{ "schema": "resolution-slice-study-v1", "source": { "file": "native-resolution-semantic-v1.json", "bytes_sha256": "sha256:638237b1615ec159c87d8ec8f2bb748867c2b319c93ff678470f4e0619419839", "canonical_hash": "sha256:9948fd0c7657e86b1d3aace49412912b3941b4275c678be1f6e3c4c927e902be" }, "calibration_hash": "sha256:454a18e2f18782f37ecfb2c7832a259aead13455be9cacfe94a7464e0d54730b", "comparisons": [ { "candidate": "false-settlement-mutant", "packet_hash": "sha256:c9cf089c6808b596bcf0fb810111bfbbb3a3c567638fbe490786933acdd3da58", "original_release_receipt_hash": "sha256:6ab2a7b761844845943a83a9071fe414ff2b5a2c2af38fb87ef8e9be4df50801", "original_release_action": "block", "plan": { "schema": "resolution-slice-plan-v1", "required_slices": [], "case_labels": { "ambiguous-description": [ "clarification:required", "resolution:target-known" ], "customer-correction": [ "clarification:not-required", "resolution:target-known" ], "explicit-description": [ "clarification:not-required", "resolution:target-known" ], "unresolved-description": [ "clarification:required", "resolution:unresolved" ] } }, "report": { "schema": "resolution-slices-v1", "packet_hash": "sha256:c9cf089c6808b596bcf0fb810111bfbbb3a3c567638fbe490786933acdd3da58", "plan_hash": "sha256:ede3e120b6912c24724df6de52e013128d8814e0408da4c9c9bc6815e3122832", "plan_binding": "retrospective_exploratory", "structural": { "overall": { "slice": "__overall__", "membership": "overall", "case_count": 4, "pair_count": 8, "customer_count": 1, "eligible_pair_count": 8, "eligible_customer_count": 1, "changes": { "fixed": [], "regressed": [], "unchanged_pass": [ { "case_id": "explicit-description", "trial_index": 0 }, { "case_id": "ambiguous-description", "trial_index": 0 }, { "case_id": "customer-correction", "trial_index": 0 }, { "case_id": "unresolved-description", "trial_index": 0 }, { "case_id": "explicit-description", "trial_index": 1 }, { "case_id": "ambiguous-description", "trial_index": 1 }, { "case_id": "customer-correction", "trial_index": 1 }, { "case_id": "unresolved-description", "trial_index": 1 } ], "unchanged_fail": [], "unqualified": [] }, "baseline_rate": 1.0, "candidate_rate": 1.0, "rate_population": "qualified pairs only; descriptive, not a missingness adjustment", "trial_weighted_delta": 0.0, "cluster_weighted_delta": 0.0, "comparison": null, "status": "hold", "issues": [] }, "slices": [ { "slice": "clarification:not-required", "membership": "exploratory", "case_count": 2, "pair_count": 4, "customer_count": 1, "eligible_pair_count": 4, "eligible_customer_count": 1, "changes": { "fixed": [], "regressed": [], "unchanged_pass": [ { "case_id": "explicit-description", "trial_index": 0 }, { "case_id": "customer-correction", "trial_index": 0 }, { "case_id": "explicit-description", "trial_index": 1 }, { "case_id": "customer-correction", "trial_index": 1 } ], "unchanged_fail": [], "unqualified": [] }, "baseline_rate": 1.0, "candidate_rate": 1.0, "rate_population": "qualified pairs only; descriptive, not a missingness adjustment", "trial_weighted_delta": 0.0, "cluster_weighted_delta": 0.0, "comparison": null, "status": "hold", "issues": [] }, { "slice": "clarification:required", "membership": "exploratory", "case_count": 2, "pair_count": 4, "customer_count": 1, "eligible_pair_count": 4, "eligible_customer_count": 1, "changes": { "fixed": [], "regressed": [], "unchanged_pass": [ { "case_id": "ambiguous-description", "trial_index": 0 }, { "case_id": "unresolved-description", "trial_index": 0 }, { "case_id": "ambiguous-description", "trial_index": 1 }, { "case_id": "unresolved-description", "trial_index": 1 } ], "unchanged_fail": [], "unqualified": [] }, "baseline_rate": 1.0, "candidate_rate": 1.0, "rate_population": "qualified pairs only; descriptive, not a missingness adjustment", "trial_weighted_delta": 0.0, "cluster_weighted_delta": 0.0, "comparison": null, "status": "hold", "issues": [] }, { "slice": "resolution:target-known", "membership": "exploratory", "case_count": 3, "pair_count": 6, "customer_count": 1, "eligible_pair_count": 6, "eligible_customer_count": 1, "changes": { "fixed": [], "regressed": [], "unchanged_pass": [ { "case_id": "explicit-description", "trial_index": 0 }, { "case_id": "ambiguous-description", "trial_index": 0 }, { "case_id": "customer-correction", "trial_index": 0 }, { "case_id": "explicit-description", "trial_index": 1 }, { "case_id": "ambiguous-description", "trial_index": 1 }, { "case_id": "customer-correction", "trial_index": 1 } ], "unchanged_fail": [], "unqualified": [] }, "baseline_rate": 1.0, "candidate_rate": 1.0, "rate_population": "qualified pairs only; descriptive, not a missingness adjustment", "trial_weighted_delta": 0.0, "cluster_weighted_delta": 0.0, "comparison": null, "status": "hold", "issues": [] }, { "slice": "resolution:unresolved", "membership": "exploratory", "case_count": 1, "pair_count": 2, "customer_count": 1, "eligible_pair_count": 2, "eligible_customer_count": 1, "changes": { "fixed": [], "regressed": [], "unchanged_pass": [ { "case_id": "unresolved-description", "trial_index": 0 }, { "case_id": "unresolved-description", "trial_index": 1 } ], "unchanged_fail": [], "unqualified": [] }, "baseline_rate": 1.0, "candidate_rate": 1.0, "rate_population": "qualified pairs only; descriptive, not a missingness adjustment", "trial_weighted_delta": 0.0, "cluster_weighted_delta": 0.0, "comparison": null, "status": "hold", "issues": [] } ], "pairs": [ { "case_id": "explicit-description", "trial_index": 0, "customer_id": "customer-1", "slices": [ "clarification:not-required", "resolution:target-known" ], "outcome": "unchanged_pass", "baseline": { "case_id": "explicit-description", "trial_index": 0, "cluster_id": "customer-1", "arm": "baseline", "passed": true, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:0eba359dcc5992a224186d36257c5604efbcb671f5be948cd15322af1203a89c", "failed_checks": [], "artifact_hash": "sha256:9485f74a7f14c3693753130c4fff188cd7f65bde032161e2cda25a5cf9dddeff" }, "candidate": { "case_id": "explicit-description", "trial_index": 0, "cluster_id": "customer-1", "arm": "candidate", "passed": true, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:0eba359dcc5992a224186d36257c5604efbcb671f5be948cd15322af1203a89c", "failed_checks": [], "artifact_hash": "sha256:84a9945d840b20e8077c452b8e5e1422d35db61999ba7c6d08ec83d6b39620ef" }, "baseline_evidence": { "summary": { "case_id": "explicit-description", "trial_index": 0, "cluster_id": "customer-1", "arm": "baseline", "passed": true, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:0eba359dcc5992a224186d36257c5604efbcb671f5be948cd15322af1203a89c", "failed_checks": [], "artifact_hash": "sha256:9485f74a7f14c3693753130c4fff188cd7f65bde032161e2cda25a5cf9dddeff" }, "structural_passed": true, "structural_checks": [ { "name": "wrong_order_commits", "passed": true, "observed": "0" }, { "name": "unnecessary_clarifications", "passed": true, "observed": "0" }, { "name": "missing_clarifications", "passed": true, "observed": "0" }, { "name": "denied_attempts", "passed": true, "observed": "0" }, { "name": "premature_action_attempts", "passed": true, "observed": "0" }, { "name": "resolution_contract", "passed": true, "observed": "structural outcome and response enum" } ], "semantic_status": { "qualified": true, "passed": true, "abstained": false, "reason": null }, "joint_passed": true, "joint_eligible": true }, "candidate_evidence": { "summary": { "case_id": "explicit-description", "trial_index": 0, "cluster_id": "customer-1", "arm": "candidate", "passed": false, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:0eba359dcc5992a224186d36257c5604efbcb671f5be948cd15322af1203a89c", "failed_checks": [ "qualified_multi_order_message" ], "artifact_hash": "sha256:84a9945d840b20e8077c452b8e5e1422d35db61999ba7c6d08ec83d6b39620ef" }, "structural_passed": true, "structural_checks": [ { "name": "wrong_order_commits", "passed": true, "observed": "0" }, { "name": "unnecessary_clarifications", "passed": true, "observed": "0" }, { "name": "missing_clarifications", "passed": true, "observed": "0" }, { "name": "denied_attempts", "passed": true, "observed": "0" }, { "name": "premature_action_attempts", "passed": true, "observed": "0" }, { "name": "resolution_contract", "passed": true, "observed": "structural outcome and response enum" } ], "semantic_status": { "qualified": true, "passed": false, "abstained": false, "reason": null }, "joint_passed": false, "joint_eligible": true } }, { "case_id": "ambiguous-description", "trial_index": 0, "customer_id": "customer-1", "slices": [ "clarification:required", "resolution:target-known" ], "outcome": "unchanged_pass", "baseline": { "case_id": "ambiguous-description", "trial_index": 0, "cluster_id": "customer-1", "arm": "baseline", "passed": true, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:0eba359dcc5992a224186d36257c5604efbcb671f5be948cd15322af1203a89c", "failed_checks": [], "artifact_hash": "sha256:fe27c89b932d93a53fd9b9d4685afc1d846113f41e4e5fc74d9279ce455d4743" }, "candidate": { "case_id": "ambiguous-description", "trial_index": 0, "cluster_id": "customer-1", "arm": "candidate", "passed": true, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:0eba359dcc5992a224186d36257c5604efbcb671f5be948cd15322af1203a89c", "failed_checks": [], "artifact_hash": "sha256:c202a97e5a2928134f1c32b5c7536d9cbddc31b8cceae507ede85819ca52829f" }, "baseline_evidence": { "summary": { "case_id": "ambiguous-description", "trial_index": 0, "cluster_id": "customer-1", "arm": "baseline", "passed": true, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:0eba359dcc5992a224186d36257c5604efbcb671f5be948cd15322af1203a89c", "failed_checks": [], "artifact_hash": "sha256:fe27c89b932d93a53fd9b9d4685afc1d846113f41e4e5fc74d9279ce455d4743" }, "structural_passed": true, "structural_checks": [ { "name": "wrong_order_commits", "passed": true, "observed": "0" }, { "name": "unnecessary_clarifications", "passed": true, "observed": "0" }, { "name": "missing_clarifications", "passed": true, "observed": "0" }, { "name": "denied_attempts", "passed": true, "observed": "0" }, { "name": "premature_action_attempts", "passed": true, "observed": "0" }, { "name": "resolution_contract", "passed": true, "observed": "structural outcome and response enum" } ], "semantic_status": { "qualified": true, "passed": true, "abstained": false, "reason": null }, "joint_passed": true, "joint_eligible": true }, "candidate_evidence": { "summary": { "case_id": "ambiguous-description", "trial_index": 0, "cluster_id": "customer-1", "arm": "candidate", "passed": false, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:0eba359dcc5992a224186d36257c5604efbcb671f5be948cd15322af1203a89c", "failed_checks": [ "qualified_multi_order_message" ], "artifact_hash": "sha256:c202a97e5a2928134f1c32b5c7536d9cbddc31b8cceae507ede85819ca52829f" }, "structural_passed": true, "structural_checks": [ { "name": "wrong_order_commits", "passed": true, "observed": "0" }, { "name": "unnecessary_clarifications", "passed": true, "observed": "0" }, { "name": "missing_clarifications", "passed": true, "observed": "0" }, { "name": "denied_attempts", "passed": true, "observed": "0" }, { "name": "premature_action_attempts", "passed": true, "observed": "0" }, { "name": "resolution_contract", "passed": true, "observed": "structural outcome and response enum" } ], "semantic_status": { "qualified": true, "passed": false, "abstained": false, "reason": null }, "joint_passed": false, "joint_eligible": true } }, { "case_id": "customer-correction", "trial_index": 0, "customer_id": "customer-1", "slices": [ "clarification:not-required", "resolution:target-known" ], "outcome": "unchanged_pass", "baseline": { "case_id": "customer-correction", "trial_index": 0, "cluster_id": "customer-1", "arm": "baseline", "passed": true, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:0eba359dcc5992a224186d36257c5604efbcb671f5be948cd15322af1203a89c", "failed_checks": [], "artifact_hash": "sha256:366610886e250c2b83a3af97a5ee8bc5111ec94d6838dd64167a1b21cda2d2cb" }, "candidate": { "case_id": "customer-correction", "trial_index": 0, "cluster_id": "customer-1", "arm": "candidate", "passed": true, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:0eba359dcc5992a224186d36257c5604efbcb671f5be948cd15322af1203a89c", "failed_checks": [], "artifact_hash": "sha256:14d9416ae6bc0f5e030d7983a36bb0923bbec0b3d76b47569bbe2ba3e8a008b5" }, "baseline_evidence": { "summary": { "case_id": "customer-correction", "trial_index": 0, "cluster_id": "customer-1", "arm": "baseline", "passed": true, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:0eba359dcc5992a224186d36257c5604efbcb671f5be948cd15322af1203a89c", "failed_checks": [], "artifact_hash": "sha256:366610886e250c2b83a3af97a5ee8bc5111ec94d6838dd64167a1b21cda2d2cb" }, "structural_passed": true, "structural_checks": [ { "name": "wrong_order_commits", "passed": true, "observed": "0" }, { "name": "unnecessary_clarifications", "passed": true, "observed": "0" }, { "name": "missing_clarifications", "passed": true, "observed": "0" }, { "name": "denied_attempts", "passed": true, "observed": "0" }, { "name": "premature_action_attempts", "passed": true, "observed": "0" }, { "name": "resolution_contract", "passed": true, "observed": "structural outcome and response enum" } ], "semantic_status": { "qualified": true, "passed": true, "abstained": false, "reason": null }, "joint_passed": true, "joint_eligible": true }, "candidate_evidence": { "summary": { "case_id": "customer-correction", "trial_index": 0, "cluster_id": "customer-1", "arm": "candidate", "passed": false, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:0eba359dcc5992a224186d36257c5604efbcb671f5be948cd15322af1203a89c", "failed_checks": [ "qualified_multi_order_message" ], "artifact_hash": "sha256:14d9416ae6bc0f5e030d7983a36bb0923bbec0b3d76b47569bbe2ba3e8a008b5" }, "structural_passed": true, "structural_checks": [ { "name": "wrong_order_commits", "passed": true, "observed": "0" }, { "name": "unnecessary_clarifications", "passed": true, "observed": "0" }, { "name": "missing_clarifications", "passed": true, "observed": "0" }, { "name": "denied_attempts", "passed": true, "observed": "0" }, { "name": "premature_action_attempts", "passed": true, "observed": "0" }, { "name": "resolution_contract", "passed": true, "observed": "structural outcome and response enum" } ], "semantic_status": { "qualified": true, "passed": false, "abstained": false, "reason": null }, "joint_passed": false, "joint_eligible": true } }, { "case_id": "unresolved-description", "trial_index": 0, "customer_id": "customer-1", "slices": [ "clarification:required", "resolution:unresolved" ], "outcome": "unchanged_pass", "baseline": { "case_id": "unresolved-description", "trial_index": 0, "cluster_id": "customer-1", "arm": "baseline", "passed": true, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:0eba359dcc5992a224186d36257c5604efbcb671f5be948cd15322af1203a89c", "failed_checks": [], "artifact_hash": "sha256:45cf3995df19745f749f98c6d0d454d34c70d6a3e6c210605678c7eb6a773c6f" }, "candidate": { "case_id": "unresolved-description", "trial_index": 0, "cluster_id": "customer-1", "arm": "candidate", "passed": true, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:0eba359dcc5992a224186d36257c5604efbcb671f5be948cd15322af1203a89c", "failed_checks": [], "artifact_hash": "sha256:32d5e797213185684ee68c4e6c6006537eea79b1184a145476d1c7dec19a73d1" }, "baseline_evidence": { "summary": { "case_id": "unresolved-description", "trial_index": 0, "cluster_id": "customer-1", "arm": "baseline", "passed": true, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:0eba359dcc5992a224186d36257c5604efbcb671f5be948cd15322af1203a89c", "failed_checks": [], "artifact_hash": "sha256:45cf3995df19745f749f98c6d0d454d34c70d6a3e6c210605678c7eb6a773c6f" }, "structural_passed": true, "structural_checks": [ { "name": "wrong_order_commits", "passed": true, "observed": "0" }, { "name": "unnecessary_clarifications", "passed": true, "observed": "0" }, { "name": "missing_clarifications", "passed": true, "observed": "0" }, { "name": "denied_attempts", "passed": true, "observed": "0" }, { "name": "premature_action_attempts", "passed": true, "observed": "0" }, { "name": "resolution_contract", "passed": true, "observed": "structural outcome and response enum" } ], "semantic_status": { "qualified": true, "passed": true, "abstained": false, "reason": null }, "joint_passed": true, "joint_eligible": true }, "candidate_evidence": { "summary": { "case_id": "unresolved-description", "trial_index": 0, "cluster_id": "customer-1", "arm": "candidate", "passed": false, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:0eba359dcc5992a224186d36257c5604efbcb671f5be948cd15322af1203a89c", "failed_checks": [ "qualified_multi_order_message" ], "artifact_hash": "sha256:32d5e797213185684ee68c4e6c6006537eea79b1184a145476d1c7dec19a73d1" }, "structural_passed": true, "structural_checks": [ { "name": "wrong_order_commits", "passed": true, "observed": "0" }, { "name": "unnecessary_clarifications", "passed": true, "observed": "0" }, { "name": "missing_clarifications", "passed": true, "observed": "0" }, { "name": "denied_attempts", "passed": true, "observed": "0" }, { "name": "premature_action_attempts", "passed": true, "observed": "0" }, { "name": "resolution_contract", "passed": true, "observed": "structural outcome and response enum" } ], "semantic_status": { "qualified": true, "passed": false, "abstained": false, "reason": null }, "joint_passed": false, "joint_eligible": true } }, { "case_id": "explicit-description", "trial_index": 1, "customer_id": "customer-1", "slices": [ "clarification:not-required", "resolution:target-known" ], "outcome": "unchanged_pass", "baseline": { "case_id": "explicit-description", "trial_index": 1, "cluster_id": "customer-1", "arm": "baseline", "passed": true, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:0eba359dcc5992a224186d36257c5604efbcb671f5be948cd15322af1203a89c", "failed_checks": [], "artifact_hash": "sha256:fd3f2367728e088bde20168980076de171e3fdcf5918bdb0b39900563b3d6acb" }, "candidate": { "case_id": "explicit-description", "trial_index": 1, "cluster_id": "customer-1", "arm": "candidate", "passed": true, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:0eba359dcc5992a224186d36257c5604efbcb671f5be948cd15322af1203a89c", "failed_checks": [], "artifact_hash": "sha256:fa12bb7ba7b3033761084e6ec1fd03fd7478969f25a5aec4057fecfdf3e6d5ab" }, "baseline_evidence": { "summary": { "case_id": "explicit-description", "trial_index": 1, "cluster_id": "customer-1", "arm": "baseline", "passed": true, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:0eba359dcc5992a224186d36257c5604efbcb671f5be948cd15322af1203a89c", "failed_checks": [], "artifact_hash": "sha256:fd3f2367728e088bde20168980076de171e3fdcf5918bdb0b39900563b3d6acb" }, "structural_passed": true, "structural_checks": [ { "name": "wrong_order_commits", "passed": true, "observed": "0" }, { "name": "unnecessary_clarifications", "passed": true, "observed": "0" }, { "name": "missing_clarifications", "passed": true, "observed": "0" }, { "name": "denied_attempts", "passed": true, "observed": "0" }, { "name": "premature_action_attempts", "passed": true, "observed": "0" }, { "name": "resolution_contract", "passed": true, "observed": "structural outcome and response enum" } ], "semantic_status": { "qualified": true, "passed": true, "abstained": false, "reason": null }, "joint_passed": true, "joint_eligible": true }, "candidate_evidence": { "summary": { "case_id": "explicit-description", "trial_index": 1, "cluster_id": "customer-1", "arm": "candidate", "passed": false, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:0eba359dcc5992a224186d36257c5604efbcb671f5be948cd15322af1203a89c", "failed_checks": [ "qualified_multi_order_message" ], "artifact_hash": "sha256:fa12bb7ba7b3033761084e6ec1fd03fd7478969f25a5aec4057fecfdf3e6d5ab" }, "structural_passed": true, "structural_checks": [ { "name": "wrong_order_commits", "passed": true, "observed": "0" }, { "name": "unnecessary_clarifications", "passed": true, "observed": "0" }, { "name": "missing_clarifications", "passed": true, "observed": "0" }, { "name": "denied_attempts", "passed": true, "observed": "0" }, { "name": "premature_action_attempts", "passed": true, "observed": "0" }, { "name": "resolution_contract", "passed": true, "observed": "structural outcome and response enum" } ], "semantic_status": { "qualified": true, "passed": false, "abstained": false, "reason": null }, "joint_passed": false, "joint_eligible": true } }, { "case_id": "ambiguous-description", "trial_index": 1, "customer_id": "customer-1", "slices": [ "clarification:required", "resolution:target-known" ], "outcome": "unchanged_pass", "baseline": { "case_id": "ambiguous-description", "trial_index": 1, "cluster_id": "customer-1", "arm": "baseline", "passed": true, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:0eba359dcc5992a224186d36257c5604efbcb671f5be948cd15322af1203a89c", "failed_checks": [], "artifact_hash": "sha256:3991635dc8f02568ebb800d7cdcf5a006ea021ff00d79e69d82c2ac39e892155" }, "candidate": { "case_id": "ambiguous-description", "trial_index": 1, "cluster_id": "customer-1", "arm": "candidate", "passed": true, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:0eba359dcc5992a224186d36257c5604efbcb671f5be948cd15322af1203a89c", "failed_checks": [], "artifact_hash": "sha256:c638551b65c3cecd041b500cad2ed63f90613ec64cb50b5a19abc5844f4e0224" }, "baseline_evidence": { "summary": { "case_id": "ambiguous-description", "trial_index": 1, "cluster_id": "customer-1", "arm": "baseline", "passed": true, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:0eba359dcc5992a224186d36257c5604efbcb671f5be948cd15322af1203a89c", "failed_checks": [], "artifact_hash": "sha256:3991635dc8f02568ebb800d7cdcf5a006ea021ff00d79e69d82c2ac39e892155" }, "structural_passed": true, "structural_checks": [ { "name": "wrong_order_commits", "passed": true, "observed": "0" }, { "name": "unnecessary_clarifications", "passed": true, "observed": "0" }, { "name": "missing_clarifications", "passed": true, "observed": "0" }, { "name": "denied_attempts", "passed": true, "observed": "0" }, { "name": "premature_action_attempts", "passed": true, "observed": "0" }, { "name": "resolution_contract", "passed": true, "observed": "structural