{
  "schema": "evalarc.independent-swe-evidence.v1",
  "attempts": 36,
  "source_tasks": 3,
  "dispositions": {
    "not_accepted": 31,
    "unavailable": 5
  },
  "changed_candidates": 8,
  "usage_incomplete": 1,
  "model_requests": 614,
  "groups": [
    {
      "condition": "none",
      "attempts": 9,
      "not_accepted": 7,
      "unavailable": 2,
      "changed_candidates": 4,
      "usage_incomplete": 1
    },
    {
      "condition": "direct",
      "attempts": 9,
      "not_accepted": 7,
      "unavailable": 2,
      "changed_candidates": 2,
      "usage_incomplete": 0
    },
    {
      "condition": "mcp",
      "attempts": 9,
      "not_accepted": 9,
      "changed_candidates": 0,
      "usage_incomplete": 0
    },
    {
      "condition": "unrelated",
      "attempts": 9,
      "not_accepted": 8,
      "unavailable": 1,
      "changed_candidates": 2,
      "usage_incomplete": 0
    }
  ],
  "model": {
    "model": "Qwen/Qwen3-8B",
    "revision": "b968826d9c46dd6066d109eabc6255188de91218",
    "device": "NVIDIA L40S",
    "dtype": "bfloat16",
    "torch": "2.11.0+cu130",
    "transformers": "5.5.4",
    "thinking_enabled": false,
    "model_files_manifest_sha256": "06188fd0f01af76aff2dfdcc87207b4977abe32a812e3802baf3f44b737515e9"
  },
  "recorder_commit": "285eadef59ec0e673e4feba6322e4c932af0b5ff",
  "config_sha256": "9c4e23bf68eab62732f0cfcd6ec36e342b439c28d4ddcb35e09feae6b006472c",
  "controls": [
    {
      "instance_id": "astropy__astropy-7606",
      "mode": "original-defect",
      "upstream_resolved": false,
      "infra_failure": false,
      "infra_failure_reason": null,
      "disposition": "not_accepted",
      "f2p_passed": 0,
      "f2p_total": 1,
      "p2p_passed": 240,
      "p2p_total": 240,
      "aggregate_pass_fraction": 0.995850622406639,
      "aggregate_95_accepted": true
    },
    {
      "instance_id": "astropy__astropy-7606",
      "mode": "original-fix",
      "upstream_resolved": true,
      "infra_failure": false,
      "infra_failure_reason": null,
      "disposition": "accepted",
      "f2p_passed": 1,
      "f2p_total": 1,
      "p2p_passed": 240,
      "p2p_total": 240,
      "aggregate_pass_fraction": 1.0,
      "aggregate_95_accepted": true
    },
    {
      "instance_id": "pytest-dev__pytest-7432",
      "mode": "original-defect",
      "upstream_resolved": false,
      "infra_failure": false,
      "infra_failure_reason": null,
      "disposition": "not_accepted",
      "f2p_passed": 0,
      "f2p_total": 1,
      "p2p_passed": 77,
      "p2p_total": 77,
      "aggregate_pass_fraction": 0.9871794871794872,
      "aggregate_95_accepted": true
    },
    {
      "instance_id": "pytest-dev__pytest-7432",
      "mode": "original-fix",
      "upstream_resolved": true,
      "infra_failure": false,
      "infra_failure_reason": null,
      "disposition": "accepted",
      "f2p_passed": 1,
      "f2p_total": 1,
      "p2p_passed": 77,
      "p2p_total": 77,
      "aggregate_pass_fraction": 1.0,
      "aggregate_95_accepted": true
    },
    {
      "instance_id": "sympy__sympy-15809",
      "mode": "original-defect",
      "upstream_resolved": false,
      "infra_failure": false,
      "infra_failure_reason": null,
      "disposition": "not_accepted",
      "f2p_passed": 0,
      "f2p_total": 2,
      "p2p_passed": 10,
      "p2p_total": 10,
      "aggregate_pass_fraction": 0.8333333333333334,
      "aggregate_95_accepted": false
    },
    {
      "instance_id": "sympy__sympy-15809",
      "mode": "original-fix",
      "upstream_resolved": true,
      "infra_failure": false,
      "infra_failure_reason": null,
      "disposition": "accepted",
      "f2p_passed": 2,
      "f2p_total": 2,
      "p2p_passed": 10,
      "p2p_total": 10,
      "aggregate_pass_fraction": 1.0,
      "aggregate_95_accepted": true
    }
  ]
}