{
  "bench_version": 1,
  "track": "dynamics",
  "model": "rnn_small",
  "metric_direction": "lower is better",
  "n_seeds": 8,
  "baseline": {
    "best_cfg": {
      "lr": 0.01,
      "gain": 0.0
    },
    "sweep": [
      {
        "cfg": {
          "lr": 0.001,
          "gain": 0.0
        },
        "mean": 0.001116242419811897
      },
      {
        "cfg": {
          "lr": 0.003,
          "gain": 0.0
        },
        "mean": 0.0005947774989181198
      },
      {
        "cfg": {
          "lr": 0.01,
          "gain": 0.0
        },
        "mean": 0.00012917288222524803
      }
    ],
    "full": {
      "mean": 0.00010188088299400988,
      "std": 6.96920114960574e-05,
      "per_seed": [
        0.00021521856251638383,
        0.00019868022354785353,
        5.157031409908086e-05,
        5.122242873767391e-05,
        2.9590093618026003e-05,
        7.646851008757949e-05,
        4.320807784097269e-05,
        0.00014908885350450873
      ],
      "n": 8
    }
  },
  "idea": {
    "mean": 5.228218060437939e-05,
    "std": 4.3368261054941446e-05,
    "per_seed": [
      0.0001497500343248248,
      9.526080248178914e-05,
      2.5922774511855096e-05,
      3.981342524639331e-05,
      1.754087134031579e-05,
      2.408532782283146e-05,
      2.590940312074963e-05,
      3.9974805986275896e-05
    ],
    "n": 8
  },
  "comparison": {
    "delta_mean": -4.959870238963049e-05,
    "idea_wins": 8,
    "n_pairs": 8,
    "per_seed_diffs": [
      -6.546852819155902e-05,
      -0.00010341942106606439,
      -2.5647539587225765e-05,
      -1.14090034912806e-05,
      -1.2049222277710214e-05,
      -5.238318226474803e-05,
      -1.7298674720223062e-05,
      -0.00010911404751823284
    ],
    "p_value": 0.0081,
    "mde": 3.3421829705425835e-05,
    "mde_rel_pct": 32.804809620064724,
    "verdict": "idea better (significant)",
    "system_worked": true
  },
  "mechanism_signature": {
    "description": "trained GRUCell hidden-state observer correction on pendulum windows",
    "prediction": "local trained recurrent map should contract latent perturbations in V=||e||^2",
    "trained_model_measurement": {
      "gain": 1.0,
      "lr": 0.01,
      "predicted_V_ratio_bound": 0.9832114113428112,
      "observed_mean_V_ratio": 0.3183068037033081,
      "observed_max_V_ratio": 0.39757975935935974,
      "certificate_contraction_predicted": true,
      "observed_contraction": true,
      "confirmed": true
    },
    "idea_sweep": [
      {
        "cfg": {
          "lr": 0.01,
          "gain": 0.25
        },
        "result": {
          "mean": 7.185172580648214e-05,
          "std": 6.221648707698256e-05,
          "per_seed": [
            0.00020611767831724137,
            0.00011917055235244334,
            3.0685481760883704e-05,
            2.7863250579684973e-05,
            1.7419906725990586e-05,
            4.1515966586302966e-05,
            2.652378134371247e-05,
            0.00010551718878559768
          ],
          "n": 8
        }
      },
      {
        "cfg": {
          "lr": 0.01,
          "gain": 0.5
        },
        "result": {
          "mean": 6.361333316817763e-05,
          "std": 5.613212133729689e-05,
          "per_seed": [
            0.00017784775991458446,
            0.00013154245971236378,
            2.7753183530876413e-05,
            3.123701753793284e-05,
            1.546565181342885e-05,
            2.7983382096863352e-05,
            2.439024865452666e-05,
            7.268696208484471e-05
          ],
          "n": 8
        }
      },
      {
        "cfg": {
          "lr": 0.01,
          "gain": 1.0
        },
        "result": {
          "mean": 5.228218060437939e-05,
          "std": 4.3368261054941446e-05,
          "per_seed": [
            0.0001497500343248248,
            9.526080248178914e-05,
            2.5922774511855096e-05,
            3.981342524639331e-05,
            1.754087134031579e-05,
            2.408532782283146e-05,
            2.590940312074963e-05,
            3.9974805986275896e-05
          ],
          "n": 8
        }
      }
    ]
  },
  "idea_best_cfg": {
    "lr": 0.01,
    "gain": 1.0
  },
  "protocol_notes": {
    "track_match": "dynamics: controlled damped pendulum",
    "epochs": 18,
    "n_train": 800,
    "n_test": 300,
    "lr_union": [
      0.001,
      0.003,
      0.01
    ],
    "paired_seeds": 8,
    "baseline_sweep_seeds": 4
  }
}