{
  "bench_version": 1,
  "track": "tabular",
  "model": "mlp_tiny",
  "metric_direction": "lower is better",
  "n_seeds": 8,
  "baseline": {
    "best_cfg": {
      "lr": 0.001
    },
    "sweep": [
      {
        "cfg": {
          "lr": 0.001
        },
        "mean": 211.45939254760742
      },
      {
        "cfg": {
          "lr": 0.003
        },
        "mean": 211.45939254760742
      },
      {
        "cfg": {
          "lr": 0.01
        },
        "mean": 211.45939254760742
      }
    ],
    "full": {
      "mean": 210.28362846374512,
      "std": 5.264942297259805,
      "per_seed": [
        216.2948760986328,
        210.3721466064453,
        217.16351318359375,
        202.0070343017578,
        208.42745971679688,
        212.19456481933594,
        202.8369903564453,
        212.97244262695312
      ],
      "n": 8
    }
  },
  "idea": {
    "mean": 210.9054012298584,
    "std": 5.309931146588209,
    "per_seed": [
      216.95388793945312,
      210.90757751464844,
      217.41943359375,
      202.13302612304688,
      210.2574005126953,
      212.32080078125,
      203.2976531982422,
      213.95343017578125
    ],
    "n": 8,
    "config": {
      "lr": 0.001
    }
  },
  "comparison": {
    "delta_mean": 0.6217727661132812,
    "idea_wins": 0,
    "n_pairs": 8,
    "per_seed_diffs": [
      0.6590118408203125,
      0.535430908203125,
      0.25592041015625,
      0.1259918212890625,
      1.8299407958984375,
      0.1262359619140625,
      0.460662841796875,
      0.980987548828125
    ],
    "p_value": 0.0081,
    "mde": 0.47391836886435124,
    "mde_rel_pct": 0.2253710249944918,
    "verdict": "idea worse (significant)",
    "system_worked": false
  },
  "mechanism_signature": {
    "predicted_transition_ess": 4.0,
    "observed_ess_at_closest_hc": 4.193504810333252,
    "predicted_high_entropy_rate": 0.05,
    "observed_high_entropy_rate": 0.049303146592420355,
    "predicted_low_entropy_rate": 0.001,
    "observed_low_entropy_rate": 0.0011907811681947413,
    "mean_entropy": 0.571972486252586,
    "confirmed": true
  },
  "idea_nearby_settings": {
    "0.001": [
      216.95388793945312,
      210.90757751464844,
      217.41943359375,
      202.13302612304688,
      210.2574005126953,
      212.32080078125,
      203.2976531982422,
      213.95343017578125
    ],
    "0.003": [
      216.95388793945312,
      210.90757751464844,
      217.41943359375,
      202.13302612304688,
      210.2574005126953,
      212.32080078125,
      203.2976531982422,
      213.95343017578125
    ],
    "0.01": [
      216.95388793945312,
      210.90757751464844,
      217.41943359375,
      202.13302612304688,
      210.2574005126953,
      212.32080078125,
      203.2976531982422,
      213.95343017578125
    ]
  },
  "protocol_notes": "Optimizer ideas are matched to tabular Friedman regression; both systems use identical mlp_tiny, population, perturbation scale, steps, data, and paired seeds. Baseline fixed exponential cooling was swept over the shared lr union.",
  "custom_track": null
}