{
  "bench_version": 1,
  "track": "tabular",
  "model": "mlp_tiny",
  "metric_direction": "lower is better",
  "n_seeds": 8,
  "baseline": {
    "best_cfg": {
      "lr": 0.01,
      "ratio": [
        100.0,
        10.0,
        1.0
      ]
    },
    "sweep": [
      {
        "cfg": {
          "lr": 0.001,
          "ratio": [
            1.0,
            1.0,
            1.0
          ]
        },
        "mean": 128.7099723815918
      },
      {
        "cfg": {
          "lr": 0.001,
          "ratio": [
            10.0,
            3.0,
            1.0
          ]
        },
        "mean": 128.4849090576172
      },
      {
        "cfg": {
          "lr": 0.001,
          "ratio": [
            100.0,
            10.0,
            1.0
          ]
        },
        "mean": 127.46754837036133
      },
      {
        "cfg": {
          "lr": 0.003,
          "ratio": [
            1.0,
            1.0,
            1.0
          ]
        },
        "mean": 16.73896622657776
      },
      {
        "cfg": {
          "lr": 0.003,
          "ratio": [
            10.0,
            3.0,
            1.0
          ]
        },
        "mean": 16.98913860321045
      },
      {
        "cfg": {
          "lr": 0.003,
          "ratio": [
            100.0,
            10.0,
            1.0
          ]
        },
        "mean": 17.95883321762085
      },
      {
        "cfg": {
          "lr": 0.01,
          "ratio": [
            1.0,
            1.0,
            1.0
          ]
        },
        "mean": 11.914008855819702
      },
      {
        "cfg": {
          "lr": 0.01,
          "ratio": [
            10.0,
            3.0,
            1.0
          ]
        },
        "mean": 11.700665712356567
      },
      {
        "cfg": {
          "lr": 0.01,
          "ratio": [
            100.0,
            10.0,
            1.0
          ]
        },
        "mean": 10.906596660614014
      }
    ],
    "full": {
      "mean": 11.41922926902771,
      "std": 1.3400197166712482,
      "per_seed": [
        11.151049613952637,
        10.485822677612305,
        10.90726375579834,
        11.082250595092773,
        9.644949913024902,
        11.029101371765137,
        12.869707107543945,
        14.18368911743164
      ],
      "n": 8
    }
  },
  "idea": {
    "mean": 11.052955508232117,
    "std": 1.160223574491651,
    "per_seed": [
      10.566213607788086,
      10.391647338867188,
      10.485791206359863,
      11.461767196655273,
      9.901324272155762,
      9.85105037689209,
      13.145121574401855,
      12.620728492736816
    ],
    "n": 8
  },
  "comparison": {
    "delta_mean": -0.36627376079559326,
    "idea_wins": 5,
    "n_pairs": 8,
    "per_seed_diffs": [
      -0.5848360061645508,
      -0.09417533874511719,
      -0.42147254943847656,
      0.3795166015625,
      0.2563743591308594,
      -1.1780509948730469,
      0.27541446685791016,
      -1.5629606246948242
    ],
    "p_value": 0.2116,
    "mde": 0.5972631267868054,
    "mde_rel_pct": 5.230327833129313,
    "verdict": "no significant win",
    "system_worked": false
  },
  "mechanism_signature": {
    "mechanism_signature": {
      "quantity": "parameter-gradient infinity stationarity residual at refreshes",
      "predicted_epsilon": 0.02,
      "observed_mean_residual": Infinity,
      "positive_margin_frequency": 0.0,
      "confirmed": false
    },
    "math_check": {
      "predicted_weights": [
        3.0,
        1.0
      ],
      "observed_weights": [
        2.999999999,
        1.000000001
      ],
      "residual": 1.000000082740371e-09,
      "passed": true
    },
    "idea_sweep": [
      {
        "cfg": {
          "lr": 0.001,
          "ratio": [
            1.0,
            1.0,
            1.0
          ]
        },
        "result": {
          "mean": 127.88450622558594,
          "std": 14.343906975103058,
          "per_seed": [
            142.4267578125,
            132.91229248046875,
            121.2231216430664,
            118.38529205322266,
            105.10956573486328,
            127.8440170288086,
            120.74082946777344,
            154.43417358398438
          ],
          "n": 8
        }
      },
      {
        "cfg": {
          "lr": 0.001,
          "ratio": [
            10.0,
            3.0,
            1.0
          ]
        },
        "result": {
          "mean": 127.65959358215332,
          "std": 14.388813821417166,
