{
  "bench_version": 1,
  "track": "tabular",
  "model": "mlp_tiny",
  "metric_direction": "lower is better",
  "n_seeds": 8,
  "baseline": {
    "best_cfg": {
      "lr": 0.006
    },
    "sweep": [
      {
        "cfg": {
          "lr": 0.001
        },
        "mean": 17.482568740844727
      },
      {
        "cfg": {
          "lr": 0.003
        },
        "mean": 12.535736083984375
      },
      {
        "cfg": {
          "lr": 0.006
        },
        "mean": 9.091599702835083
      }
    ],
    "full": {
      "mean": 9.481420636177063,
      "std": 0.9072998646400672,
      "per_seed": [
        8.627572059631348,
        8.736048698425293,
        8.82056999206543,
        10.182208061218262,
        8.84079647064209,
        9.23827075958252,
        11.380922317504883,
        10.02497673034668
      ],
      "n": 8
    }
  },
  "idea": {
    "mean": 9.610056400299072,
    "std": 1.0356528912793441,
    "per_seed": [
      9.995634078979492,
      8.813212394714355,
      8.717215538024902,
      9.493178367614746,
      8.498827934265137,
      9.12738037109375,
      11.815202713012695,
      10.4197998046875
    ],
    "n": 8
  },
  "comparison": {
    "delta_mean": 0.12863576412200928,
    "idea_wins": 4,
    "n_pairs": 8,
    "per_seed_diffs": [
      1.3680620193481445,
      0.0771636962890625,
      -0.10335445404052734,
      -0.6890296936035156,
      -0.3419685363769531,
      -0.11089038848876953,
      0.4342803955078125,
      0.3948230743408203
    ],
    "p_value": 0.61945,
    "mde": 0.5197178870759055,
    "mde_rel_pct": 5.481434766145522,
    "verdict": "no significant win",
    "system_worked": false
  },
  "mechanism_signature": {
    "signature": {
      "quantity": "family allocation variance proxy sum(u_c^2/n_c)",
      "predicted_mean": 0.0982159124687314,
      "observed_mean": 0.09136932715773582,
      "relative_error": 0.06970953218171544,
      "confirmed": true,
      "note": "u and n are computed from trained model residuals; observed proxy uses held-out family residual variances."
    },
    "idea_sweep": [
      {
        "cfg": {
          "lr": 0.001,
          "epsilon": 0.05
        },
        "result": {
          "mean": 18.78651201725006,
          "std": 2.1210090381527644,
          "per_seed": [
            17.013912200927734,
            15.361128807067871,
            21.161161422729492,
            16.918540954589844,
            19.89978790283203,
            17.97165870666504,
            20.517269134521484,
            21.448637008666992
          ],
          "n": 8
        }
      },
      {
        "cfg": {
          "lr": 0.003,
          "epsilon": 0.15
        },
        "result": {
          "mean": 13.162447214126587,
          "std": 1.3889931613668718,
          "per_seed": [
            13.541401863098145,
            11.67493724822998,
            12.949481964111328,
            12.458735466003418,
            11.737956047058105,
            12.227039337158203,
            15.25829792022705,
            15.451727867126465
          ],
          "n": 8
        }
      },
      {
        "cfg": {
          "lr": 0.006,
          "epsilon": 0.3
        },
        "result": {
          "mean": 9.610056400299072,
          "std": 1.0356528912793441,
          "per_seed": [
            9.995634078979492,
            8.813212394714355,
            8.717215538024902,
            9.493178367614746,
            8.498827934265137,
            9.12738037109375,
            11.815202713012695,
            10.4197998046875
          ],
          "n": 8
        }
      }
    ],
    "track_rationale": "Family-conditioned minibatch selection is an optimizer/training-data mechanism; tabular regression is the mandated structural track for optimizer/data mechanics."
  },
  "stage2_protocol": {
    "paired_seeds": [
      0,
      1,
      2,
      3,
      4,
      5,
      6,
      7
    ],
    "epochs": 12,
    "batch": 64,
    "baseline_grid": [
      {
        "lr": 0.001
      },
      {
        "lr": 0.003
      },
      {
        "lr": 0.006
      }
    ],
    "idea_grid": [
      {
        "lr": 0.001,
        "epsilon": 0.05
      },
      {
        "lr": 0.003,
        "epsilon": 0.15
      },
      {
        "lr": 0.006,
        "epsilon": 0.3
      }
    ]
  }
}