{
  "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": 17.2008535861969
      },
      {
        "cfg": {
          "lr": 0.003
        },
        "mean": 17.275779485702515
      },
      {
        "cfg": {
          "lr": 0.01
        },
        "mean": 18.10535430908203
      }
    ],
    "full": {
      "mean": 17.477807760238647,
      "std": 1.5610336422417328,
      "per_seed": [
        18.36859893798828,
        17.082815170288086,
        17.65996551513672,
        15.692034721374512,
        14.8045015335083,
        18.059467315673828,
        17.911405563354492,
        20.24367332458496
      ],
      "n": 8
    }
  },
  "idea": {
    "mean": 22.206230878829956,
    "std": 1.2034773425543663,
    "per_seed": [
      22.064266204833984,
      22.58884048461914,
      20.288148880004883,
      21.367250442504883,
      21.962797164916992,
      21.877431869506836,
      22.725826263427734,
      24.775285720825195
    ],
    "n": 8,
    "selected_cfg": {
      "lr": 0.01
    }
  },
  "comparison": {
    "delta_mean": 4.728423118591309,
    "idea_wins": 0,
    "n_pairs": 8,
    "per_seed_diffs": [
      3.695667266845703,
      5.506025314331055,
      2.628183364868164,
      5.675215721130371,
      7.158295631408691,
      3.817964553833008,
      4.814420700073242,
      4.531612396240234
    ],
    "p_value": 0.0081,
    "mde": 1.172236051535733,
    "mde_rel_pct": 6.70699705372962,
    "verdict": "idea worse (significant)",
    "system_worked": false
  },
  "mechanism_signature": {
    "mechanism_signature": {
      "prediction": "factorized Cholesky-preconditioned outer product equals matrix update",
      "per_seed": [
        {
          "factorization_abs_err": 0.0021419525146484375,
          "factorization_rel_err": 1.424667971150484e-05,
          "raw_update_norm": 99.73847961425781,
          "conditioned_update_norm": 471.4458923339844,
          "predicted_norm_ratio": 4.726820468902588,
          "trained_activation_trace": 3.3609566688537598,
          "trained_error_trace": 22.064266204833984,
          "confirmed": false
        },
        {
          "factorization_abs_err": 0.0013217926025390625,
          "factorization_rel_err": 9.378113645652775e-06,
          "raw_update_norm": 119.99061584472656,
          "conditioned_update_norm": 431.77825927734375,
          "predicted_norm_ratio": 3.598433494567871,
          "trained_activation_trace": 3.3852806091308594,
          "trained_error_trace": 22.588836669921875,
          "confirmed": true
        },
        {
          "factorization_abs_err": 0.0012187957763671875,
          "factorization_rel_err": 7.294798706425354e-06,
          "raw_update_norm": 106.99031829833984,
          "conditioned_update_norm": 435.8670654296875,
          "predicted_norm_ratio": 4.073892593383789,
          "trained_activation_trace": 3.3690006732940674,
          "trained_error_trace": 20.288150787353516,
          "confirmed": true
        },
        {
          "factorization_abs_err": 0.0012369155883789062,
          "factorization_rel_err": 8.22768470243318e-06,
          "raw_update_norm": 88.57902526855469,
          "conditioned_update_norm": 449.8793640136719,
          "predicted_norm_ratio": 5.078847408294678,
          "trained_activation_trace": 3.282832145690918,
          "trained_error_trace": 21.367252349853516,
          "confirmed": true
        },
        {
          "factorization_abs_err": 0.0022106170654296875,
          "factorization_rel_err": 1.4202377315086778e-05,
          "raw_update_norm": 126.09120178222656,
          "conditioned_update_norm": 424.4380798339844,
          "predicted_norm_ratio": 3.366119623184204,
          "trained_activation_trace": 3.368903636932373,
          "trained_error_trace": 21.96279525756836,
          "confirmed": false
        },
        {
          "factorization_abs_err": 0.001743316650390625,
          "factorization_rel_err": 9.334406058769673e-06,
          "raw_update_norm": 113.74923706054688,
          "conditioned_update_norm": 444.450927734375,
          "predicted_norm_ratio": 3.907287120819092,
          "trained_activation_trace": 3.3790221214294434,
          "trained_error_trace": 21.877431869506836,
          "confirmed": true
        },
        {
          "factorization_abs_err": 0.002315521240234375,
          "factorization_rel_err": 1.217806766362628e-05,
          "raw_update_norm": 90.95088195800781,
          "conditioned_update_norm": 474.05853271484375,
          "predicted_norm_ratio": 5.212247848510742,
          "trained_activation_trace": 3.336045265197754,
          "trained_error_trace": 22.7258243560791,
          "confirmed": false
        },
        {
          "factorization_abs_err": 0.0021305084228515625,
          "factorization_rel_err": 1.0202298653894104e-05,
          "raw_update_norm": 137.70616149902344,
          "conditioned_update_norm": 439.57916259765625,
          "predicted_norm_ratio": 3.192153215408325,
          "trained_activation_trace": 3.364262580871582,
          "trained_error_trace": 24.775285720825195,
          "confirmed": false
        }
      ],
      "confirmed": false,
      "mean_observed_norm_ratio": 4.144475221633911
    },
    "track_rationale": "Two-sided conditioning is an optimizer/update-rule idea; tabular is the prescribed optimizer track.",
    "idea_sweep": [
      {
        "cfg": {
          "lr": 0.001
        },
        "mean": 200.41951751708984
      },
      {
        "cfg": {
          "lr": 0.003
        },
        "mean": 104.71292400360107
      },
      {
        "cfg": {
          "lr": 0.01
        },
        "mean": 22.206230878829956
      }
    ],
    "selected_idea_cfg": {
      "lr": 0.01
    },
    "elapsed_sec": 17.28634238243103
  }
}