{
  "bench_version": 1,
  "track": "dynamics",
  "model": "rnn_small",
  "metric_direction": "lower is better",
  "n_seeds": 8,
  "baseline": {
    "best_cfg": {
      "lr": 0.01,
      "epochs": 10
    },
    "sweep": [
      {
        "cfg": {
          "lr": 0.001,
          "epochs": 10
        },
        "mean": 0.02325876720715314,
        "std": 0.009097045470607702,
        "per_seed": [
          0.006040765903890133,
          0.03010248765349388,
          0.02543320134282112,
          0.024085573852062225,
          0.010201698169112206,
          0.030222969129681587,
          0.031349945813417435,
          0.028633495792746544
        ],
        "n": 8
      },
      {
        "cfg": {
          "lr": 0.003,
          "epochs": 10
        },
        "mean": 0.003875804533890914,
        "std": 0.0019874791140290942,
        "per_seed": [
          0.0009708286379463971,
          0.003958100453019142,
          0.004776353016495705,
          0.005382170435041189,
          0.0012958660954609513,
          0.002271931618452072,
          0.006541829090565443,
          0.005809356924146414
        ],
        "n": 8
      },
      {
        "cfg": {
          "lr": 0.01,
          "epochs": 10
        },
        "mean": 0.002775895278318785,
        "std": 0.0017084856346187775,
        "per_seed": [
          0.0014615225372835994,
          0.004334237892180681,
          0.0012712624156847596,
          0.006652707699686289,
          0.0020110509358346462,
          0.0024756770581007004,
          0.0018914247630164027,
          0.0021092789247632027
        ],
        "n": 8
      }
    ],
    "harness_tuning": {
      "best_cfg": {
        "lr": 0.01,
        "epochs": 10
      },
      "sweep": [
        {
          "cfg": {
            "lr": 0.001,
            "epochs": 10
          },
          "mean": 0.02141550718806684
        },
        {
          "cfg": {
            "lr": 0.003,
            "epochs": 10
          },
          "mean": 0.0037718631356256083
        },
        {
          "cfg": {
            "lr": 0.01,
            "epochs": 10
          },
          "mean": 0.0034299326362088323
        }
      ],
      "full": {
        "mean": 0.002775895278318785,
        "std": 0.0017084856346187775,
        "per_seed": [
          0.0014615225372835994,
          0.004334237892180681,
          0.0012712624156847596,
          0.006652707699686289,
          0.0020110509358346462,
          0.0024756770581007004,
          0.0018914247630164027,
          0.0021092789247632027
        ],
        "n": 8
      }
    },
    "full": {
      "mean": 0.002775895278318785,
      "std": 0.0017084856346187775,
      "per_seed": [
        0.0014615225372835994,
        0.004334237892180681,
        0.0012712624156847596,
        0.006652707699686289,
        0.0020110509358346462,
        0.0024756770581007004,
        0.0018914247630164027,
        0.0021092789247632027
      ],
      "n": 8
    }
  },
  "idea": {
    "per_seed": [
      0.0018240498611703515,
      0.0038136225193738937,
      0.001709936186671257,
      0.005960390437394381,
      0.001420309068635106,
      0.0025334814563393593,
      0.002282342640683055,
      0.0019062625942751765
    ],
    "mean": 0.0026812993455678225,
    "std": 0.0014179568156557768,
    "n": 8
  },
  "comparison": {
    "delta_mean": -9.459593275096267e-05,
    "idea_wins": 4,
    "n_pairs": 8,
    "per_seed_diffs": [
      0.00036252732388675213,
      -0.0005206153728067875,
      0.0004386737709864974,
      -0.0006923172622919083,
      -0.0005907418671995401,
      5.7804398238658905e-05,
      0.0003909178776666522,
      -0.00020301633048802614
    ],
    "p_value": 0.58435,
    "mde": 0.00039333102068697407,
    "mde_rel_pct": 14.16951942528589,
    "verdict": "no significant win",
    "system_worked": false
  },
  "mechanism_signature": {
    "idea_sweep": [
      {
        "cfg": {
          "lr": 0.001,
          "epochs": 10,
          "shield_lambda": 0.01
        },
        "mean": 0.022998466796707362,
        "std": 0.009109715346858763,
        "per_seed": [
          0.005741507280617952,
          0.029806293547153473,
          0.02514277957379818,
          0.02393505349755287,
          0.00990633387118578,
          0.029928065836429596,
          0.031049689278006554,
          0.02847801148891449
        ],
        "n": 8
      },
      {
        "cfg": {
          "lr": 0.003,
          "epochs": 10,
          "shield_lambda": 0.03
        },
        "mean": 0.0036644875799538568,
        "std": 0.001954599449464149,
        "per_seed": [
          0.0011084122816100717,
          0.0033393101766705513,
          0.005045472178608179,
          0.0046833050437271595,
          0.0011171806836500764,
          0.0019260895205661654,
          0.006277866195887327,
          0.0058182645589113235
        ],
        "n": 8
      },
      {
        "cfg": {
          "lr": 0.01,
          "epochs": 10,
          "shield_lambda": 0.06
        },
        "mean": 0.0026812993455678225,
        "std": 0.0014179568156557768,
        "per_seed": [
          0.0018240498611703515,
          0.0038136225193738937,
          0.001709936186671257,
          0.005960390437394381,
          0.001420309068635106,
          0.0025334814563393593,
          0.002282342640683055,
          0.0019062625942751765
        ],
        "n": 8
      }
    ],
    "mechanism_signature": {
      "prediction": "online residual zonotope contracts after burn-in while held-out residual coverage remains high",
      "trained_test_mse": 0.0011084122816100717,
      "observed_residual_center": 1.96993350982666e-05,
      "observed_residual_radius_q": 3.6635516929626464,
      "heldout_residual_coverage": 1.0,
      "unsafe_point_prediction_rate": 0.05999999865889549,
      "unsafe_reachable_interval_rate": 1.0,
      "confirmed": false
    }
  },
  "math_check": {
    "cases": 1000,
    "containment_disagreements": 0,
    "max_sample_support_gap": 0.0
  },
  "protocol_note": "Eight paired seeds; baseline and idea share rnn_small, data, Adam, epochs, batch, and union learning rates. Lower test MSE is primary."
}