{
  "bench_version": 1,
  "track": "tabular",
  "model": "mlp_tiny",
  "metric_direction": "lower is better",
  "n_seeds": 8,
  "baseline": {
    "best_cfg": {
      "lr": 0.008
    },
    "sweep": [
      {
        "cfg": {
          "lr": 0.001
        },
        "mean": 61.77081108093262
      },
      {
        "cfg": {
          "lr": 0.002
        },
        "mean": 17.307271003723145
      },
      {
        "cfg": {
          "lr": 0.003
        },
        "mean": 15.343324422836304
      },
      {
        "cfg": {
          "lr": 0.005
        },
        "mean": 12.93674373626709
      },
      {
        "cfg": {
          "lr": 0.008
        },
        "mean": 10.50437068939209
      }
    ],
    "full": {
      "mean": 10.8468017578125,
      "std": 0.9460715154353679,
      "per_seed": [
        10.433314323425293,
        9.383343696594238,
        12.219518661499023,
        9.981306076049805,
        10.63367748260498,
        10.9376859664917,
        10.86903190612793,
        12.316535949707031
      ],
      "n": 8
    }
  },
  "idea": {
    "mean": 8.208642661571503,
    "std": 0.5742573828718772,
    "per_seed": [
      7.412168025970459,
      7.734030246734619,
      8.057981491088867,
      7.578333377838135,
      8.25620174407959,
      8.907169342041016,
      8.773388862609863,
      8.949868202209473
    ],
    "n": 8
  },
  "comparison": {
    "delta_mean": -2.6381590962409973,
    "idea_wins": 8,
    "n_pairs": 8,
    "per_seed_diffs": [
      -3.021146297454834,
      -1.6493134498596191,
      -4.161537170410156,
      -2.40297269821167,
      -2.3774757385253906,
      -2.0305166244506836,
      -2.0956430435180664,
      -3.3666677474975586
    ],
    "p_value": 0.0081,
    "mde": 0.6897071822966465,
    "mde_rel_pct": 6.3586225478849485,
    "verdict": "idea better (significant)",
    "system_worked": true
  },
  "track_choice": "tabular: the idea is an optimizer/black-box hyperparameter infill method; no architecture or sequence/control structure is required.",
  "shared_lr_union": [
    0.001,
    0.002,
    0.003,
    0.005,
    0.008
  ],
  "epochs": 12,
  "n_train": 800,
  "mechanism_signature": {
    "pilot_validation_mean": 7.560150101780891,
    "gn_predicted_validation_mean": 8.995661415256667,
    "gn_observed_validation_mean": 8.700061082839966,
    "absolute_prediction_error": 0.5079499775828902,
    "condition_estimate": 1.0,
    "confirmed": true
  },
  "math_check": {
    "exact_linear_jacobian_relative_error": 0.8606629657653821
  }
}