{
  "bench_version": 1,
  "track": "dynamics",
  "model": "rnn_small",
  "metric_direction": "lower is better",
  "n_seeds": 8,
  "baseline": {
    "best_cfg": {
      "lr": 0.01,
      "reg": 0.03
    },
    "sweep": [
      {
        "cfg": {
          "lr": 0.001,
          "reg": 0.03
        },
        "mean": 0.000516663167218212
      },
      {
        "cfg": {
          "lr": 0.001,
          "reg": 0.1
        },
        "mean": 0.0005222226318437606
      },
      {
        "cfg": {
          "lr": 0.001,
          "reg": 0.3
        },
        "mean": 0.0005512491334229708
      },
      {
        "cfg": {
          "lr": 0.003,
          "reg": 0.03
        },
        "mean": 5.068586688139476e-05
      },
      {
        "cfg": {
          "lr": 0.003,
          "reg": 0.1
        },
        "mean": 5.551056983676972e-05
      },
      {
        "cfg": {
          "lr": 0.003,
          "reg": 0.3
        },
        "mean": 8.239948328991886e-05
      },
      {
        "cfg": {
          "lr": 0.01,
          "reg": 0.03
        },
        "mean": 1.4466989341599401e-05
      },
      {
        "cfg": {
          "lr": 0.01,
          "reg": 0.1
        },
        "mean": 1.7228220713150222e-05
      },
      {
        "cfg": {
          "lr": 0.01,
          "reg": 0.3
        },
        "mean": 3.9540235775348265e-05
      }
    ],
    "full": {
      "mean": 1.3864131801710755e-05,
      "std": 2.917053139718998e-06,
      "per_seed": [
        1.2164169675088488e-05,
        1.9545861505321227e-05,
        1.0918912266788539e-05,
        1.5239013919199351e-05,
        1.629439248063136e-05,
        1.428560062777251e-05,
        1.23286408779677e-05,
        1.0136463060916867e-05
      ],
      "n": 8
    }
  },
  "idea": {
    "mean": 1.4849264971417142e-05,
    "std": 2.2583458302260994e-06,
    "per_seed": [
      1.3598252735391725e-05,
      1.7523587303003296e-05,
      1.194078413391253e-05,
      1.765600791259203e-05,
      1.7466716599301435e-05,
      1.4715534234710503e-05,
      1.3851561561750714e-05,
      1.2041675290674902e-05
    ],
    "n": 8
  },
  "comparison": {
    "delta_mean": 9.85133169706387e-07,
    "idea_wins": 1,
    "n_pairs": 8,
    "per_seed_diffs": [
      1.4340830603032373e-06,
      -2.0222742023179308e-06,
      1.0218718671239913e-06,
      2.41699399339268e-06,
      1.1723241186700761e-06,
      4.299336069379933e-07,
      1.5229206837830134e-06,
      1.9052122297580354e-06
    ],
    "p_value": 0.0956,
    "mde": 1.1296557285196537e-06,
    "mde_rel_pct": 8.148045219681629,
    "verdict": "no significant win",
    "system_worked": false
  },
  "mechanism_signature": {
    "selected_cfg": {
      "lr": 0.01,
      "reg": 0.03
    },
    "idea_grid": [
      {
        "cfg": {
          "lr": 0.001,
          "reg": 0.03
        },
        "result": {
          "mean": 0.00052486258209683,
          "std": 5.698033059267576e-05,
          "per_seed": [
            0.00045051053166389465,
            0.0006266050622798502,
            0.0004562842077575624,
            0.0005525902379304171,
            0.0005077053210698068,
            0.0005103448056615889,
            0.0005080681294202805,
            0.0005867923609912395
          ],
          "n": 8
        }
      },
      {
        "cfg": {
          "lr": 0.001,
          "reg": 0.1
        },
        "result": {
          "mean": 0.0005531014503503684,
          "std": 6.433722525463851e-05,
          "per_seed": [
            0.0004721346776932478,
            0.00065583735704422,
            0.00045898332609795034,
            0.0005830824957229197,
            0.0005418244400061667,
            0.000540129782166332,
            0.0005432094330899417,
            0.0006296100909821689
          ],
          "n": 8
        }
      },
      {
        "cfg": {
          "lr": 0.001,
          "reg": 0.3
        },
        "result": {
          "mean": 0.0006540729009429924,
          "std": 8.133966965134085e-05,
          "per_seed": [
            0.0005586238112300634,
            0.0007529730210080743,
            0.0005105234449729323,
