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    "sweep": {
  9      "best_cfg": {
 10        "lr": 0.001
 11      },
 12      "sweep": [
 13        {
 14          "cfg": {
 15            "lr": 0.001
 16          },
 17          "mean": 0.7862499803304672
 18        },
 19        {
 20          "cfg": {
 21            "lr": 0.003
 22          },
 23          "mean": 0.8006249815225601
 24        },
 25        {
 26          "cfg": {
 27            "lr": 0.006
 28          },
 29          "mean": 0.8237500041723251
 30        }
 31      ],
 32      "full": {
 33        "mean": 0.7849999815225601,
 34        "std": 0.03665719838879184,
 35        "per_seed": [
 36          0.7874999642372131,
 37          0.75,
 38          0.8349999785423279,
 39          0.7724999785423279,
 40          0.8274999856948853,
 41          0.8174999952316284,
 42          0.7249999642372131,
 43          0.7649999856948853
 44        ],
 45        "n": 8
 46      }
 47    },
 48    "full": {
 49      "mean": 0.7849999815225601,
 50      "std": 0.03665719838879184,
 51      "per_seed": [
 52        0.7874999642372131,
 53        0.75,
 54        0.8349999785423279,
 55        0.7724999785423279,
 56        0.8274999856948853,
 57        0.8174999952316284,
 58        0.7249999642372131,
 59        0.7649999856948853
 60      ],
 61      "n": 8
 62    }
 63  },
 64  "idea": {
 65    "lr": 0.001,
 66    "per_seed": [
 67      0.7899999618530273,
 68      0.7699999809265137,
 69      0.8399999737739563,
 70      0.762499988079071,
 71      0.824999988079071,
 72      0.8075000047683716,
 73      0.7475000023841858,
 74      0.7549999952316284
 75    ],
 76    "mean": 0.7871874868869781
 77  },
 78  "comparison": {
 79    "delta_mean": 0.00218750536441803,
 80    "idea_wins": 4,
 81    "n_pairs": 8,
 82    "per_seed_diffs": [
 83      0.002499997615814209,
 84      0.019999980926513672,
 85      0.004999995231628418,
 86      -0.009999990463256836,
 87      -0.002499997615814209,
 88      -0.009999990463256836,
 89      0.022500038146972656,
 90      -0.009999990463256836
 91    ],
 92    "p_value": 0.64985,
 93    "mde": 0.01097274108925552,
 94    "mde_rel_pct": 1.3978014455456615,
 95    "verdict": "no measurable effect",
 96    "system_worked": false
 97  },
 98  "math_check": {
 99    "matrix_frobenius_residual": 0.0,
100    "max_abs_energy_error": 1.4210854715202004e-14,
101    "relative_energy_rmse": 1.4996245561724451e-16,
102    "random_relative_rmse": 0.32222904703868205,
103    "random_mean_relative_std": 0.3208364891796913
104  },
105  "idea_grid": [
106    {
107      "lr": 0.001,
108      "mean": 0.7871874868869781
109    },
110    {
111      "lr": 0.003,
112      "mean": 0.7893749848008156
113    },
114    {
115      "lr": 0.006,
116      "mean": 0.8271874785423279
117    }
118  ],
119  "mechanism_signature": {
120    "prediction": "exact modeled energy has zero estimator variance; random signs have nonzero variance",
121    "predicted_exact_relative_rmse": 0.0,
122    "observed_mean_exact_relative_rmse": 5.16787475035585e-08,
123    "observed_mean_random_relative_std": 0.3647565320134163,
124    "per_seed": [
125      {
126        "seed": 0,
127        "exact_rel_rmse": 0.0,
128        "random_rel_std": 0.33480483293533325
129      },
130      {
131        "seed": 1,
132        "exact_rel_rmse": 1.1523898280074718e-07,
133        "random_rel_std": 0.3534063696861267
134      },
135      {
136        "seed": 2,
137        "exact_rel_rmse": 0.0,
138        "random_rel_std": 0.44275858998298645
139      },
140      {
141        "seed": 3,
142        "exact_rel_rmse": 0.0,
143        "random_rel_std": 0.4704761505126953
144      },
145      {
146        "seed": 4,
147        "exact_rel_rmse": 0.0,
148        "random_rel_std": 0.3139871060848236
149      },
150      {
151        "seed": 5,
152        "exact_rel_rmse": 1.5191113789114752e-07,
153        "random_rel_std": 0.28184041380882263
154      },
155      {
156        "seed": 6,
157        "exact_rel_rmse": 1.4627985933657328e-07,
158        "random_rel_std": 0.2919972240924835
159      },
160      {
161        "seed": 7,
162        "exact_rel_rmse": 0.0,
163        "random_rel_std": 0.42878156900405884
164      }
165    ],
166    "confirmed": true
167  },
168  "protocol_note": "Vision selected because the intervention is activation/architecture normalization; duplicate CNN layers and identical trainer settings, with only energy statistic changed."
169}