{
  "bench_version": 1,
  "track": "dynamics",
  "model": "rnn_small",
  "metric_direction": "lower is better",
  "n_seeds": 8,
  "baseline": {
    "best_cfg": {
      "lr": 0.006,
      "epochs": 24
    },
    "sweep": [
      {
        "cfg": {
          "lr": 0.0015,
          "epochs": 24
        },
        "mean": 0.0027078510902356356
      },
      {
        "cfg": {
          "lr": 0.003,
          "epochs": 24
        },
        "mean": 0.002131043642293662
      },
      {
        "cfg": {
          "lr": 0.006,
          "epochs": 24
        },
        "mean": 0.0017603125161258504
      }
    ],
    "full": {
      "mean": 0.0017603125161258504,
      "std": 0.0003135227734284653,
      "per_seed": [
        0.001380653353407979,
        0.0020586533937603235,
        0.00220737443305552,
        0.0021458740811794996,
        0.0016572552267462015,
        0.00137868314050138,
        0.0015466997865587473,
        0.001707306713797152
      ],
      "n": 8
    }
  },
  "idea": {
    "mean": 0.0019414099806454033,
    "std": 0.000537525437857345,
    "per_seed": [
      0.001175302779302001,
      0.0021009868942201138,
      0.0018062262097373605,
      0.0030378331430256367,
      0.0023114001378417015,
      0.001854351139627397,
      0.0013667570892721415,
      0.0018784224521368742
    ],
    "n": 8,
    "selected_cfg": {
      "lr": 0.006,
      "epochs": 24
    }
  },
  "comparison": {
    "delta_mean": 0.0001810974645195529,
    "idea_wins": 3,
    "n_pairs": 8,
    "per_seed_diffs": [
      -0.000205350574105978,
      4.233350045979023e-05,
      -0.0004011482233181596,
      0.000891959061846137,
      0.0006541449110955,
      0.0004756679991260171,
      -0.00017994269728660583,
      0.00017111573833972216
    ],
    "p_value": 0.2882,
    "mde": 0.00038107323489683164,
    "mde_rel_pct": 21.648044390180743,
    "verdict": "no significant win",
    "system_worked": false
  },
  "mechanism_signature": {
    "quantity": "trained-model mean absolute 1^T B P",
    "predicted": "zero internal net exchange",
    "observed_max": 0.0,
    "observed_mean": 0.0,
    "per_seed": [
      {
        "seed": 0,
        "mean_abs_internal_residual": 0.0,
        "mean_abs_predicted_output": 0.7605881690979004
      },
      {
        "seed": 1,
        "mean_abs_internal_residual": 0.0,
        "mean_abs_predicted_output": 0.7429919838905334
      },
      {
        "seed": 2,
        "mean_abs_internal_residual": 0.0,
        "mean_abs_predicted_output": 0.7579967379570007
      },
      {
        "seed": 3,
        "mean_abs_internal_residual": 0.0,
        "mean_abs_predicted_output": 0.8042439818382263
      },
      {
        "seed": 4,
        "mean_abs_internal_residual": 0.0,
        "mean_abs_predicted_output": 0.7317724227905273
      },
      {
        "seed": 5,
        "mean_abs_internal_residual": 0.0,
        "mean_abs_predicted_output": 0.7670498490333557
      },
      {
        "seed": 6,
        "mean_abs_internal_residual": 0.0,
        "mean_abs_predicted_output": 0.756341278553009
      },
      {
        "seed": 7,
        "mean_abs_internal_residual": 0.0,
        "mean_abs_predicted_output": 0.6808968782424927
      }
    ],
    "confirmed": true
  },
  "protocol": {
    "paired_seeds": [
      0,
      1,
      2,
      3,
      4,
      5,
      6,
      7
    ],
    "grid": [
      {
        "lr": 0.0015,
        "epochs": 24
      },
      {
        "lr": 0.003,
        "epochs": 24
      },
      {
        "lr": 0.006,
        "epochs": 24
      }
    ],
    "baseline_best_cfg": {
      "lr": 0.006,
      "epochs": 24
    },
    "idea_grid_results": [
      {
        "cfg": {
          "lr": 0.0015,
          "epochs": 24
        },
        "result": {
          "mean": 0.003114120219834149,
          "std": 0.0007247547558258706,
          "per_seed": [
            0.002160337520763278,
            0.003606663551181555,
            0.003154655685648322,
            0.003743839683011174,
            0.0032033741008490324,
            0.0022424152120947838,
            0.002473633736371994,
            0.004328042268753052
          ],
          "n": 8
        }
      },
      {
        "cfg": {
          "lr": 0.003,
          "epochs": 24
        },
        "result": {
          "mean": 0.002424486941890791,
          "std": 0.0005536617718879092,
          "per_seed": [
            0.0017574743833392859,
            0.0028110698331147432,
            0.00233089760877192,
            0.0031846617348492146,
            0.002762784017249942,
            0.0018244534730911255,
            0.0017341335769742727,
            0.0029904209077358246
          ],
          "n": 8
        }
      },
      {
        "cfg": {
          "lr": 0.006,
          "epochs": 24
        },
        "result": {
          "mean": 0.0019414099806454033,
          "std": 0.000537525437857345,
          "per_seed": [
            0.001175302779302001,
            0.0021009868942201138,
            0.0018062262097373605,
            0.0030378331430256367,
            0.0023114001378417015,
            0.001854351139627397,
            0.0013667570892721415,
            0.0018784224521368742
          ],
          "n": 8,
          "selected_cfg": {
            "lr": 0.006,
            "epochs": 24
          }
        }
      }
    ]
  }
}