{
  "status": "passed",
  "source_metric_functions_executed": true,
  "full_model_executed": false,
  "levenshtein_pairs": 961,
  "gold_answers_rebuilt": 80,
  "numeric_rows_independently_checked": 225,
  "invalid_submission_checks": 3,
  "reordered_submission_invariance": true,
  "chart_edges": [
    {
      "gold": "100",
      "prediction": "105.2",
      "source_chart": 1.0,
      "local_gold_denominator": 0.0
    },
    {
      "gold": "100",
      "prediction": "95",
      "source_chart": 0.0,
      "local_gold_denominator": 1.0
    },
    {
      "gold": "0",
      "prediction": "0.0",
      "source_chart": 0.0,
      "local_gold_denominator": 1.0
    },
    {
      "gold": "50%",
      "prediction": "0.5",
      "source_chart": 1.0,
      "local_gold_denominator": 1.0
    },
    {
      "gold": "50%",
      "prediction": "50",
      "source_chart": 0.0,
      "local_gold_denominator": 0.0
    },
    {
      "gold": "-100",
      "prediction": "-105.2",
      "source_chart": 1.0,
      "local_gold_denominator": 0.0
    },
    {
      "gold": "100",
      "prediction": "The answer is 100",
      "source_chart": 0.0,
      "local_gold_denominator": 0.0
    }
  ],
  "anls_edges": [
    {
      "gold": "ab",
      "prediction": "ac",
      "source_anls": 0.5,
      "local_normalized_strict": 0.0
    },
    {
      "gold": "abcd",
      "prediction": "abxx",
      "source_anls": 0.5,
      "local_normalized_strict": 0.0
    },
    {
      "gold": "abcd",
      "prediction": "abxx                ",
      "source_anls": 0.9,
      "local_normalized_strict": 0.0
    },
    {
      "gold": "AB123456",
      "prediction": "AB123450",
      "source_anls": 0.875,
      "local_normalized_strict": 0.875
    }
  ],
  "warning": "Snapshot-specific metric behavior; not benchmark impact or model quality."
}