best_model/best_model_metadata.json

{
  "iteration": 226,
  "training_reward": 0.9533333333333334,
  "timestamp": "2026-01-02T01:35:36.597389",
  "experiment": "EXP-TRAIN-CDG",
  "cdg_metrics": {
    "total_episodes": 832,
    "wrong_to_right": 188,
    "right_to_wrong": 82,
    "right_to_right": 295,
    "wrong_to_wrong": 267,
    "filtered_episodes": 388,
    "avg_episode_quality": 0.9533333333333334
  },
  "note": "Best model by training reward (EuroEval evaluation pending)"
}