{
  "bench_version": 1,
  "track": "tabular",
  "model": "mlp_tiny",
  "metric_direction": "lower is better",
  "n_seeds": 8,
  "baseline": {
    "best_cfg": {
      "lr": 0.01,
      "epochs": 1,
      "wd": 0.0001
    },
    "sweep": [
      {
        "cfg": {
          "lr": 0.001,
          "epochs": 1,
          "wd": 0.0
        },
        "mean": 236.9449577331543
      },
      {
        "cfg": {
          "lr": 0.001,
          "epochs": 1,
          "wd": 0.0001
        },
        "mean": 236.94502639770508
      },
      {
        "cfg": {
          "lr": 0.003,
          "epochs": 1,
          "wd": 0.0
        },
        "mean": 235.13161849975586
      },
      {
        "cfg": {
          "lr": 0.003,
          "epochs": 1,
          "wd": 0.0001
        },
        "mean": 235.1318359375
      },
      {
        "cfg": {
          "lr": 0.01,
          "epochs": 1,
          "wd": 0.0
        },
        "mean": 228.67240142822266
      },
      {
        "cfg": {
          "lr": 0.01,
          "epochs": 1,
          "wd": 0.0001
        },
        "mean": 228.67166900634766
      }
    ],
    "full": {
      "mean": 231.52675819396973,
      "std": 21.300660878391202,
      "per_seed": [
        256.47283935546875,
        229.34703063964844,
        246.06045532226562,
        182.8063507080078,
        247.07217407226562,
        224.0729522705078,
        240.7747039794922,
        225.60755920410156
      ],
      "n": 8
    }
  },
  "idea": {
    "mean": 489.4525787830353,
    "std": 991.4184272998954,
    "per_seed": [
      29.48086166381836,
      76.03532409667969,
      67.20081329345703,
      22.497228622436523,
      3082.7841796875,
      505.4776916503906,
      78.81880950927734,
      53.325721740722656
    ],
    "n": 8
  },
  "comparison": {
    "delta_mean": 257.92582058906555,
    "idea_wins": 6,
    "n_pairs": 8,
    "per_seed_diffs": [
      -226.9919776916504,
      -153.31170654296875,
      -178.8596420288086,
      -160.3091220855713,
      2835.7120056152344,
      281.4047393798828,
      -161.95589447021484,
      -172.2818374633789
    ],
    "p_value": 0.9404,
    "mde": 881.3323499422643,
    "mde_rel_pct": 380.66111961188386,
    "verdict": "no significant win",
    "system_worked": false
  },
  "mechanism_signature": {
    "prediction": "larger cross-block certificate predicts larger observed projected Fisher spectral distortion",
    "predicted_values": [
      1728.5964137448777,
      1842.2349990275347,
      580.0484882837169,
      667.8731875822858,
      3872.9491724735503,
      1046.0315549770417,
      1335.0321904102336,
      1300.5161090698366
    ],
    "observed_values": [
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0,
      0.0
    ],
    "correlation": 0.0,
    "confirmed": false
  },
  "idea_sweep": [
    {
      "cfg": {
        "lr": 0.001,
        "epochs": 1,
        "block": 64,
        "tau": 0.02,
        "damp": 0.01,
        "sketch": 3
      },
      "mean": 228.61152267456055
    },
    {
      "cfg": {
        "lr": 0.003,
        "epochs": 1,
        "block": 64,
        "tau": 0.02,
        "damp": 0.01,
        "sketch": 3
      },
      "mean": 200.8146095275879
    },
    {
      "cfg": {
        "lr": 0.01,
        "epochs": 1,
        "block": 64,
        "tau": 0.02,
        "damp": 0.01,
        "sketch": 3
      },
      "mean": 48.8035569190979
    }
  ],
  "chosen_idea_cfg": {
    "lr": 0.01,
    "epochs": 1,
    "block": 64,
    "tau": 0.02,
    "damp": 0.01,
    "sketch": 3
  },
  "protocol_notes": "Tabular matches an optimizer/preconditioning intervention; both systems use identical mlp_tiny and paired seeds. Fisher uses a small per-example gradient sketch and block-local factors."
}