llm/eval.json

{
  "accuracy": 0.9452,
  "ece": 0.0548,
  "brier": 0.0548,
  "test_n": 73,
  "dataset": "llm held-out"
}