outcome and response enum" } ], "semantic_status": { "qualified": true, "passed": true, "abstained": false, "reason": null }, "joint_passed": true, "joint_eligible": true }, "candidate_evidence": { "summary": { "case_id": "ambiguous-description", "trial_index": 1, "cluster_id": "customer-1", "arm": "candidate", "passed": false, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:0eba359dcc5992a224186d36257c5604efbcb671f5be948cd15322af1203a89c", "failed_checks": [ "qualified_multi_order_message" ], "artifact_hash": "sha256:c638551b65c3cecd041b500cad2ed63f90613ec64cb50b5a19abc5844f4e0224" }, "structural_passed": true, "structural_checks": [ { "name": "wrong_order_commits", "passed": true, "observed": "0" }, { "name": "unnecessary_clarifications", "passed": true, "observed": "0" }, { "name": "missing_clarifications", "passed": true, "observed": "0" }, { "name": "denied_attempts", "passed": true, "observed": "0" }, { "name": "premature_action_attempts", "passed": true, "observed": "0" }, { "name": "resolution_contract", "passed": true, "observed": "structural outcome and response enum" } ], "semantic_status": { "qualified": true, "passed": false, "abstained": false, "reason": null }, "joint_passed": false, "joint_eligible": true } }, { "case_id": "customer-correction", "trial_index": 1, "customer_id": "customer-1", "slices": [ "clarification:not-required", "resolution:target-known" ], "outcome": "unchanged_pass", "baseline": { "case_id": "customer-correction", "trial_index": 1, "cluster_id": "customer-1", "arm": "baseline", "passed": true, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:0eba359dcc5992a224186d36257c5604efbcb671f5be948cd15322af1203a89c", "failed_checks": [], "artifact_hash": "sha256:e092d1239d3e65a1d64a817c619461004a702f22863e23960f11b86fae8be7db" }, "candidate": { "case_id": "customer-correction", "trial_index": 1, "cluster_id": "customer-1", "arm": "candidate", "passed": true, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:0eba359dcc5992a224186d36257c5604efbcb671f5be948cd15322af1203a89c", "failed_checks": [], "artifact_hash": "sha256:5acb97836817899df50db297348a0f49e86e069ebe1646defa7364bb56c3ba1f" }, "baseline_evidence": { "summary": { "case_id": "customer-correction", "trial_index": 1, "cluster_id": "customer-1", "arm": "baseline", "passed": true, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:0eba359dcc5992a224186d36257c5604efbcb671f5be948cd15322af1203a89c", "failed_checks": [], "artifact_hash": "sha256:e092d1239d3e65a1d64a817c619461004a702f22863e23960f11b86fae8be7db" }, "structural_passed": true, "structural_checks": [ { "name": "wrong_order_commits", "passed": true, "observed": "0" }, { "name": "unnecessary_clarifications", "passed": true, "observed": "0" }, { "name": "missing_clarifications", "passed": true, "observed": "0" }, { "name": "denied_attempts", "passed": true, "observed": "0" }, { "name": "premature_action_attempts", "passed": true, "observed": "0" }, { "name": "resolution_contract", "passed": true, "observed": "structural outcome and response enum" } ], "semantic_status": { "qualified": true, "passed": true, "abstained": false, "reason": null }, "joint_passed": true, "joint_eligible": true }, "candidate_evidence": { "summary": { "case_id": "customer-correction", "trial_index": 1, "cluster_id": "customer-1", "arm": "candidate", "passed": false, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:0eba359dcc5992a224186d36257c5604efbcb671f5be948cd15322af1203a89c", "failed_checks": [ "qualified_multi_order_message" ], "artifact_hash": "sha256:5acb97836817899df50db297348a0f49e86e069ebe1646defa7364bb56c3ba1f" }, "structural_passed": true, "structural_checks": [ { "name": "wrong_order_commits", "passed": true, "observed": "0" }, { "name": "unnecessary_clarifications", "passed": true, "observed": "0" }, { "name": "missing_clarifications", "passed": true, "observed": "0" }, { "name": "denied_attempts", "passed": true, "observed": "0" }, { "name": "premature_action_attempts", "passed": true, "observed": "0" }, { "name": "resolution_contract", "passed": true, "observed": "structural outcome and response enum" } ], "semantic_status": { "qualified": true, "passed": false, "abstained": false, "reason": null }, "joint_passed": false, "joint_eligible": true } }, { "case_id": "unresolved-description", "trial_index": 1, "customer_id": "customer-1", "slices": [ "clarification:required", "resolution:unresolved" ], "outcome": "unchanged_pass", "baseline": { "case_id": "unresolved-description", "trial_index": 1, "cluster_id": "customer-1", "arm": "baseline", "passed": true, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:0eba359dcc5992a224186d36257c5604efbcb671f5be948cd15322af1203a89c", "failed_checks": [], "artifact_hash": "sha256:232078391b76eee68dbdebf0a509e31ce82edaec620c0ed54d7bed0e2c64bed3" }, "candidate": { "case_id": "unresolved-description", "trial_index": 1, "cluster_id": "customer-1", "arm": "candidate", "passed": true, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:0eba359dcc5992a224186d36257c5604efbcb671f5be948cd15322af1203a89c", "failed_checks": [], "artifact_hash": "sha256:62870031fe62a31233981a4f5ea49fc6f4682891ecd24681f68cb3ef17e7fb17" }, "baseline_evidence": { "summary": { "case_id": "unresolved-description", "trial_index": 1, "cluster_id": "customer-1", "arm": "baseline", "passed": true, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:0eba359dcc5992a224186d36257c5604efbcb671f5be948cd15322af1203a89c", "failed_checks": [], "artifact_hash": "sha256:232078391b76eee68dbdebf0a509e31ce82edaec620c0ed54d7bed0e2c64bed3" }, "structural_passed": true, "structural_checks": [ { "name": "wrong_order_commits", "passed": true, "observed": "0" }, { "name": "unnecessary_clarifications", "passed": true, "observed": "0" }, { "name": "missing_clarifications", "passed": true, "observed": "0" }, { "name": "denied_attempts", "passed": true, "observed": "0" }, { "name": "premature_action_attempts", "passed": true, "observed": "0" }, { "name": "resolution_contract", "passed": true, "observed": "structural outcome and response enum" } ], "semantic_status": { "qualified": true, "passed": true, "abstained": false, "reason": null }, "joint_passed": true, "joint_eligible": true }, "candidate_evidence": { "summary": { "case_id": "unresolved-description", "trial_index": 1, "cluster_id": "customer-1", "arm": "candidate", "passed": false, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:0eba359dcc5992a224186d36257c5604efbcb671f5be948cd15322af1203a89c", "failed_checks": [ "qualified_multi_order_message" ], "artifact_hash": "sha256:62870031fe62a31233981a4f5ea49fc6f4682891ecd24681f68cb3ef17e7fb17" }, "structural_passed": true, "structural_checks": [ { "name": "wrong_order_commits", "passed": true, "observed": "0" }, { "name": "unnecessary_clarifications", "passed": true, "observed": "0" }, { "name": "missing_clarifications", "passed": true, "observed": "0" }, { "name": "denied_attempts", "passed": true, "observed": "0" }, { "name": "premature_action_attempts", "passed": true, "observed": "0" }, { "name": "resolution_contract", "passed": true, "observed": "structural outcome and response enum" } ], "semantic_status": { "qualified": true, "passed": false, "abstained": false, "reason": null }, "joint_passed": false, "joint_eligible": true } } ], "status": "hold", "estimand_scope": "descriptive_structural_projection" }, "joint": { "overall": { "slice": "__overall__", "membership": "overall", "case_count": 4, "pair_count": 8, "customer_count": 1, "eligible_pair_count": 8, "eligible_customer_count": 1, "changes": { "fixed": [], "regressed": [ { "case_id": "explicit-description", "trial_index": 0 }, { "case_id": "ambiguous-description", "trial_index": 0 }, { "case_id": "customer-correction", "trial_index": 0 }, { "case_id": "unresolved-description", "trial_index": 0 }, { "case_id": "explicit-description", "trial_index": 1 }, { "case_id": "ambiguous-description", "trial_index": 1 }, { "case_id": "customer-correction", "trial_index": 1 }, { "case_id": "unresolved-description", "trial_index": 1 } ], "unchanged_pass": [], "unchanged_fail": [], "unqualified": [] }, "baseline_rate": 1.0, "candidate_rate": 0.0, "rate_population": "qualified pairs only; descriptive, not a missingness adjustment", "trial_weighted_delta": -1.0, "cluster_weighted_delta": -1.0, "comparison": null, "status": "hold", "issues": [] }, "slices": [ { "slice": "clarification:not-required", "membership": "exploratory", "case_count": 2, "pair_count": 4, "customer_count": 1, "eligible_pair_count": 4, "eligible_customer_count": 1, "changes": { "fixed": [], "regressed": [ { "case_id": "explicit-description", "trial_index": 0 }, { "case_id": "customer-correction", "trial_index": 0 }, { "case_id": "explicit-description", "trial_index": 1 }, { "case_id": "customer-correction", "trial_index": 1 } ], "unchanged_pass": [], "unchanged_fail": [], "unqualified": [] }, "baseline_rate": 1.0, "candidate_rate": 0.0, "rate_population": "qualified pairs only; descriptive, not a missingness adjustment", "trial_weighted_delta": -1.0, "cluster_weighted_delta": -1.0, "comparison": null, "status": "hold", "issues": [] }, { "slice": "clarification:required", "membership": "exploratory", "case_count": 2, "pair_count": 4, "customer_count": 1, "eligible_pair_count": 4, "eligible_customer_count": 1, "changes": { "fixed": [], "regressed": [ { "case_id": "ambiguous-description", "trial_index": 0 }, { "case_id": "unresolved-description", "trial_index": 0 }, { "case_id": "ambiguous-description", "trial_index": 1 }, { "case_id": "unresolved-description", "trial_index": 1 } ], "unchanged_pass": [], "unchanged_fail": [], "unqualified": [] }, "baseline_rate": 1.0, "candidate_rate": 0.0, "rate_population": "qualified pairs only; descriptive, not a missingness adjustment", "trial_weighted_delta": -1.0, "cluster_weighted_delta": -1.0, "comparison": null, "status": "hold", "issues": [] }, { "slice": "resolution:target-known", "membership": "exploratory", "case_count": 3, "pair_count": 6, "customer_count": 1, "eligible_pair_count": 6, "eligible_customer_count": 1, "changes": { "fixed": [], "regressed": [ { "case_id": "explicit-description", "trial_index": 0 }, { "case_id": "ambiguous-description", "trial_index": 0 }, { "case_id": "customer-correction", "trial_index": 0 }, { "case_id": "explicit-description", "trial_index": 1 }, { "case_id": "ambiguous-description", "trial_index": 1 }, { "case_id": "customer-correction", "trial_index": 1 } ], "unchanged_pass": [], "unchanged_fail": [], "unqualified": [] }, "baseline_rate": 1.0, "candidate_rate": 0.0, "rate_population": "qualified pairs only; descriptive, not a missingness adjustment", "trial_weighted_delta": -1.0, "cluster_weighted_delta": -1.0, "comparison": null, "status": "hold", "issues": [] }, { "slice": "resolution:unresolved", "membership": "exploratory", "case_count": 1, "pair_count": 2, "customer_count": 1, "eligible_pair_count": 2, "eligible_customer_count": 1, "changes": { "fixed": [], "regressed": [ { "case_id": "unresolved-description", "trial_index": 0 }, { "case_id": "unresolved-description", "trial_index": 1 } ], "unchanged_pass": [], "unchanged_fail": [], "unqualified": [] }, "baseline_rate": 1.0, "candidate_rate": 0.0, "rate_population": "qualified pairs only; descriptive, not a missingness adjustment", "trial_weighted_delta": -1.0, "cluster_weighted_delta": -1.0, "comparison": null, "status": "hold", "issues": [] } ], "pairs": [ { "case_id": "explicit-description", "trial_index": 0, "customer_id": "customer-1", "slices": [ "clarification:not-required", "resolution:target-known" ], "outcome": "regressed", "baseline": { "case_id": "explicit-description", "trial_index": 0, "cluster_id": "customer-1", "arm": "baseline", "passed": true, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:0eba359dcc5992a224186d36257c5604efbcb671f5be948cd15322af1203a89c", "failed_checks": [], "artifact_hash": "sha256:9485f74a7f14c3693753130c4fff188cd7f65bde032161e2cda25a5cf9dddeff" }, "candidate": { "case_id": "explicit-description", "trial_index": 0, "cluster_id": "customer-1", "arm": "candidate", "passed": false, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:0eba359dcc5992a224186d36257c5604efbcb671f5be948cd15322af1203a89c", "failed_checks": [ "qualified_multi_order_message" ], "artifact_hash": "sha256:84a9945d840b20e8077c452b8e5e1422d35db61999ba7c6d08ec83d6b39620ef" }, "baseline_evidence": { "summary": { "case_id": "explicit-description", "trial_index": 0, "cluster_id": "customer-1", "arm": "baseline", "passed": true, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:0eba359dcc5992a224186d36257c5604efbcb671f5be948cd15322af1203a89c", "failed_checks": [], "artifact_hash": "sha256:9485f74a7f14c3693753130c4fff188cd7f65bde032161e2cda25a5cf9dddeff" }, "structural_passed": true, "structural_checks": [ { "name": "wrong_order_commits", "passed": true, "observed": "0" }, { "name": "unnecessary_clarifications", "passed": true, "observed": "0" }, { "name": "missing_clarifications", "passed": true, "observed": "0" }, { "name": "denied_attempts", "passed": true, "observed": "0" }, { "name": "premature_action_attempts", "passed": true, "observed": "0" }, { "name": "resolution_contract", "passed": true, "observed": "structural outcome and response enum" } ], "semantic_status": { "qualified": true, "passed": true, "abstained": false, "reason": null }, "joint_passed": true, "joint_eligible": true }, "candidate_evidence": { "summary": { "case_id": "explicit-description", "trial_index": 0, "cluster_id": "customer-1", "arm": "candidate", "passed": false, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:0eba359dcc5992a224186d36257c5604efbcb671f5be948cd15322af1203a89c", "failed_checks": [ "qualified_multi_order_message" ], "artifact_hash": "sha256:84a9945d840b20e8077c452b8e5e1422d35db61999ba7c6d08ec83d6b39620ef" }, "structural_passed": true, "structural_checks": [ { "name": "wrong_order_commits", "passed": true, "observed": "0" }, { "name": "unnecessary_clarifications", "passed": true, "observed": "0" }, { "name": "missing_clarifications", "passed": true, "observed": "0" }, { "name": "denied_attempts", "passed": true, "observed": "0" }, { "name": "premature_action_attempts", "passed": true, "observed": "0" }, { "name": "resolution_contract", "passed": true, "observed": "structural outcome and response enum" } ], "semantic_status": { "qualified": true, "passed": false, "abstained": false, "reason": null }, "joint_passed": false, "joint_eligible": true } }, { "case_id": "ambiguous-description", "trial_index": 0, "customer_id": "customer-1", "slices": [ "clarification:required", "resolution:target-known" ], "outcome": "regressed", "baseline": { "case_id": "ambiguous-description", "trial_index": 0, "cluster_id": "customer-1", "arm": "baseline", "passed": true, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:0eba359dcc5992a224186d36257c5604efbcb671f5be948cd15322af1203a89c", "failed_checks": [], "artifact_hash": "sha256:fe27c89b932d93a53fd9b9d4685afc1d846113f41e4e5fc74d9279ce455d4743" }, "candidate": { "case_id": "ambiguous-description", "trial_index": 0, "cluster_id": "customer-1", "arm": "candidate", "passed": false, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:0eba359dcc5992a224186d36257c5604efbcb671f5be948cd15322af1203a89c", "failed_checks": [ "qualified_multi_order_message" ], "artifact_hash": "sha256:c202a97e5a2928134f1c32b5c7536d9cbddc31b8cceae507ede85819ca52829f" }, "baseline_evidence": { "summary": { "case_id": "ambiguous-description", "trial_index": 0, "cluster_id": "customer-1", "arm": "baseline", "passed": true, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:0eba359dcc5992a224186d36257c5604efbcb671f5be948cd15322af1203a89c", "failed_checks": [], "artifact_hash": "sha256:fe27c89b932d93a53fd9b9d4685afc1d846113f41e4e5fc74d9279ce455d4743" }, "structural_passed": true, "structural_checks": [ { "name": "wrong_order_commits", "passed": true, "observed": "0" }, { "name": "unnecessary_clarifications", "passed": true, "observed": "0" }, { "name": "missing_clarifications", "passed": true, "observed": "0" }, { "name": "denied_attempts", "passed": true, "observed": "0" }, { "name": "premature_action_attempts", "passed": true, "observed": "0" }, { "name": "resolution_contract", "passed": true, "observed": "structural outcome and response enum" } ], "semantic_status": { "qualified": true, "passed": true, "abstained": false, "reason": null }, "joint_passed": true, "joint_eligible": true }, "candidate_evidence": { "summary": { "case_id": "ambiguous-description", "trial_index": 0, "cluster_id": "customer-1", "arm": "candidate", "passed": false, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:0eba359dcc5992a224186d36257c5604efbcb671f5be948cd15322af1203a89c", "failed_checks": [ "qualified_multi_order_message" ], "artifact_hash": "sha256:c202a97e5a2928134f1c32b5c7536d9cbddc31b8cceae507ede85819ca52829f" }, "structural_passed": true, "structural_checks": [ { "name": "wrong_order_commits", "passed": true, "observed": "0" }, { "name": "unnecessary_clarifications", "passed": true, "observed": "0" }, { "name": "missing_clarifications", "passed": true, "observed": "0" }, { "name": "denied_attempts", "passed": true, "observed": "0" }, { "name": "premature_action_attempts", "passed": true, "observed": "0" }, { "name": "resolution_contract", "passed": true, "observed": "structural outcome and response enum" } ], "semantic_status": { "qualified": true, "passed": false, "abstained": false, "reason": null }, "joint_passed": false, "joint_eligible": true } }, { "case_id": "customer-correction", "trial_index": 0, "customer_id": "customer-1", "slices": [ "clarification:not-required", "resolution:target-known" ], "outcome": "regressed", "baseline": { "case_id": "customer-correction", "trial_index": 0, "cluster_id": "customer-1", "arm": "baseline", "passed": true, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:0eba359dcc5992a224186d36257c5604efbcb671f5be948cd15322af1203a89c", "failed_checks": [], "artifact_hash": "sha256:366610886e250c2b83a3af97a5ee8bc5111ec94d6838dd64167a1b21cda2d2cb" }, "candidate": { "case_id": "customer-correction", "trial_index": 0, "cluster_id": "customer-1", "arm": "candidate", "passed": false, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:0eba359dcc5992a224186d36257c5604efbcb671f5be948cd15322af1203a89c", "failed_checks": [ "qualified_multi_order_message" ], "artifact_hash": "sha256:14d9416ae6bc0f5e030d7983a36bb0923bbec0b3d76b47569bbe2ba3e8a008b5" }, "baseline_evidence": { "summary": { "case_id": "customer-correction", "trial_index": 0, "cluster_id": "customer-1", "arm": "baseline", "passed": true, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:0eba359dcc5992a224186d36257c5604efbcb671f5be948cd15322af1203a89c", "failed_checks": [], "artifact_hash": "sha256:366610886e250c2b83a3af97a5ee8bc5111ec94d6838dd64167a1b21cda2d2cb" }, "structural_passed": true, "structural_checks": [ { "name": "wrong_order_commits", "passed": true, "observed": "0" }, { "name": "unnecessary_clarifications", "passed": true, "observed": "0" }, { "name": "missing_clarifications", "passed": true, "observed": "0" }, { "name": "denied_attempts", "passed": true, "observed": "0" }, { "name": "premature_action_attempts", "passed": true, "observed": "0" }, { "name": "resolution_contract", "passed": true, "observed": "structural outcome and response enum" } ], "semantic_status": { "qualified": true, "passed": true, "abstained": false, "reason": null }, "joint_passed": true, "joint_eligible": true }, "candidate_evidence": { "summary": { "case_id": "customer-correction", "trial_index": 0, "cluster_id": "customer-1", "arm": "candidate", "passed": false, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:0eba359dcc5992a224186d36257c5604efbcb671f5be948cd15322af1203a89c", "failed_checks": [ "qualified_multi_order_message" ], "artifact_hash": "sha256:14d9416ae6bc0f5e030d7983a36bb0923bbec0b3d76b47569bbe2ba3e8a008b5" }, "structural_passed": true, "structural_checks": [ { "name": "wrong_order_commits", "passed": true, "observed": "0" }, { "name": "unnecessary_clarifications", "passed": true, "observed": "0" }, { "name": "missing_clarifications", "passed": true, "observed": "0" }, { "name": "denied_attempts", "passed": true, "observed": "0" }, { "name": "premature_action_attempts", "passed": true, "observed": "0" }, { "name": "resolution_contract", "passed": true, "observed": "structural outcome and response enum" } ], "semantic_status": { "qualified": true, "passed": false, "abstained": false, "reason": null }, "joint_passed": false, "joint_eligible": true } }, { "case_id": "unresolved-description", "trial_index": 0, "customer_id": "customer-1", "slices": [ "clarification:required", "resolution:unresolved" ], "outcome": "regressed", "baseline": { "case_id": "unresolved-description", "trial_index": 0, "cluster_id": "customer-1", "arm": "baseline", "passed": true, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:0eba359dcc5992a224186d36257c5604efbcb671f5be948cd15322af1203a89c", "failed_checks": [], "artifact_hash": "sha256:45cf3995df19745f749f98c6d0d454d34c70d6a3e6c210605678c7eb6a773c6f" }, "candidate": { "case_id": "unresolved-description", "trial_index": 0, "cluster_id": "customer-1", "arm": "candidate", "passed": false, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:0eba359dcc5992a224186d36257c5604efbcb671f5be948cd15322af1203a89c", "failed_checks": [ "qualified_multi_order_message" ], "artifact_hash": "sha256:32d5e797213185684ee68c4e6c6006537eea79b1184a145476d1c7dec19a73d1" }, "baseline_evidence": { "summary": { "case_id": "unresolved-description", "trial_index": 0, "cluster_id": "customer-1", "arm": "baseline", "passed": true, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:0eba359dcc5992a224186d36257c5604efbcb671f5be948cd15322af1203a89c", "failed_checks": [], "artifact_hash": "sha256:45cf3995df19745f749f98c6d0d454d34c70d6a3e6c210605678c7eb6a773c6f" }, "structural_passed": true, "structural_checks": [ { "name": "wrong_order_commits", "passed": true, "observed": "0" }, { "name": "unnecessary_clarifications", "passed": true, "observed": "0" }, { "name": "missing_clarifications", "passed": true, "observed": "0" }, { "name": "denied_attempts", "passed": true, "observed": "0" }, { "name": "premature_action_attempts", "passed": true, "observed": "0" }, { "name": "resolution_contract", "passed": true, "observed": "structural outcome and response enum" } ], "semantic_status": { "qualified": true, "passed": true, "abstained": false, "reason": null }, "joint_passed": true, "joint_eligible": true }, "candidate_evidence": { "summary": { "case_id": "unresolved-description", "trial_index": 0, "cluster_id": "customer-1", "arm": "candidate", "passed": false, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:0eba359dcc5992a224186d36257c5604efbcb671f5be948cd15322af1203a89c", "failed_checks": [ "qualified_multi_order_message" ], "artifact_hash": "sha256:32d5e797213185684ee68c4e6c6006537eea79b1184a145476d1c7dec19a73d1" }, "structural_passed": true, "structural_checks": [ { "name": "wrong_order_commits", "passed": true, "observed": "0" }, { "name": "unnecessary_clarifications", "passed": true, "observed": "0" }, { "name": "missing_clarifications", "passed": true, "observed": "0" }, { "name": "denied_attempts", "passed": true, "observed": "0" }, { "name": "premature_action_attempts", "passed": true, "observed": "0" }, { "name": "resolution_contract", "passed": true, "observed": "structural outcome and response enum" } ], "semantic_status": { "qualified": true, "passed": false, "abstained": false, "reason": null }, "joint_passed": false, "joint_eligible": true } }, { "case_id": "explicit-description", "trial_index": 1, "customer_id": "customer-1", "slices": [ "clarification:not-required", "resolution:target-known" ], "outcome": "regressed", "baseline": { "case_id": "explicit-description", "trial_index": 1, "cluster_id": "customer-1", "arm": "baseline", "passed": true, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:0eba359dcc5992a224186d36257c5604efbcb671f5be948cd15322af1203a89c", "failed_checks": [], "artifact_hash": "sha256:fd3f2367728e088bde20168980076de171e3fdcf5918bdb0b39900563b3d6acb" }, "candidate": { "case_id": "explicit-description", "trial_index": 1, "cluster_id": "customer-1", "arm": "candidate", "passed": false, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:0eba359dcc5992a224186d36257c5604efbcb671f5be948cd15322af1203a89c", "failed_checks": [ "qualified_multi_order_message" ], "artifact_hash": "sha256:fa12bb7ba7b3033761084e6ec1fd03fd7478969f25a5aec4057fecfdf3e6d5ab" }, "baseline_evidence": { "summary": { "case_id": "explicit-description", "trial_index": 1, "cluster_id": "customer-1", "arm": "baseline", "passed": true, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:0eba359dcc5992a224186d36257c5604efbcb671f5be948cd15322af1203a89c", "failed_checks": [], "artifact_hash": "sha256:fd3f2367728e088bde20168980076de171e3fdcf5918bdb0b39900563b3d6acb" }, "structural_passed": true, "structural_checks": [ { "name": "wrong_order_commits", "passed": true, "observed": "0" }, { "name": "unnecessary_clarifications", "passed": true, "observed": "0" }, { "name": "missing_clarifications", "passed": true, "observed": "0" }, { "name": "denied_attempts", "passed": true, "observed": "0" }, { "name": "premature_action_attempts", "passed": true, "observed": "0" }, { "name": "resolution_contract", "passed": true, "observed": "structural outcome and response enum" } ], "semantic_status": { "qualified": true, "passed": true, "abstained": false, "reason": null }, "joint_passed": true, "joint_eligible": true }, "candidate_evidence": { "summary": { "case_id": "explicit-description", "trial_index": 1, "cluster_id": "customer-1", "arm": "candidate", "passed": false, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:0eba359dcc5992a224186d36257c5604efbcb671f5be948cd15322af1203a89c", "failed_checks": [ "qualified_multi_order_message" ], "artifact_hash": "sha256:fa12bb7ba7b3033761084e6ec1fd03fd7478969f25a5aec4057fecfdf3e6d5ab" }, "structural_passed": true, "structural_checks": [ { "name": "wrong_order_commits", "passed": true, "observed": "0" }, { "name": "unnecessary_clarifications", "passed": true, "observed": "0" }, { "name": "missing_clarifications", "passed": true, "observed": "0" }, { "name": "denied_attempts", "passed": true, "observed": "0" }, { "name": "premature_action_attempts", "passed": true, "observed": "0" }, { "name": "resolution_contract", "passed": true, "observed": "structural outcome and response enum" } ], "semantic_status": { "qualified": true, "passed": false, "abstained": false, "reason": null }, "joint_passed": false, "joint_eligible": true } }, { "case_id": "ambiguous-description", "trial_index": 1, "customer_id": "customer-1", "slices": [ "clarification:required", "resolution:target-known" ], "outcome": "regressed", "baseline": { "case_id": "ambiguous-description", "trial_index": 1, "cluster_id": "customer-1", "arm": "baseline", "passed": true, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:0eba359dcc5992a224186d36257c5604efbcb671f5be948cd15322af1203a89c", "failed_checks": [], "artifact_hash": "sha256:3991635dc8f02568ebb800d7cdcf5a006ea021ff00d79e69d82c2ac39e892155" }, "candidate": { "case_id": "ambiguous-description", "trial_index": 1, "cluster_id": "customer-1", "arm": "candidate", "passed": false, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:0eba359dcc5992a224186d36257c5604efbcb671f5be948cd15322af1203a89c", "failed_checks": [ "qualified_multi_order_message" ], "artifact_hash": "sha256:c638551b65c3cecd041b500cad2ed63f90613ec64cb50b5a19abc5844f4e0224" }, "baseline_evidence": { "summary": { "case_id": "ambiguous-description", "trial_index": 1, "cluster_id": "customer-1", "arm": "baseline", "passed": true, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:0eba359dcc5992a224186d36257c5604efbcb671f5be948cd15322af1203a89c", "failed_checks": [], "artifact_hash": "sha256:3991635dc8f02568ebb800d7cdcf5a006ea021ff00d79e69d82c2ac39e892155" }, "structural_passed": true, "structural_checks": [ { "name": "wrong_order_commits", "passed": true, "observed": "0" }, { "name": "unnecessary_clarifications", "passed": true, "observed": "0" }, { "name": "missing_clarifications", "passed": true, "observed": "0" }, { "name": "denied_attempts", "passed": true, "observed": "0" }, { "name": "premature_action_attempts", "passed": true, "observed": "0" }, { "name": "resolution_contract", "passed": true, "observed": "structural outcome and response enum" } ], "semantic_status": { "qualified": true, "passed": true, "abstained": false, "reason": null }, "joint_passed": true, "joint_eligible": true }, "candidate_evidence": { "summary": { "case_id": "ambiguous-description", "trial_index": 1, "cluster_id": "customer-1", "arm": "candidate", "passed": false, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:0eba359dcc5992a224186d36257c5604efbcb671f5be948cd15322af1203a89c", "failed_checks": [ "qualified_multi_order_message" ], "artifact_hash": "sha256:c638551b65c3cecd041b500cad2ed63f90613ec64cb50b5a19abc5844f4e0224" }, "structural_passed": true, "structural_checks": [ { "name": "wrong_order_commits", "passed": true, "observed": "0" }, { "name": "unnecessary_clarifications", "passed": true, "observed": "0" }, { "name": "missing_clarifications", "passed": true, "observed": "0" }, { "name": "denied_attempts", "passed": true, "observed": "0" }, { "name": "premature_action_attempts", "passed": true, "observed": "0" }, { "name": "resolution_contract", "passed": true, "observed": "structural outcome and response enum" } ], "semantic_status": { "qualified": true, "passed": false, "abstained": false, "reason": null }, "joint_passed": false, "joint_eligible": true } }, { "case_id": "customer-correction", "trial_index": 1, "customer_id": "customer-1", "slices": [ "clarification:not-required", "resolution:target-known" ], "outcome": "regressed", "baseline": { "case_id": "customer-correction", "trial_index": 1, "cluster_id": "customer-1", "arm": "baseline", "passed": true, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:0eba359dcc5992a224186d36257c5604efbcb671f5be948cd15322af1203a89c", "failed_checks": [], "artifact_hash": "sha256:e092d1239d3e65a1d64a817c619461004a702f22863e23960f11b86fae8be7db" }, "candidate": { "case_id": "customer-correction", "trial_index": 1, "cluster_id": "customer-1", "arm": "candidate", "passed": false, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:0eba359dcc5992a224186d36257c5604efbcb671f5be948cd15322af1203a89c", "failed_checks": [ "qualified_multi_order_message" ], "artifact_hash": "sha256:5acb97836817899df50db297348a0f49e86e069ebe1646defa7364bb56c3ba1f" }, "baseline_evidence": { "summary": { "case_id": "customer-correction", "trial_index": 1, "cluster_id": "customer-1", "arm": "baseline", "passed": true, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:0eba359dcc5992a224186d36257c5604efbcb671f5be948cd15322af1203a89c", "failed_checks": [], "artifact_hash": "sha256:e092d1239d3e65a1d64a817c619461004a702f22863e23960f11b86fae8be7db" }, "structural_passed": true, "structural_checks": [ { "name": "wrong_order_commits", "passed": true, "observed": "0" }, { "name": "unnecessary_clarifications", "passed": true, "observed": "0" }, { "name": "missing_clarifications", "passed": true, "observed": "0" }, { "name": "denied_attempts", "passed": true, "observed": "0" }, { "name": "premature_action_attempts", "passed": true, "observed": "0" }, { "name": "resolution_contract", "passed": true, "observed": "structural outcome and response enum" } ], "semantic_status": { "qualified": true, "passed": true, "abstained": false, "reason": null }, "joint_passed": true, "joint_eligible": true }, "candidate_evidence": { "summary": { "case_id": "customer-correction", "trial_index": 1, "cluster_id": "customer-1", "arm": "candidate", "passed": false, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:0eba359dcc5992a224186d36257c5604efbcb671f5be948cd15322af1203a89c", "failed_checks": [ "qualified_multi_order_message" ], "artifact_hash": "sha256:5acb97836817899df50db297348a0f49e86e069ebe1646defa7364bb56c3ba1f" }, "structural_passed": true, "structural_checks": [ { "name": "wrong_order_commits", "passed": true, "observed": "0" }, { "name": "unnecessary_clarifications", "passed": true, "observed": "0" }, { "name": "missing_clarifications", "passed": true, "observed": "0" }, { "name": "denied_attempts", "passed": true, "observed": "0" }, { "name": "premature_action_attempts", "passed": true, "observed": "0" }, { "name": "resolution_contract", "passed": true, "observed": "structural outcome and response enum" } ], "semantic_status": { "qualified": true, "passed": false, "abstained": false, "reason": null }, "joint_passed": false, "joint_eligible": true } }, { "case_id": "unresolved-description", "trial_index": 1, "customer_id": "customer-1", "slices": [ "clarification:required", "resolution:unresolved" ], "outcome": "regressed", "baseline": { "case_id": "unresolved-description", "trial_index": 1, "cluster_id": "customer-1", "arm": "baseline", "passed": true, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:0eba359dcc5992a224186d36257c5604efbcb671f5be948cd15322af1203a89c", "failed_checks": [], "artifact_hash": "sha256:232078391b76eee68dbdebf0a509e31ce82edaec620c0ed54d7bed0e2c64bed3" }, "candidate": { "case_id": "unresolved-description", "trial_index": 1, "cluster_id": "customer-1", "arm": "candidate", "passed": false, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:0eba359dcc5992a224186d36257c5604efbcb671f5be948cd15322af1203a89c", "failed_checks": [ "qualified_multi_order_message" ], "artifact_hash": "sha256:62870031fe62a31233981a4f5ea49fc6f4682891ecd24681f68cb3ef17e7fb17" }, "baseline_evidence": { "summary": { "case_id": "unresolved-description", "trial_index": 1, "cluster_id": "customer-1", "arm": "baseline", "passed": true, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:0eba359dcc5992a224186d36257c5604efbcb671f5be948cd15322af1203a89c", "failed_checks": [], "artifact_hash": "sha256:232078391b76eee68dbdebf0a509e31ce82edaec620c0ed54d7bed0e2c64bed3" }, "structural_passed": true, "structural_checks": [ { "name": "wrong_order_commits", "passed": true, "observed": "0" }, { "name": "unnecessary_clarifications", "passed": true, "observed": "0" }, { "name": "missing_clarifications", "passed": true, "observed": "0" }, { "name": "denied_attempts", "passed": true, "observed": "0" }, { "name": "premature_action_attempts", "passed": true, "observed": "0" }, { "name": "resolution_contract", "passed": true, "observed": "structural outcome and response enum" } ], "semantic_status": { "qualified": true, "passed": true, "abstained": false, "reason": null }, "joint_passed": true, "joint_eligible": true }, "candidate_evidence": { "summary": { "case_id": "unresolved-description", "trial_index": 1, "cluster_id": "customer-1", "arm": "candidate", "passed": false, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:0eba359dcc5992a224186d36257c5604efbcb671f5be948cd15322af1203a89c", "failed_checks": [ "qualified_multi_order_message" ], "artifact_hash": "sha256:62870031fe62a31233981a4f5ea49fc6f4682891ecd24681f68cb3ef17e7fb17" }, "structural_passed": true, "structural_checks": [ { "name": "wrong_order_commits", "passed": true, "observed": "0" }, { "name": "unnecessary_clarifications", "passed": true, "observed": "0" }, { "name": "missing_clarifications", "passed": true, "observed": "0" }, { "name": "denied_attempts", "passed": true, "observed": "0" }, { "name": "premature_action_attempts", "passed": true, "observed": "0" }, { "name": "resolution_contract", "passed": true, "observed": "structural outcome and response enum" } ], "semantic_status": { "qualified": true, "passed": false, "abstained": false, "reason": null }, "joint_passed": false, "joint_eligible": true } } ], "status": "hold", "estimand_scope": "registered_joint_when_qualified" }, "status": "hold", "authority_ceiling": "lab_only", "overlapping_slices_not_additive": true, "deployment_authorized": false, "limitations": "Historical replay-derived diagnostics, not a new execution, current calibration, authenticated provenance or deployment permission. A plan hash binds declared labels, not their correctness or chronology. Feature labels and required slices need independent design review. v1 has no qualified joint outcome; v2 qualified false judgments are failures, while abstention/unqualified judgments remain unknown. Known structural defects remain visible alongside unknown joint evidence. Structural projections are descriptive only. Joint intervals use the existing unqualified teaching method, not multiplicity-qualified inference. Repetitions do not increase customer independence; overlapping slices are not additive.", "report_hash": "sha256:c06d879ec41c28ca957d663391dcfccc43ebf5e798ad92c713b26b7a0c5c5c84" } }, { "candidate": "first-record-mutant", "packet_hash": "sha256:998bcc1975f76a73f63952b6758539ebfb55182e5791257a71b94a62d1f22e9d", "original_release_receipt_hash": "sha256:4f56edcbdb94c7b4cf461103155f52dfd8179bbfb90820ce8d91423a5df383a2", "original_release_action": "block", "plan": { "schema": "resolution-slice-plan-v1", "required_slices": [], "case_labels": { "ambiguous-description": [ "clarification:required", "resolution:target-known" ], "customer-correction": [ "clarification:not-required", "resolution:target-known" ], "explicit-description": [ "clarification:not-required", "resolution:target-known" ], "unresolved-description": [ "clarification:required", "resolution:unresolved" ] } }, "report": { "schema": "resolution-slices-v1", "packet_hash": "sha256:998bcc1975f76a73f63952b6758539ebfb55182e5791257a71b94a62d1f22e9d", "plan_hash": "sha256:ede3e120b6912c24724df6de52e013128d8814e0408da4c9c9bc6815e3122832", "plan_binding": "retrospective_exploratory", "structural": { "overall": { "slice": "__overall__", "membership": "overall", "case_count": 4, "pair_count": 8, "customer_count": 1, "eligible_pair_count": 8, "eligible_customer_count": 1, "changes": { "fixed": [], "regressed": [ { "case_id": "explicit-description", "trial_index": 0 }, { "case_id": "ambiguous-description", "trial_index": 0 }, { "case_id": "unresolved-description", "trial_index": 0 }, { "case_id": "ambiguous-description", "trial_index": 1 }, { "case_id": "customer-correction", "trial_index": 1 }, { "case_id": "unresolved-description", "trial_index": 1 } ], "unchanged_pass": [ { "case_id": "customer-correction", "trial_index": 0 }, { "case_id": "explicit-description", "trial_index": 1 } ], "unchanged_fail": [], "unqualified": [] }, "baseline_rate": 1.0, "candidate_rate": 0.25, "rate_population": "qualified pairs only; descriptive, not a missingness adjustment", "trial_weighted_delta": -0.75, "cluster_weighted_delta": -0.75, "comparison": null, "status": "hold", "issues": [] }, "slices": [ { "slice": "clarification:not-required", "membership": "exploratory", "case_count": 2, "pair_count": 4, "customer_count": 1, "eligible_pair_count": 4, "eligible_customer_count": 1, "changes": { "fixed": [], "regressed": [ { "case_id": "explicit-description", "trial_index": 0 }, { "case_id": "customer-correction", "trial_index": 1 } ], "unchanged_pass": [ { "case_id": "customer-correction", "trial_index": 0 }, { "case_id": "explicit-description", "trial_index": 1 } ], "unchanged_fail": [], "unqualified": [] }, "baseline_rate": 1.0, "candidate_rate": 0.5, "rate_population": "qualified pairs only; descriptive, not a missingness adjustment", "trial_weighted_delta": -0.5, "cluster_weighted_delta": -0.5, "comparison": null, "status": "hold", "issues": [] }, { "slice": "clarification:required", "membership": "exploratory", "case_count": 2, "pair_count": 4, "customer_count": 1, "eligible_pair_count": 4, "eligible_customer_count": 1, "changes": { "fixed": [], "regressed": [ { "case_id": "ambiguous-description", "trial_index": 0 }, { "case_id": "unresolved-description", "trial_index": 0 }, { "case_id": "ambiguous-description", "trial_index": 1 }, { "case_id": "unresolved-description", "trial_index": 1 } ], "unchanged_pass": [], "unchanged_fail": [], "unqualified": [] }, "baseline_rate": 1.0, "candidate_rate": 0.0, "rate_population": "qualified pairs only; descriptive, not a missingness adjustment", "trial_weighted_delta": -1.0, "cluster_weighted_delta": -1.0, "comparison": null, "status": "hold", "issues": [] }, { "slice": "resolution:target-known", "membership": "exploratory", "case_count": 3, "pair_count": 6, "customer_count": 1, "eligible_pair_count": 6, "eligible_customer_count": 1, "changes": { "fixed": [], "regressed": [ { "case_id": "explicit-description", "trial_index": 0 }, { "case_id": "ambiguous-description", "trial_index": 0 }, { "case_id": "ambiguous-description", "trial_index": 1 }, { "case_id": "customer-correction", "trial_index": 1 } ], "unchanged_pass": [ { "case_id": "customer-correction", "trial_index": 0 }, { "case_id": "explicit-description", "trial_index": 1 } ], "unchanged_fail": [], "unqualified": [] }, "baseline_rate": 1.0, "candidate_rate": 0.3333333333333333, "rate_population": "qualified pairs only; descriptive, not a missingness adjustment", "trial_weighted_delta": -0.6666666666666667, "cluster_weighted_delta": -0.6666666666666666, "comparison": null, "status": "hold", "issues": [] }, { "slice": "resolution:unresolved", "membership": "exploratory", "case_count": 1, "pair_count": 2, "customer_count": 1, "eligible_pair_count": 2, "eligible_customer_count": 1, "changes": { "fixed": [], "regressed": [ { "case_id": "unresolved-description", "trial_index": 0 }, { "case_id": "unresolved-description", "trial_index": 1 } ], "unchanged_pass": [], "unchanged_fail": [], "unqualified": [] }, "baseline_rate": 1.0, "candidate_rate": 0.0, "rate_population": "qualified pairs only; descriptive, not a missingness adjustment", "trial_weighted_delta": -1.0, "cluster_weighted_delta": -1.0, "comparison": null, "status": "hold", "issues": [] } ], "pairs": [ { "case_id": "explicit-description", "trial_index": 0, "customer_id": "customer-1", "slices": [ "clarification:not-required", "resolution:target-known" ], "outcome": "regressed", "baseline": { "case_id": "explicit-description", "trial_index": 0, "cluster_id": "customer-1", "arm": "baseline", "passed": true, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:0d40666f55f47274554e2b4da2ad4a8f446797d0139b4d31f9f5afada82640e8", "failed_checks": [], "artifact_hash": "sha256:5922528a80a571ee7a2b513ab79c963b06efbc1ff76135856135a113b2a07d82" }, "candidate": { "case_id": "explicit-description", "trial_index": 0, "cluster_id": "customer-1", "arm": "candidate", "passed": false, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:0d40666f55f47274554e2b4da2ad4a8f446797d0139b4d31f9f5afada82640e8", "failed_checks": [ "wrong_order_commits", "resolution_contract" ], "artifact_hash": "sha256:14512659b6ee21994a74c536d90825c43a4d9b7a3ea930ea1cd7022dd2beef15" }, "baseline_evidence": { "summary": { "case_id": "explicit-description", "trial_index": 0, "cluster_id": "customer-1", "arm": "baseline", "passed": true, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:0d40666f55f47274554e2b4da2ad4a8f446797d0139b4d31f9f5afada82640e8", "failed_checks": [], "artifact_hash": "sha256:5922528a80a571ee7a2b513ab79c963b06efbc1ff76135856135a113b2a07d82" }, "structural_passed": true, "structural_checks": [ { "name": "wrong_order_commits", "passed": true, "observed": "0" }, { "name": "unnecessary_clarifications", "passed": true, "observed": "0" }, { "name": "missing_clarifications", "passed": true, "observed": "0" }, { "name": "denied_attempts", "passed": true, "observed": "0" }, { "name": "premature_action_attempts", "passed": true, "observed": "0" }, { "name": "resolution_contract", "passed": true, "observed": "structural outcome and response enum" } ], "semantic_status": { "qualified": true, "passed": true, "abstained": false, "reason": null }, "joint_passed": true, "joint_eligible": true }, "candidate_evidence": { "summary": { "case_id": "explicit-description", "trial_index": 0, "cluster_id": "customer-1", "arm": "candidate", "passed": false, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:0d40666f55f47274554e2b4da2ad4a8f446797d0139b4d31f9f5afada82640e8", "failed_checks": [ "wrong_order_commits", "resolution_contract" ], "artifact_hash": "sha256:14512659b6ee21994a74c536d90825c43a4d9b7a3ea930ea1cd7022dd2beef15" }, "structural_passed": false, "structural_checks": [ { "name": "wrong_order_commits", "passed": false, "observed": "1" }, { "name": "unnecessary_clarifications", "passed": true, "observed": "0" }, { "name": "missing_clarifications", "passed": true, "observed": "0" }, { "name": "denied_attempts", "passed": true, "observed": "0" }, { "name": "premature_action_attempts", "passed": true, "observed": "0" }, { "name": "resolution_contract", "passed": false, "observed": "structural outcome and response enum" } ], "semantic_status": { "qualified": true, "passed": true, "abstained": false, "reason": null }, "joint_passed": false, "joint_eligible": true } }, { "case_id": "ambiguous-description", "trial_index": 0, "customer_id": "customer-1", "slices": [ "clarification:required", "resolution:target-known" ], "outcome": "regressed", "baseline": { "case_id": "ambiguous-description", "trial_index": 0, "cluster_id": "customer-1", "arm": "baseline", "passed": true, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:0d40666f55f47274554e2b4da2ad4a8f446797d0139b4d31f9f5afada82640e8", "failed_checks": [], "artifact_hash": "sha256:6d45684da7bb243e7c77a59c798c980170c47a7d9e1fccd98767e70b7682647c" }, "candidate": { "case_id": "ambiguous-description", "trial_index": 0, "cluster_id": "customer-1", "arm": "candidate", "passed": false, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:0d40666f55f47274554e2b4da2ad4a8f446797d0139b4d31f9f5afada82640e8", "failed_checks": [ "missing_clarifications", "premature_action_attempts", "resolution_contract" ], "artifact_hash": "sha256:aa6906ede862bcfe7928eda57a3455b764b9e4211ec7adf56ce3f8b94eaa0620" }, "baseline_evidence": { "summary": { "case_id": "ambiguous-description", "trial_index": 0, "cluster_id": "customer-1", "arm": "baseline", "passed": true, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:0d40666f55f47274554e2b4da2ad4a8f446797d0139b4d31f9f5afada82640e8", "failed_checks": [], "artifact_hash": "sha256:6d45684da7bb243e7c77a59c798c980170c47a7d9e1fccd98767e70b7682647c" }, "structural_passed": true, "structural_checks": [ { "name": "wrong_order_commits", "passed": true, "observed": "0" }, { "name": "unnecessary_clarifications", "passed": true, "observed": "0" }, { "name": "missing_clarifications", "passed": true, "observed": "0" }, { "name": "denied_attempts", "passed": true, "observed": "0" }, { "name": "premature_action_attempts", "passed": true, "observed": "0" }, { "name": "resolution_contract", "passed": true, "observed": "structural outcome and response enum" } ], "semantic_status": { "qualified": true, "passed": true, "abstained": false, "reason": null }, "joint_passed": true, "joint_eligible": true }, "candidate_evidence": { "summary": { "case_id": "ambiguous-description", "trial_index": 0, "cluster_id": "customer-1", "arm": "candidate", "passed": false, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:0d40666f55f47274554e2b4da2ad4a8f446797d0139b4d31f9f5afada82640e8", "failed_checks": [ "missing_clarifications", "premature_action_attempts", "resolution_contract" ], "artifact_hash": "sha256:aa6906ede862bcfe7928eda57a3455b764b9e4211ec7adf56ce3f8b94eaa0620" }, "structural_passed": false, "structural_checks": [ { "name": "wrong_order_commits", "passed": true, "observed": "0" }, { "name": "unnecessary_clarifications", "passed": true, "observed": "0" }, { "name": "missing_clarifications", "passed": false, "observed": "1" }, { "name": "denied_attempts", "passed": true, "observed": "0" }, { "name": "premature_action_attempts", "passed": false, "observed": "1" }, { "name": "resolution_contract", "passed": false, "observed": "structural outcome and response enum" } ], "semantic_status": { "qualified": true, "passed": true, "abstained": false, "reason": null }, "joint_passed": false, "joint_eligible": true } }, { "case_id": "customer-correction", "trial_index": 0, "customer_id": "customer-1", "slices": [ "clarification:not-required", "resolution:target-known" ], "outcome": "unchanged_pass", "baseline": { "case_id": "customer-correction", "trial_index": 0, "cluster_id": "customer-1", "arm": "baseline", "passed": true, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:0d40666f55f47274554e2b4da2ad4a8f446797d0139b4d31f9f5afada82640e8", "failed_checks": [], "artifact_hash": "sha256:eb40603ec515fe9c49dc09662f9cd7902188daab9ff6a3a2e463d0041b2cfe08" }, "candidate": { "case_id": "customer-correction", "trial_index": 0, "cluster_id": "customer-1", "arm": "candidate", "passed": true, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:0d40666f55f47274554e2b4da2ad4a8f446797d0139b4d31f9f5afada82640e8", "failed_checks": [], "artifact_hash": "sha256:bb548d5446c026f131c530d1d2219e9f67c656516e5bcc6e99c2534157f997b5" }, "baseline_evidence": { "summary": { "case_id": "customer-correction", "trial_index": 0, "cluster_id": "customer-1", "arm": "baseline", "passed": true, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:0d40666f55f47274554e2b4da2ad4a8f446797d0139b4d31f9f5afada82640e8", "failed_checks": [], "artifact_hash": "sha256:eb40603ec515fe9c49dc09662f9cd7902188daab9ff6a3a2e463d0041b2cfe08" }, "structural_passed": true, "structural_checks": [ { "name": "wrong_order_commits", "passed": true, "observed": "0" }, { "name": "unnecessary_clarifications", "passed": true, "observed": "0" }, { "name": "missing_clarifications", "passed": true, "observed": "0" }, { "name": "denied_attempts", "passed": true, "observed": "0" }, { "name": "premature_action_attempts", "passed": true, "observed": "0" }, { "name": "resolution_contract", "passed": true, "observed": "structural outcome and response enum" } ], "semantic_status": { "qualified": true, "passed": true, "abstained": false, "reason": null }, "joint_passed": true, "joint_eligible": true }, "candidate_evidence": { "summary": { "case_id": "customer-correction", "trial_index": 0, "cluster_id": "customer-1", "arm": "candidate", "passed": true, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:0d40666f55f47274554e2b4da2ad4a8f446797d0139b4d31f9f5afada82640e8", "failed_checks": [], "artifact_hash": "sha256:bb548d5446c026f131c530d1d2219e9f67c656516e5bcc6e99c2534157f997b5" }, "structural_passed": true, "structural_checks": [ { "name": "wrong_order_commits", "passed": true, "observed": "0" }, { "name": "unnecessary_clarifications", "passed": true, "observed": "0" }, { "name": "missing_clarifications", "passed": true, "observed": "0" }, { "name": "denied_attempts", "passed": true, "observed": "0" }, { "name": "premature_action_attempts", "passed": true, "observed": "0" }, { "name": "resolution_contract", "passed": true, "observed": "structural outcome and response enum" } ], "semantic_status": { "qualified": true, "passed": true, "abstained": false, "reason": null }, "joint_passed": true, "joint_eligible": true } }, { "case_id": "unresolved-description", "trial_index": 0, "customer_id": "customer-1", "slices": [ "clarification:required", "resolution:unresolved" ], "outcome": "regressed", "baseline": { "case_id": "unresolved-description", "trial_index": 0, "cluster_id": "customer-1", "arm": "baseline", "passed": true, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:0d40666f55f47274554e2b4da2ad4a8f446797d0139b4d31f9f5afada82640e8", "failed_checks": [], "artifact_hash": "sha256:15e7d1c41ed2b4eaa56065e907d1fa65e2cb291a967e188b9c8208dfea41f645" }, "candidate": { "case_id": "unresolved-description", "trial_index": 0, "cluster_id": "customer-1", "arm": "candidate", "passed": false, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:0d40666f55f47274554e2b4da2ad4a8f446797d0139b4d31f9f5afada82640e8", "failed_checks": [ "wrong_order_commits", "missing_clarifications", "premature_action_attempts", "resolution_contract" ], "artifact_hash": "sha256:c1b2e50e86a4e86c86f23a11395560533a7562f170e34b75066f8fe77fcf36a8" }, "baseline_evidence": { "summary": { "case_id": "unresolved-description", "trial_index": 0, "cluster_id": "customer-1", "arm": "baseline", "passed": true, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:0d40666f55f47274554e2b4da2ad4a8f446797d0139b4d31f9f5afada82640e8", "failed_checks": [], "artifact_hash": "sha256:15e7d1c41ed2b4eaa56065e907d1fa65e2cb291a967e188b9c8208dfea41f645" }, "structural_passed": true, "structural_checks": [ { "name": "wrong_order_commits", "passed": true, "observed": "0" }, { "name": "unnecessary_clarifications", "passed": true, "observed": "0" }, { "name": "missing_clarifications", "passed": true, "observed": "0" }, { "name": "denied_attempts", "passed": true, "observed": "0" }, { "name": "premature_action_attempts", "passed": true, "observed": "0" }, { "name": "resolution_contract", "passed": true, "observed": "structural outcome and response enum" } ], "semantic_status": { "qualified": true, "passed": true, "abstained": false, "reason": null }, "joint_passed": true, "joint_eligible": true }, "candidate_evidence": { "summary": { "case_id": "unresolved-description", "trial_index": 0, "cluster_id": "customer-1", "arm": "candidate", "passed": false, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:0d40666f55f47274554e2b4da2ad4a8f446797d0139b4d31f9f5afada82640e8", "failed_checks": [ "wrong_order_commits", "missing_clarifications", "premature_action_attempts", "resolution_contract" ], "artifact_hash": "sha256:c1b2e50e86a4e86c86f23a11395560533a7562f170e34b75066f8fe77fcf36a8" }, "structural_passed": false, "structural_checks": [ { "name": "wrong_order_commits", "passed": false, "observed": "1" }, { "name": "unnecessary_clarifications", "passed": true, "observed": "0" }, { "name": "missing_clarifications", "passed": false, "observed": "1" }, { "name": "denied_attempts", "passed": true, "observed": "0" }, { "name": "premature_action_attempts", "passed": false, "observed": "1" }, { "name": "resolution_contract", "passed": false, "observed": "structural outcome and response enum" } ], "semantic_status": { "qualified": true, "passed": true, "abstained": false, "reason": null }, "joint_passed": false, "joint_eligible": true } }, { "case_id": "explicit-description", "trial_index": 1, "customer_id": "customer-1", "slices": [ "clarification:not-required", "resolution:target-known" ], "outcome": "unchanged_pass", "baseline": { "case_id": "explicit-description", "trial_index": 1, "cluster_id": "customer-1", "arm": "baseline", "passed": true, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:0d40666f55f47274554e2b4da2ad4a8f446797d0139b4d31f9f5afada82640e8", "failed_checks": [], "artifact_hash": "sha256:a7952628f956845eb43264e29c92091e5eff677fc7986521bc8cff6a6ee938d4" }, "candidate": { "case_id": "explicit-description", "trial_index": 1, "cluster_id": "customer-1", "arm": "candidate", "passed": true, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:0d40666f55f47274554e2b4da2ad4a8f446797d0139b4d31f9f5afada82640e8", "failed_checks": [], "artifact_hash": "sha256:5411d7547463c862e041214d5a7165644106b8c5ad7cd712e96099f1d2dac6af" }, "baseline_evidence": { "summary": { "case_id": "explicit-description", "trial_index": 1, "cluster_id": "customer-1", "arm": "baseline", "passed": true, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:0d40666f55f47274554e2b4da2ad4a8f446797d0139b4d31f9f5afada82640e8", "failed_checks": [], "artifact_hash": "sha256:a7952628f956845eb43264e29c92091e5eff677fc7986521bc8cff6a6ee938d4" }, "structural_passed": true, "structural_checks": [ { "name": "wrong_order_commits", "passed": true, "observed": "0" }, { "name": "unnecessary_clarifications", "passed": true, "observed": "0" }, { "name": "missing_clarifications", "passed": true, "observed": "0" }, { "name": "denied_attempts", "passed": true, "observed": "0" }, { "name": "premature_action_attempts", "passed": true, "observed": "0" }, { "name": "resolution_contract", "passed": true, "observed": "structural outcome and response enum" } ], "semantic_status": { "qualified": true, "passed": true, "abstained": false, "reason": null }, "joint_passed": true, "joint_eligible": true }, "candidate_evidence": { "summary": { "case_id": "explicit-description", "trial_index": 1, "cluster_id": "customer-1", "arm": "candidate", "passed": true, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:0d40666f55f47274554e2b4da2ad4a8f446797d0139b4d31f9f5afada82640e8", "failed_checks": [], "artifact_hash": "sha256:5411d7547463c862e041214d5a7165644106b8c5ad7cd712e96099f1d2dac6af" }, "structural_passed": true, "structural_checks": [ { "name": "wrong_order_commits", "passed": true, "observed": "0" }, { "name": "unnecessary_clarifications", "passed": true, "observed": "0" }, { "name": "missing_clarifications", "passed": true, "observed": "0" }, { "name": "denied_attempts", "passed": true, "observed": "0" }, { "name": "premature_action_attempts", "passed": true, "observed": "0" }, { "name": "resolution_contract", "passed": true, "observed": "structural outcome and response enum" } ], "semantic_status": { "qualified": true, "passed": true, "abstained": false, "reason": null }, "joint_passed": true, "joint_eligible": true } }, { "case_id": "ambiguous-description", "trial_index": 1, "customer_id": "customer-1", "slices": [ "clarification:required", "resolution:target-known" ], "outcome": "regressed", "baseline": { "case_id": "ambiguous-description", "trial_index": 1, "cluster_id": "customer-1", "arm": "baseline", "passed": true, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:0d40666f55f47274554e2b4da2ad4a8f446797d0139b4d31f9f5afada82640e8", "failed_checks": [], "artifact_hash": "sha256:4c12b88f3b3918be42a735c4146ec32bed03350af134ca04753217751e5814a1" }, "candidate": { "case_id": "ambiguous-description", "trial_index": 1, "cluster_id": "customer-1", "arm": "candidate", "passed": false, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:0d40666f55f47274554e2b4da2ad4a8f446797d0139b4d31f9f5afada82640e8", "failed_checks": [ "wrong_order_commits", "missing_clarifications", "premature_action_attempts", "resolution_contract" ], "artifact_hash": "sha256:ce31da700076e49e3f550ec99fea4505c23e2081848ac0e41478fa475abd2a42" }, "baseline_evidence": { "summary": { "case_id": "ambiguous-description", "trial_index": 1, "cluster_id": "customer-1", "arm": "baseline", "passed": true, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:0d40666f55f47274554e2b4da2ad4a8f446797d0139b4d31f9f5afada82640e8", "failed_checks": [], "artifact_hash": "sha256:4c12b88f3b3918be42a735c4146ec32bed03350af134ca04753217751e5814a1" }, "structural_passed": true, "structural_checks": [ { "name": "wrong_order_commits", "passed": true, "observed": "0" }, { "name": "unnecessary_clarifications", "passed": true, "observed": "0" }, { "name": "missing_clarifications", "passed": true, "observed": "0" }, { "name": "denied_attempts", "passed": true, "observed": "0" }, { "name": "premature_action_attempts", "passed": true, "observed": "0" }, { "name": "resolution_contract", "passed": true, "observed": "structural outcome and response enum" } ], "semantic_status": { "qualified": true, "passed": true, "abstained": false, "reason": null }, "joint_passed": true, "joint_eligible": true }, "candidate_evidence": { "summary": { "case_id": "ambiguous-description", "trial_index": 1, "cluster_id": "customer-1", "arm": "candidate", "passed": false, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:0d40666f55f47274554e2b4da2ad4a8f446797d0139b4d31f9f5afada82640e8", "failed_checks": [ "wrong_order_commits", "missing_clarifications", "premature_action_attempts", "resolution_contract" ], "artifact_hash": "sha256:ce31da700076e49e3f550ec99fea4505c23e2081848ac0e41478fa475abd2a42" }, "structural_passed": false, "structural_checks": [ { "name": "wrong_order_commits", "passed": false, "observed": "1" }, { "name": "unnecessary_clarifications", "passed": true, "observed": "0" }, { "name": "missing_clarifications", "passed": false, "observed": "1" }, { "name": "denied_attempts", "passed": true, "observed": "0" }, { "name": "premature_action_attempts", "passed": false, "observed": "1" }, { "name": "resolution_contract", "passed": false, "observed": "structural outcome and response enum" } ], "semantic_status": { "qualified": true, "passed": true, "abstained": false, "reason": null }, "joint_passed": false, "joint_eligible": true } }, { "case_id": "customer-correction", "trial_index": 1, "customer_id": "customer-1", "slices": [ "clarification:not-required", "resolution:target-known" ], "outcome": "regressed", "baseline": { "case_id": "customer-correction", "trial_index": 1, "cluster_id": "customer-1", "arm": "baseline", "passed": true, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:0d40666f55f47274554e2b4da2ad4a8f446797d0139b4d31f9f5afada82640e8", "failed_checks": [], "artifact_hash": "sha256:ab4a00a2a644f51ca2fe91f9245503aae887e93241ffd1559bfc56733e472802" }, "candidate": { "case_id": "customer-correction", "trial_index": 1, "cluster_id": "customer-1", "arm": "candidate", "passed": false, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:0d40666f55f47274554e2b4da2ad4a8f446797d0139b4d31f9f5afada82640e8", "failed_checks": [ "wrong_order_commits", "resolution_contract" ], "artifact_hash": "sha256:abd38a3079744c9fe26206e773df710806071290ace8d7cff411290a9d3efd94" }, "baseline_evidence": { "summary": { "case_id": "customer-correction", "trial_index": 1, "cluster_id": "customer-1", "arm": "baseline", "passed": true, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:0d40666f55f47274554e2b4da2ad4a8f446797d0139b4d31f9f5afada82640e8", "failed_checks": [], "artifact_hash": "sha256:ab4a00a2a644f51ca2fe91f9245503aae887e93241ffd1559bfc56733e472802" }, "structural_passed": true, "structural_checks": [ { "name": "wrong_order_commits", "passed": true, "observed": "0" }, { "name": "unnecessary_clarifications", "passed": true, "observed": "0" }, { "name": "missing_clarifications", "passed": true, "observed": "0" }, { "name": "denied_attempts", "passed": true, "observed": "0" }, { "name": "premature_action_attempts", "passed": true, "observed": "0" }, { "name": "resolution_contract", "passed": true, "observed": "structural outcome and response enum" } ], "semantic_status": { "qualified": true, "passed": true, "abstained": false, "reason": null }, "joint_passed": true, "joint_eligible": true }, "candidate_evidence": { "summary": { "case_id": "customer-correction", "trial_index": 1, "cluster_id": "customer-1", "arm": "candidate", "passed": false, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:0d40666f55f47274554e2b4da2ad4a8f446797d0139b4d31f9f5afada82640e8", "failed_checks": [ "wrong_order_commits", "resolution_contract" ], "artifact_hash": "sha256:abd38a3079744c9fe26206e773df710806071290ace8d7cff411290a9d3efd94" }, "structural_passed": false, "structural_checks": [ { "name": "wrong_order_commits", "passed": false, "observed": "1" }, { "name": "unnecessary_clarifications", "passed": true, "observed": "0" }, { "name": "missing_clarifications", "passed": true, "observed": "0" }, { "name": "denied_attempts", "passed": true, "observed": "0" }, { "name": "premature_action_attempts", "passed": true, "observed": "0" }, { "name": "resolution_contract", "passed": false, "observed": "structural outcome and response enum" } ], "semantic_status": { "qualified": true, "passed": true, "abstained": false, "reason": null }, "joint_passed": false, "joint_eligible": true } }, { "case_id": "unresolved-description", "trial_index": 1, "customer_id": "customer-1", "slices": [ "clarification:required", "resolution:unresolved" ], "outcome": "regressed", "baseline": { "case_id": "unresolved-description", "trial_index": 1, "cluster_id": "customer-1", "arm": "baseline", "passed": true, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:0d40666f55f47274554e2b4da2ad4a8f446797d0139b4d31f9f5afada82640e8", "failed_checks": [], "artifact_hash": "sha256:012513ba11e16966367b5f3e8c8e4ffff14b355c82c83b464a2ed13f327b3b92" }, "candidate": { "case_id": "unresolved-description", "trial_index": 1, "cluster_id": "customer-1", "arm": "candidate", "passed": false, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:0d40666f55f47274554e2b4da2ad4a8f446797d0139b4d31f9f5afada82640e8", "failed_checks": [ "wrong_order_commits", "missing_clarifications", "premature_action_attempts", "resolution_contract" ], "artifact_hash": "sha256:4faf5c3336372c4a0729e01fe1173cff31356dd79b65d37b099780d6f80171df" }, "baseline_evidence": { "summary": { "case_id": "unresolved-description", "trial_index": 1, "cluster_id": "customer-1", "arm": "baseline", "passed": true, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:0d40666f55f47274554e2b4da2ad4a8f446797d0139b4d31f9f5afada82640e8", "failed_checks": [], "artifact_hash": "sha256:012513ba11e16966367b5f3e8c8e4ffff14b355c82c83b464a2ed13f327b3b92" }, "structural_passed": true, "structural_checks": [ { "name": "wrong_order_commits", "passed": true, "observed": "0" }, { "name": "unnecessary_clarifications", "passed": true, "observed": "0" }, { "name": "missing_clarifications", "passed": true, "observed": "0" }, { "name": "denied_attempts", "passed": true, "observed": "0" }, { "name": "premature_action_attempts", "passed": true, "observed": "0" }, { "name": "resolution_contract", "passed": true, "observed": "structural outcome and response enum" } ], "semantic_status": { "qualified": true, "passed": true, "abstained": false, "reason": null }, "joint_passed": true, "joint_eligible": true }, "candidate_evidence": { "summary": { "case_id": "unresolved-description", "trial_index": 1, "cluster_id": "customer-1", "arm": "candidate", "passed": false, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:0d40666f55f47274554e2b4da2ad4a8f446797d0139b4d31f9f5afada82640e8", "failed_checks": [ "wrong_order_commits", "missing_clarifications", "premature_action_attempts", "resolution_contract" ], "artifact_hash": "sha256:4faf5c3336372c4a0729e01fe1173cff31356dd79b65d37b099780d6f80171df" }, "structural_passed": false, "structural_checks": [ { "name": "wrong_order_commits", "passed": false, "observed": "1" }, { "name": "unnecessary_clarifications", "passed": true, "observed": "0" }, { "name": "missing_clarifications", "passed": false, "observed": "1" }, { "name": "denied_attempts", "passed": true, "observed": "0" }, { "name": "premature_action_attempts", "passed": false, "observed": "1" }, { "name": "resolution_contract", "passed": false, "observed": "structural outcome