{
  "bench_version": 1,
  "track": "dynamics",
  "model": "rnn_small",
  "metric_direction": "lower is better",
  "n_seeds": 8,
  "baseline": {
    "best_cfg": {
      "lr": 0.006
    },
    "sweep": [
      {
        "cfg": {
          "lr": 0.001
        },
        "mean": 0.0014035784406587481
      },
      {
        "cfg": {
          "lr": 0.003
        },
        "mean": 0.000857145045301877
      },
      {
        "cfg": {
          "lr": 0.006
        },
        "mean": 0.0005846045824000612
      }
    ],
    "full": {
      "mean": 0.0006132683593023103,
      "std": 0.00012424553172221234,
      "per_seed": [
        0.0005031590699218214,
        0.0007117226487025619,
        0.0004876864841207862,
        0.0006358501268550754,
        0.0004713028611149639,
        0.0005687447264790535,
        0.0006691908929497004,
        0.0008584900642745197
      ],
      "n": 8
    }
  },
  "idea": {
    "mean": 0.0004085009823029395,
    "std": 0.00011041109434057678,
    "per_seed": [
      0.00036431028274819255,
      0.00038561588735319674,
      0.00044469136628322303,
      0.0003898605646099895,
      0.00032400351483374834,
      0.000476225366583094,
      0.00024231788120232522,
      0.0006409829948097467
    ],
    "n": 8
  },
  "comparison": {
    "delta_mean": -0.00020476737699937075,
    "idea_wins": 8,
    "n_pairs": 8,
    "per_seed_diffs": [
      -0.0001388487871736288,
      -0.0003261067613493651,
      -4.299511783756316e-05,
      -0.00024598956224508584,
      -0.00014729934628121555,
      -9.25193598959595e-05,
      -0.00042687301174737513,
      -0.00021750706946477294
    ],
    "p_value": 0.0081,
    "mde": 0.00010592334035123618,
    "mde_rel_pct": 17.27193955868532,
    "verdict": "idea better (significant)",
    "system_worked": true
  },
  "mechanism_signature": {
    "alpha": 0.1,
    "calibration_q_mean": 0.13871399220079184,
    "observed_test_abs_residual_mean": 0.06220734538510442,
    "predicted_coverage": 0.9,
    "observed_calibration_fraction_below_q": 1.0,
    "confirmed": false,
    "note": "trained-model residual signature; marginal coverage is not directly testable without deployment labels"
  },
  "idea_sweep": [
    {
      "cfg": {
        "lr": 0.001
      },
      "mean": 0.0020116496889386326,
      "per_seed": [
        0.001484628184698522,
        0.0012722134124487638,
        0.001204720581881702,
        0.0028885509818792343,
        0.0018111122772097588,
        0.0019790560472756624,
        0.002846983028575778,
        0.0026059329975396395
      ]
    },
    {
      "cfg": {
        "lr": 0.003
      },
      "mean": 0.0007203747300081886,
      "per_seed": [
        0.0005618855939246714,
        0.0005979266134090722,
        0.0006078000296838582,
        0.0009234340395778418,
        0.000840717984829098,
        0.0006266810232773423,
        0.0007128350553102791,
        0.0008917175000533462
      ]
    },
    {
      "cfg": {
        "lr": 0.006
      },
      "mean": 0.0004085009823029395,
      "per_seed": [
        0.00036431028274819255,
        0.00038561588735319674,
        0.00044469136628322303,
        0.0003898605646099895,
        0.00032400351483374834,
        0.000476225366583094,
        0.00024231788120232522,
        0.0006409829948097467
      ]
    }
  ],
  "protocol_notes": {
    "matched_structure": "actuated pendulum rollout prediction",
    "intervention": "conformal-calibrated residual downweighting",
    "baseline_union_lr": [
      0.001,
      0.003,
      0.006
    ],
    "idea_union_lr": [
      0.001,
      0.003,
      0.006
    ],
    "n_seeds": 8
  }
}