{
  "bench_version": 1,
  "track": "tabular",
  "model": "mlp_med",
  "metric_direction": "lower is better",
  "n_seeds": 8,
  "baseline": {
    "best_cfg": {
      "lr": 0.006,
      "beta": 0.002
    },
    "sweep": [
      {
        "cfg": {
          "lr": 0.0015,
          "beta": 0.0
        },
        "mean": 7.332403898239136
      },
      {
        "cfg": {
          "lr": 0.0015,
          "beta": 0.002
        },
        "mean": 6.974947094917297
      },
      {
        "cfg": {
          "lr": 0.0015,
          "beta": 0.005
        },
        "mean": 6.950373411178589
      },
      {
        "cfg": {
          "lr": 0.0015,
          "beta": 0.01
        },
        "mean": 6.856605052947998
      },
      {
        "cfg": {
          "lr": 0.0015,
          "beta": 0.02
        },
        "mean": 6.820917129516602
      },
      {
        "cfg": {
          "lr": 0.003,
          "beta": 0.0
        },
        "mean": 6.740176439285278
      },
      {
        "cfg": {
          "lr": 0.003,
          "beta": 0.002
        },
        "mean": 6.5683335065841675
      },
      {
        "cfg": {
          "lr": 0.003,
          "beta": 0.005
        },
        "mean": 6.506619572639465
      },
      {
        "cfg": {
          "lr": 0.003,
          "beta": 0.01
        },
        "mean": 6.451507329940796
      },
      {
        "cfg": {
          "lr": 0.003,
          "beta": 0.02
        },
        "mean": 6.507548213005066
      },
      {
        "cfg": {
          "lr": 0.006,
          "beta": 0.0
        },
        "mean": 6.459080219268799
      },
      {
        "cfg": {
          "lr": 0.006,
          "beta": 0.002
        },
        "mean": 6.255969762802124
      },
      {
        "cfg": {
          "lr": 0.006,
          "beta": 0.005
        },
        "mean": 6.3951462507247925
      },
      {
        "cfg": {
          "lr": 0.006,
          "beta": 0.01
        },
        "mean": 6.444561243057251
      },
      {
        "cfg": {
          "lr": 0.006,
          "beta": 0.02
        },
        "mean": 6.318377375602722
      }
    ],
    "full": {
      "mean": 6.492175281047821,
      "std": 0.4696585684033177,
      "per_seed": [
        5.431238651275635,
        6.641201019287109,
        6.379562854766846,
        6.571876525878906,
        6.9486002922058105,
        6.291627407073975,
        6.600069522857666,
        7.073225975036621
      ],
      "n": 8
    }
  },
  "idea": {
    "mean": 6.48457407951355,
    "std": 0.4566508873984473,
    "per_seed": [
      5.4823994636535645,
      6.262659549713135,
      6.524853229522705,
      6.588247776031494,
      6.916043758392334,
      6.340798377990723,
      6.678991317749023,
      7.08259916305542
    ],
    "n": 8,
    "best_cfg": {
      "lr": 0.006,
      "beta_start": 0.005,
      "beta_end": 5e-05
    }
  },
  "comparison": {
    "delta_mean": -0.00760120153427124,
    "idea_wins": 2,
    "n_pairs": 8,
    "per_seed_diffs": [
      0.05116081237792969,
      -0.3785414695739746,
      0.14529037475585938,
      0.01637125015258789,
      -0.03255653381347656,
      0.04917097091674805,
      0.07892179489135742,
      0.009373188018798828
    ],
    "p_value": 0.96975,
    "mde": 0.1328181829200986,
    "mde_rel_pct": 2.045819423696491,
    "verdict": "no measurable effect",
    "system_worked": false
  },
  "math_check": {
    "predicted_exponent": 1.3333333333333333,
    "observed_exponent": 1.3333333333333335,
    "median_relative_boundary_error": 0.30000000000000004,
    "pass": false
  },
  "idea_sweep": {
    "configs": [
      {
        "lr": 0.0015,
        "beta_start": 0.02,
        "beta_end": 0.0002
      },
      {
        "lr": 0.003,
        "beta_start": 0.01,
        "beta_end": 0.0001
      },
      {
        "lr": 0.006,
        "beta_start": 0.005,
        "beta_end": 5e-05
      }
    ],
    "selected": {
      "lr": 0.006,
      "beta_start": 0.005,
      "beta_end": 5e-05
    }
  },
  "budget": {
    "epochs": 12,
    "batch": 128,
    "train_samples": 1200
  },
  "mechanism_signature": {
    "kind": "trained_end_to_end_spectrum_vs_crosscov",
    "measurements": [
      {
        "epoch": 1,
        "beta": 0.005,
        "teacher_crosscov_singular_values": [
          1.3328860998153687
        ],
        "end_to_end_singular_values": [
          8.559873580932617
        ],
        "predicted_active_count": 1,
        "observed_active_count": 1
      },
      {
        "epoch": 7,
        "beta": 0.0004055654153948436,
        "teacher_crosscov_singular_values": [
          1.3328860998153687
        ],
        "end_to_end_singular_values": [
          18.78715705871582
        ],
        "predicted_active_count": 1,
        "observed_active_count": 1
      },
      {
        "epoch": 12,
        "beta": 5e-05,
        "teacher_crosscov_singular_values": [
          1.3328860998153687
        ],
        "end_to_end_singular_values": [
          21.196590423583984
        ],
        "predicted_active_count": 1,
        "observed_active_count": 1
      }
    ],
    "confirmed": true,
    "note": "Signature is measured from trained MLP weights and benchmark data; it is not an analytic identity."
  }
}