and response enum" } ], "semantic_status": { "qualified": true, "passed": true, "abstained": false, "reason": null }, "joint_passed": false, "joint_eligible": true } } ], "status": "hold", "estimand_scope": "descriptive_structural_projection" }, "joint": { "overall": { "slice": "__overall__", "membership": "overall", "case_count": 4, "pair_count": 8, "customer_count": 1, "eligible_pair_count": 8, "eligible_customer_count": 1, "changes": { "fixed": [], "regressed": [ { "case_id": "explicit-description", "trial_index": 0 }, { "case_id": "ambiguous-description", "trial_index": 0 }, { "case_id": "unresolved-description", "trial_index": 0 }, { "case_id": "ambiguous-description", "trial_index": 1 }, { "case_id": "customer-correction", "trial_index": 1 }, { "case_id": "unresolved-description", "trial_index": 1 } ], "unchanged_pass": [ { "case_id": "customer-correction", "trial_index": 0 }, { "case_id": "explicit-description", "trial_index": 1 } ], "unchanged_fail": [], "unqualified": [] }, "baseline_rate": 1.0, "candidate_rate": 0.25, "rate_population": "qualified pairs only; descriptive, not a missingness adjustment", "trial_weighted_delta": -0.75, "cluster_weighted_delta": -0.75, "comparison": null, "status": "hold", "issues": [] }, "slices": [ { "slice": "clarification:not-required", "membership": "exploratory", "case_count": 2, "pair_count": 4, "customer_count": 1, "eligible_pair_count": 4, "eligible_customer_count": 1, "changes": { "fixed": [], "regressed": [ { "case_id": "explicit-description", "trial_index": 0 }, { "case_id": "customer-correction", "trial_index": 1 } ], "unchanged_pass": [ { "case_id": "customer-correction", "trial_index": 0 }, { "case_id": "explicit-description", "trial_index": 1 } ], "unchanged_fail": [], "unqualified": [] }, "baseline_rate": 1.0, "candidate_rate": 0.5, "rate_population": "qualified pairs only; descriptive, not a missingness adjustment", "trial_weighted_delta": -0.5, "cluster_weighted_delta": -0.5, "comparison": null, "status": "hold", "issues": [] }, { "slice": "clarification:required", "membership": "exploratory", "case_count": 2, "pair_count": 4, "customer_count": 1, "eligible_pair_count": 4, "eligible_customer_count": 1, "changes": { "fixed": [], "regressed": [ { "case_id": "ambiguous-description", "trial_index": 0 }, { "case_id": "unresolved-description", "trial_index": 0 }, { "case_id": "ambiguous-description", "trial_index": 1 }, { "case_id": "unresolved-description", "trial_index": 1 } ], "unchanged_pass": [], "unchanged_fail": [], "unqualified": [] }, "baseline_rate": 1.0, "candidate_rate": 0.0, "rate_population": "qualified pairs only; descriptive, not a missingness adjustment", "trial_weighted_delta": -1.0, "cluster_weighted_delta": -1.0, "comparison": null, "status": "hold", "issues": [] }, { "slice": "resolution:target-known", "membership": "exploratory", "case_count": 3, "pair_count": 6, "customer_count": 1, "eligible_pair_count": 6, "eligible_customer_count": 1, "changes": { "fixed": [], "regressed": [ { "case_id": "explicit-description", "trial_index": 0 }, { "case_id": "ambiguous-description", "trial_index": 0 }, { "case_id": "ambiguous-description", "trial_index": 1 }, { "case_id": "customer-correction", "trial_index": 1 } ], "unchanged_pass": [ { "case_id": "customer-correction", "trial_index": 0 }, { "case_id": "explicit-description", "trial_index": 1 } ], "unchanged_fail": [], "unqualified": [] }, "baseline_rate": 1.0, "candidate_rate": 0.3333333333333333, "rate_population": "qualified pairs only; descriptive, not a missingness adjustment", "trial_weighted_delta": -0.6666666666666667, "cluster_weighted_delta": -0.6666666666666666, "comparison": null, "status": "hold", "issues": [] }, { "slice": "resolution:unresolved", "membership": "exploratory", "case_count": 1, "pair_count": 2, "customer_count": 1, "eligible_pair_count": 2, "eligible_customer_count": 1, "changes": { "fixed": [], "regressed": [ { "case_id": "unresolved-description", "trial_index": 0 }, { "case_id": "unresolved-description", "trial_index": 1 } ], "unchanged_pass": [], "unchanged_fail": [], "unqualified": [] }, "baseline_rate": 1.0, "candidate_rate": 0.0, "rate_population": "qualified pairs only; descriptive, not a missingness adjustment", "trial_weighted_delta": -1.0, "cluster_weighted_delta": -1.0, "comparison": null, "status": "hold", "issues": [] } ], "pairs": [ { "case_id": "explicit-description", "trial_index": 0, "customer_id": "customer-1", "slices": [ "clarification:not-required", "resolution:target-known" ], "outcome": "regressed", "baseline": { "case_id": "explicit-description", "trial_index": 0, "cluster_id": "customer-1", "arm": "baseline", "passed": true, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:0d40666f55f47274554e2b4da2ad4a8f446797d0139b4d31f9f5afada82640e8", "failed_checks": [], "artifact_hash": "sha256:5922528a80a571ee7a2b513ab79c963b06efbc1ff76135856135a113b2a07d82" }, "candidate": { "case_id": "explicit-description", "trial_index": 0, "cluster_id": "customer-1", "arm": "candidate", "passed": false, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:0d40666f55f47274554e2b4da2ad4a8f446797d0139b4d31f9f5afada82640e8", "failed_checks": [ "wrong_order_commits", "resolution_contract" ], "artifact_hash": "sha256:14512659b6ee21994a74c536d90825c43a4d9b7a3ea930ea1cd7022dd2beef15" }, "baseline_evidence": { "summary": { "case_id": "explicit-description", "trial_index": 0, "cluster_id": "customer-1", "arm": "baseline", "passed": true, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:0d40666f55f47274554e2b4da2ad4a8f446797d0139b4d31f9f5afada82640e8", "failed_checks": [], "artifact_hash": "sha256:5922528a80a571ee7a2b513ab79c963b06efbc1ff76135856135a113b2a07d82" }, "structural_passed": true, "structural_checks": [ { "name": "wrong_order_commits", "passed": true, "observed": "0" }, { "name": "unnecessary_clarifications", "passed": true, "observed": "0" }, { "name": "missing_clarifications", "passed": true, "observed": "0" }, { "name": "denied_attempts", "passed": true, "observed": "0" }, { "name": "premature_action_attempts", "passed": true, "observed": "0" }, { "name": "resolution_contract", "passed": true, "observed": "structural outcome and response enum" } ], "semantic_status": { "qualified": true, "passed": true, "abstained": false, "reason": null }, "joint_passed": true, "joint_eligible": true }, "candidate_evidence": { "summary": { "case_id": "explicit-description", "trial_index": 0, "cluster_id": "customer-1", "arm": "candidate", "passed": false, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:0d40666f55f47274554e2b4da2ad4a8f446797d0139b4d31f9f5afada82640e8", "failed_checks": [ "wrong_order_commits", "resolution_contract" ], "artifact_hash": "sha256:14512659b6ee21994a74c536d90825c43a4d9b7a3ea930ea1cd7022dd2beef15" }, "structural_passed": false, "structural_checks": [ { "name": "wrong_order_commits", "passed": false, "observed": "1" }, { "name": "unnecessary_clarifications", "passed": true, "observed": "0" }, { "name": "missing_clarifications", "passed": true, "observed": "0" }, { "name": "denied_attempts", "passed": true, "observed": "0" }, { "name": "premature_action_attempts", "passed": true, "observed": "0" }, { "name": "resolution_contract", "passed": false, "observed": "structural outcome and response enum" } ], "semantic_status": { "qualified": true, "passed": true, "abstained": false, "reason": null }, "joint_passed": false, "joint_eligible": true } }, { "case_id": "ambiguous-description", "trial_index": 0, "customer_id": "customer-1", "slices": [ "clarification:required", "resolution:target-known" ], "outcome": "regressed", "baseline": { "case_id": "ambiguous-description", "trial_index": 0, "cluster_id": "customer-1", "arm": "baseline", "passed": true, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:0d40666f55f47274554e2b4da2ad4a8f446797d0139b4d31f9f5afada82640e8", "failed_checks": [], "artifact_hash": "sha256:6d45684da7bb243e7c77a59c798c980170c47a7d9e1fccd98767e70b7682647c" }, "candidate": { "case_id": "ambiguous-description", "trial_index": 0, "cluster_id": "customer-1", "arm": "candidate", "passed": false, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:0d40666f55f47274554e2b4da2ad4a8f446797d0139b4d31f9f5afada82640e8", "failed_checks": [ "missing_clarifications", "premature_action_attempts", "resolution_contract" ], "artifact_hash": "sha256:aa6906ede862bcfe7928eda57a3455b764b9e4211ec7adf56ce3f8b94eaa0620" }, "baseline_evidence": { "summary": { "case_id": "ambiguous-description", "trial_index": 0, "cluster_id": "customer-1", "arm": "baseline", "passed": true, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:0d40666f55f47274554e2b4da2ad4a8f446797d0139b4d31f9f5afada82640e8", "failed_checks": [], "artifact_hash": "sha256:6d45684da7bb243e7c77a59c798c980170c47a7d9e1fccd98767e70b7682647c" }, "structural_passed": true, "structural_checks": [ { "name": "wrong_order_commits", "passed": true, "observed": "0" }, { "name": "unnecessary_clarifications", "passed": true, "observed": "0" }, { "name": "missing_clarifications", "passed": true, "observed": "0" }, { "name": "denied_attempts", "passed": true, "observed": "0" }, { "name": "premature_action_attempts", "passed": true, "observed": "0" }, { "name": "resolution_contract", "passed": true, "observed": "structural outcome and response enum" } ], "semantic_status": { "qualified": true, "passed": true, "abstained": false, "reason": null }, "joint_passed": true, "joint_eligible": true }, "candidate_evidence": { "summary": { "case_id": "ambiguous-description", "trial_index": 0, "cluster_id": "customer-1", "arm": "candidate", "passed": false, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:0d40666f55f47274554e2b4da2ad4a8f446797d0139b4d31f9f5afada82640e8", "failed_checks": [ "missing_clarifications", "premature_action_attempts", "resolution_contract" ], "artifact_hash": "sha256:aa6906ede862bcfe7928eda57a3455b764b9e4211ec7adf56ce3f8b94eaa0620" }, "structural_passed": false, "structural_checks": [ { "name": "wrong_order_commits", "passed": true, "observed": "0" }, { "name": "unnecessary_clarifications", "passed": true, "observed": "0" }, { "name": "missing_clarifications", "passed": false, "observed": "1" }, { "name": "denied_attempts", "passed": true, "observed": "0" }, { "name": "premature_action_attempts", "passed": false, "observed": "1" }, { "name": "resolution_contract", "passed": false, "observed": "structural outcome and response enum" } ], "semantic_status": { "qualified": true, "passed": true, "abstained": false, "reason": null }, "joint_passed": false, "joint_eligible": true } }, { "case_id": "customer-correction", "trial_index": 0, "customer_id": "customer-1", "slices": [ "clarification:not-required", "resolution:target-known" ], "outcome": "unchanged_pass", "baseline": { "case_id": "customer-correction", "trial_index": 0, "cluster_id": "customer-1", "arm": "baseline", "passed": true, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:0d40666f55f47274554e2b4da2ad4a8f446797d0139b4d31f9f5afada82640e8", "failed_checks": [], "artifact_hash": "sha256:eb40603ec515fe9c49dc09662f9cd7902188daab9ff6a3a2e463d0041b2cfe08" }, "candidate": { "case_id": "customer-correction", "trial_index": 0, "cluster_id": "customer-1", "arm": "candidate", "passed": true, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:0d40666f55f47274554e2b4da2ad4a8f446797d0139b4d31f9f5afada82640e8", "failed_checks": [], "artifact_hash": "sha256:bb548d5446c026f131c530d1d2219e9f67c656516e5bcc6e99c2534157f997b5" }, "baseline_evidence": { "summary": { "case_id": "customer-correction", "trial_index": 0, "cluster_id": "customer-1", "arm": "baseline", "passed": true, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:0d40666f55f47274554e2b4da2ad4a8f446797d0139b4d31f9f5afada82640e8", "failed_checks": [], "artifact_hash": "sha256:eb40603ec515fe9c49dc09662f9cd7902188daab9ff6a3a2e463d0041b2cfe08" }, "structural_passed": true, "structural_checks": [ { "name": "wrong_order_commits", "passed": true, "observed": "0" }, { "name": "unnecessary_clarifications", "passed": true, "observed": "0" }, { "name": "missing_clarifications", "passed": true, "observed": "0" }, { "name": "denied_attempts", "passed": true, "observed": "0" }, { "name": "premature_action_attempts", "passed": true, "observed": "0" }, { "name": "resolution_contract", "passed": true, "observed": "structural outcome and response enum" } ], "semantic_status": { "qualified": true, "passed": true, "abstained": false, "reason": null }, "joint_passed": true, "joint_eligible": true }, "candidate_evidence": { "summary": { "case_id": "customer-correction", "trial_index": 0, "cluster_id": "customer-1", "arm": "candidate", "passed": true, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:0d40666f55f47274554e2b4da2ad4a8f446797d0139b4d31f9f5afada82640e8", "failed_checks": [], "artifact_hash": "sha256:bb548d5446c026f131c530d1d2219e9f67c656516e5bcc6e99c2534157f997b5" }, "structural_passed": true, "structural_checks": [ { "name": "wrong_order_commits", "passed": true, "observed": "0" }, { "name": "unnecessary_clarifications", "passed": true, "observed": "0" }, { "name": "missing_clarifications", "passed": true, "observed": "0" }, { "name": "denied_attempts", "passed": true, "observed": "0" }, { "name": "premature_action_attempts", "passed": true, "observed": "0" }, { "name": "resolution_contract", "passed": true, "observed": "structural outcome and response enum" } ], "semantic_status": { "qualified": true, "passed": true, "abstained": false, "reason": null }, "joint_passed": true, "joint_eligible": true } }, { "case_id": "unresolved-description", "trial_index": 0, "customer_id": "customer-1", "slices": [ "clarification:required", "resolution:unresolved" ], "outcome": "regressed", "baseline": { "case_id": "unresolved-description", "trial_index": 0, "cluster_id": "customer-1", "arm": "baseline", "passed": true, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:0d40666f55f47274554e2b4da2ad4a8f446797d0139b4d31f9f5afada82640e8", "failed_checks": [], "artifact_hash": "sha256:15e7d1c41ed2b4eaa56065e907d1fa65e2cb291a967e188b9c8208dfea41f645" }, "candidate": { "case_id": "unresolved-description", "trial_index": 0, "cluster_id": "customer-1", "arm": "candidate", "passed": false, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:0d40666f55f47274554e2b4da2ad4a8f446797d0139b4d31f9f5afada82640e8", "failed_checks": [ "wrong_order_commits", "missing_clarifications", "premature_action_attempts", "resolution_contract" ], "artifact_hash": "sha256:c1b2e50e86a4e86c86f23a11395560533a7562f170e34b75066f8fe77fcf36a8" }, "baseline_evidence": { "summary": { "case_id": "unresolved-description", "trial_index": 0, "cluster_id": "customer-1", "arm": "baseline", "passed": true, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:0d40666f55f47274554e2b4da2ad4a8f446797d0139b4d31f9f5afada82640e8", "failed_checks": [], "artifact_hash": "sha256:15e7d1c41ed2b4eaa56065e907d1fa65e2cb291a967e188b9c8208dfea41f645" }, "structural_passed": true, "structural_checks": [ { "name": "wrong_order_commits", "passed": true, "observed": "0" }, { "name": "unnecessary_clarifications", "passed": true, "observed": "0" }, { "name": "missing_clarifications", "passed": true, "observed": "0" }, { "name": "denied_attempts", "passed": true, "observed": "0" }, { "name": "premature_action_attempts", "passed": true, "observed": "0" }, { "name": "resolution_contract", "passed": true, "observed": "structural outcome and response enum" } ], "semantic_status": { "qualified": true, "passed": true, "abstained": false, "reason": null }, "joint_passed": true, "joint_eligible": true }, "candidate_evidence": { "summary": { "case_id": "unresolved-description", "trial_index": 0, "cluster_id": "customer-1", "arm": "candidate", "passed": false, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:0d40666f55f47274554e2b4da2ad4a8f446797d0139b4d31f9f5afada82640e8", "failed_checks": [ "wrong_order_commits", "missing_clarifications", "premature_action_attempts", "resolution_contract" ], "artifact_hash": "sha256:c1b2e50e86a4e86c86f23a11395560533a7562f170e34b75066f8fe77fcf36a8" }, "structural_passed": false, "structural_checks": [ { "name": "wrong_order_commits", "passed": false, "observed": "1" }, { "name": "unnecessary_clarifications", "passed": true, "observed": "0" }, { "name": "missing_clarifications", "passed": false, "observed": "1" }, { "name": "denied_attempts", "passed": true, "observed": "0" }, { "name": "premature_action_attempts", "passed": false, "observed": "1" }, { "name": "resolution_contract", "passed": false, "observed": "structural outcome and response enum" } ], "semantic_status": { "qualified": true, "passed": true, "abstained": false, "reason": null }, "joint_passed": false, "joint_eligible": true } }, { "case_id": "explicit-description", "trial_index": 1, "customer_id": "customer-1", "slices": [ "clarification:not-required", "resolution:target-known" ], "outcome": "unchanged_pass", "baseline": { "case_id": "explicit-description", "trial_index": 1, "cluster_id": "customer-1", "arm": "baseline", "passed": true, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:0d40666f55f47274554e2b4da2ad4a8f446797d0139b4d31f9f5afada82640e8", "failed_checks": [], "artifact_hash": "sha256:a7952628f956845eb43264e29c92091e5eff677fc7986521bc8cff6a6ee938d4" }, "candidate": { "case_id": "explicit-description", "trial_index": 1, "cluster_id": "customer-1", "arm": "candidate", "passed": true, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:0d40666f55f47274554e2b4da2ad4a8f446797d0139b4d31f9f5afada82640e8", "failed_checks": [], "artifact_hash": "sha256:5411d7547463c862e041214d5a7165644106b8c5ad7cd712e96099f1d2dac6af" }, "baseline_evidence": { "summary": { "case_id": "explicit-description", "trial_index": 1, "cluster_id": "customer-1", "arm": "baseline", "passed": true, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:0d40666f55f47274554e2b4da2ad4a8f446797d0139b4d31f9f5afada82640e8", "failed_checks": [], "artifact_hash": "sha256:a7952628f956845eb43264e29c92091e5eff677fc7986521bc8cff6a6ee938d4" }, "structural_passed": true, "structural_checks": [ { "name": "wrong_order_commits", "passed": true, "observed": "0" }, { "name": "unnecessary_clarifications", "passed": true, "observed": "0" }, { "name": "missing_clarifications", "passed": true, "observed": "0" }, { "name": "denied_attempts", "passed": true, "observed": "0" }, { "name": "premature_action_attempts", "passed": true, "observed": "0" }, { "name": "resolution_contract", "passed": true, "observed": "structural outcome and response enum" } ], "semantic_status": { "qualified": true, "passed": true, "abstained": false, "reason": null }, "joint_passed": true, "joint_eligible": true }, "candidate_evidence": { "summary": { "case_id": "explicit-description", "trial_index": 1, "cluster_id": "customer-1", "arm": "candidate", "passed": true, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:0d40666f55f47274554e2b4da2ad4a8f446797d0139b4d31f9f5afada82640e8", "failed_checks": [], "artifact_hash": "sha256:5411d7547463c862e041214d5a7165644106b8c5ad7cd712e96099f1d2dac6af" }, "structural_passed": true, "structural_checks": [ { "name": "wrong_order_commits", "passed": true, "observed": "0" }, { "name": "unnecessary_clarifications", "passed": true, "observed": "0" }, { "name": "missing_clarifications", "passed": true, "observed": "0" }, { "name": "denied_attempts", "passed": true, "observed": "0" }, { "name": "premature_action_attempts", "passed": true, "observed": "0" }, { "name": "resolution_contract", "passed": true, "observed": "structural outcome and response enum" } ], "semantic_status": { "qualified": true, "passed": true, "abstained": false, "reason": null }, "joint_passed": true, "joint_eligible": true } }, { "case_id": "ambiguous-description", "trial_index": 1, "customer_id": "customer-1", "slices": [ "clarification:required", "resolution:target-known" ], "outcome": "regressed", "baseline": { "case_id": "ambiguous-description", "trial_index": 1, "cluster_id": "customer-1", "arm": "baseline", "passed": true, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:0d40666f55f47274554e2b4da2ad4a8f446797d0139b4d31f9f5afada82640e8", "failed_checks": [], "artifact_hash": "sha256:4c12b88f3b3918be42a735c4146ec32bed03350af134ca04753217751e5814a1" }, "candidate": { "case_id": "ambiguous-description", "trial_index": 1, "cluster_id": "customer-1", "arm": "candidate", "passed": false, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:0d40666f55f47274554e2b4da2ad4a8f446797d0139b4d31f9f5afada82640e8", "failed_checks": [ "wrong_order_commits", "missing_clarifications", "premature_action_attempts", "resolution_contract" ], "artifact_hash": "sha256:ce31da700076e49e3f550ec99fea4505c23e2081848ac0e41478fa475abd2a42" }, "baseline_evidence": { "summary": { "case_id": "ambiguous-description", "trial_index": 1, "cluster_id": "customer-1", "arm": "baseline", "passed": true, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:0d40666f55f47274554e2b4da2ad4a8f446797d0139b4d31f9f5afada82640e8", "failed_checks": [], "artifact_hash": "sha256:4c12b88f3b3918be42a735c4146ec32bed03350af134ca04753217751e5814a1" }, "structural_passed": true, "structural_checks": [ { "name": "wrong_order_commits", "passed": true, "observed": "0" }, { "name": "unnecessary_clarifications", "passed": true, "observed": "0" }, { "name": "missing_clarifications", "passed": true, "observed": "0" }, { "name": "denied_attempts", "passed": true, "observed": "0" }, { "name": "premature_action_attempts", "passed": true, "observed": "0" }, { "name": "resolution_contract", "passed": true, "observed": "structural outcome and response enum" } ], "semantic_status": { "qualified": true, "passed": true, "abstained": false, "reason": null }, "joint_passed": true, "joint_eligible": true }, "candidate_evidence": { "summary": { "case_id": "ambiguous-description", "trial_index": 1, "cluster_id": "customer-1", "arm": "candidate", "passed": false, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:0d40666f55f47274554e2b4da2ad4a8f446797d0139b4d31f9f5afada82640e8", "failed_checks": [ "wrong_order_commits", "missing_clarifications", "premature_action_attempts", "resolution_contract" ], "artifact_hash": "sha256:ce31da700076e49e3f550ec99fea4505c23e2081848ac0e41478fa475abd2a42" }, "structural_passed": false, "structural_checks": [ { "name": "wrong_order_commits", "passed": false, "observed": "1" }, { "name": "unnecessary_clarifications", "passed": true, "observed": "0" }, { "name": "missing_clarifications", "passed": false, "observed": "1" }, { "name": "denied_attempts", "passed": true, "observed": "0" }, { "name": "premature_action_attempts", "passed": false, "observed": "1" }, { "name": "resolution_contract", "passed": false, "observed": "structural outcome and response enum" } ], "semantic_status": { "qualified": true, "passed": true, "abstained": false, "reason": null }, "joint_passed": false, "joint_eligible": true } }, { "case_id": "customer-correction", "trial_index": 1, "customer_id": "customer-1", "slices": [ "clarification:not-required", "resolution:target-known" ], "outcome": "regressed", "baseline": { "case_id": "customer-correction", "trial_index": 1, "cluster_id": "customer-1", "arm": "baseline", "passed": true, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:0d40666f55f47274554e2b4da2ad4a8f446797d0139b4d31f9f5afada82640e8", "failed_checks": [], "artifact_hash": "sha256:ab4a00a2a644f51ca2fe91f9245503aae887e93241ffd1559bfc56733e472802" }, "candidate": { "case_id": "customer-correction", "trial_index": 1, "cluster_id": "customer-1", "arm": "candidate", "passed": false, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:0d40666f55f47274554e2b4da2ad4a8f446797d0139b4d31f9f5afada82640e8", "failed_checks": [ "wrong_order_commits", "resolution_contract" ], "artifact_hash": "sha256:abd38a3079744c9fe26206e773df710806071290ace8d7cff411290a9d3efd94" }, "baseline_evidence": { "summary": { "case_id": "customer-correction", "trial_index": 1, "cluster_id": "customer-1", "arm": "baseline", "passed": true, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:0d40666f55f47274554e2b4da2ad4a8f446797d0139b4d31f9f5afada82640e8", "failed_checks": [], "artifact_hash": "sha256:ab4a00a2a644f51ca2fe91f9245503aae887e93241ffd1559bfc56733e472802" }, "structural_passed": true, "structural_checks": [ { "name": "wrong_order_commits", "passed": true, "observed": "0" }, { "name": "unnecessary_clarifications", "passed": true, "observed": "0" }, { "name": "missing_clarifications", "passed": true, "observed": "0" }, { "name": "denied_attempts", "passed": true, "observed": "0" }, { "name": "premature_action_attempts", "passed": true, "observed": "0" }, { "name": "resolution_contract", "passed": true, "observed": "structural outcome and response enum" } ], "semantic_status": { "qualified": true, "passed": true, "abstained": false, "reason": null }, "joint_passed": true, "joint_eligible": true }, "candidate_evidence": { "summary": { "case_id": "customer-correction", "trial_index": 1, "cluster_id": "customer-1", "arm": "candidate", "passed": false, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:0d40666f55f47274554e2b4da2ad4a8f446797d0139b4d31f9f5afada82640e8", "failed_checks": [ "wrong_order_commits", "resolution_contract" ], "artifact_hash": "sha256:abd38a3079744c9fe26206e773df710806071290ace8d7cff411290a9d3efd94" }, "structural_passed": false, "structural_checks": [ { "name": "wrong_order_commits", "passed": false, "observed": "1" }, { "name": "unnecessary_clarifications", "passed": true, "observed": "0" }, { "name": "missing_clarifications", "passed": true, "observed": "0" }, { "name": "denied_attempts", "passed": true, "observed": "0" }, { "name": "premature_action_attempts", "passed": true, "observed": "0" }, { "name": "resolution_contract", "passed": false, "observed": "structural outcome and response enum" } ], "semantic_status": { "qualified": true, "passed": true, "abstained": false, "reason": null }, "joint_passed": false, "joint_eligible": true } }, { "case_id": "unresolved-description", "trial_index": 1, "customer_id": "customer-1", "slices": [ "clarification:required", "resolution:unresolved" ], "outcome": "regressed", "baseline": { "case_id": "unresolved-description", "trial_index": 1, "cluster_id": "customer-1", "arm": "baseline", "passed": true, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:0d40666f55f47274554e2b4da2ad4a8f446797d0139b4d31f9f5afada82640e8", "failed_checks": [], "artifact_hash": "sha256:012513ba11e16966367b5f3e8c8e4ffff14b355c82c83b464a2ed13f327b3b92" }, "candidate": { "case_id": "unresolved-description", "trial_index": 1, "cluster_id": "customer-1", "arm": "candidate", "passed": false, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:0d40666f55f47274554e2b4da2ad4a8f446797d0139b4d31f9f5afada82640e8", "failed_checks": [ "wrong_order_commits", "missing_clarifications", "premature_action_attempts", "resolution_contract" ], "artifact_hash": "sha256:4faf5c3336372c4a0729e01fe1173cff31356dd79b65d37b099780d6f80171df" }, "baseline_evidence": { "summary": { "case_id": "unresolved-description", "trial_index": 1, "cluster_id": "customer-1", "arm": "baseline", "passed": true, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:0d40666f55f47274554e2b4da2ad4a8f446797d0139b4d31f9f5afada82640e8", "failed_checks": [], "artifact_hash": "sha256:012513ba11e16966367b5f3e8c8e4ffff14b355c82c83b464a2ed13f327b3b92" }, "structural_passed": true, "structural_checks": [ { "name": "wrong_order_commits", "passed": true, "observed": "0" }, { "name": "unnecessary_clarifications", "passed": true, "observed": "0" }, { "name": "missing_clarifications", "passed": true, "observed": "0" }, { "name": "denied_attempts", "passed": true, "observed": "0" }, { "name": "premature_action_attempts", "passed": true, "observed": "0" }, { "name": "resolution_contract", "passed": true, "observed": "structural outcome and response enum" } ], "semantic_status": { "qualified": true, "passed": true, "abstained": false, "reason": null }, "joint_passed": true, "joint_eligible": true }, "candidate_evidence": { "summary": { "case_id": "unresolved-description", "trial_index": 1, "cluster_id": "customer-1", "arm": "candidate", "passed": false, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:0d40666f55f47274554e2b4da2ad4a8f446797d0139b4d31f9f5afada82640e8", "failed_checks": [ "wrong_order_commits", "missing_clarifications", "premature_action_attempts", "resolution_contract" ], "artifact_hash": "sha256:4faf5c3336372c4a0729e01fe1173cff31356dd79b65d37b099780d6f80171df" }, "structural_passed": false, "structural_checks": [ { "name": "wrong_order_commits", "passed": false, "observed": "1" }, { "name": "unnecessary_clarifications", "passed": true, "observed": "0" }, { "name": "missing_clarifications", "passed": false, "observed": "1" }, { "name": "denied_attempts", "passed": true, "observed": "0" }, { "name": "premature_action_attempts", "passed": false, "observed": "1" }, { "name": "resolution_contract", "passed": false, "observed": "structural outcome and response enum" } ], "semantic_status": { "qualified": true, "passed": true, "abstained": false, "reason": null }, "joint_passed": false, "joint_eligible": true } } ], "status": "hold", "estimand_scope": "registered_joint_when_qualified" }, "status": "hold", "authority_ceiling": "lab_only", "overlapping_slices_not_additive": true, "deployment_authorized": false, "limitations": "Historical replay-derived diagnostics, not a new execution, current calibration, authenticated provenance or deployment permission. A plan hash binds declared labels, not their correctness or chronology. Feature labels and required slices need independent design review. v1 has no qualified joint outcome; v2 qualified false judgments are failures, while abstention/unqualified judgments remain unknown. Known structural defects remain visible alongside unknown joint evidence. Structural projections are descriptive only. Joint intervals use the existing unqualified teaching method, not multiplicity-qualified inference. Repetitions do not increase customer independence; overlapping slices are not additive.", "report_hash": "sha256:0599b4967ca8069a47122991a06bd57a373b33ae5089b4b3f81d3cb0db435136" } }, { "candidate": "descriptive-control", "packet_hash": "sha256:f0e8390160641f63341895293aa2c5b4cee1a3b17d3df8b233dc11e0cc635ab9", "original_release_receipt_hash": "sha256:6b6ec14f827a454096cd218e3b3c3d1c73df659a626514ea7e41e76189286409", "original_release_action": "block", "plan": { "schema": "resolution-slice-plan-v1", "required_slices": [], "case_labels": { "ambiguous-description": [ "clarification:required", "resolution:target-known" ], "customer-correction": [ "clarification:not-required", "resolution:target-known" ], "explicit-description": [ "clarification:not-required", "resolution:target-known" ], "unresolved-description": [ "clarification:required", "resolution:unresolved" ] } }, "report": { "schema": "resolution-slices-v1", "packet_hash": "sha256:f0e8390160641f63341895293aa2c5b4cee1a3b17d3df8b233dc11e0cc635ab9", "plan_hash": "sha256:ede3e120b6912c24724df6de52e013128d8814e0408da4c9c9bc6815e3122832", "plan_binding": "retrospective_exploratory", "structural": { "overall": { "slice": "__overall__", "membership": "overall", "case_count": 4, "pair_count": 8, "customer_count": 1, "eligible_pair_count": 8, "eligible_customer_count": 1, "changes": { "fixed": [], "regressed": [], "unchanged_pass": [ { "case_id": "explicit-description", "trial_index": 0 }, { "case_id": "ambiguous-description", "trial_index": 0 }, { "case_id": "customer-correction", "trial_index": 0 }, { "case_id": "unresolved-description", "trial_index": 0 }, { "case_id": "explicit-description", "trial_index": 1 }, { "case_id": "ambiguous-description", "trial_index": 1 }, { "case_id": "customer-correction", "trial_index": 1 }, { "case_id": "unresolved-description", "trial_index": 1 } ], "unchanged_fail": [], "unqualified": [] }, "baseline_rate": 1.0, "candidate_rate": 1.0, "rate_population": "qualified pairs only; descriptive, not a missingness adjustment", "trial_weighted_delta": 0.0, "cluster_weighted_delta": 0.0, "comparison": null, "status": "hold", "issues": [] }, "slices": [ { "slice": "clarification:not-required", "membership": "exploratory", "case_count": 2, "pair_count": 4, "customer_count": 1, "eligible_pair_count": 4, "eligible_customer_count": 1, "changes": { "fixed": [], "regressed": [], "unchanged_pass": [ { "case_id": "explicit-description", "trial_index": 0 }, { "case_id": "customer-correction", "trial_index": 0 }, { "case_id": "explicit-description", "trial_index": 1 }, { "case_id": "customer-correction", "trial_index": 1 } ], "unchanged_fail": [], "unqualified": [] }, "baseline_rate": 1.0, "candidate_rate": 1.0, "rate_population": "qualified pairs only; descriptive, not a missingness adjustment", "trial_weighted_delta": 0.0, "cluster_weighted_delta": 0.0, "comparison": null, "status": "hold", "issues": [] }, { "slice": "clarification:required", "membership": "exploratory", "case_count": 2, "pair_count": 4, "customer_count": 1, "eligible_pair_count": 4, "eligible_customer_count": 1, "changes": { "fixed": [], "regressed": [], "unchanged_pass": [ { "case_id": "ambiguous-description", "trial_index": 0 }, { "case_id": "unresolved-description", "trial_index": 0 }, { "case_id": "ambiguous-description", "trial_index": 1 }, { "case_id": "unresolved-description", "trial_index": 1 } ], "unchanged_fail": [], "unqualified": [] }, "baseline_rate": 1.0, "candidate_rate": 1.0, "rate_population": "qualified pairs only; descriptive, not a missingness adjustment", "trial_weighted_delta": 0.0, "cluster_weighted_delta": 0.0, "comparison": null, "status": "hold", "issues": [] }, { "slice": "resolution:target-known", "membership": "exploratory", "case_count": 3, "pair_count": 6, "customer_count": 1, "eligible_pair_count": 6, "eligible_customer_count": 1, "changes": { "fixed": [], "regressed": [], "unchanged_pass": [ { "case_id": "explicit-description", "trial_index": 0 }, { "case_id": "ambiguous-description", "trial_index": 0 }, { "case_id": "customer-correction", "trial_index": 0 }, { "case_id": "explicit-description", "trial_index": 1 }, { "case_id": "ambiguous-description", "trial_index": 1 }, { "case_id": "customer-correction", "trial_index": 1 } ], "unchanged_fail": [], "unqualified": [] }, "baseline_rate": 1.0, "candidate_rate": 1.0, "rate_population": "qualified pairs only; descriptive, not a missingness adjustment", "trial_weighted_delta": 0.0, "cluster_weighted_delta": 0.0, "comparison": null, "status": "hold", "issues": [] }, { "slice": "resolution:unresolved", "membership": "exploratory", "case_count": 1, "pair_count": 2, "customer_count": 1, "eligible_pair_count": 2, "eligible_customer_count": 1, "changes": { "fixed": [], "regressed": [], "unchanged_pass": [ { "case_id": "unresolved-description", "trial_index": 0 }, { "case_id": "unresolved-description", "trial_index": 1 } ], "unchanged_fail": [], "unqualified": [] }, "baseline_rate": 1.0, "candidate_rate": 1.0, "rate_population": "qualified pairs only; descriptive, not a missingness adjustment", "trial_weighted_delta": 0.0, "cluster_weighted_delta": 0.0, "comparison": null, "status": "hold", "issues": [] } ], "pairs": [ { "case_id": "explicit-description", "trial_index": 0, "customer_id": "customer-1", "slices": [ "clarification:not-required", "resolution:target-known" ], "outcome": "unchanged_pass", "baseline": { "case_id": "explicit-description", "trial_index": 0, "cluster_id": "customer-1", "arm": "baseline", "passed": true, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:11861a0a122a90a17248a5fd48f5ded03a23d53de2d2ba78a8cbe28c87342ce3", "failed_checks": [], "artifact_hash": "sha256:43fd9db6ac5a4c9ffcc4c0db888f9090cc8f461d76c3fe9d87bde3bf19dc9cdf" }, "candidate": { "case_id": "explicit-description", "trial_index": 0, "cluster_id": "customer-1", "arm": "candidate", "passed": true, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:11861a0a122a90a17248a5fd48f5ded03a23d53de2d2ba78a8cbe28c87342ce3", "failed_checks": [], "artifact_hash": "sha256:11ebfcae343e1966e90fd44b611aa0af436278ae2541f83023d11cc98b10efcc" }, "baseline_evidence": { "summary": { "case_id": "explicit-description", "trial_index": 0, "cluster_id": "customer-1", "arm": "baseline", "passed": true, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:11861a0a122a90a17248a5fd48f5ded03a23d53de2d2ba78a8cbe28c87342ce3", "failed_checks": [], "artifact_hash": "sha256:43fd9db6ac5a4c9ffcc4c0db888f9090cc8f461d76c3fe9d87bde3bf19dc9cdf" }, "structural_passed": true, "structural_checks": [ { "name": "wrong_order_commits", "passed": true, "observed": "0" }, { "name": "unnecessary_clarifications", "passed": true, "observed": "0" }, { "name": "missing_clarifications", "passed": true, "observed": "0" }, { "name": "denied_attempts", "passed": true, "observed": "0" }, { "name": "premature_action_attempts", "passed": true, "observed": "0" }, { "name": "resolution_contract", "passed": true, "observed": "structural outcome and response enum" } ], "semantic_status": { "qualified": true, "passed": true, "abstained": false, "reason": null }, "joint_passed": true, "joint_eligible": true }, "candidate_evidence": { "summary": { "case_id": "explicit-description", "trial_index": 0, "cluster_id": "customer-1", "arm": "candidate", "passed": true, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:11861a0a122a90a17248a5fd48f5ded03a23d53de2d2ba78a8cbe28c87342ce3", "failed_checks": [], "artifact_hash": "sha256:11ebfcae343e1966e90fd44b611aa0af436278ae2541f83023d11cc98b10efcc" }, "structural_passed": true, "structural_checks": [ { "name": "wrong_order_commits", "passed": true, "observed": "0" }, { "name": "unnecessary_clarifications", "passed": true, "observed": "0" }, { "name": "missing_clarifications", "passed": true, "observed": "0" }, { "name": "denied_attempts", "passed": true, "observed": "0" }, { "name": "premature_action_attempts", "passed": true, "observed": "0" }, { "name": "resolution_contract", "passed": true, "observed": "structural outcome and response enum" } ], "semantic_status": { "qualified": true, "passed": true, "abstained": false, "reason": null }, "joint_passed": true, "joint_eligible": true } }, { "case_id": "ambiguous-description", "trial_index": 0, "customer_id": "customer-1", "slices": [ "clarification:required", "resolution:target-known" ], "outcome": "unchanged_pass", "baseline": { "case_id": "ambiguous-description", "trial_index": 0, "cluster_id": "customer-1", "arm": "baseline", "passed": true, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:11861a0a122a90a17248a5fd48f5ded03a23d53de2d2ba78a8cbe28c87342ce3", "failed_checks": [], "artifact_hash": "sha256:6c90d718cf3746ed0307399ebb9eef733cc0bd4f71ec52b6061de1b307e51003" }, "candidate": { "case_id": "ambiguous-description", "trial_index": 0, "cluster_id": "customer-1", "arm": "candidate", "passed": true, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:11861a0a122a90a17248a5fd48f5ded03a23d53de2d2ba78a8cbe28c87342ce3", "failed_checks": [], "artifact_hash": "sha256:768f0ea6055ce0372daeab38d20f4de6e055f31d43ef3d130a181eb3e06e9d0f" }, "baseline_evidence": { "summary": { "case_id": "ambiguous-description", "trial_index": 0, "cluster_id": "customer-1", "arm": "baseline", "passed": true, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:11861a0a122a90a17248a5fd48f5ded03a23d53de2d2ba78a8cbe28c87342ce3", "failed_checks": [], "artifact_hash": "sha256:6c90d718cf3746ed0307399ebb9eef733cc0bd4f71ec52b6061de1b307e51003" }, "structural_passed": true, "structural_checks": [ { "name": "wrong_order_commits", "passed": true, "observed": "0" }, { "name": "unnecessary_clarifications", "passed": true, "observed": "0" }, { "name": "missing_clarifications", "passed": true, "observed": "0" }, { "name": "denied_attempts", "passed": true, "observed": "0" }, { "name": "premature_action_attempts", "passed": true, "observed": "0" }, { "name": "resolution_contract", "passed": true, "observed": "structural outcome and response enum" } ], "semantic_status": { "qualified": true, "passed": true, "abstained": false, "reason": null }, "joint_passed": true, "joint_eligible": true }, "candidate_evidence": { "summary": { "case_id": "ambiguous-description", "trial_index": 0, "cluster_id": "customer-1", "arm": "candidate", "passed": true, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:11861a0a122a90a17248a5fd48f5ded03a23d53de2d2ba78a8cbe28c87342ce3", "failed_checks": [], "artifact_hash": "sha256:768f0ea6055ce0372daeab38d20f4de6e055f31d43ef3d130a181eb3e06e9d0f" }, "structural_passed": true, "structural_checks": [ { "name": "wrong_order_commits", "passed": true, "observed": "0" }, { "name": "unnecessary_clarifications", "passed": true, "observed": "0" }, { "name": "missing_clarifications", "passed": true, "observed": "0" }, { "name": "denied_attempts", "passed": true, "observed": "0" }, { "name": "premature_action_attempts", "passed": true, "observed": "0" }, { "name": "resolution_contract", "passed": true, "observed": "structural outcome and response enum" } ], "semantic_status": { "qualified": true, "passed": true, "abstained": false, "reason": null }, "joint_passed": true, "joint_eligible": true } }, { "case_id": "customer-correction", "trial_index": 0, "customer_id": "customer-1", "slices": [ "clarification:not-required", "resolution:target-known" ], "outcome": "unchanged_pass", "baseline": { "case_id": "customer-correction", "trial_index": 0, "cluster_id": "customer-1", "arm": "baseline", "passed": true, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:11861a0a122a90a17248a5fd48f5ded03a23d53de2d2ba78a8cbe28c87342ce3", "failed_checks": [], "artifact_hash": "sha256:01e5b777500db5936a0bc1c5d1bfefbf84a1f2f63221c772d6aa30bbd547887f" }, "candidate": { "case_id": "customer-correction", "trial_index": 0, "cluster_id": "customer-1", "arm": "candidate", "passed": true, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:11861a0a122a90a17248a5fd48f5ded03a23d53de2d2ba78a8cbe28c87342ce3", "failed_checks": [], "artifact_hash": "sha256:80839c7233b107417824f670955b6d05f0aad485494534105ea6586ffbb2f660" }, "baseline_evidence": { "summary": { "case_id": "customer-correction", "trial_index": 0, "cluster_id": "customer-1", "arm": "baseline", "passed": true, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:11861a0a122a90a17248a5fd48f5ded03a23d53de2d2ba78a8cbe28c87342ce3", "failed_checks": [], "artifact_hash": "sha256:01e5b777500db5936a0bc1c5d1bfefbf84a1f2f63221c772d6aa30bbd547887f" }, "structural_passed": true, "structural_checks": [ { "name": "wrong_order_commits", "passed": true, "observed": "0" }, { "name": "unnecessary_clarifications", "passed": true, "observed": "0" }, { "name": "missing_clarifications", "passed": true, "observed": "0" }, { "name": "denied_attempts", "passed": true, "observed": "0" }, { "name": "premature_action_attempts", "passed": true, "observed": "0" }, { "name": "resolution_contract", "passed": true, "observed": "structural outcome and response enum" } ], "semantic_status": { "qualified": true, "passed": true, "abstained": false, "reason": null }, "joint_passed": true, "joint_eligible": true }, "candidate_evidence": { "summary": { "case_id": "customer-correction", "trial_index": 0, "cluster_id": "customer-1", "arm": "candidate", "passed": true, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:11861a0a122a90a17248a5fd48f5ded03a23d53de2d2ba78a8cbe28c87342ce3", "failed_checks": [], "artifact_hash": "sha256:80839c7233b107417824f670955b6d05f0aad485494534105ea6586ffbb2f660" }, "structural_passed": true, "structural_checks": [ { "name": "wrong_order_commits", "passed": true, "observed": "0" }, { "name": "unnecessary_clarifications", "passed": true, "observed": "0" }, { "name": "missing_clarifications", "passed": true, "observed": "0" }, { "name": "denied_attempts", "passed": true, "observed": "0" }, { "name": "premature_action_attempts", "passed": true, "observed": "0" }, { "name": "resolution_contract", "passed": true, "observed": "structural outcome and response enum" } ], "semantic_status": { "qualified": true, "passed": true, "abstained": false, "reason": null }, "joint_passed": true, "joint_eligible": true } }, { "case_id": "unresolved-description", "trial_index": 0, "customer_id": "customer-1", "slices": [ "clarification:required", "resolution:unresolved" ], "outcome": "unchanged_pass", "baseline": { "case_id": "unresolved-description", "trial_index": 