{
  "device": "CUDA",
  "precision": "FP32 unmerged",
  "compared_with": "original evaluated GPU checkpoint, development split, raw logits",
  "cases": [
    {
      "case": 4222,
      "qid": "injection",
      "qtype": "noul",
      "tokens": 140,
      "expected_label": 0,
      "max_probability_difference": 3.9981387089937925e-09,
      "max_logit_difference": 0.0003533363342285156,
      "exported_prediction": 0,
      "reference_prediction": 0
    },
    {
      "case": 4654,
      "qid": "injection",
      "qtype": "noul",
      "tokens": 152,
      "expected_label": 0,
      "max_probability_difference": 0.00026088953018188477,
      "max_logit_difference": 0.0010182857513427734,
      "exported_prediction": 0,
      "reference_prediction": 0
    },
    {
      "case": 4222,
      "qid": "boundary_action",
      "qtype": "choice",
      "tokens": 192,
      "expected_label": 0,
      "max_probability_difference": 1.1920928955078125e-07,
      "max_logit_difference": 0.002742290496826172,
      "exported_prediction": 0,
      "reference_prediction": 0
    },
    {
      "case": 4654,
      "qid": "boundary_action",
      "qtype": "choice",
      "tokens": 204,
      "expected_label": 0,
      "max_probability_difference": 3.2186508178710938e-06,
      "max_logit_difference": 0.0013228654861450195,
      "exported_prediction": 0,
      "reference_prediction": 0
    },
    {
      "case": 2334,
      "qid": "evidence_level",
      "qtype": "score",
      "tokens": 234,
      "expected_label": 1,
      "max_probability_difference": 9.085852070711553e-10,
      "max_logit_difference": 0.0014123916625976562,
      "exported_prediction": 1,
      "reference_prediction": 1
    },
    {
      "case": 1790,
      "qid": "evidence_level",
      "qtype": "score",
      "tokens": 250,
      "expected_label": 1,
      "max_probability_difference": 1.3608314475277439e-09,
      "max_logit_difference": 0.0016317367553710938,
      "exported_prediction": 1,
      "reference_prediction": 1
    }
  ],
  "all_predictions_match": true,
  "max_probability_difference": 0.00026088953018188477,
  "tolerance": 0.001,
  "sample_selection": "two shortest development rows for each of three types; packaging parity only",
  "handwritten_example": {
    "expected_action": "reject",
    "actual_prediction": "reject",
    "matches_label": true,
    "purpose": "Interface smoke example; not a benchmark result"
  }
}