          "per_seed": [
            142.22030639648438,
            132.71397399902344,
            120.96118927001953,
            118.2004623413086,
            104.7640380859375,
            127.63045501708984,
            120.5007553100586,
            154.2855682373047
          ],
          "n": 8
        }
      },
      {
        "cfg": {
          "lr": 0.001,
          "ratio": [
            100.0,
            10.0,
            1.0
          ]
        },
        "result": {
          "mean": 126.63898754119873,
          "std": 14.61229098741295,
          "per_seed": [
            141.26730346679688,
            131.85458374023438,
            119.78008270263672,
            117.36537170410156,
            103.09697723388672,
            126.7028579711914,
            119.42959594726562,
            153.61512756347656
          ],
          "n": 8
        }
      },
      {
        "cfg": {
          "lr": 0.003,
          "ratio": [
            1.0,
            1.0,
            1.0
          ]
        },
        "result": {
          "mean": 17.716270089149475,
          "std": 2.1854981702675587,
          "per_seed": [
            17.7928466796875,
            14.453100204467773,
            20.243223190307617,
            15.165732383728027,
            18.23601722717285,
            15.786874771118164,
            19.99270248413086,
            20.059663772583008
          ],
          "n": 8
        }
      },
      {
        "cfg": {
          "lr": 0.003,
          "ratio": [
            10.0,
            3.0,
            1.0
          ]
        },
        "result": {
          "mean": 17.921056747436523,
          "std": 2.208602464634322,
          "per_seed": [
            18.224416732788086,
            14.801833152770996,
            20.551616668701172,
            15.228721618652344,
            18.417009353637695,
            15.783102989196777,
            20.219045639038086,
            20.14270782470703
          ],
          "n": 8
        }
      },
      {
        "cfg": {
          "lr": 0.003,
          "ratio": [
            100.0,
            10.0,
            1.0
          ]
        },
        "result": {
          "mean": 18.67002558708191,
          "std": 2.200020788096735,
          "per_seed": [
            19.961009979248047,
            16.091012954711914,
            21.057720184326172,
            15.669567108154297,
            18.654521942138672,
            16.20878791809082,
            21.031444549560547,
            20.686140060424805
          ],
          "n": 8
        }
      },
      {
        "cfg": {
          "lr": 0.01,
          "ratio": [
            1.0,
            1.0,
            1.0
          ]
        },
        "result": {
          "mean": 11.539970397949219,
          "std": 1.112516442321061,
          "per_seed": [
            10.973492622375488,
            9.973017692565918,
            10.890155792236328,
            11.609981536865234,
            10.801332473754883,
            11.49035358428955,
            13.19100284576416,
            13.390426635742188
          ],
          "n": 8
        }
      },
      {
        "cfg": {
          "lr": 0.01,
          "ratio": [
            10.0,
            3.0,
            1.0
          ]
        },
        "result": {
          "mean": 11.519728541374207,
          "std": 0.9986832975547216,
          "per_seed": [
            10.72226333618164,
            10.168411254882812,
            11.434860229492188,
            11.341761589050293,
            11.246848106384277,
            11.001540184020996,
            13.17798900604248,
            13.064154624938965
          ],
          "n": 8
        }
      },
      {
        "cfg": {
          "lr": 0.01,
          "ratio": [
            100.0,
            10.0,
            1.0
          ]
        },
        "result": {
          "mean": 11.052955508232117,
          "std": 1.160223574491651,
          "per_seed": [
            10.566213607788086,
            10.391647338867188,
            10.485791206359863,
            11.461767196655273,
            9.901324272155762,
            9.85105037689209,
            13.145121574401855,
            12.620728492736816
          ],
          "n": 8
        }
      }
    ],
    "protocol_notes": "Matched tabular Friedman regression and mlp_tiny; baseline and controller share all lr and ratio configurations, Adam, epochs, batch, and data seeds."
  }
}