{
  "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": 0.0014785313687752932
      },
      {
        "cfg": {
          "lr": 0.001,
          "weight_decay": 0.0001
        },
        "mean": 0.0014645301271229982
      },
      {
        "cfg": {
          "lr": 0.003,
          "weight_decay": 0.0
        },
        "mean": 0.0008221298194257542
      },
      {
        "cfg": {
          "lr": 0.003,
          "weight_decay": 0.0001
        },
        "mean": 0.0008076394879026338
      },
      {
        "cfg": {
          "lr": 0.01,
          "weight_decay": 0.0
        },
        "mean": 0.00032767555967438966
      },
      {
        "cfg": {
          "lr": 0.01,
          "weight_decay": 0.0001
        },
        "mean": 0.00036696675670100376
      }
    ],
    "full": {
      "mean": 0.0002959608482342446,
      "std": 0.0001316782688333712,
      "per_seed": [
        0.0004682304570451379,
        0.0003980887704528868,
        0.00017163399024866521,
        0.0002727490209508687,
        0.00015086510393302888,
        0.00030392719781957567,
        0.00012582281487993896,
        0.00047636943054385483
      ],
      "n": 8
    }
  },
  "idea": {
    "mean": 0.001754267912474461,
    "std": 0.00043015553066540066,
    "per_seed": [
      0.0012427795445546508,
      0.001973426202312112,
      0.0013205275172367692,
      0.0024030047934502363,
      0.002149099251255393,
      0.0012187783140689135,
      0.0020814703311771154,
      0.0016450573457404971
    ],
    "n": 8,
    "sweep": [
      {
        "cfg": {
          "lr": 0.001,
          "weight_decay": 0.0
        },
        "mean": 0.024796552257612348,
        "std": 0.004088271419227938,
        "per_seed": [
          0.022675156593322754,
          0.02744518220424652,
          0.023942390456795692,
          0.03346814215183258,
          0.025026222690939903,
          0.019205331802368164,
          0.021013041958212852,
          0.025596950203180313
        ]
      },
      {
        "cfg": {
          "lr": 0.001,
          "weight_decay": 0.0001
        },
        "mean": 0.024904589634388685,
        "std": 0.004089925380114752,
        "per_seed": [
          0.022724661976099014,
          0.02740549109876156,
          0.024043483659625053,
          0.03363170102238655,
          0.025220675393939018,
          0.01937870867550373,
          0.021087665110826492,
          0.025744330137968063
        ]
      },
      {
        "cfg": {
          "lr": 0.003,
          "weight_decay": 0.0
        },
        "mean": 0.006038541585439816,
        "std": 0.002237967748799504,
        "per_seed": [
          0.00579321151599288,
          0.006784805096685886,
          0.005182903725653887,
          0.011018984019756317,
          0.005261050537228584,
          0.0027950687799602747,
          0.00687775295227766,
          0.004594556055963039
        ]
      },
      {
        "cfg": {
          "lr": 0.003,
          "weight_decay": 0.0001
        },
        "mean": 0.006389912683516741,
        "std": 0.0023475474959639103,
        "per_seed": [
          0.006135317496955395,
          0.007102395873516798,
          0.005393301136791706,
          0.011719641275703907,
          0.005564150866121054,
          0.0029893447645008564,
          0.007032498717308044,
          0.005182651337236166
        ]
      },
      {
        "cfg": {
          "lr": 0.01,
          "weight_decay": 0.0
        },
        "mean": 0.001754267912474461,
        "std": 0.00043015553066540066,
        "per_seed": [
          0.0012427795445546508,
          0.001973426202312112,
          0.0013205275172367692,
          0.0024030047934502363,
          0.002149099251255393,
