prompts/subclaim_support_valid.txt

You are an expert medical fact verification judge.

Input:
1) A medical document
2) A list of subclaims extracted from the document
3) A model-predicted label for each subclaim

Label definitions:
- supported: The document explicitly supports the subclaim.
- refuted: The document explicitly contradicts the subclaim.
- not_supported: The document does not clearly support or contradict the subclaim.

Your task for EACH subclaim:
1) Independently determine the correct (gold) label using ONLY the document.
2) Compare it with the model-predicted label.

Rules:
- Use ONLY the provided document.
- Do NOT use external medical knowledge.
- Be conservative: if evidence is unclear, choose not_supported.
- Judge each subclaim independently.

Return your response STRICTLY in valid JSON.
Do NOT include any text outside the JSON.

JSON output format:
{
  "results": [
    {
      "subclaim_index": "<string>",
      "gold_label": "supported | refuted | not_supported",
      "model_label": "supported | refuted | not_supported",
      "model_label_correct": true | false
    }
  ],
  "accuracy": <float between 0 and 1>
}

Accuracy definition:
accuracy = (number of subclaims where model_label_correct = true) / (total number of subclaims)

Document:
<<<DOCUMENT>>>

Subclaims with predicted model results:
<<<SUBCLAIMS>>>