            0.0006941424799151719,
            0.0006594074657186866,
            0.0006338193779811263,
            0.000662658188957721,
            0.0007604354177601635
          ],
          "n": 8
        }
      },
      {
        "cfg": {
          "lr": 0.003,
          "reg": 0.03
        },
        "result": {
          "mean": 5.004885133530479e-05,
          "std": 9.772504653528356e-06,
          "per_seed": [
            6.087642759666778e-05,
            6.89755252096802e-05,
            4.271534999134019e-05,
            3.5638498957268894e-05,
            4.9467758799437433e-05,
            4.593151970766485e-05,
            4.691871799877845e-05,
            4.986701242160052e-05
          ],
          "n": 8
        }
      },
      {
        "cfg": {
          "lr": 0.003,
          "reg": 0.1
        },
        "result": {
          "mean": 6.530662221848615e-05,
          "std": 1.2555145162208372e-05,
          "per_seed": [
            7.812787953298539e-05,
            8.879921369953081e-05,
            4.93601146445144e-05,
            4.9300262617180124e-05,
            6.575142469955608e-05,
            6.285968993324786e-05,
            6.057017526472919e-05,
            6.768421735614538e-05
          ],
          "n": 8
        }
      },
      {
        "cfg": {
          "lr": 0.003,
          "reg": 0.3
        },
        "result": {
          "mean": 0.0001406980663887225,
          "std": 2.26067156523767e-05,
          "per_seed": [
            0.00016091138240881264,
            0.0001822679041652009,
            0.00010793711408041418,
            0.00011437860666774213,
            0.00014717398153152317,
            0.00013510147982742637,
            0.00013161970127839595,
            0.00014619436115026474
          ],
          "n": 8
        }
      },
      {
        "cfg": {
          "lr": 0.01,
          "reg": 0.03
        },
        "result": {
          "mean": 1.4849264971417142e-05,
          "std": 2.2583458302260994e-06,
          "per_seed": [
            1.3598252735391725e-05,
            1.7523587303003296e-05,
            1.194078413391253e-05,
            1.765600791259203e-05,
            1.7466716599301435e-05,
            1.4715534234710503e-05,
            1.3851561561750714e-05,
            1.2041675290674902e-05
          ],
          "n": 8
        }
      },
      {
        "cfg": {
          "lr": 0.01,
          "reg": 0.1
        },
        "result": {
          "mean": 2.5755787191883428e-05,
          "std": 4.61634426197786e-06,
          "per_seed": [
            2.8114078304497525e-05,
            1.8590118997963145e-05,
            2.2178555809659883e-05,
            3.419041968300007e-05,
            2.8991913495701738e-05,
            2.144245991075877e-05,
            2.614234108477831e-05,
            2.639641024870798e-05
          ],
          "n": 8
        }
      },
      {
        "cfg": {
          "lr": 0.01,
          "reg": 0.3
        },
        "result": {
          "mean": 9.178056825476233e-05,
          "std": 1.33129719536729e-05,
          "per_seed": [
            0.00011019065277650952,
            7.42549600545317e-05,
            7.907469989731908e-05,
            0.00011509206524351612,
            8.817380148684606e-05,
            8.460405661026016e-05,
            9.247064735973254e-05,
            9.038366260938346e-05
          ],
          "n": 8
        }
      }
    ],
    "signature": {
      "prediction": "inverse covariance weights low-noise drift more than high-noise drift",
      "low_high_weighted_cost_ratio_observed": 0.8153622015675527,
      "models": [
        {
          "kind": "baseline",
          "raw_drift_cost": 0.0008920590626075864,
          "whitened_drift_cost": 0.002272645477205515
        },
        {
          "kind": "idea",
          "raw_drift_cost": 0.000797593267634511,
          "whitened_drift_cost": 0.00203404831700027
        }
      ],
      "confirmed": false
    },
    "track_justification": "Dynamics is structurally matched: controlled pendulum rollout forecasting and drift/control regularization."
  }
}