data/benchmark_summary.json

{
  "summary": {
    "timestamp": "2026-06-16T21:50:28.207Z",
    "total_processed": 6,
    "reproducibility_rate": 50,
    "specificity_rate": 100,
    "overall_ground_truth_rate": 50,
    "average_iterations": 13.333333333333334
  },
  "details": [
    {
      "id": "008",
      "reproducible": false,
      "specific": false,
      "iterations": 30
    },
    {
      "id": "020",
      "reproducible": true,
      "specific": true,
      "iterations": 9
    },
    {
      "id": "041",
      "reproducible": false,
      "specific": false,
      "iterations": 13
    },
    {
      "id": "054",
      "reproducible": true,
      "specific": true,
      "iterations": 7
    },
    {
      "id": "070",
      "reproducible": false,
      "specific": false,
      "iterations": 13
    },
    {
      "id": "077",
      "reproducible": true,
      "specific": true,
      "iterations": 8
    }
  ]
}