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      "weight_decay": 0.0
 11    },
 12    "sweep": [
 13      {
 14        "cfg": {
 15          "lr": 0.001,
 16          "weight_decay": 0.0
 17        },
 18        "mean": 16.985830068588257
 19      },
 20      {
 21        "cfg": {
 22          "lr": 0.003,
 23          "weight_decay": 0.0
 24        },
 25        "mean": 11.592020750045776
 26      },
 27      {
 28        "cfg": {
 29          "lr": 0.01,
 30          "weight_decay": 0.0
 31        },
 32        "mean": 7.1914142370224
 33      }
 34    ],
 35    "full": {
 36      "mean": 7.462258458137512,
 37      "std": 0.633509839885441,
 38      "per_seed": [
 39        6.857808589935303,
 40        7.703956604003906,
 41        7.029021263122559,
 42        7.174870491027832,
 43        6.467283725738525,
 44        7.960057258605957,
 45        8.130547523498535,
 46        8.37452220916748
 47      ],
 48      "n": 8
 49    }
 50  },
 51  "idea": {
 52    "mean": 7.855136156082153,
 53    "std": 0.8610502821731341,
 54    "per_seed": [
 55      7.060885906219482,
 56      8.29471206665039,
 57      7.688426494598389,
 58      7.102557182312012,
 59      6.474584579467773,
 60      9.063955307006836,
 61      8.353429794311523,
 62      8.80253791809082
 63    ],
 64    "n": 8
 65  },
 66  "comparison": {
 67    "delta_mean": 0.3928776979446411,
 68    "idea_wins": 1,
 69    "n_pairs": 8,
 70    "per_seed_diffs": [
 71      0.2030773162841797,
 72      0.5907554626464844,
 73      0.6594052314758301,
 74      -0.07231330871582031,
 75      0.007300853729248047,
 76      1.103898048400879,
 77      0.22288227081298828,
 78      0.42801570892333984
 79    ],
 80    "p_value": 0.0235,
 81    "mde": 0.3233736709994596,
 82    "mde_rel_pct": 4.333455787058997,
 83    "verdict": "idea worse (significant)",
 84    "system_worked": false
 85  },
 86  "mechanism_signature": {
 87    "idea_config": {
 88      "lr": 0.01,
 89      "tail_weight": 4.0
 90    },
 91    "idea_sweep": [
 92      {
 93        "cfg": {
 94          "lr": 0.001,
 95          "tail_weight": 4.0
 96        },
 97        "result": {
 98          "mean": 22.177693843841553,
 99          "std": 0.9958955479883927,
100          "per_seed": [
101            22.019962310791016,
102            20.866891860961914,
103            22.155040740966797,
104            23.668880462646484
105          ],
106          "n": 4
107        }
108      },
109      {
110        "cfg": {
111          "lr": 0.003,
112          "tail_weight": 4.0
113        },
114        "result": {
115          "mean": 11.351011991500854,
116          "std": 0.7376803433578396,
117          "per_seed": [
118            10.290882110595703,
119            11.659903526306152,
120            12.307673454284668,
121            11.145588874816895
122          ],
123          "n": 4
124        }
125      },
126      {
127        "cfg": {
128          "lr": 0.01,
129          "tail_weight": 4.0
130        },
131        "result": {
132          "mean": 7.536645412445068,
133          "std": 0.5031108845030456,
134          "per_seed": [
135            7.060885906219482,
136            8.29471206665039,
137            7.688426494598389,
138            7.102557182312012
139          ],
140          "n": 4
141        }
142      }
143    ],
144    "mechanism_signature": {
145      "prediction": "tail residual correction should preferentially reduce upper-tail prediction error relative to global error",
146      "baseline_behavior": {
147        "global_mse": 7.462258815765381,
148        "tail_mse": 14.653417646884918,
149        "tail_to_global": 1.9636705196993307,
150        "cvar_abs_error": 2.055535450577737
151      },
152      "idea_behavior": {
153        "global_mse": 7.855136334896088,
154        "tail_mse": 8.973360180854797,
155        "tail_to_global": 1.1423557527564545,
156        "cvar_abs_error": 0.8575644314289099
157      },
158      "predicted_tail_focus": true,
159      "observed_tail_to_global_ratio_change": -0.8213147669428762,
160      "confirmed": true
161    }
162  },
163  "protocol_notes": {
164    "epochs": 24,
165    "batch": 128,
166    "baseline_grid": [
167      {
168        "lr": 0.001,
169        "weight_decay": 0.0
170      },
171      {
172        "lr": 0.003,
173        "weight_decay": 0.0
174      },
175      {
176        "lr": 0.01,
177        "weight_decay": 0.0
178      }
179    ],
180    "idea_grid": [
181      {
182        "lr": 0.001,
183        "tail_weight": 4.0
184      },
185      {
186        "lr": 0.003,
187        "tail_weight": 4.0
188      },
189      {
190        "lr": 0.01,
191        "tail_weight": 4.0
192      }
193    ],
194    "intervention": "same mlp_tiny trained with global-MSE warmup followed by upper-tail weighted residual correction"
195  }
196}