{
  "arms": [
    {
      "arm": "arm-1",
      "jobs": {
        "attempted": {
          "count": 3,
          "known_count": 3,
          "unknown_cells": 0
        },
        "cancelled": {
          "count": 0,
          "known_count": 0,
          "unknown_cells": 0
        },
        "completed": {
          "count": 2,
          "known_count": 2,
          "unknown_cells": 0
        },
        "failed": {
          "count": 1,
          "known_count": 1,
          "unknown_cells": 0
        },
        "planned": {
          "count": 3,
          "known_count": 3,
          "unknown_cells": 0
        },
        "scored": {
          "count": 2,
          "known_count": 2,
          "unknown_cells": 0
        },
        "submitted": {
          "count": 3,
          "known_count": 3,
          "unknown_cells": 0
        },
        "unfinished": {
          "count": 0,
          "known_count": 0,
          "unknown_cells": 0
        }
      },
      "samples": {
        "attempted": {
          "count": null,
          "known_count": 2,
          "unknown_cells": 1
        },
        "cancelled": {
          "count": 0,
          "known_count": 0,
          "unknown_cells": 0
        },
        "failed": {
          "count": null,
          "known_count": 0,
          "unknown_cells": 1
        },
        "planned": {
          "count": 3,
          "known_count": 3,
          "unknown_cells": 0
        },
        "scored": {
          "count": 2,
          "known_count": 2,
          "unknown_cells": 0
        },
        "submitted": {
          "count": 3,
          "known_count": 3,
          "unknown_cells": 0
        },
        "unfinished": {
          "count": 1,
          "known_count": 1,
          "unknown_cells": 0
        }
      }
    }
  ],
  "cells": [
    {
      "arm": "arm-1",
      "cache": {
        "proven": null,
        "requested": "uncontrolled"
      },
      "cell": "cell-1",
      "evidence_status": "retained",
      "identity_status": "verified",
      "matching": {
        "cache_sha256": "085931b8e2281b6f2b9b6106a89daf2a6ed108c79630bba93ea4aae627c01130",
        "controls_sha256": "ffd04e617211ad465fe0a068b1cb698bc88b7c90172132422df06f297aa05c1a",
        "oracle_sha256": "e30fa5b8a89b72aab501582d7c6d101d5ddbdfc9a697434524fd40b062bb2b66",
        "workload_sha256": "b5c0f5a5398051c7a683982f6555c835a40b984ab8d021a04b0f7e0e14311ce2"
      },
      "measurement_complete": false,
      "native_aggregate_metrics": {},
      "outcomes": [
        {
          "attempted": null,
          "failed": null,
          "latency_sum_ms": null,
          "passed": null
        }
      ],
      "raw_task_grade_diagnostics": null,
      "raw_task_grades": null,
      "repeat": 1,
      "requested_concurrency": 1,
      "samples": {
        "attempted": null,
        "cancelled": 0,
        "failed": null,
        "planned": 1,
        "scored": 0,
        "submitted": 1,
        "unfinished": 1
      },
      "spec_sha256": "a063a82faddbb37002a1f88e0f864239f5161675cfa38d7f2d8208ca82064a73",
      "state": "failed",
      "suite": "aider",
      "task": "task-1",
      "telemetry_coverage": {
        "cgroup": "measured",
        "engine": "measured",
        "gpu": "measured",
        "host": "measured"
      }
    },
    {
      "arm": "arm-1",
      "cache": {
        "proven": null,
        "requested": "uncontrolled"
      },
      "cell": "cell-2",
      "evidence_status": "retained",
      "identity_status": "verified",
      "matching": {
        "cache_sha256": "085931b8e2281b6f2b9b6106a89daf2a6ed108c79630bba93ea4aae627c01130",
        "controls_sha256": "ffd04e617211ad465fe0a068b1cb698bc88b7c90172132422df06f297aa05c1a",
        "oracle_sha256": "e30fa5b8a89b72aab501582d7c6d101d5ddbdfc9a697434524fd40b062bb2b66",
        "workload_sha256": "3d3313c0a390fce13c8572aea94ec54402f5784459b2803c8bf5336f993e5456"
      },
      "measurement_complete": true,
      "native_aggregate_metrics": {},
      "outcomes": [
        {
          "attempted": true,
          "failed": false,
          "latency_sum_ms": null,
          "passed": true
        }
      ],
      "raw_task_grade_diagnostics": null,
      "raw_task_grades": null,
      "repeat": 1,
      "requested_concurrency": 1,
      "samples": {
        "attempted": 1,
        "cancelled": 0,
        "failed": 0,
        "planned": 1,
        "scored": 1,
        "submitted": 1,
        "unfinished": 0
      },
      "spec_sha256": "197839f20cdd058bff2ceaf69fa223f6fb0e46df9297a124e5128acbe78a7bc2",
      "state": "completed",
      "suite": "aider",
      "task": "task-2",
      "telemetry_coverage": {
        "cgroup": "measured",
        "engine": "measured",
