{
  "bench_version": 1,
  "track": "dynamics",
  "model": "rnn_small",
  "metric_direction": "lower is better",
  "n_seeds": 8,
  "baseline": {
    "best_cfg": {
      "lr": 0.01
    },
    "sweep": [
      {
        "cfg": {
          "lr": 0.001
        },
        "mean": 0.001143058019806631
      },
      {
        "cfg": {
          "lr": 0.003
        },
        "mean": 0.0005966055832686834
      },
      {
        "cfg": {
          "lr": 0.01
        },
        "mean": 0.0001800259415176697
      }
    ],
    "full": {
      "mean": 0.00013833412685926305,
      "std": 8.74152934908656e-05,
      "per_seed": [
        0.00026837096083909273,
        0.0002815499610733241,
        7.163507689256221e-05,
        9.85477672656998e-05,
        0.00010150683374376968,
        5.219123704591766e-05,
        5.341393261915073e-05,
        0.00017945724539458752
      ],
      "n": 8
    },
    "idea_sweep_same_union": [
      {
        "cfg": {
          "lr": 0.001
        },
        "mean": 0.0003716218598128762
      },
      {
        "cfg": {
          "lr": 0.003
        },
        "mean": 0.0001124112732213689
      },
      {
        "cfg": {
          "lr": 0.01
        },
        "mean": 0.0001620453786017606
      }
    ]
  },
  "idea": {
    "mean": 0.00010261458010063507,
    "std": 2.643143893782112e-05,
    "per_seed": [
      7.763464964227751e-05,
      0.0001378427550662309,
      9.273155592381954e-05,
      0.00014143613225314766,
      7.422755152219906e-05,
      7.65531585784629e-05,
      9.41569305723533e-05,
      0.00012633390724658966
    ],
    "n": 8
  },
  "comparison": {
    "delta_mean": -3.5719546758627985e-05,
    "idea_wins": 4,
    "n_pairs": 8,
    "per_seed_diffs": [
      -0.00019073631119681522,
      -0.0001437072060070932,
      2.109647903125733e-05,
      4.288836498744786e-05,
      -2.727928222157061e-05,
      2.436192153254524e-05,
      4.0742997953202575e-05,
      -5.3123338147997856e-05
    ],
    "p_value": 0.3492,
    "mde": 7.406941029805223e-05,
    "mde_rel_pct": 53.543844877416404,
    "verdict": "no significant win",
    "system_worked": false
  },
  "mechanism_signature": {
    "prediction_metric": "test MSE under independent lognormal recurrent gain perturbations",
    "baseline_gain_mse": {
      "clean_mse": 1.45693838596344,
      "0.05": 1.4605102241039276,
      "0.15": 1.4719401597976685,
      "0.3": 1.5009662508964539
    },
    "idea_gain_mse": {
      "clean_mse": 1.4612292051315308,
      "0.05": 1.46112659573555,
      "0.15": 1.461238443851471,
      "0.3": 1.4622179567813873
    },
    "predicted_effect": "cyclic loop should preserve recurrent response under small independent gains",
    "observed_effect": "compare perturbed-vs-clean MSE ratios on trained dynamics models",
    "confirmed": true
  },
  "selected_idea_cfg": {
    "lr": 0.003
  },
  "protocol_notes": {
    "structural_match": "dynamics: controlled pendulum rollout and recurrent stability",
    "paired_seeds": [
      0,
      1,
      2,
      3,
      4,
      5,
      6,
      7
    ],
    "shared_architecture": "GRU(3,64)+linear head; only hidden-to-hidden topology differs",
    "budget": {
      "epochs": 12,
      "batch": 128,
      "n_train": 1200,
      "n_test": 300
    }
  }
}