{
  "bench_version": 1,
  "track": "dynamics",
  "model": "rnn_small",
  "metric_direction": "lower is better",
  "n_seeds": 8,
  "baseline": {
    "best_cfg": {
      "lr": 0.003,
      "weight_decay": 0.0,
      "kappa": 0.0
    },
    "sweep": [
      {
        "cfg": {
          "lr": 0.001,
          "weight_decay": 0.0,
          "kappa": 0.0
        },
        "mean": 0.4700177311897278
      },
      {
        "cfg": {
          "lr": 0.001,
          "weight_decay": 0.0001,
          "kappa": 0.0
        },
        "mean": 0.47006765753030777
      },
      {
        "cfg": {
          "lr": 0.003,
          "weight_decay": 0.0,
          "kappa": 0.0
        },
        "mean": 0.02822397626005113
      },
      {
        "cfg": {
          "lr": 0.003,
          "weight_decay": 0.0001,
          "kappa": 0.0
        },
        "mean": 0.02829078072682023
      },
      {
        "cfg": {
          "lr": 0.01,
          "weight_decay": 0.0,
          "kappa": 0.0
        },
        "mean": 0.03205677215009928
      },
      {
        "cfg": {
          "lr": 0.01,
          "weight_decay": 0.0001,
          "kappa": 0.0
        },
        "mean": 0.03197645582258701
      }
    ],
    "full": {
      "mean": 0.03168170212302357,
      "std": 0.018327317088355676,
      "per_seed": [
        0.06044498085975647,
        0.014531154185533524,
        0.01185071561485529,
        0.026069054380059242,
        0.05299823358654976,
        0.023797301575541496,
        0.014214038848876953,
        0.04954813793301582
      ],
      "n": 8
    }
  },
  "idea": {
    "mean": 0.030110739055089653,
    "std": 0.018290232592247106,
    "per_seed": [
      0.062178634107112885,
      0.010466991923749447,
      0.017325958237051964,
      0.034337423741817474,
      0.055281877517700195,
      0.03069481998682022,
      0.01328994520008564,
      0.017310261726379395
    ],
    "n": 8,
    "best_cfg": {
      "lr": 0.003,
      "weight_decay": 0.0,
      "kappa": 0.002
    },
    "sweep": [
      {
        "cfg": {
          "lr": 0.003,
          "weight_decay": 0.0,
          "kappa": 0.002
        },
        "mean": 0.030110739055089653,
        "per_seed": [
          0.062178634107112885,
          0.010466991923749447,
          0.017325958237051964,
          0.034337423741817474,
          0.055281877517700195,
          0.03069481998682022,
          0.01328994520008564,
          0.017310261726379395
        ]
      },
      {
        "cfg": {
          "lr": 0.003,
          "weight_decay": 0.0,
          "kappa": 0.005
        },
        "mean": 0.030580853228457272,
        "per_seed": [
          0.06277664005756378,
          0.010813195258378983,
          0.017922919243574142,
          0.03458359092473984,
          0.056420501321554184,
          0.030933773145079613,
          0.013734272681176662,
          0.017461933195590973
        ]
      },
      {
        "cfg": {
          "lr": 0.003,
          "weight_decay": 0.0,
          "kappa": 0.01
        },
        "mean": 0.03123972751200199,
        "per_seed": [
          0.06403753161430359,
          0.010984514839947224,
          0.01874014548957348,
          0.034955769777297974,
          0.05827799811959267,
          0.031186075881123543,
          0.014287841506302357,
          0.0174479428678751
        ]
      }
    ]
  },
  "comparison": {
    "delta_mean": -0.001570963067933917,
    "idea_wins": 3,
    "n_pairs": 8,
    "per_seed_diffs": [
      0.0017336532473564148,
      -0.004064162261784077,
      0.005475242622196674,
      0.008268369361758232,
      0.0022836439311504364,
      0.006897518411278725,
      -0.0009240936487913132,
      -0.03223787620663643
    ],
    "p_value": 0.93175,
    "mde": 0.010906916277571396,
    "mde_rel_pct": 34.426547649550606,
    "verdict": "no significant win",
    "system_worked": false
  },
  "mechanism_signature": {
    "math_sanity": {
      "predicted_rate": 1.4,
      "observed_rate": 1.288026525213435,
      "relative_error": 0.07998105341897488,
      "passed": true
    },
    "nn_scale": {
      "trained_model_metric": 0.062178634107112885,
      "predicted_rate": 0.008,
      "observed_rate": 0.009309132871376703,
      "relative_error": 0.16364160892208784,
      "confirmed": true,
      "mean_gap": 0.03696886791537205
    }
  },
  "track_justification": "Dynamics is the structural match because the idea targets recurrent hidden-state stability and controlled relaxation."
}