1{
  2  "bench_version": 1,
  3  "track": "vision",
  4  "model": "cnn_small",
  5  "metric_direction": "lower is better",
  6  "n_seeds": 8,
  7  "baseline": {
  8    "best_cfg": {
  9      "lr": 0.003
 10    },
 11    "sweep": [
 12      {
 13        "cfg": {
 14          "lr": 0.001
 15        },
 16        "mean": 0.8724999874830246
 17      },
 18      {
 19        "cfg": {
 20          "lr": 0.003
 21        },
 22        "mean": 0.8224999606609344
 23      },
 24      {
 25        "cfg": {
 26          "lr": 0.01
 27        },
 28        "mean": 0.8424999862909317
 29      }
 30    ],
 31    "full": {
 32      "mean": 0.8362499698996544,
 33      "std": 0.05694021657567501,
 34      "per_seed": [
 35        0.7999999523162842,
 36        0.8499999642372131,
 37        0.9099999666213989,
 38        0.7299999594688416,
 39        0.8499999642372131,
 40        0.8999999761581421,
 41        0.8650000095367432,
 42        0.7849999666213989
 43      ],
 44      "n": 8
 45    }
 46  },
 47  "idea": {
 48    "mean": 0.7549999803304672,
 49    "std": 0.04500000261598574,
 50    "per_seed": [
 51      0.7149999737739563,
 52      0.8100000023841858,
 53      0.7899999618530273,
 54      0.7649999856948853,
 55      0.7299999594688416,
 56      0.73499995470047,
 57      0.8149999976158142,
 58      0.6800000071525574
 59    ],
 60    "n": 8
 61  },
 62  "comparison": {
 63    "delta_mean": -0.08124998956918716,
 64    "idea_wins": 7,
 65    "n_pairs": 8,
 66    "per_seed_diffs": [
 67      -0.08499997854232788,
 68      -0.039999961853027344,
 69      -0.12000000476837158,
 70      0.0350000262260437,
 71      -0.12000000476837158,
 72      -0.16500002145767212,
 73      -0.050000011920928955,
 74      -0.10499995946884155
 75    ],
 76    "p_value": 0.0152,
 77    "mde": 0.051725334839130825,
 78    "mde_rel_pct": 6.185391533746494,
 79    "verdict": "idea better (significant)",
 80    "system_worked": true
 81  },
 82  "mechanism_signature": {
 83    "math_check": {
 84      "analytic_jacobian": [
 85        0.28400640000000005,
 86        0.018753524999999997
 87      ],
 88      "finite_difference": [
 89        0.28400640005137034,
 90        0.01875352500860239
 91      ],
 92      "max_abs_error": 5.137029690516215e-11,
 93      "cubic_gain_ratio": 0.5120000000000001
 94    },
 95    "trained_behavior": {
 96      "prediction": "local Jacobian controls perturbation amplification; cubic neutral derivative scales as lambda^3",
 97      "observed_trained_pool_perturbation_gain_mean": 2.372135490179062,
 98      "observed_trained_pool_perturbation_gain_std": 0.5063665608536517,
 99      "trained_coefficients": [
100        [
101          0.31366875767707825,
102          0.6198698282241821
103        ],
104        [
105          0.3111434876918793,
106          0.6182912588119507
107        ],
108        [
109          0.3147978186607361,
110          0.6183661222457886
111        ],
112        [
113          0.3124944567680359,
114          0.6156984567642212
115        ],
116        [
117          0.3157919943332672,
118          0.6252565383911133
119        ],
120        [
121          0.3116231858730316,
122          0.6227443814277649
123        ],
124        [
125          0.3117213547229767,
126          0.6220788955688477
127        ],
128        [
129          0.3104568421840668,
130          0.6197155117988586
131        ]
132      ],
133      "predicted_lambda_cubic_gain": true,
134      "confirmed": false
135    },
136    "baseline_full_by_lr": {
137      "0.001": {
138        "mean": 0.8631249815225601,
139        "std": 0.03141830539156022,
140        "per_seed": [
141          0.8650000095367432,
142          0.8700000047683716,
143          0.9099999666213989,
144          0.8449999690055847,
145          0.7999999523162842,
146          0.8899999856948853,
147          0.8449999690055847,
148          0.8799999952316284
149        ],
150        "n": 8
151      },
152      "0.003": {
153        "mean": 0.8362499698996544,
154        "std": 0.05694021657567501,
155        "per_seed": [
156          0.7999999523162842,
157          0.8499999642372131,
158          0.9099999666213989,
159          0.7299999594688416,
160          0.8499999642372131,
161          0.8999999761581421,
162          0.8650000095367432,
163          0.7849999666213989
164        ],
165        "n": 8
166      },
167      "0.01": {
168        "mean": 0.8606249839067459,
169        "std": 0.049588159207269375,
170        "per_seed": [
171          0.8849999904632568,
172          0.8799999952316284,
173          0.8700000047683716,
174          0.73499995470047,
175          0.8449999690055847,
176          0.8899999856948853,
177          0.8899999856948853,
178          0.8899999856948853
179        ],
180        "n": 8
181      }
182    },
183    "idea_sweep": [
184      {
185        "lr": 0.001,
186        "mean": 0.7549999803304672
187      },
188      {
189        "lr": 0.003,
190        "mean": 0.8287499770522118
191      },
192      {
193        "lr": 0.01,
194        "mean": 0.9049999788403511
195      }
196    ]
197  },
198  "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."
199}