1{
  2  "bench_version": 1,
  3  "track": "tabular",
  4  "model": "mlp_med",
  5  "metric_direction": "lower is better",
  6  "n_seeds": 8,
  7  "baseline": {
  8    "best_cfg": {
  9      "lr": 0.006,
 10      "beta": 0.002
 11    },
 12    "sweep": [
 13      {
 14        "cfg": {
 15          "lr": 0.0015,
 16          "beta": 0.0
 17        },
 18        "mean": 7.332403898239136
 19      },
 20      {
 21        "cfg": {
 22          "lr": 0.0015,
 23          "beta": 0.002
 24        },
 25        "mean": 6.974947094917297
 26      },
 27      {
 28        "cfg": {
 29          "lr": 0.0015,
 30          "beta": 0.005
 31        },
 32        "mean": 6.950373411178589
 33      },
 34      {
 35        "cfg": {
 36          "lr": 0.0015,
 37          "beta": 0.01
 38        },
 39        "mean": 6.856605052947998
 40      },
 41      {
 42        "cfg": {
 43          "lr": 0.0015,
 44          "beta": 0.02
 45        },
 46        "mean": 6.820917129516602
 47      },
 48      {
 49        "cfg": {
 50          "lr": 0.003,
 51          "beta": 0.0
 52        },
 53        "mean": 6.740176439285278
 54      },
 55      {
 56        "cfg": {
 57          "lr": 0.003,
 58          "beta": 0.002
 59        },
 60        "mean": 6.5683335065841675
 61      },
 62      {
 63        "cfg": {
 64          "lr": 0.003,
 65          "beta": 0.005
 66        },
 67        "mean": 6.506619572639465
 68      },
 69      {
 70        "cfg": {
 71          "lr": 0.003,
 72          "beta": 0.01
 73        },
 74        "mean": 6.451507329940796
 75      },
 76      {
 77        "cfg": {
 78          "lr": 0.003,
 79          "beta": 0.02
 80        },
 81        "mean": 6.507548213005066
 82      },
 83      {
 84        "cfg": {
 85          "lr": 0.006,
 86          "beta": 0.0
 87        },
 88        "mean": 6.459080219268799
 89      },
 90      {
 91        "cfg": {
 92          "lr": 0.006,
 93          "beta": 0.002
 94        },
 95        "mean": 6.255969762802124
 96      },
 97      {
 98        "cfg": {
 99          "lr": 0.006,
100          "beta": 0.005
101        },
102        "mean": 6.3951462507247925
103      },
104      {
105        "cfg": {
106          "lr": 0.006,
107          "beta": 0.01
108        },
109        "mean": 6.444561243057251
110      },
111      {
112        "cfg": {
113          "lr": 0.006,
114          "beta": 0.02
115        },
116        "mean": 6.318377375602722
117      }
118    ],
119    "full": {
120      "mean": 6.492175281047821,
121      "std": 0.4696585684033177,
122      "per_seed": [
123        5.431238651275635,
124        6.641201019287109,
125        6.379562854766846,
126        6.571876525878906,
127        6.9486002922058105,
128        6.291627407073975,
129        6.600069522857666,
130        7.073225975036621
131      ],
132      "n": 8
133    }
134  },
135  "idea": {
136    "mean": 6.48457407951355,
137    "std": 0.4566508873984473,
138    "per_seed": [
139      5.4823994636535645,
140      6.262659549713135,
141      6.524853229522705,
142      6.588247776031494,
143      6.916043758392334,
144      6.340798377990723,
145      6.678991317749023,
146      7.08259916305542
147    ],
148    "n": 8,
149    "best_cfg": {
150      "lr": 0.006,
151      "beta_start": 0.005,
152      "beta_end": 5e-05
153    }
154  },
155  "comparison": {
156    "delta_mean": -0.00760120153427124,
157    "idea_wins": 2,
158    "n_pairs": 8,
159    "per_seed_diffs": [
160      0.05116081237792969,
161      -0.3785414695739746,
162      0.14529037475585938,
163      0.01637125015258789,
164      -0.03255653381347656,
165      0.04917097091674805,
166      0.07892179489135742,
167      0.009373188018798828
168    ],
169    "p_value": 0.96975,
170    "mde": 0.1328181829200986,
171    "mde_rel_pct": 2.045819423696491,
172    "verdict": "no measurable effect",
173    "system_worked": false
174  },
175  "math_check": {
176    "predicted_exponent": 1.3333333333333333,
177    "observed_exponent": 1.3333333333333335,
178    "median_relative_boundary_error": 0.30000000000000004,
179    "pass": false
180  },
181  "idea_sweep": {
182    "configs": [
183      {
184        "lr": 0.0015,
185        "beta_start": 0.02,
186        "beta_end": 0.0002
187      },
188      {
189        "lr": 0.003,
190        "beta_start": 0.01,
191        "beta_end": 0.0001
192      },
193      {
194        "lr": 0.006,
195        "beta_start": 0.005,
196        "beta_end": 5e-05
197      }
198    ],
199    "selected": {
200      "lr": 0.006,
201      "beta_start": 0.005,
202      "beta_end": 5e-05
203    }
204  },
205  "budget": {
206    "epochs": 12,
207    "batch": 128,
208    "train_samples": 1200
209  },
210  "mechanism_signature": {
211    "kind": "trained_end_to_end_spectrum_vs_crosscov",
212    "measurements": [
213      {
214        "epoch": 1,
215        "beta": 0.005,
216        "teacher_crosscov_singular_values": [
217          1.3328860998153687
218        ],
219        "end_to_end_singular_values": [
220          8.559873580932617
221        ],
222        "predicted_active_count": 1,
223        "observed_active_count": 1
224      },
225      {
226        "epoch": 7,
227        "beta": 0.0004055654153948436,
228        "teacher_crosscov_singular_values": [
229          1.3328860998153687
230        ],
231        "end_to_end_singular_values": [
232          18.78715705871582
233        ],
234        "predicted_active_count": 1,
235        "observed_active_count": 1
236      },
237      {
238        "epoch": 12,
239        "beta": 5e-05,
240        "teacher_crosscov_singular_values": [
241          1.3328860998153687
242        ],
243        "end_to_end_singular_values": [
244          21.196590423583984
245        ],
246        "predicted_active_count": 1,
247        "observed_active_count": 1
248      }
249    ],
250    "confirmed": true,
251    "note": "Signature is measured from trained MLP weights and benchmark data; it is not an analytic identity."
252  }
253}