{
  "arms": [
    {
      "accuracy_pct": 100.0,
      "correct": 5,
      "cost_is_lower_bound": false,
      "cost_usd": 0.203786,
      "max_wall_s": 126.745,
      "mean_wall_s": 69.715,
      "name": "ALEXIS Fusion",
      "timeout_count": 0,
      "total": 5
    },
    {
      "accuracy_pct": 60.0,
      "correct": 3,
      "cost_is_lower_bound": true,
      "cost_usd": 1.043435,
      "max_wall_s": 486.369,
      "mean_wall_s": 226.259,
      "name": "GPT-5.6 Sol",
      "timeout_count": 0,
      "total": 5
    },
    {
      "accuracy_pct": 60.0,
      "correct": 3,
      "cost_is_lower_bound": false,
      "cost_usd": 2.49154,
      "max_wall_s": 267.331,
      "mean_wall_s": 87.823,
      "name": "Fable 5",
      "timeout_count": 0,
      "total": 5
    },
    {
      "accuracy_pct": 60.0,
      "correct": 3,
      "cost_is_lower_bound": false,
      "cost_usd": 0.883398,
      "max_wall_s": 547.146,
      "mean_wall_s": 255.281,
      "name": "Grok 4.5",
      "timeout_count": 0,
      "total": 5
    },
    {
      "accuracy_pct": 60.0,
      "correct": 3,
      "cost_is_lower_bound": false,
      "cost_usd": 1.094415,
      "max_wall_s": 247.149,
      "mean_wall_s": 98.111,
      "name": "GPT-5.5",
      "timeout_count": 0,
      "total": 5
    },
    {
      "accuracy_pct": 60.0,
      "correct": 3,
      "cost_is_lower_bound": false,
      "cost_usd": 2.154045,
      "max_wall_s": 327.236,
      "mean_wall_s": 161.147,
      "name": "Opus 4.8",
      "timeout_count": 0,
      "total": 5
    },
    {
      "accuracy_pct": 60.0,
      "correct": 3,
      "cost_is_lower_bound": true,
      "cost_usd": 1.91337,
      "max_wall_s": 900.0,
      "mean_wall_s": 495.093,
      "name": "Fugu Ultra",
      "timeout_count": 2,
      "total": 5
    },
    {
      "accuracy_pct": 40.0,
      "correct": 2,
      "cost_is_lower_bound": false,
      "cost_usd": 0.092749,
      "max_wall_s": 483.657,
      "mean_wall_s": 245.678,
      "name": "DeepSeek V4 Pro",
      "timeout_count": 0,
      "total": 5
    },
    {
      "accuracy_pct": 60.0,
      "correct": 3,
      "cost_is_lower_bound": true,
      "cost_usd": 0.406857,
      "max_wall_s": 900.0,
      "mean_wall_s": 556.083,
      "name": "Qwen3.7 Max",
      "timeout_count": 2,
      "total": 5
    },
    {
      "accuracy_pct": 40.0,
      "correct": 2,
      "cost_is_lower_bound": true,
      "cost_usd": 0.296594,
      "max_wall_s": 900.0,
      "mean_wall_s": 444.398,
      "name": "GLM-5.2",
      "timeout_count": 1,
      "total": 5
    },
    {
      "accuracy_pct": 60.0,
      "correct": 3,
      "cost_is_lower_bound": false,
      "cost_usd": 1.485566,
      "max_wall_s": 338.64,
      "mean_wall_s": 162.915,
      "name": "Gemini 3.1 Pro",
      "timeout_count": 0,
      "total": 5
    },
    {
      "accuracy_pct": 40.0,
      "correct": 2,
      "cost_is_lower_bound": false,
      "cost_usd": 0.280271,
      "max_wall_s": 455.878,
      "mean_wall_s": 208.228,
      "name": "Grok 4.3",
      "timeout_count": 0,
      "total": 5
    }
  ],
  "benchmark": "postcutoff_pilot_v1",
  "dataset_sha256": "b5ef06357d5c0077ed5dbf35508eeee46a73dfb22dbcc81f7510989601ca4595",
  "freeze_tag": "postcutoff-pilot-v1-freeze-20260705",
  "task_count": 5
}