceff20ae97fefe0be873f23004b270d07d0c88ac6a44ea43",
          "status": "block"
        },
        "receipt_hash": "sha256:40d45855525bc3b0b66b8f943bc4d19658bbeeec98fd1ecab2ddbd315ca63f88"
      }
    }
  ],
  "limitations": "synthetic judges/prices/calibration/prerequisites; pinned snapshot consistency only; not current registry qualification, invoice or execution attestation"
}