1{
  2  "bench_version": 1,
  3  "track": "tabular",
  4  "model": "mlp_tiny",
  5  "metric_direction": "lower is better",
  6  "n_seeds": 8,
  7  "baseline": {
  8    "best_cfg": {
  9      "lr": 0.01
 10    },
 11    "sweep": [
 12      {
 13        "cfg": {
 14          "lr": 0.001
 15        },
 16        "mean": 16.10005497932434
 17      },
 18      {
 19        "cfg": {
 20          "lr": 0.003
 21        },
 22        "mean": 11.262676000595093
 23      },
 24      {
 25        "cfg": {
 26          "lr": 0.01
 27        },
 28        "mean": 7.429885745048523
 29      }
 30    ],
 31    "full": {
 32      "mean": 7.599956035614014,
 33      "std": 0.8051347687842139,
 34      "per_seed": [
 35        6.844658851623535,
 36        6.109804630279541,
 37        8.30047607421875,
 38        8.464603424072266,
 39        7.705488681793213,
 40        6.993262767791748,
 41        8.389989852905273,
 42        7.991364002227783
 43      ],
 44      "n": 8
 45    }
 46  },
 47  "idea": {
 48    "mean": 8.154384911060333,
 49    "std": 1.0389464855334478,
 50    "per_seed": [
 51      7.118060111999512,
 52      6.544161319732666,
 53      9.722457885742188,
 54      8.549818992614746,
 55      7.820003986358643,
 56      7.930368423461914,
 57      9.617923736572266,
 58      7.932284832000732
 59    ],
 60    "n": 8
 61  },
 62  "comparison": {
 63    "delta_mean": 0.5544288754463196,
 64    "idea_wins": 1,
 65    "n_pairs": 8,
 66    "per_seed_diffs": [
 67      0.27340126037597656,
 68      0.434356689453125,
 69      1.4219818115234375,
 70      0.08521556854248047,
 71      0.11451530456542969,
 72      0.937105655670166,
 73      1.2279338836669922,
 74      -0.05907917022705078
 75    ],
 76    "p_value": 0.0156,
 77    "mde": 0.4725954183814587,
 78    "mde_rel_pct": 6.218396740281628,
 79    "verdict": "idea worse (significant)",
 80    "system_worked": false
 81  },
 82  "track_justification": "ReLU hidden-axis pruning is structurally matched to the tabular MLP track; both systems use the identical trained mlp_tiny and optimizer budgets.",
 83  "pruning_fraction": 0.5,
 84  "idea_config": {
 85    "lr": 0.01
 86  },
 87  "idea_sweep": [
 88    {
 89      "cfg": {
 90        "lr": 0.001
 91      },
 92      "result": {
 93        "mean": 22.81420075893402,
 94        "std": 5.572511935733092,
 95        "per_seed": [
 96          23.423412322998047,
 97          15.931241035461426,
 98          33.59147262573242,
 99          17.00380516052246,
100          26.614585876464844,
101          17.325483322143555,
102          25.237001419067383,
103          23.38660430908203
104        ],
105        "n": 8
106      }
107    },
108    {
109      "cfg": {
110        "lr": 0.003
111      },
112      "result": {
113        "mean": 14.472419738769531,
114        "std": 2.867801864411425,
115        "per_seed": [
116          12.962059020996094,
117          10.092577934265137,
118          17.148317337036133,
119          11.091870307922363,
120          15.618391036987305,
121          13.338394165039062,
122          16.90511131286621,
123          18.622636795043945
124        ],
125        "n": 8
126      }
127    },
128    {
129      "cfg": {
130        "lr": 0.01
131      },
132      "result": {
133        "mean": 8.154384911060333,
134        "std": 1.0389464855334478,
135        "per_seed": [
136          7.118060111999512,
137          6.544161319732666,
138          9.722457885742188,
139          8.549818992614746,
140          7.820003986358643,
141          7.930368423461914,
142          9.617923736572266,
143          7.932284832000732
144        ],
145        "n": 8
146      }
147    }
148  ],
149  "parameter_fraction_estimate": 0.25,
150  "mechanism_signature": {
151    "prediction": "strictly one-sided selected ReLU axes are affine or zero on the test patch",
152    "predicted_max_replacement_error": 0.0,
153    "observed_max_replacement_error": 0.1555086076259613,
154    "tested_one_sided_axes": 53,
155    "downstream_norm_sum": 24.949480190873146,
156    "confirmed": false
157  }
158}