{
  "calibration_config_count": 96,
  "calibration_passes": 96,
  "device": "Tesla T4",
  "heldout_config_count": 72,
  "heldout_passes": 72,
  "heldout_runs": 90,
  "limitations": [
    "hand authored finite kNN code variants",
    "no LLM code generation",
    "cached executable CPU profile used in RL rollouts",
    "six train/evaluation seeds",
    "shared Colab timing noise",
    "run measurements are machine- and time-dependent"
  ],
  "paired": [
    {
      "mean_log_gain_vs_random": 0.19110180517757902,
      "method": "online_only",
      "p_holm": 0.25,
      "p_raw": 0.0625
    },
    {
      "mean_log_gain_vs_random": 0.18143698832589936,
      "method": "dqn_hard",
      "p_holm": 0.25,
      "p_raw": 0.0625
    },
    {
      "mean_log_gain_vs_random": 0.1771087510893763,
      "method": "double_dqn_hard",
      "p_holm": 0.25,
      "p_raw": 0.0625
    },
    {
      "mean_log_gain_vs_random": 0.1616316751476158,
      "method": "dueling_double_soft",
      "p_holm": 0.25,
      "p_raw": 0.15625
    }
  ],
  "seed_count": 6,
  "summary": [
    {
      "geometric_speedup": 12.377281,
      "mean_accepted_edits": 3.666667,
      "mean_speedup": 15.469313,
      "method": "online_only"
    },
    {
      "geometric_speedup": 12.258233,
      "mean_accepted_edits": 3.5,
      "mean_speedup": 15.459313,
      "method": "dqn_hard"
    },
    {
      "geometric_speedup": 12.205291,
      "mean_accepted_edits": 3.222222,
      "mean_speedup": 15.37464,
      "method": "double_dqn_hard"
    },
    {
      "geometric_speedup": 12.017843,
      "mean_accepted_edits": 3.444444,
      "mean_speedup": 14.900204,
      "method": "dueling_double_soft"
    },
    {
      "geometric_speedup": 10.224234,
      "mean_accepted_edits": 2.833333,
      "mean_speedup": 13.037474,
      "method": "random"
    }
  ],
  "train_episodes_each": 75
}
