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": 0.001
 10    },
 11    "sweep": [
 12      {
 13        "cfg": {
 14          "lr": 0.001
 15        },
 16        "mean": 0.8887499868869781
 17      },
 18      {
 19        "cfg": {
 20          "lr": 0.003
 21        },
 22        "mean": 0.8899999856948853
 23      },
 24      {
 25        "cfg": {
 26          "lr": 0.01
 27        },
 28        "mean": 0.8912499845027924
 29      }
 30    ],
 31    "full": {
 32      "mean": 0.8762499839067459,
 33      "std": 0.024843258193087,
 34      "per_seed": [
 35        0.8949999809265137,
 36        0.8849999904632568,
 37        0.8999999761581421,
 38        0.875,
 39        0.8299999833106995,
 40        0.9049999713897705,
 41        0.8449999690055847,
 42        0.875
 43      ],
 44      "n": 8
 45    }
 46  },
 47  "idea": {
 48    "mean": 0.8762499839067459,
 49    "std": 0.024843258193087,
 50    "per_seed": [
 51      0.8949999809265137,
 52      0.8849999904632568,
 53      0.8999999761581421,
 54      0.875,
 55      0.8299999833106995,
 56      0.9049999713897705,
 57      0.8449999690055847,
 58      0.875
 59    ],
 60    "n": 8
 61  },
 62  "comparison": {
 63    "delta_mean": 0.0,
 64    "idea_wins": 0,
 65    "n_pairs": 8,
 66    "per_seed_diffs": [
 67      0.0,
 68      0.0,
 69      0.0,
 70      0.0,
 71      0.0,
 72      0.0,
 73      0.0,
 74      0.0
 75    ],
 76    "p_value": 1.0,
 77    "mde": 0.0,
 78    "mde_rel_pct": 0.0,
 79    "verdict": "no measurable effect",
 80    "system_worked": false
 81  },
 82  "idea_sweep": [
 83    {
 84      "cfg": {
 85        "lr": 0.001
 86      },
 87      "result": {
 88        "mean": 0.8762499839067459,
 89        "std": 0.024843258193087,
 90        "per_seed": [
 91          0.8949999809265137,
 92          0.8849999904632568,
 93          0.8999999761581421,
 94          0.875,
 95          0.8299999833106995,
 96          0.9049999713897705,
 97          0.8449999690055847,
 98          0.875
 99        ],
100        "n": 8
101      }
102    },
103    {
104      "cfg": {
105        "lr": 0.003
106      },
107      "result": {
108        "mean": 0.877499982714653,
109        "std": 0.029895645776799453,
110        "per_seed": [
111          0.9049999713897705,
112          0.8799999952316284,
113          0.9049999713897705,
114          0.8700000047683716,
115          0.8499999642372131,
116          0.9049999713897705,
117          0.8899999856948853,
118          0.8149999976158142
119        ],
120        "n": 8
121      }
122    },
123    {
124      "cfg": {
125        "lr": 0.01
126      },
127      "result": {
128        "mean": 0.877499982714653,
129        "std": 0.02958039211486051,
130        "per_seed": [
131          0.8949999809265137,
132          0.8799999952316284,
133          0.9149999618530273,
134          0.875,
135          0.824999988079071,
136          0.9149999618530273,
137          0.8449999690055847,
138          0.8700000047683716
139        ],
140        "n": 8
141      }
142    }
143  ],
144  "protocol_note": "Verifier pruning is inference-time only; training and standard test metric are unchanged systems.",
145  "mechanism_signature": {
146    "fixed_sign_fraction": 0.5423773306387443,
147    "max_fixed_substitution_error": 0.0,
148    "predicted_reduction_fraction": 0.5423773306387443,
149    "fixed_sign_mean": 959721.5,
150    "unstable_mean": 809750.5,
151    "total_relu_instances_mean": 1769472.0,
152    "confirmed": true
153  },
154  "experiment_settings": {
155    "epochs": 3,
156    "n_train": 400,
157    "n_test": 200,
158    "batch": 128,
159    "lr_union": [
160      0.001,
161      0.003,
162      0.01
163    ],
164    "seeds": [
165      0,
166      1,
167      2,
168      3,
169      4,
170      5,
171      6,
172      7
173    ]
174  }
175}