{
  "schemaVersion": 1,
  "kind": "cross-package-confirmation-action-replay",
  "task": "vue-type-impact-heldout-grading-v2",
  "repeat": 1,
  "arms": [
    "native",
    "impact"
  ],
  "runs": [
    {
      "id": "vue-type-impact-heldout-grading-v2-r1-native",
      "evidence": "evidence/cross-package/confirmation-002/results/runs/vue-type-impact-heldout-grading-v2-r1-native/public-evidence.json",
      "evidenceSha256": "eae4c54c0533e76212110455d2169f27dd9203222e50d350992e3f46bf01592a",
      "toolOutputs": "evidence/cross-package/confirmation-002/results/runs/vue-type-impact-heldout-grading-v2-r1-native/tool-outputs.json",
      "toolOutputsSha256": "cc04ac72ffd29ce24972da6863301a0153e69de0ed40aac2bbb9463766e807c8"
    },
    {
      "id": "vue-type-impact-heldout-grading-v2-r1-impact",
      "evidence": "evidence/cross-package/confirmation-002/results/runs/vue-type-impact-heldout-grading-v2-r1-impact/public-evidence.json",
      "evidenceSha256": "44b4935a65f48702f18cbb04ad8b714567d0e0b0a298a12f139c4898df5ed8ec",
      "toolOutputs": "evidence/cross-package/confirmation-002/results/runs/vue-type-impact-heldout-grading-v2-r1-impact/tool-outputs.json",
      "toolOutputsSha256": "e0bd2a6fb3873547bb8fc20a80f50f891a21a8a8673df9e28666b00c35c67d75"
    }
  ],
  "protocolSha256": "cb597d1d19d5f813ce8327f74b34bdfa74cbea42e6cfa24e363b355e31b1ea63",
  "report": {
    "path": "evidence/cross-package/confirmation-002/results/report.json",
    "sha256": "0559fc46e82f124f4d882cfec3e028d2f2f0687730216a881d3ca06f3748e469"
  },
  "adjudication": {
    "path": "evidence/cross-package/confirmation-002/results/adjudicated-result.json",
    "sha256": "c149c39fc104a262de553048de90178c86e1cca79f46d2470e626948662f6c0a"
  },
  "protocol": {
    "path": "evidence/cross-package/confirmation-002/protocol.json",
    "sha256": "2464bdb2404b1247d4aeea9aecec5b24e9b8702fffd709cda00eb3afb8262e8c"
  },
  "workflowAudit": {
    "path": "evidence/cross-package/confirmation-002/results/workflow-audit.json",
    "sha256": "5dc73d9de6236d1123ed2112db1896258029ded69bdf7f8698ccf33ab973a92a"
  },
  "exportManifest": {
    "path": "evidence/cross-package/confirmation-002/results/manifest.json",
    "sha256": "cc541404a3005463cfd739ab2cd0f8a179044c9a31584f9ab43224f432832086"
  },
  "codegraphChallenge": {
    "path": "evidence/cross-package/confirmation-002/results/codegraph-challenge.json",
    "sha256": "113ba8f68cf40d1b418bb33deb0fdb2e61e8a5f39fcd9b6d2e1a3782e14d9d66"
  },
  "codegraphChallengePolicy": {
    "path": "evidence/cross-package/confirmation-002/results/codegraph-challenge-policy.json",
    "sha256": "7f0c6e2419b494a79f42d9b88835a69173c8b23c31b18d59a86e9a0d84534a67"
  },
  "occurrenceAssessment": {
    "path": "evidence/cross-package/confirmation-002/results/occurrence-assessment-v1.json",
    "sha256": "11772608865210b5fc2f1651ef9432a365da45abddf4b1c572888cf2b1064611"
  },
  "controlAudit": {
    "path": "evidence/cross-package/confirmation-002/results/control-audit.json",
    "sha256": "fdd7cd9bfca51ba9a1df2bc78ed1f8ebb3736cc0b6025e63d53661349e70bcef"
  },
  "fullStudy": {
    "sessionCount": 16,
    "taskCount": 2,
    "repeats": 2,
    "qualifiedEdge": false,
    "minimumReductionPercent": 20,
    "arms": {
      "native": {
        "sessions": 4,
        "totalTokens": 886587,
        "chargedTimeMs": 408042.488166,
        "originalStrictPass": 2
      },
      "impact": {
        "sessions": 4,
        "totalTokens": 749471,
        "chargedTimeMs": 317935.1662910002,
        "originalStrictPass": 4
      },
      "competitor": {
        "sessions": 4,
        "totalTokens": 1597731,
        "chargedTimeMs": 487053.06250199955,
        "originalStrictPass": 2
      },
      "codegraph": {
        "sessions": 4,
        "totalTokens": 1526064,
        "chargedTimeMs": 476734.14204199985,
        "originalStrictPass": 3
      }
    },
    "comparisons": {
      "native": {
        "totalTokenReductionPercent": 15.465600104671061,
        "chargedTimeReductionPercent": 22.08282825643939
      },
      "competitor": {
        "totalTokenReductionPercent": 53.09154044078759,
        "chargedTimeReductionPercent": 34.72268408338055
      },
      "codegraph": {
        "totalTokenReductionPercent": 50.88862590297655,
        "chargedTimeReductionPercent": 33.3097552171981
      }
    },
    "nativeTokenGate": {
      "aggregateReductionPercent": 15.465600104671061,
      "medianPairedReductionPercent": 15.189077164118027,
      "passed": false
    },
    "originalStrictPass": 11,
    "occurrenceAssessedPass": 15,
    "scope": "Original strict grading and a separate occurrence-contract assessment are distinct. The overall failed gate is unchanged."