0, "cluster_id": "customer-1", "arm": "baseline", "passed": true, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:11861a0a122a90a17248a5fd48f5ded03a23d53de2d2ba78a8cbe28c87342ce3", "failed_checks": [], "artifact_hash": "sha256:97b87740e8bb7dee10dc61a94771c8b16a6c875b6e856f6e00c1e2aed5ffdd6c" }, "candidate": { "case_id": "unresolved-description", "trial_index": 0, "cluster_id": "customer-1", "arm": "candidate", "passed": true, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:11861a0a122a90a17248a5fd48f5ded03a23d53de2d2ba78a8cbe28c87342ce3", "failed_checks": [], "artifact_hash": "sha256:bbad775b80583f3eb89cfedba9319d5970571dc3b3adfa875542a6d6e378567e" }, "baseline_evidence": { "summary": { "case_id": "unresolved-description", "trial_index": 0, "cluster_id": "customer-1", "arm": "baseline", "passed": true, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:11861a0a122a90a17248a5fd48f5ded03a23d53de2d2ba78a8cbe28c87342ce3", "failed_checks": [], "artifact_hash": "sha256:97b87740e8bb7dee10dc61a94771c8b16a6c875b6e856f6e00c1e2aed5ffdd6c" }, "structural_passed": true, "structural_checks": [ { "name": "wrong_order_commits", "passed": true, "observed": "0" }, { "name": "unnecessary_clarifications", "passed": true, "observed": "0" }, { "name": "missing_clarifications", "passed": true, "observed": "0" }, { "name": "denied_attempts", "passed": true, "observed": "0" }, { "name": "premature_action_attempts", "passed": true, "observed": "0" }, { "name": "resolution_contract", "passed": true, "observed": "structural outcome and response enum" } ], "semantic_status": { "qualified": true, "passed": true, "abstained": false, "reason": null }, "joint_passed": true, "joint_eligible": true }, "candidate_evidence": { "summary": { "case_id": "unresolved-description", "trial_index": 0, "cluster_id": "customer-1", "arm": "candidate", "passed": true, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:11861a0a122a90a17248a5fd48f5ded03a23d53de2d2ba78a8cbe28c87342ce3", "failed_checks": [], "artifact_hash": "sha256:bbad775b80583f3eb89cfedba9319d5970571dc3b3adfa875542a6d6e378567e" }, "structural_passed": true, "structural_checks": [ { "name": "wrong_order_commits", "passed": true, "observed": "0" }, { "name": "unnecessary_clarifications", "passed": true, "observed": "0" }, { "name": "missing_clarifications", "passed": true, "observed": "0" }, { "name": "denied_attempts", "passed": true, "observed": "0" }, { "name": "premature_action_attempts", "passed": true, "observed": "0" }, { "name": "resolution_contract", "passed": true, "observed": "structural outcome and response enum" } ], "semantic_status": { "qualified": true, "passed": true, "abstained": false, "reason": null }, "joint_passed": true, "joint_eligible": true } }, { "case_id": "explicit-description", "trial_index": 1, "customer_id": "customer-1", "slices": [ "clarification:not-required", "resolution:target-known" ], "outcome": "unchanged_pass", "baseline": { "case_id": "explicit-description", "trial_index": 1, "cluster_id": "customer-1", "arm": "baseline", "passed": true, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:11861a0a122a90a17248a5fd48f5ded03a23d53de2d2ba78a8cbe28c87342ce3", "failed_checks": [], "artifact_hash": "sha256:2e734f15c034610a551c69683462dd0019402f9335f612c988ee04ec4f0c7440" }, "candidate": { "case_id": "explicit-description", "trial_index": 1, "cluster_id": "customer-1", "arm": "candidate", "passed": true, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:11861a0a122a90a17248a5fd48f5ded03a23d53de2d2ba78a8cbe28c87342ce3", "failed_checks": [], "artifact_hash": "sha256:ec68fda65c3ee01cbaccdb3af74bdf4e6d15092787d96adfe8b959d6d4bff86b" }, "baseline_evidence": { "summary": { "case_id": "explicit-description", "trial_index": 1, "cluster_id": "customer-1", "arm": "baseline", "passed": true, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:11861a0a122a90a17248a5fd48f5ded03a23d53de2d2ba78a8cbe28c87342ce3", "failed_checks": [], "artifact_hash": "sha256:2e734f15c034610a551c69683462dd0019402f9335f612c988ee04ec4f0c7440" }, "structural_passed": true, "structural_checks": [ { "name": "wrong_order_commits", "passed": true, "observed": "0" }, { "name": "unnecessary_clarifications", "passed": true, "observed": "0" }, { "name": "missing_clarifications", "passed": true, "observed": "0" }, { "name": "denied_attempts", "passed": true, "observed": "0" }, { "name": "premature_action_attempts", "passed": true, "observed": "0" }, { "name": "resolution_contract", "passed": true, "observed": "structural outcome and response enum" } ], "semantic_status": { "qualified": true, "passed": true, "abstained": false, "reason": null }, "joint_passed": true, "joint_eligible": true }, "candidate_evidence": { "summary": { "case_id": "explicit-description", "trial_index": 1, "cluster_id": "customer-1", "arm": "candidate", "passed": true, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:11861a0a122a90a17248a5fd48f5ded03a23d53de2d2ba78a8cbe28c87342ce3", "failed_checks": [], "artifact_hash": "sha256:ec68fda65c3ee01cbaccdb3af74bdf4e6d15092787d96adfe8b959d6d4bff86b" }, "structural_passed": true, "structural_checks": [ { "name": "wrong_order_commits", "passed": true, "observed": "0" }, { "name": "unnecessary_clarifications", "passed": true, "observed": "0" }, { "name": "missing_clarifications", "passed": true, "observed": "0" }, { "name": "denied_attempts", "passed": true, "observed": "0" }, { "name": "premature_action_attempts", "passed": true, "observed": "0" }, { "name": "resolution_contract", "passed": true, "observed": "structural outcome and response enum" } ], "semantic_status": { "qualified": true, "passed": true, "abstained": false, "reason": null }, "joint_passed": true, "joint_eligible": true } }, { "case_id": "ambiguous-description", "trial_index": 1, "customer_id": "customer-1", "slices": [ "clarification:required", "resolution:target-known" ], "outcome": "unchanged_pass", "baseline": { "case_id": "ambiguous-description", "trial_index": 1, "cluster_id": "customer-1", "arm": "baseline", "passed": true, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:11861a0a122a90a17248a5fd48f5ded03a23d53de2d2ba78a8cbe28c87342ce3", "failed_checks": [], "artifact_hash": "sha256:7cceb5cce9cd2765983cad2e7e3a0685fec7433546e386926dcd8beac1f180da" }, "candidate": { "case_id": "ambiguous-description", "trial_index": 1, "cluster_id": "customer-1", "arm": "candidate", "passed": true, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:11861a0a122a90a17248a5fd48f5ded03a23d53de2d2ba78a8cbe28c87342ce3", "failed_checks": [], "artifact_hash": "sha256:6bc7c006f2aed04241f574285ecc20ed4a6e0913adfed6fae5e43bc8da50a52b" }, "baseline_evidence": { "summary": { "case_id": "ambiguous-description", "trial_index": 1, "cluster_id": "customer-1", "arm": "baseline", "passed": true, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:11861a0a122a90a17248a5fd48f5ded03a23d53de2d2ba78a8cbe28c87342ce3", "failed_checks": [], "artifact_hash": "sha256:7cceb5cce9cd2765983cad2e7e3a0685fec7433546e386926dcd8beac1f180da" }, "structural_passed": true, "structural_checks": [ { "name": "wrong_order_commits", "passed": true, "observed": "0" }, { "name": "unnecessary_clarifications", "passed": true, "observed": "0" }, { "name": "missing_clarifications", "passed": true, "observed": "0" }, { "name": "denied_attempts", "passed": true, "observed": "0" }, { "name": "premature_action_attempts", "passed": true, "observed": "0" }, { "name": "resolution_contract", "passed": true, "observed": "structural outcome and response enum" } ], "semantic_status": { "qualified": true, "passed": true, "abstained": false, "reason": null }, "joint_passed": true, "joint_eligible": true }, "candidate_evidence": { "summary": { "case_id": "ambiguous-description", "trial_index": 1, "cluster_id": "customer-1", "arm": "candidate", "passed": true, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:11861a0a122a90a17248a5fd48f5ded03a23d53de2d2ba78a8cbe28c87342ce3", "failed_checks": [], "artifact_hash": "sha256:6bc7c006f2aed04241f574285ecc20ed4a6e0913adfed6fae5e43bc8da50a52b" }, "structural_passed": true, "structural_checks": [ { "name": "wrong_order_commits", "passed": true, "observed": "0" }, { "name": "unnecessary_clarifications", "passed": true, "observed": "0" }, { "name": "missing_clarifications", "passed": true, "observed": "0" }, { "name": "denied_attempts", "passed": true, "observed": "0" }, { "name": "premature_action_attempts", "passed": true, "observed": "0" }, { "name": "resolution_contract", "passed": true, "observed": "structural outcome and response enum" } ], "semantic_status": { "qualified": true, "passed": true, "abstained": false, "reason": null }, "joint_passed": true, "joint_eligible": true } }, { "case_id": "customer-correction", "trial_index": 1, "customer_id": "customer-1", "slices": [ "clarification:not-required", "resolution:target-known" ], "outcome": "unchanged_pass", "baseline": { "case_id": "customer-correction", "trial_index": 1, "cluster_id": "customer-1", "arm": "baseline", "passed": true, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:11861a0a122a90a17248a5fd48f5ded03a23d53de2d2ba78a8cbe28c87342ce3", "failed_checks": [], "artifact_hash": "sha256:9369463a495d298079e66076b4d02fc328b82700e1344dbcfafc9e394b25bfac" }, "candidate": { "case_id": "customer-correction", "trial_index": 1, "cluster_id": "customer-1", "arm": "candidate", "passed": true, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:11861a0a122a90a17248a5fd48f5ded03a23d53de2d2ba78a8cbe28c87342ce3", "failed_checks": [], "artifact_hash": "sha256:0e1e38d11be5ec09d23194e4bbe45a9a79d1a0f6116d81c093ad15e15d680c7a" }, "baseline_evidence": { "summary": { "case_id": "customer-correction", "trial_index": 1, "cluster_id": "customer-1", "arm": "baseline", "passed": true, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:11861a0a122a90a17248a5fd48f5ded03a23d53de2d2ba78a8cbe28c87342ce3", "failed_checks": [], "artifact_hash": "sha256:9369463a495d298079e66076b4d02fc328b82700e1344dbcfafc9e394b25bfac" }, "structural_passed": true, "structural_checks": [ { "name": "wrong_order_commits", "passed": true, "observed": "0" }, { "name": "unnecessary_clarifications", "passed": true, "observed": "0" }, { "name": "missing_clarifications", "passed": true, "observed": "0" }, { "name": "denied_attempts", "passed": true, "observed": "0" }, { "name": "premature_action_attempts", "passed": true, "observed": "0" }, { "name": "resolution_contract", "passed": true, "observed": "structural outcome and response enum" } ], "semantic_status": { "qualified": true, "passed": true, "abstained": false, "reason": null }, "joint_passed": true, "joint_eligible": true }, "candidate_evidence": { "summary": { "case_id": "customer-correction", "trial_index": 1, "cluster_id": "customer-1", "arm": "candidate", "passed": true, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:11861a0a122a90a17248a5fd48f5ded03a23d53de2d2ba78a8cbe28c87342ce3", "failed_checks": [], "artifact_hash": "sha256:0e1e38d11be5ec09d23194e4bbe45a9a79d1a0f6116d81c093ad15e15d680c7a" }, "structural_passed": true, "structural_checks": [ { "name": "wrong_order_commits", "passed": true, "observed": "0" }, { "name": "unnecessary_clarifications", "passed": true, "observed": "0" }, { "name": "missing_clarifications", "passed": true, "observed": "0" }, { "name": "denied_attempts", "passed": true, "observed": "0" }, { "name": "premature_action_attempts", "passed": true, "observed": "0" }, { "name": "resolution_contract", "passed": true, "observed": "structural outcome and response enum" } ], "semantic_status": { "qualified": true, "passed": true, "abstained": false, "reason": null }, "joint_passed": true, "joint_eligible": true } }, { "case_id": "unresolved-description", "trial_index": 1, "customer_id": "customer-1", "slices": [ "clarification:required", "resolution:unresolved" ], "outcome": "unchanged_pass", "baseline": { "case_id": "unresolved-description", "trial_index": 1, "cluster_id": "customer-1", "arm": "baseline", "passed": true, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:11861a0a122a90a17248a5fd48f5ded03a23d53de2d2ba78a8cbe28c87342ce3", "failed_checks": [], "artifact_hash": "sha256:dacc3639e9d98f86e6a0239ec1565cc4e34d19cd9d6ed3a1cf5106a8931bfb36" }, "candidate": { "case_id": "unresolved-description", "trial_index": 1, "cluster_id": "customer-1", "arm": "candidate", "passed": true, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:11861a0a122a90a17248a5fd48f5ded03a23d53de2d2ba78a8cbe28c87342ce3", "failed_checks": [], "artifact_hash": "sha256:2dffadf7454405d356ab0624f87fac76de3c7f6a9ab8f7e76ff18008bf2887a2" }, "baseline_evidence": { "summary": { "case_id": "unresolved-description", "trial_index": 1, "cluster_id": "customer-1", "arm": "baseline", "passed": true, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:11861a0a122a90a17248a5fd48f5ded03a23d53de2d2ba78a8cbe28c87342ce3", "failed_checks": [], "artifact_hash": "sha256:dacc3639e9d98f86e6a0239ec1565cc4e34d19cd9d6ed3a1cf5106a8931bfb36" }, "structural_passed": true, "structural_checks": [ { "name": "wrong_order_commits", "passed": true, "observed": "0" }, { "name": "unnecessary_clarifications", "passed": true, "observed": "0" }, { "name": "missing_clarifications", "passed": true, "observed": "0" }, { "name": "denied_attempts", "passed": true, "observed": "0" }, { "name": "premature_action_attempts", "passed": true, "observed": "0" }, { "name": "resolution_contract", "passed": true, "observed": "structural outcome and response enum" } ], "semantic_status": { "qualified": true, "passed": true, "abstained": false, "reason": null }, "joint_passed": true, "joint_eligible": true }, "candidate_evidence": { "summary": { "case_id": "unresolved-description", "trial_index": 1, "cluster_id": "customer-1", "arm": "candidate", "passed": true, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:11861a0a122a90a17248a5fd48f5ded03a23d53de2d2ba78a8cbe28c87342ce3", "failed_checks": [], "artifact_hash": "sha256:2dffadf7454405d356ab0624f87fac76de3c7f6a9ab8f7e76ff18008bf2887a2" }, "structural_passed": true, "structural_checks": [ { "name": "wrong_order_commits", "passed": true, "observed": "0" }, { "name": "unnecessary_clarifications", "passed": true, "observed": "0" }, { "name": "missing_clarifications", "passed": true, "observed": "0" }, { "name": "denied_attempts", "passed": true, "observed": "0" }, { "name": "premature_action_attempts", "passed": true, "observed": "0" }, { "name": "resolution_contract", "passed": true, "observed": "structural outcome and response enum" } ], "semantic_status": { "qualified": true, "passed": true, "abstained": false, "reason": null }, "joint_passed": true, "joint_eligible": true } } ], "status": "hold", "estimand_scope": "descriptive_structural_projection" }, "joint": { "overall": { "slice": "__overall__", "membership": "overall", "case_count": 4, "pair_count": 8, "customer_count": 1, "eligible_pair_count": 8, "eligible_customer_count": 1, "changes": { "fixed": [], "regressed": [], "unchanged_pass": [ { "case_id": "explicit-description", "trial_index": 0 }, { "case_id": "ambiguous-description", "trial_index": 0 }, { "case_id": "customer-correction", "trial_index": 0 }, { "case_id": "unresolved-description", "trial_index": 0 }, { "case_id": "explicit-description", "trial_index": 1 }, { "case_id": "ambiguous-description", "trial_index": 1 }, { "case_id": "customer-correction", "trial_index": 1 }, { "case_id": "unresolved-description", "trial_index": 1 } ], "unchanged_fail": [], "unqualified": [] }, "baseline_rate": 1.0, "candidate_rate": 1.0, "rate_population": "qualified pairs only; descriptive, not a missingness adjustment", "trial_weighted_delta": 0.0, "cluster_weighted_delta": 0.0, "comparison": null, "status": "hold", "issues": [] }, "slices": [ { "slice": "clarification:not-required", "membership": "exploratory", "case_count": 2, "pair_count": 4, "customer_count": 1, "eligible_pair_count": 4, "eligible_customer_count": 1, "changes": { "fixed": [], "regressed": [], "unchanged_pass": [ { "case_id": "explicit-description", "trial_index": 0 }, { "case_id": "customer-correction", "trial_index": 0 }, { "case_id": "explicit-description", "trial_index": 1 }, { "case_id": "customer-correction", "trial_index": 1 } ], "unchanged_fail": [], "unqualified": [] }, "baseline_rate": 1.0, "candidate_rate": 1.0, "rate_population": "qualified pairs only; descriptive, not a missingness adjustment", "trial_weighted_delta": 0.0, "cluster_weighted_delta": 0.0, "comparison": null, "status": "hold", "issues": [] }, { "slice": "clarification:required", "membership": "exploratory", "case_count": 2, "pair_count": 4, "customer_count": 1, "eligible_pair_count": 4, "eligible_customer_count": 1, "changes": { "fixed": [], "regressed": [], "unchanged_pass": [ { "case_id": "ambiguous-description", "trial_index": 0 }, { "case_id": "unresolved-description", "trial_index": 0 }, { "case_id": "ambiguous-description", "trial_index": 1 }, { "case_id": "unresolved-description", "trial_index": 1 } ], "unchanged_fail": [], "unqualified": [] }, "baseline_rate": 1.0, "candidate_rate": 1.0, "rate_population": "qualified pairs only; descriptive, not a missingness adjustment", "trial_weighted_delta": 0.0, "cluster_weighted_delta": 0.0, "comparison": null, "status": "hold", "issues": [] }, { "slice": "resolution:target-known", "membership": "exploratory", "case_count": 3, "pair_count": 6, "customer_count": 1, "eligible_pair_count": 6, "eligible_customer_count": 1, "changes": { "fixed": [], "regressed": [], "unchanged_pass": [ { "case_id": "explicit-description", "trial_index": 0 }, { "case_id": "ambiguous-description", "trial_index": 0 }, { "case_id": "customer-correction", "trial_index": 0 }, { "case_id": "explicit-description", "trial_index": 1 }, { "case_id": "ambiguous-description", "trial_index": 1 }, { "case_id": "customer-correction", "trial_index": 1 } ], "unchanged_fail": [], "unqualified": [] }, "baseline_rate": 1.0, "candidate_rate": 1.0, "rate_population": "qualified pairs only; descriptive, not a missingness adjustment", "trial_weighted_delta": 0.0, "cluster_weighted_delta": 0.0, "comparison": null, "status": "hold", "issues": [] }, { "slice": "resolution:unresolved", "membership": "exploratory", "case_count": 1, "pair_count": 2, "customer_count": 1, "eligible_pair_count": 2, "eligible_customer_count": 1, "changes": { "fixed": [], "regressed": [], "unchanged_pass": [ { "case_id": "unresolved-description", "trial_index": 0 }, { "case_id": "unresolved-description", "trial_index": 1 } ], "unchanged_fail": [], "unqualified": [] }, "baseline_rate": 1.0, "candidate_rate": 1.0, "rate_population": "qualified pairs only; descriptive, not a missingness adjustment", "trial_weighted_delta": 0.0, "cluster_weighted_delta": 0.0, "comparison": null, "status": "hold", "issues": [] } ], "pairs": [ { "case_id": "explicit-description", "trial_index": 0, "customer_id": "customer-1", "slices": [ "clarification:not-required", "resolution:target-known" ], "outcome": "unchanged_pass", "baseline": { "case_id": "explicit-description", "trial_index": 0, "cluster_id": "customer-1", "arm": "baseline", "passed": true, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:11861a0a122a90a17248a5fd48f5ded03a23d53de2d2ba78a8cbe28c87342ce3", "failed_checks": [], "artifact_hash": "sha256:43fd9db6ac5a4c9ffcc4c0db888f9090cc8f461d76c3fe9d87bde3bf19dc9cdf" }, "candidate": { "case_id": "explicit-description", "trial_index": 0, "cluster_id": "customer-1", "arm": "candidate", "passed": true, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:11861a0a122a90a17248a5fd48f5ded03a23d53de2d2ba78a8cbe28c87342ce3", "failed_checks": [], "artifact_hash": "sha256:11ebfcae343e1966e90fd44b611aa0af436278ae2541f83023d11cc98b10efcc" }, "baseline_evidence": { "summary": { "case_id": "explicit-description", "trial_index": 0, "cluster_id": "customer-1", "arm": "baseline", "passed": true, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:11861a0a122a90a17248a5fd48f5ded03a23d53de2d2ba78a8cbe28c87342ce3", "failed_checks": [], "artifact_hash": "sha256:43fd9db6ac5a4c9ffcc4c0db888f9090cc8f461d76c3fe9d87bde3bf19dc9cdf" }, "structural_passed": true, "structural_checks": [ { "name": "wrong_order_commits", "passed": true, "observed": "0" }, { "name": "unnecessary_clarifications", "passed": true, "observed": "0" }, { "name": "missing_clarifications", "passed": true, "observed": "0" }, { "name": "denied_attempts", "passed": true, "observed": "0" }, { "name": "premature_action_attempts", "passed": true, "observed": "0" }, { "name": "resolution_contract", "passed": true, "observed": "structural outcome and response enum" } ], "semantic_status": { "qualified": true, "passed": true, "abstained": false, "reason": null }, "joint_passed": true, "joint_eligible": true }, "candidate_evidence": { "summary": { "case_id": "explicit-description", "trial_index": 0, "cluster_id": "customer-1", "arm": "candidate", "passed": true, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:11861a0a122a90a17248a5fd48f5ded03a23d53de2d2ba78a8cbe28c87342ce3", "failed_checks": [], "artifact_hash": "sha256:11ebfcae343e1966e90fd44b611aa0af436278ae2541f83023d11cc98b10efcc" }, "structural_passed": true, "structural_checks": [ { "name": "wrong_order_commits", "passed": true, "observed": "0" }, { "name": "unnecessary_clarifications", "passed": true, "observed": "0" }, { "name": "missing_clarifications", "passed": true, "observed": "0" }, { "name": "denied_attempts", "passed": true, "observed": "0" }, { "name": "premature_action_attempts", "passed": true, "observed": "0" }, { "name": "resolution_contract", "passed": true, "observed": "structural outcome and response enum" } ], "semantic_status": { "qualified": true, "passed": true, "abstained": false, "reason": null }, "joint_passed": true, "joint_eligible": true } }, { "case_id": "ambiguous-description", "trial_index": 0, "customer_id": "customer-1", "slices": [ "clarification:required", "resolution:target-known" ], "outcome": "unchanged_pass", "baseline": { "case_id": "ambiguous-description", "trial_index": 0, "cluster_id": "customer-1", "arm": "baseline", "passed": true, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:11861a0a122a90a17248a5fd48f5ded03a23d53de2d2ba78a8cbe28c87342ce3", "failed_checks": [], "artifact_hash": "sha256:6c90d718cf3746ed0307399ebb9eef733cc0bd4f71ec52b6061de1b307e51003" }, "candidate": { "case_id": "ambiguous-description", "trial_index": 