baseline_scores.json

{
  "easy": 0.9,
  "medium": 0.62,
  "hard": 0.7239,
  "expert": 0.6996,
  "_metadata": {
    "model": "qwen/qwen3-32b",
    "provider": "OpenRouter",
    "temperature": 0.0,
    "max_tokens": 4096,
    "timestamp": "2026-04-12T18:00:00Z"
  },
  "_oracle": {
    "easy": 0.90,
    "medium": 0.63,
    "hard": 0.78,
    "expert": 0.95
  },
  "_heuristic": {
    "easy": 0.90,
    "medium": 0.63,
    "hard": 0.07,
    "expert": 0.78
  }
}