{
  "bench_version": 1,
  "track": "vision",
  "model": "cnn_small",
  "metric_direction": "lower is better",
  "n_seeds": 8,
  "baseline": {
    "best_cfg": {
      "lr": 0.003
    },
    "sweep": [
      {
        "cfg": {
          "lr": 0.001
        },
        "mean": 0.8724999874830246
      },
      {
        "cfg": {
          "lr": 0.003
        },
        "mean": 0.8224999606609344
      },
      {
        "cfg": {
          "lr": 0.01
        },
        "mean": 0.8424999862909317
      }
    ],
    "full": {
      "mean": 0.8362499698996544,
      "std": 0.05694021657567501,
      "per_seed": [
        0.7999999523162842,
        0.8499999642372131,
        0.9099999666213989,
        0.7299999594688416,
        0.8499999642372131,
        0.8999999761581421,
        0.8650000095367432,
        0.7849999666213989
      ],
      "n": 8
    }
  },
  "idea": {
    "mean": 0.7549999803304672,
    "std": 0.04500000261598574,
    "per_seed": [
      0.7149999737739563,
      0.8100000023841858,
      0.7899999618530273,
      0.7649999856948853,
      0.7299999594688416,
      0.73499995470047,
      0.8149999976158142,
      0.6800000071525574
    ],
    "n": 8
  },
  "comparison": {
    "delta_mean": -0.08124998956918716,
    "idea_wins": 7,
    "n_pairs": 8,
    "per_seed_diffs": [
      -0.08499997854232788,
      -0.039999961853027344,
      -0.12000000476837158,
      0.0350000262260437,
      -0.12000000476837158,
      -0.16500002145767212,
      -0.050000011920928955,
      -0.10499995946884155
    ],
    "p_value": 0.0152,
    "mde": 0.051725334839130825,
    "mde_rel_pct": 6.185391533746494,
    "verdict": "idea better (significant)",
    "system_worked": true
  },
  "mechanism_signature": {
    "math_check": {
      "analytic_jacobian": [
        0.28400640000000005,
        0.018753524999999997
      ],
      "finite_difference": [
        0.28400640005137034,
        0.01875352500860239
      ],
      "max_abs_error": 5.137029690516215e-11,
      "cubic_gain_ratio": 0.5120000000000001
    },
    "trained_behavior": {
      "prediction": "local Jacobian controls perturbation amplification; cubic neutral derivative scales as lambda^3",
      "observed_trained_pool_perturbation_gain_mean": 2.372135490179062,
      "observed_trained_pool_perturbation_gain_std": 0.5063665608536517,
      "trained_coefficients": [
        [
          0.31366875767707825,
          0.6198698282241821
        ],
        [
          0.3111434876918793,
          0.6182912588119507
        ],
        [
          0.3147978186607361,
          0.6183661222457886
        ],
        [
          0.3124944567680359,
          0.6156984567642212
        ],
        [
          0.3157919943332672,
          0.6252565383911133
        ],
        [
          0.3116231858730316,
          0.6227443814277649
        ],
        [
          0.3117213547229767,
          0.6220788955688477
        ],
        [
          0.3104568421840668,
          0.6197155117988586
        ]
      ],
      "predicted_lambda_cubic_gain": true,
      "confirmed": false
    },
    "baseline_full_by_lr": {
      "0.001": {
        "mean": 0.8631249815225601,
        "std": 0.03141830539156022,
        "per_seed": [
          0.8650000095367432,
          0.8700000047683716,
          0.9099999666213989,
          0.8449999690055847,
          0.7999999523162842,
          0.8899999856948853,
          0.8449999690055847,
          0.8799999952316284
        ],
        "n": 8
      },
      "0.003": {
        "mean": 0.8362499698996544,
        "std": 0.05694021657567501,
        "per_seed": [
          0.7999999523162842,
          0.8499999642372131,
          0.9099999666213989,
          0.7299999594688416,
          0.8499999642372131,
          0.8999999761581421,
          0.8650000095367432,
          0.7849999666213989
        ],
        "n": 8
      },
      "0.01": {
        "mean": 0.8606249839067459,
        "std": 0.049588159207269375,
        "per_seed": [
          0.8849999904632568,
          0.8799999952316284,
          0.8700000047683716,
          0.73499995470047,
          0.8449999690055847,
          0.8899999856948853,
          0.8899999856948853,
          0.8899999856948853
        ],
        "n": 8
      }
    },
    "idea_sweep": [
      {
        "lr": 0.001,
        "mean": 0.7549999803304672
      },
      {
        "lr": 0.003,
        "mean": 0.8287499770522118
      },
      {
        "lr": 0.01,
        "mean": 0.9049999788403511
      }
    ]
  },
  "protocol_note": "Vision selected because the idea changes convolutional hierarchical feature propagation; baseline and idea share convolution/classifier stages and differ only in pooling mechanism."
}