{
  "bench_version": 1,
  "track": "dynamics",
  "model": "rnn_small",
  "metric_direction": "lower is better",
  "n_seeds": 8,
  "baseline": {
    "best_cfg": {
      "lr": 0.01,
      "weight_decay": 0.0
    },
    "sweep": [
      {
        "cfg": {
          "lr": 0.001,
          "weight_decay": 0.0
        },
        "mean": 0.0014035784406587481
      },
      {
        "cfg": {
          "lr": 0.001,
          "weight_decay": 0.0001
        },
        "mean": 0.0013894051662646234
      },
      {
        "cfg": {
          "lr": 0.003,
          "weight_decay": 0.0
        },
        "mean": 0.000857145045301877
      },
      {
        "cfg": {
          "lr": 0.003,
          "weight_decay": 0.0001
        },
        "mean": 0.0008355810859939083
      },
      {
        "cfg": {
          "lr": 0.01,
          "weight_decay": 0.0
        },
        "mean": 0.0004466182872420177
      },
      {
        "cfg": {
          "lr": 0.01,
          "weight_decay": 0.0001
        },
        "mean": 0.0004611926560755819
      }
    ],
    "full": {
      "mean": 0.00039566773739352357,
      "std": 0.00014846719173678902,
      "per_seed": [
        0.0003142073401249945,
        0.00040627806447446346,
        0.0003180095227435231,
        0.0007479782216250896,
        0.00039004243444651365,
        0.00019843246263917536,
        0.000411077169701457,
        0.00037931668339297175
      ],
      "n": 8
    }
  },
  "idea": {
    "mean": 0.004026755806989968,
    "std": 0.001575483042595438,
    "per_seed": [
      0.0026446538977324963,
      0.003989342134445906,
      0.004552504979074001,
      0.006674116477370262,
      0.0018010623753070831,
      0.0026006274856626987,
      0.004008540417999029,
      0.005943198688328266
    ],
    "n": 8,
    "cfg": {
      "lr": 0.01,
      "weight_decay": 0.0,
      "beta": 0.1
    },
    "settings": [
      {
        "cfg": {
          "lr": 0.01,
          "weight_decay": 0.0,
          "beta": 0.1
        },
        "mean": 0.004026755806989968,
        "std": 0.001575483042595438,
        "per_seed": [
          0.0026446538977324963,
          0.003989342134445906,
          0.004552504979074001,
          0.006674116477370262,
          0.0018010623753070831,
          0.0026006274856626987,
          0.004008540417999029,
          0.005943198688328266
        ],
        "n": 8
      },
      {
        "cfg": {
          "lr": 0.01,
          "weight_decay": 0.0,
          "beta": 0.5
        },
        "mean": 0.004029476782307029,
        "std": 0.0015759746225326617,
        "per_seed": [
          0.002647238317877054,
          0.003994931932538748,
          0.004556045867502689,
          0.006677755154669285,
          0.0018026318866759539,
          0.002601723885163665,
          0.0040104272775352,
          0.005945059936493635
        ],
        "n": 8
      },
      {
        "cfg": {
          "lr": 0.01,
          "weight_decay": 0.0,
          "beta": 1.0
        },
        "mean": 0.004029815681860782,
        "std": 0.001576035937708343,
        "per_seed": [
          0.002647562650963664,
          0.003995629493147135,
          0.004556490574032068,
          0.006678206846117973,
          0.001802825485356152,
          0.002601858926936984,
          0.004010660108178854,
          0.005945291370153427
        ],
        "n": 8
      }
    ]
  },
  "comparison": {
    "delta_mean": 0.0036310880695964443,
    "idea_wins": 0,
    "n_pairs": 8,
    "per_seed_diffs": [
      0.0023304465576075017,
      0.0035830640699714422,
      0.004234495456330478,
      0.0059261382557451725,
      0.0014110199408605695,
      0.0024021950230235234,
      0.003597463248297572,
      0.005563882004935294
    ],
    "p_value": 0.0081,
    "mde": 0.0013221142535230909,
    "mde_rel_pct": 334.1476012759012,
    "verdict": "idea worse (significant)",
    "system_worked": false
  },
  "mechanism_signature": {
    "prediction": "Lyapunov drift penalty lowers positive one-step drift events on trained pendulum forecasts",
    "baseline_trained": {
      "positive_drift_fraction": 0.6109374836087227,
      "mean_drift_plus_lambdaV": 0.49601201340556145,
      "mean_V": 0.4504404291510582
    },
    "idea_trained": {
      "positive_drift_fraction": 0.47468749061226845,
      "mean_drift_plus_lambdaV": 0.19337605126202106,
      "mean_V": 0.4504404291510582
    },
    "relative_positive_drift_reduction": 0.22301789733320088,
    "confirmed": true
  },
  "protocol": {
    "paired_seeds": [
      0,
      1,
      2,
      3,
      4,
      5,
      6,
      7
    ],
    "sweep_seeds": [
      0,
      1,
      2,
      3
    ],
    "epochs": 20,
    "batch": 128,
    "loss": "MSE + beta*[finite_difference_dV + lambda*V]_+^2",
    "structural_match": "control/stability -> dynamics"
  }
}