{
  "bench_version": 1,
  "track": "dynamics",
  "model": "rnn_small",
  "metric_direction": "lower is better",
  "n_seeds": 8,
  "baseline": {
    "best_cfg": {
      "lr": 0.006
    },
    "sweep": [
      {
        "cfg": {
          "lr": 0.001
        },
        "mean": 3.534919142111903e-05
      },
      {
        "cfg": {
          "lr": 0.003
        },
        "mean": 7.330269568228687e-06
      },
      {
        "cfg": {
          "lr": 0.006
        },
        "mean": 5.007404297430185e-06
      }
    ],
    "full": {
      "mean": 6.034935353227411e-06,
      "std": 2.324192087412265e-06,
      "per_seed": [
        3.5480811675370205e-06,
        5.396596407081233e-06,
        4.0187219383369666e-06,
        7.066217676765518e-06,
        3.6791009279113496e-06,
        7.341893251577858e-06,
        6.276848125708057e-06,
        1.0952023330901284e-05
      ],
      "n": 8
    }
  },
  "idea": {
    "mean": 7.2727560791463475e-06,
    "std": 3.341130715564463e-06,
    "per_seed": [
      6.330245923891198e-06,
      5.862209491169779e-06,
      3.913731234206352e-06,
      5.612296263279859e-06,
      6.7712248892348725e-06,
      1.0451367415953428e-05,
      4.5703227442572825e-06,
      1.467065067117801e-05
    ],
    "n": 8
  },
  "comparison": {
    "delta_mean": 1.2378207259189367e-06,
    "idea_wins": 3,
    "n_pairs": 8,
    "per_seed_diffs": [
      2.7821647563541774e-06,
      4.656130840885453e-07,
      -1.0499070413061418e-07,
      -1.4539214134856593e-06,
      3.092123961323523e-06,
      3.1094741643755697e-06,
      -1.7065253814507741e-06,
      3.7186273402767256e-06
    ],
    "p_value": 0.164,
    "mde": 1.8372299258490038e-06,
    "mde_rel_pct": 30.443241200031668,
    "verdict": "no significant win",
    "system_worked": false
  },
  "mechanism_signature": {
    "critical_fraction": 0.2639999985694885,
    "baseline_critical_bad_rate": 0.0,
    "idea_critical_bad_rate": 0.0,
    "observed_bad_rate_reduction": 0.0,
    "required_rollouts_r_0.02_delta_0.05": 149,
    "confirmed": false
  },
  "idea_sweep": [
    {
      "cfg": {
        "lr": 0.001
      },
      "result": {
        "mean": 4.6562476200051606e-05,
        "std": 7.0209403088518815e-06,
        "per_seed": [
          3.4763717849273235e-05,
          5.860313831362873e-05,
          5.2090246754232794e-05,
          4.800773604074493e-05,
          3.818494951701723e-05,
          4.759346120408736e-05,
          4.485895624384284e-05,
          4.839760367758572e-05
        ],
        "n": 8
      }
    },
    {
      "cfg": {
        "lr": 0.003
      },
      "result": {
        "mean": 1.1251939099565789e-05,
        "std": 2.436469153192465e-06,
        "per_seed": [
          9.606156709196512e-06,
          1.1305033694952726e-05,
          6.192657565406989e-06,
          1.2527690159913618e-05,
          1.1936091141251381e-05,
          1.4925769391993526e-05,
          1.3000466424273327e-05,
          1.0521647709538229e-05
        ],
        "n": 8
      }
    },
    {
      "cfg": {
        "lr": 0.006
      },
      "result": {
        "mean": 7.2727560791463475e-06,
        "std": 3.341130715564463e-06,
        "per_seed": [
          6.330245923891198e-06,
          5.862209491169779e-06,
          3.913731234206352e-06,
          5.612296263279859e-06,
          6.7712248892348725e-06,
          1.0451367415953428e-05,
          4.5703227442572825e-06,
          1.467065067117801e-05
        ],
        "n": 8
      }
    }
  ],
  "protocol_note": "Official registered dynamics track; baseline uses bench.train_model; idea changes only the training loss and shares model, data, epochs, batch, and lr grid."
}