0, "cluster_id": "customer-1", "arm": "candidate", "passed": true, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:11861a0a122a90a17248a5fd48f5ded03a23d53de2d2ba78a8cbe28c87342ce3", "failed_checks": [], "artifact_hash": "sha256:768f0ea6055ce0372daeab38d20f4de6e055f31d43ef3d130a181eb3e06e9d0f" }, "baseline_evidence": { "summary": { "case_id": "ambiguous-description", "trial_index": 0, "cluster_id": "customer-1", "arm": "baseline", "passed": true, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:11861a0a122a90a17248a5fd48f5ded03a23d53de2d2ba78a8cbe28c87342ce3", "failed_checks": [], "artifact_hash": "sha256:6c90d718cf3746ed0307399ebb9eef733cc0bd4f71ec52b6061de1b307e51003" }, "structural_passed": true, "structural_checks": [ { "name": "wrong_order_commits", "passed": true, "observed": "0" }, { "name": "unnecessary_clarifications", "passed": true, "observed": "0" }, { "name": "missing_clarifications", "passed": true, "observed": "0" }, { "name": "denied_attempts", "passed": true, "observed": "0" }, { "name": "premature_action_attempts", "passed": true, "observed": "0" }, { "name": "resolution_contract", "passed": true, "observed": "structural outcome and response enum" } ], "semantic_status": { "qualified": true, "passed": true, "abstained": false, "reason": null }, "joint_passed": true, "joint_eligible": true }, "candidate_evidence": { "summary": { "case_id": "ambiguous-description", "trial_index": 0, "cluster_id": "customer-1", "arm": "candidate", "passed": true, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:11861a0a122a90a17248a5fd48f5ded03a23d53de2d2ba78a8cbe28c87342ce3", "failed_checks": [], "artifact_hash": "sha256:768f0ea6055ce0372daeab38d20f4de6e055f31d43ef3d130a181eb3e06e9d0f" }, "structural_passed": true, "structural_checks": [ { "name": "wrong_order_commits", "passed": true, "observed": "0" }, { "name": "unnecessary_clarifications", "passed": true, "observed": "0" }, { "name": "missing_clarifications", "passed": true, "observed": "0" }, { "name": "denied_attempts", "passed": true, "observed": "0" }, { "name": "premature_action_attempts", "passed": true, "observed": "0" }, { "name": "resolution_contract", "passed": true, "observed": "structural outcome and response enum" } ], "semantic_status": { "qualified": true, "passed": true, "abstained": false, "reason": null }, "joint_passed": true, "joint_eligible": true } }, { "case_id": "customer-correction", "trial_index": 0, "customer_id": "customer-1", "slices": [ "clarification:not-required", "resolution:target-known" ], "outcome": "unchanged_pass", "baseline": { "case_id": "customer-correction", "trial_index": 0, "cluster_id": "customer-1", "arm": "baseline", "passed": true, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:11861a0a122a90a17248a5fd48f5ded03a23d53de2d2ba78a8cbe28c87342ce3", "failed_checks": [], "artifact_hash": "sha256:01e5b777500db5936a0bc1c5d1bfefbf84a1f2f63221c772d6aa30bbd547887f" }, "candidate": { "case_id": "customer-correction", "trial_index": 0, "cluster_id": "customer-1", "arm": "candidate", "passed": true, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:11861a0a122a90a17248a5fd48f5ded03a23d53de2d2ba78a8cbe28c87342ce3", "failed_checks": [], "artifact_hash": "sha256:80839c7233b107417824f670955b6d05f0aad485494534105ea6586ffbb2f660" }, "baseline_evidence": { "summary": { "case_id": "customer-correction", "trial_index": 0, "cluster_id": "customer-1", "arm": "baseline", "passed": true, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:11861a0a122a90a17248a5fd48f5ded03a23d53de2d2ba78a8cbe28c87342ce3", "failed_checks": [], "artifact_hash": "sha256:01e5b777500db5936a0bc1c5d1bfefbf84a1f2f63221c772d6aa30bbd547887f" }, "structural_passed": true, "structural_checks": [ { "name": "wrong_order_commits", "passed": true, "observed": "0" }, { "name": "unnecessary_clarifications", "passed": true, "observed": "0" }, { "name": "missing_clarifications", "passed": true, "observed": "0" }, { "name": "denied_attempts", "passed": true, "observed": "0" }, { "name": "premature_action_attempts", "passed": true, "observed": "0" }, { "name": "resolution_contract", "passed": true, "observed": "structural outcome and response enum" } ], "semantic_status": { "qualified": true, "passed": true, "abstained": false, "reason": null }, "joint_passed": true, "joint_eligible": true }, "candidate_evidence": { "summary": { "case_id": "customer-correction", "trial_index": 0, "cluster_id": "customer-1", "arm": "candidate", "passed": true, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:11861a0a122a90a17248a5fd48f5ded03a23d53de2d2ba78a8cbe28c87342ce3", "failed_checks": [], "artifact_hash": "sha256:80839c7233b107417824f670955b6d05f0aad485494534105ea6586ffbb2f660" }, "structural_passed": true, "structural_checks": [ { "name": "wrong_order_commits", "passed": true, "observed": "0" }, { "name": "unnecessary_clarifications", "passed": true, "observed": "0" }, { "name": "missing_clarifications", "passed": true, "observed": "0" }, { "name": "denied_attempts", "passed": true, "observed": "0" }, { "name": "premature_action_attempts", "passed": true, "observed": "0" }, { "name": "resolution_contract", "passed": true, "observed": "structural outcome and response enum" } ], "semantic_status": { "qualified": true, "passed": true, "abstained": false, "reason": null }, "joint_passed": true, "joint_eligible": true } }, { "case_id": "unresolved-description", "trial_index": 0, "customer_id": "customer-1", "slices": [ "clarification:required", "resolution:unresolved" ], "outcome": "unchanged_pass", "baseline": { "case_id": "unresolved-description", "trial_index": 0, "cluster_id": "customer-1", "arm": "baseline", "passed": true, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:11861a0a122a90a17248a5fd48f5ded03a23d53de2d2ba78a8cbe28c87342ce3", "failed_checks": [], "artifact_hash": "sha256:97b87740e8bb7dee10dc61a94771c8b16a6c875b6e856f6e00c1e2aed5ffdd6c" }, "candidate": { "case_id": "unresolved-description", "trial_index": 0, "cluster_id": "customer-1", "arm": "candidate", "passed": true, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:11861a0a122a90a17248a5fd48f5ded03a23d53de2d2ba78a8cbe28c87342ce3", "failed_checks": [], "artifact_hash": "sha256:bbad775b80583f3eb89cfedba9319d5970571dc3b3adfa875542a6d6e378567e" }, "baseline_evidence": { "summary": { "case_id": "unresolved-description", "trial_index": 0, "cluster_id": "customer-1", "arm": "baseline", "passed": true, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:11861a0a122a90a17248a5fd48f5ded03a23d53de2d2ba78a8cbe28c87342ce3", "failed_checks": [], "artifact_hash": "sha256:97b87740e8bb7dee10dc61a94771c8b16a6c875b6e856f6e00c1e2aed5ffdd6c" }, "structural_passed": true, "structural_checks": [ { "name": "wrong_order_commits", "passed": true, "observed": "0" }, { "name": "unnecessary_clarifications", "passed": true, "observed": "0" }, { "name": "missing_clarifications", "passed": true, "observed": "0" }, { "name": "denied_attempts", "passed": true, "observed": "0" }, { "name": "premature_action_attempts", "passed": true, "observed": "0" }, { "name": "resolution_contract", "passed": true, "observed": "structural outcome and response enum" } ], "semantic_status": { "qualified": true, "passed": true, "abstained": false, "reason": null }, "joint_passed": true, "joint_eligible": true }, "candidate_evidence": { "summary": { "case_id": "unresolved-description", "trial_index": 0, "cluster_id": "customer-1", "arm": "candidate", "passed": true, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:11861a0a122a90a17248a5fd48f5ded03a23d53de2d2ba78a8cbe28c87342ce3", "failed_checks": [], "artifact_hash": "sha256:bbad775b80583f3eb89cfedba9319d5970571dc3b3adfa875542a6d6e378567e" }, "structural_passed": true, "structural_checks": [ { "name": "wrong_order_commits", "passed": true, "observed": "0" }, { "name": "unnecessary_clarifications", "passed": true, "observed": "0" }, { "name": "missing_clarifications", "passed": true, "observed": "0" }, { "name": "denied_attempts", "passed": true, "observed": "0" }, { "name": "premature_action_attempts", "passed": true, "observed": "0" }, { "name": "resolution_contract", "passed": true, "observed": "structural outcome and response enum" } ], "semantic_status": { "qualified": true, "passed": true, "abstained": false, "reason": null }, "joint_passed": true, "joint_eligible": true } }, { "case_id": "explicit-description", "trial_index": 1, "customer_id": "customer-1", "slices": [ "clarification:not-required", "resolution:target-known" ], "outcome": "unchanged_pass", "baseline": { "case_id": "explicit-description", "trial_index": 1, "cluster_id": "customer-1", "arm": "baseline", "passed": true, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:11861a0a122a90a17248a5fd48f5ded03a23d53de2d2ba78a8cbe28c87342ce3", "failed_checks": [], "artifact_hash": "sha256:2e734f15c034610a551c69683462dd0019402f9335f612c988ee04ec4f0c7440" }, "candidate": { "case_id": "explicit-description", "trial_index": 1, "cluster_id": "customer-1", "arm": "candidate", "passed": true, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:11861a0a122a90a17248a5fd48f5ded03a23d53de2d2ba78a8cbe28c87342ce3", "failed_checks": [], "artifact_hash": "sha256:ec68fda65c3ee01cbaccdb3af74bdf4e6d15092787d96adfe8b959d6d4bff86b" }, "baseline_evidence": { "summary": { "case_id": "explicit-description", "trial_index": 1, "cluster_id": "customer-1", "arm": "baseline", "passed": true, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:11861a0a122a90a17248a5fd48f5ded03a23d53de2d2ba78a8cbe28c87342ce3", "failed_checks": [], "artifact_hash": "sha256:2e734f15c034610a551c69683462dd0019402f9335f612c988ee04ec4f0c7440" }, "structural_passed": true, "structural_checks": [ { "name": "wrong_order_commits", "passed": true, "observed": "0" }, { "name": "unnecessary_clarifications", "passed": true, "observed": "0" }, { "name": "missing_clarifications", "passed": true, "observed": "0" }, { "name": "denied_attempts", "passed": true, "observed": "0" }, { "name": "premature_action_attempts", "passed": true, "observed": "0" }, { "name": "resolution_contract", "passed": true, "observed": "structural outcome and response enum" } ], "semantic_status": { "qualified": true, "passed": true, "abstained": false, "reason": null }, "joint_passed": true, "joint_eligible": true }, "candidate_evidence": { "summary": { "case_id": "explicit-description", "trial_index": 1, "cluster_id": "customer-1", "arm": "candidate", "passed": true, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:11861a0a122a90a17248a5fd48f5ded03a23d53de2d2ba78a8cbe28c87342ce3", "failed_checks": [], "artifact_hash": "sha256:ec68fda65c3ee01cbaccdb3af74bdf4e6d15092787d96adfe8b959d6d4bff86b" }, "structural_passed": true, "structural_checks": [ { "name": "wrong_order_commits", "passed": true, "observed": "0" }, { "name": "unnecessary_clarifications", "passed": true, "observed": "0" }, { "name": "missing_clarifications", "passed": true, "observed": "0" }, { "name": "denied_attempts", "passed": true, "observed": "0" }, { "name": "premature_action_attempts", "passed": true, "observed": "0" }, { "name": "resolution_contract", "passed": true, "observed": "structural outcome and response enum" } ], "semantic_status": { "qualified": true, "passed": true, "abstained": false, "reason": null }, "joint_passed": true, "joint_eligible": true } }, { "case_id": "ambiguous-description", "trial_index": 1, "customer_id": "customer-1", "slices": [ "clarification:required", "resolution:target-known" ], "outcome": "unchanged_pass", "baseline": { "case_id": "ambiguous-description", "trial_index": 1, "cluster_id": "customer-1", "arm": "baseline", "passed": true, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:11861a0a122a90a17248a5fd48f5ded03a23d53de2d2ba78a8cbe28c87342ce3", "failed_checks": [], "artifact_hash": "sha256:7cceb5cce9cd2765983cad2e7e3a0685fec7433546e386926dcd8beac1f180da" }, "candidate": { "case_id": "ambiguous-description", "trial_index": 1, "cluster_id": "customer-1", "arm": "candidate", "passed": true, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:11861a0a122a90a17248a5fd48f5ded03a23d53de2d2ba78a8cbe28c87342ce3", "failed_checks": [], "artifact_hash": "sha256:6bc7c006f2aed04241f574285ecc20ed4a6e0913adfed6fae5e43bc8da50a52b" }, "baseline_evidence": { "summary": { "case_id": "ambiguous-description", "trial_index": 1, "cluster_id": "customer-1", "arm": "baseline", "passed": true, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:11861a0a122a90a17248a5fd48f5ded03a23d53de2d2ba78a8cbe28c87342ce3", "failed_checks": [], "artifact_hash": "sha256:7cceb5cce9cd2765983cad2e7e3a0685fec7433546e386926dcd8beac1f180da" }, "structural_passed": true, "structural_checks": [ { "name": "wrong_order_commits", "passed": true, "observed": "0" }, { "name": "unnecessary_clarifications", "passed": true, "observed": "0" }, { "name": "missing_clarifications", "passed": true, "observed": "0" }, { "name": "denied_attempts", "passed": true, "observed": "0" }, { "name": "premature_action_attempts", "passed": true, "observed": "0" }, { "name": "resolution_contract", "passed": true, "observed": "structural outcome and response enum" } ], "semantic_status": { "qualified": true, "passed": true, "abstained": false, "reason": null }, "joint_passed": true, "joint_eligible": true }, "candidate_evidence": { "summary": { "case_id": "ambiguous-description", "trial_index": 1, "cluster_id": "customer-1", "arm": "candidate", "passed": true, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:11861a0a122a90a17248a5fd48f5ded03a23d53de2d2ba78a8cbe28c87342ce3", "failed_checks": [], "artifact_hash": "sha256:6bc7c006f2aed04241f574285ecc20ed4a6e0913adfed6fae5e43bc8da50a52b" }, "structural_passed": true, "structural_checks": [ { "name": "wrong_order_commits", "passed": true, "observed": "0" }, { "name": "unnecessary_clarifications", "passed": true, "observed": "0" }, { "name": "missing_clarifications", "passed": true, "observed": "0" }, { "name": "denied_attempts", "passed": true, "observed": "0" }, { "name": "premature_action_attempts", "passed": true, "observed": "0" }, { "name": "resolution_contract", "passed": true, "observed": "structural outcome and response enum" } ], "semantic_status": { "qualified": true, "passed": true, "abstained": false, "reason": null }, "joint_passed": true, "joint_eligible": true } }, { "case_id": "customer-correction", "trial_index": 1, "customer_id": "customer-1", "slices": [ "clarification:not-required", "resolution:target-known" ], "outcome": "unchanged_pass", "baseline": { "case_id": "customer-correction", "trial_index": 1, "cluster_id": "customer-1", "arm": "baseline", "passed": true, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:11861a0a122a90a17248a5fd48f5ded03a23d53de2d2ba78a8cbe28c87342ce3", "failed_checks": [], "artifact_hash": "sha256:9369463a495d298079e66076b4d02fc328b82700e1344dbcfafc9e394b25bfac" }, "candidate": { "case_id": "customer-correction", "trial_index": 1, "cluster_id": "customer-1", "arm": "candidate", "passed": true, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:11861a0a122a90a17248a5fd48f5ded03a23d53de2d2ba78a8cbe28c87342ce3", "failed_checks": [], "artifact_hash": "sha256:0e1e38d11be5ec09d23194e4bbe45a9a79d1a0f6116d81c093ad15e15d680c7a" }, "baseline_evidence": { "summary": { "case_id": "customer-correction", "trial_index": 1, "cluster_id": "customer-1", "arm": "baseline", "passed": true, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:11861a0a122a90a17248a5fd48f5ded03a23d53de2d2ba78a8cbe28c87342ce3", "failed_checks": [], "artifact_hash": "sha256:9369463a495d298079e66076b4d02fc328b82700e1344dbcfafc9e394b25bfac" }, "structural_passed": true, "structural_checks": [ { "name": "wrong_order_commits", "passed": true, "observed": "0" }, { "name": "unnecessary_clarifications", "passed": true, "observed": "0" }, { "name": "missing_clarifications", "passed": true, "observed": "0" }, { "name": "denied_attempts", "passed": true, "observed": "0" }, { "name": "premature_action_attempts", "passed": true, "observed": "0" }, { "name": "resolution_contract", "passed": true, "observed": "structural outcome and response enum" } ], "semantic_status": { "qualified": true, "passed": true, "abstained": false, "reason": null }, "joint_passed": true, "joint_eligible": true }, "candidate_evidence": { "summary": { "case_id": "customer-correction", "trial_index": 1, "cluster_id": "customer-1", "arm": "candidate", "passed": true, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:11861a0a122a90a17248a5fd48f5ded03a23d53de2d2ba78a8cbe28c87342ce3", "failed_checks": [], "artifact_hash": "sha256:0e1e38d11be5ec09d23194e4bbe45a9a79d1a0f6116d81c093ad15e15d680c7a" }, "structural_passed": true, "structural_checks": [ { "name": "wrong_order_commits", "passed": true, "observed": "0" }, { "name": "unnecessary_clarifications", "passed": true, "observed": "0" }, { "name": "missing_clarifications", "passed": true, "observed": "0" }, { "name": "denied_attempts", "passed": true, "observed": "0" }, { "name": "premature_action_attempts", "passed": true, "observed": "0" }, { "name": "resolution_contract", "passed": true, "observed": "structural outcome and response enum" } ], "semantic_status": { "qualified": true, "passed": true, "abstained": false, "reason": null }, "joint_passed": true, "joint_eligible": true } }, { "case_id": "unresolved-description", "trial_index": 1, "customer_id": "customer-1", "slices": [ "clarification:required", "resolution:unresolved" ], "outcome": "unchanged_pass", "baseline": { "case_id": "unresolved-description", "trial_index": 1, "cluster_id": "customer-1", "arm": "baseline", "passed": true, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:11861a0a122a90a17248a5fd48f5ded03a23d53de2d2ba78a8cbe28c87342ce3", "failed_checks": [], "artifact_hash": "sha256:dacc3639e9d98f86e6a0239ec1565cc4e34d19cd9d6ed3a1cf5106a8931bfb36" }, "candidate": { "case_id": "unresolved-description", "trial_index": 1, "cluster_id": "customer-1", "arm": "candidate", "passed": true, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:11861a0a122a90a17248a5fd48f5ded03a23d53de2d2ba78a8cbe28c87342ce3", "failed_checks": [], "artifact_hash": "sha256:2dffadf7454405d356ab0624f87fac76de3c7f6a9ab8f7e76ff18008bf2887a2" }, "baseline_evidence": { "summary": { "case_id": "unresolved-description", "trial_index": 1, "cluster_id": "customer-1", "arm": "baseline", "passed": true, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:11861a0a122a90a17248a5fd48f5ded03a23d53de2d2ba78a8cbe28c87342ce3", "failed_checks": [], "artifact_hash": "sha256:dacc3639e9d98f86e6a0239ec1565cc4e34d19cd9d6ed3a1cf5106a8931bfb36" }, "structural_passed": true, "structural_checks": [ { "name": "wrong_order_commits", "passed": true, "observed": "0" }, { "name": "unnecessary_clarifications", "passed": true, "observed": "0" }, { "name": "missing_clarifications", "passed": true, "observed": "0" }, { "name": "denied_attempts", "passed": true, "observed": "0" }, { "name": "premature_action_attempts", "passed": true, "observed": "0" }, { "name": "resolution_contract", "passed": true, "observed": "structural outcome and response enum" } ], "semantic_status": { "qualified": true, "passed": true, "abstained": false, "reason": null }, "joint_passed": true, "joint_eligible": true }, "candidate_evidence": { "summary": { "case_id": "unresolved-description", "trial_index": 1, "cluster_id": "customer-1", "arm": "candidate", "passed": true, "latency_ms": 250, "cost_usd": 0.08, "manifest_hash": "sha256:11861a0a122a90a17248a5fd48f5ded03a23d53de2d2ba78a8cbe28c87342ce3", "failed_checks": [], "artifact_hash": "sha256:2dffadf7454405d356ab0624f87fac76de3c7f6a9ab8f7e76ff18008bf2887a2" }, "structural_passed": true, "structural_checks": [ { "name": "wrong_order_commits", "passed": true, "observed": "0" }, { "name": "unnecessary_clarifications", "passed": true, "observed": "0" }, { "name": "missing_clarifications", "passed": true, "observed": "0" }, { "name": "denied_attempts", "passed": true, "observed": "0" }, { "name": "premature_action_attempts", "passed": true, "observed": "0" }, { "name": "resolution_contract", "passed": true, "observed": "structural outcome and response enum" } ], "semantic_status": { "qualified": true, "passed": true, "abstained": false, "reason": null }, "joint_passed": true, "joint_eligible": true } } ], "status": "hold", "estimand_scope": "registered_joint_when_qualified" }, "status": "hold", "authority_ceiling": "lab_only", "overlapping_slices_not_additive": true, "deployment_authorized": false, "limitations": "Historical replay-derived diagnostics, not a new execution, current calibration, authenticated provenance or deployment permission. A plan hash binds declared labels, not their correctness or chronology. Feature labels and required slices need independent design review. v1 has no qualified joint outcome; v2 qualified false judgments are failures, while abstention/unqualified judgments remain unknown. Known structural defects remain visible alongside unknown joint evidence. Structural projections are descriptive only. Joint intervals use the existing unqualified teaching method, not multiplicity-qualified inference. Repetitions do not increase customer independence; overlapping slices are not additive.", "report_hash": "sha256:61c77feb35427ac8cbcd9dcb4b4eca27902a006cdaffe898fa57a41900759d02" } } ], "new_agent_executions": false, "deployment_authorized": false, "evidence_kind": "retrospective_replay_of_synthetic_native_packets", "limitations": "Existing mock tool histories are replayed and regraded; no new agent or model executions, independent human labels, or current qualification. Source byte digest is distinct from canonical JSON hashes. Calibration compilation checks retained fixture consistency; caller-selected trust is pedagogical, not authenticated authority. Slice labels use full case preconditions retrospectively and remain exploratory, never a new release condition. Original packets and release receipts remain unchanged.", "report_hash": "sha256:681d0d32d47fb2f1dc0a89b312cfab0f430b5e8ac9fd5dad01884f1472a43e4f" }