          0.0012187783140689135,
          0.0020814703311771154,
          0.0016450573457404971
        ]
      },
      {
        "cfg": {
          "lr": 0.01,
          "weight_decay": 0.0001
        },
        "mean": 0.0017732195847202092,
        "std": 0.0004718425245521136,
        "per_seed": [
          0.001313930144533515,
          0.002031040843576193,
          0.0011662384495139122,
          0.002488275757059455,
          0.0022803107276558876,
          0.0012248011771589518,
          0.002009331015869975,
          0.0016718285623937845
        ]
      }
    ]
  },
  "comparison": {
    "delta_mean": 0.0014583070642402163,
    "idea_wins": 0,
    "n_pairs": 8,
    "per_seed_diffs": [
      0.0007745490875095129,
      0.001575337431859225,
      0.001148893526988104,
      0.0021302557724993676,
      0.001998234147322364,
      0.0009148511162493378,
      0.0019556475162971765,
      0.0011686879151966423
    ],
    "p_value": 0.0081,
    "mde": 0.0004409319997531043,
    "mde_rel_pct": 148.98321936289327,
    "verdict": "idea worse (significant)",
    "system_worked": false
  },
  "mechanism_signature": {
    "basis_size": 25,
    "trained_model_observations": [
      {
        "seed": 0,
        "metric": 0.0012427795445546508,
        "mean_coeff_norm": 0.28018829226493835,
        "max_coeff_norm": 0.6576406955718994,
        "mass_error": 2.384185791015625e-07
      },
      {
        "seed": 1,
        "metric": 0.001973426202312112,
        "mean_coeff_norm": 0.30172738432884216,
        "max_coeff_norm": 0.6713038086891174,
        "mass_error": 2.384185791015625e-07
      },
      {
        "seed": 2,
        "metric": 0.0013205275172367692,
        "mean_coeff_norm": 0.31470319628715515,
        "max_coeff_norm": 0.7678529024124146,
        "mass_error": 1.7881393432617188e-07
      },
      {
        "seed": 3,
        "metric": 0.0024030047934502363,
        "mean_coeff_norm": 0.3005197048187256,
        "max_coeff_norm": 0.6498276591300964,
        "mass_error": 2.384185791015625e-07
      },
      {
        "seed": 4,
        "metric": 0.002149099251255393,
        "mean_coeff_norm": 0.304610013961792,
        "max_coeff_norm": 0.6628603935241699,
        "mass_error": 2.384185791015625e-07
      },
      {
        "seed": 5,
        "metric": 0.0012187783140689135,
        "mean_coeff_norm": 0.28839680552482605,
        "max_coeff_norm": 0.5926362872123718,
        "mass_error": 1.7881393432617188e-07
      },
      {
        "seed": 6,
        "metric": 0.0020814703311771154,
        "mean_coeff_norm": 0.30479809641838074,
        "max_coeff_norm": 0.6872127056121826,
        "mass_error": 2.384185791015625e-07
      },
      {
        "seed": 7,
        "metric": 0.0016450573457404971,
        "mean_coeff_norm": 0.288260281085968,
        "max_coeff_norm": 0.6056498885154724,
        "mass_error": 1.7881393432617188e-07
      }
    ],
    "predicted": "nonnegative unit-mass coefficients remain bounded",
    "observed_mean_mass_error": 2.1606683731079102e-07,
    "observed_max_coeff_norm": 0.7678529024124146,
    "confirmed": true,
    "math_sanity": {
      "mass_error": 2.220446049250313e-16,
      "rho": 1.0000000000000004,
      "observed_log_norm_slope": -3.223008136576186e-16,
      "predicted_log_rho": 4.440892098500625e-16,
      "slope_abs_error": 7.663900235076811e-16
    }
  },
  "protocol_notes": {
    "n_train": 800,
    "n_test": 400,
    "epochs": 12,
    "baseline_and_idea_lr_union": [
      0.001,
      0.003,
      0.01
    ],
    "paired_seeds": [
      0,
      1,
      2,
      3,
      4,
      5,
      6,
      7
    ],
    "architecture": "same 64-unit GRU encoder; scalar linear head vs PF RBF expectation head"
  }
}