{
  "submission": "random_HIPE-2026-v1.0-surprise-test-fr_run1.jsonl",
  "reference": "HIPE-2026-v1.0-surprise-test-fr.jsonl",
  "team": "random",
  "dataset": "surprise",
  "split": "test",
  "language": "fr",
  "run": "run1",
  "at_label_mode": "TERNARY",
  "confusion_matrices": {
    "at": {
      "labels": [
        "TRUE",
        "PROBABLE",
        "FALSE"
      ],
      "matrix": [
        [
          38,
          52,
          40
        ],
        [
          16,
          25,
          19
        ],
        [
          98,
          82,
          110
        ]
      ],
      "rows": "gold",
      "columns": "prediction",
      "total": 480,
      "table": [
        {
          "gold": "TRUE",
          "pred_TRUE": 38,
          "pred_PROBABLE": 52,
          "pred_FALSE": 40
        },
        {
          "gold": "PROBABLE",
          "pred_TRUE": 16,
          "pred_PROBABLE": 25,
          "pred_FALSE": 19
        },
        {
          "gold": "FALSE",
          "pred_TRUE": 98,
          "pred_PROBABLE": 82,
          "pred_FALSE": 110
        }
      ],
      "accuracy": 0.3604,
      "per_label": {
        "TRUE": {
          "precision": 0.25,
          "recall": 0.2923,
          "f1": 0.2695,
          "support": 130,
          "true_positive": 38,
          "false_positive": 114,
          "false_negative": 92
        },
        "PROBABLE": {
          "precision": 0.1572,
          "recall": 0.4167,
          "f1": 0.2283,
          "support": 60,
          "true_positive": 25,
          "false_positive": 134,
          "false_negative": 35
        },
        "FALSE": {
          "precision": 0.6509,
          "recall": 0.3793,
          "f1": 0.4793,
          "support": 290,
          "true_positive": 110,
          "false_positive": 59,
          "false_negative": 180
        }
      }
    }
  }
}