{
  "analysis": "Strict means raw response parses as JSON and decision AND amount match the answer key. Fence normalization was added after observing the outputs; it removes only a complete outer Markdown code fence before parsing. It does not score customer-facing prose or the full schema.",
  "models": {
    "gpt-6-luna": {
      "attempts": 12,
      "strict_pass": 12,
      "fenced_json": 0,
      "normalized_pass": 12,
      "api_errors": 0,
      "strict_flag_mismatches": 0
    },
    "gemini-3.8-flash": {
      "attempts": 12,
      "strict_pass": 0,
      "fenced_json": 12,
      "normalized_pass": 12,
      "api_errors": 0,
      "strict_flag_mismatches": 0
    }
  }
}