  },
  "renderer": {
    "path": "scripts/cross-package-confirmation-gif.py",
    "sha256": "368c8465be6a30944333dee97bd8f54586a9d746c9999ea9d35e628105d5549c"
  },
  "assets": {
    "gif": {
      "path": "assets/cross-package-confirmation.gif",
      "sha256": "a2075aab3a59bd6aac1aa2971bd5af85aae9bd541c7926311f308cbb423f0c5b",
      "width": 1080,
      "height": 680,
      "bytes": 427760
    },
    "poster": {
      "path": "assets/cross-package-confirmation.png",
      "sha256": "bf9af03f473bc05724798db034e131db2f0578a5c6a4671ca526c502d66e6156",
      "width": 1080,
      "height": 680,
      "bytes": 72591
    }
  },
  "dimensions": {
    "width": 1080,
    "height": 680
  },
  "playback": {
    "durationMs": 32000,
    "durationSeconds": 32.0,
    "accelerated": true,
    "separateSessionsAlignedAtStart": true,
    "eventTimes": [
      0,
      6153.533083000002,
      6218.517791000002,
      9695.388291000003,
      10809.341750000021,
      14846.469250000024,
      14917.695958000026,
      15705.606,
      22069.444875000016,
      29601.836499999998,
      37727.001625000004,
      52932.421,
      54047.8135,
      55508.779541000025,
      61968.855791,
      67952.675625,
      69607.35241600001,
      74849.16191600001,
      76635.562916,
      116703.32725,
      119170.45754099998
    ],
    "readingHolds": true,
    "renderedActionIds": {
      "vue-type-impact-heldout-grading-v2-r1-native": [
        "6a9722cdf589a2a068832c955a8253b9df87983501c03dd92e95debae1f409da",
        "4a317b6972e2f1834050485420bd71fe6fc2d87cd25d572cb4b8b327b37bc030",
        "57dbfc296665f072254d15345fb3a2bbf457ae03cbb6ee8f6020e8f11d4b056b",
        "9462965f2002ffb6e5ebf84d633936910b584441932a9dee26f8497eb9d82c75",
        "919299b8b8860441036a08b8ce4a6486ee1712366afaec2286c8579387359f27",
        "f4fdd034b54297294baf2c6977028dba2f4e8949c52e921b04118670cad8192e",
        "cfc5c391b5ba915c081c7d09c3edfc599f5369196d03ab00aa96b73369f89b5f",
        "a26c6de8780e4e792e8ad6df3fb6ac31fe229b7b2f8bc2aa29c602e8d647918c",
        "00bf463c9d57c76acacd039e798555d6101dbddf9708e68d232d71a4c30e0633",
        "b40ee9c64c8028e7786be4b1266661a398dfa60613358262a0ce9e36bddd365d"
      ],
      "vue-type-impact-heldout-grading-v2-r1-impact": [
        "6a9722cdf589a2a068832c955a8253b9df87983501c03dd92e95debae1f409da",
        "4a317b6972e2f1834050485420bd71fe6fc2d87cd25d572cb4b8b327b37bc030",
        "57dbfc296665f072254d15345fb3a2bbf457ae03cbb6ee8f6020e8f11d4b056b",
        "9462965f2002ffb6e5ebf84d633936910b584441932a9dee26f8497eb9d82c75",
        "3f21b4ea2fbe67cd553dbcfae108e6632bdb6c655218a1e4ad5a543cc24800f4",
        "f4fdd034b54297294baf2c6977028dba2f4e8949c52e921b04118670cad8192e",
        "cfc5c391b5ba915c081c7d09c3edfc599f5369196d03ab00aa96b73369f89b5f",
        "00bf463c9d57c76acacd039e798555d6101dbddf9708e68d232d71a4c30e0633",
        "fcf995ce3ce15e435101e7a49adbafdc79fcf9203eefbab40b5d124cae39b03f"
      ]
    },
    "timing": "Completed actions use measured receipt timestamps. One shared timeline aligns process starts; reading holds change playback spacing. Provider setup is included only in labelled charged-time totals.",
    "tokens": "Only terminal totals are shown. Cached input is a subset already included in total input; no interpolated token counters.",
    "selection": "First task and repeat, fixed before outcomes. Retain failure and disclose accelerated replay."
  },
  "caption": "Actual public Codex action replay of the preselected first task and repeat. Native tools and experimental impact query, shared workflow with explicit provider inventories. This pair is not the full study. The overall 16 session confirmation gate failed. Accelerated separate sessions, observed failures retained, final session totals only. Not a screen recording."
}