        "gpu": "measured",
        "host": "measured"
      }
    },
    {
      "arm": "arm-1",
      "cache": {
        "proven": null,
        "requested": "uncontrolled"
      },
      "cell": "cell-3",
      "evidence_status": "retained",
      "identity_status": "verified",
      "matching": {
        "cache_sha256": "085931b8e2281b6f2b9b6106a89daf2a6ed108c79630bba93ea4aae627c01130",
        "controls_sha256": "ffd04e617211ad465fe0a068b1cb698bc88b7c90172132422df06f297aa05c1a",
        "oracle_sha256": "e30fa5b8a89b72aab501582d7c6d101d5ddbdfc9a697434524fd40b062bb2b66",
        "workload_sha256": "19be18ade96ec853ec6ad1f8ebf6a3c47efccd95625a9cb44000c65e234320b0"
      },
      "measurement_complete": true,
      "native_aggregate_metrics": {},
      "outcomes": [
        {
          "attempted": true,
          "failed": false,
          "latency_sum_ms": null,
          "passed": true
        }
      ],
      "raw_task_grade_diagnostics": null,
      "raw_task_grades": null,
      "repeat": 1,
      "requested_concurrency": 1,
      "samples": {
        "attempted": 1,
        "cancelled": 0,
        "failed": 0,
        "planned": 1,
        "scored": 1,
        "submitted": 1,
        "unfinished": 0
      },
      "spec_sha256": "99965232d18db773bc5de204ecb737c5e7c4a2d49148b66fe3d3b311299182c4",
      "state": "completed",
      "suite": "aider",
      "task": "task-3",
      "telemetry_coverage": {
        "cgroup": "measured",
        "engine": "measured",
        "gpu": "measured",
        "host": "measured"
      }
    }
  ],
  "comparisons": [],
  "jobs": {
    "attempted": {
      "count": 3,
      "known_count": 3,
      "unknown_cells": 0
    },
    "cancelled": {
      "count": 0,
      "known_count": 0,
      "unknown_cells": 0
    },
    "completed": {
      "count": 2,
      "known_count": 2,
      "unknown_cells": 0
    },
    "failed": {
      "count": 1,
      "known_count": 1,
      "unknown_cells": 0
    },
    "planned": {
      "count": 3,
      "known_count": 3,
      "unknown_cells": 0
    },
    "scored": {
      "count": 2,
      "known_count": 2,
      "unknown_cells": 0
    },
    "submitted": {
      "count": 3,
      "known_count": 3,
      "unknown_cells": 0
    },
    "unfinished": {
      "count": 0,
      "known_count": 0,
      "unknown_cells": 0
    }
  },
  "labels": "synthetic labels in manifest order; private identities omitted",
  "manifest_sha256": "0f5942ede3c4a79c2cc14e66df6bd0e3af6f51c5fc142dffd73aebf3b7224e91",
  "promotion_authorized": false,
  "repeats": [
    {
      "agreement": null,
      "arm": "arm-1",
      "inference": "descriptive_only; no confidence interval or significance claim",
      "latency_sum_ms_range": null,
      "planned_repeats": 1,
      "requested_concurrency": 1,
      "score_range": null,
      "scored_repeats": 0,
      "task": "task-1"
    },
    {
      "agreement": null,
      "arm": "arm-1",
      "inference": "descriptive_only; no confidence interval or significance claim",
      "latency_sum_ms_range": null,
      "planned_repeats": 1,
      "requested_concurrency": 1,
      "score_range": null,
      "scored_repeats": 1,
      "task": "task-2"
    },
    {
      "agreement": null,
      "arm": "arm-1",
      "inference": "descriptive_only; no confidence interval or significance claim",
      "latency_sum_ms_range": null,
      "planned_repeats": 1,
      "requested_concurrency": 1,
      "score_range": null,
      "scored_repeats": 1,
      "task": "task-3"
    }
  ],
  "samples": {
    "attempted": {
      "count": null,
      "known_count": 2,
      "unknown_cells": 1
    },
    "cancelled": {
      "count": 0,
      "known_count": 0,
      "unknown_cells": 0
    },
    "failed": {
      "count": null,
      "known_count": 0,
      "unknown_cells": 1
    },
    "planned": {
      "count": 3,
      "known_count": 3,
      "unknown_cells": 0
    },
    "scored": {
      "count": 2,
      "known_count": 2,
      "unknown_cells": 0
    },
    "submitted": {
      "count": 3,
      "known_count": 3,
      "unknown_cells": 0
    },
    "unfinished": {
      "count": 1,
      "known_count": 1,
      "unknown_cells": 0
    }
  },
  "schema": "anvil-serving.benchmark-campaign-report/v1",
  "timing": {
    "active_wall_seconds": 455.5326589846518,
    "child_active_wall_seconds": 410.6106981278863,
    "known_active_wall_seconds": 455.5326589846518,
    "observed_job_overlap": 1,
    "observed_request_overlap": null,
    "request_overlap_status": "unknown",
    "scope": "active wall includes release, verified drain and bounded telemetry finalization; child overlap is not request concurrency",
    "unknown_batch_intervals": 0,
    "unknown_intervals": 0
  }
}
