1{
  2  "bench_version": 1,
  3  "track": "dynamics",
  4  "model": "rnn_small",
  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": 0.0014035784406587481
 19      },
 20      {
 21        "cfg": {
 22          "lr": 0.001,
 23          "weight_decay": 0.0001
 24        },
 25        "mean": 0.0013894051662646234
 26      },
 27      {
 28        "cfg": {
 29          "lr": 0.003,
 30          "weight_decay": 0.0
 31        },
 32        "mean": 0.000857145045301877
 33      },
 34      {
 35        "cfg": {
 36          "lr": 0.003,
 37          "weight_decay": 0.0001
 38        },
 39        "mean": 0.0008355810859939083
 40      },
 41      {
 42        "cfg": {
 43          "lr": 0.01,
 44          "weight_decay": 0.0
 45        },
 46        "mean": 0.0004466182872420177
 47      },
 48      {
 49        "cfg": {
 50          "lr": 0.01,
 51          "weight_decay": 0.0001
 52        },
 53        "mean": 0.0004611926560755819
 54      }
 55    ],
 56    "full": {
 57      "mean": 0.00039566773739352357,
 58      "std": 0.00014846719173678902,
 59      "per_seed": [
 60        0.0003142073401249945,
 61        0.00040627806447446346,
 62        0.0003180095227435231,
 63        0.0007479782216250896,
 64        0.00039004243444651365,
 65        0.00019843246263917536,
 66        0.000411077169701457,
 67        0.00037931668339297175
 68      ],
 69      "n": 8
 70    }
 71  },
 72  "idea": {
 73    "mean": 0.0003887512521032477,
 74    "std": 0.0001616191005487337,
 75    "per_seed": [
 76      0.0002601368469186127,
 77      0.00042557879351079464,
 78      0.00030667721875943244,
 79      0.0007812847616150975,
 80      0.0003697451902553439,
 81      0.00021702404774259776,
 82      0.0003767137532122433,
 83      0.00037284940481185913
 84    ],
 85    "n": 8
 86  },
 87  "comparison": {
 88    "delta_mean": -6.916485290275887e-06,
 89    "idea_wins": 5,
 90    "n_pairs": 8,
 91    "per_seed_diffs": [
 92      -5.40704932063818e-05,
 93      1.9300729036331177e-05,
 94      -1.1332303984090686e-05,
 95      3.330653999000788e-05,
 96      -2.029724419116974e-05,
 97      1.8591585103422403e-05,
 98      -3.4363416489213705e-05,
 99      -6.467278581112623e-06
100    ],
101    "p_value": 0.51045,
102    "mde": 2.474683454228356e-05,
103    "mde_rel_pct": 6.254448418085408,
104    "verdict": "no significant win",
105    "system_worked": false
106  },
107  "mechanism_signature": {
108    "predicted": "feasibility weighting should reduce terminal-set error",
109    "observed_event_mse_mean": 4.5046147988614393e-05,
110    "observed_nonevent_mse_mean": 0.0004491949330258649,
111    "confirmed": true,
112    "idea_sweep": [
113      {
114        "cfg": {
115          "lr": 0.001,
116          "weight_decay": 0.0,
117          "alpha": 0.5
118        },
119        "mean": 0.0013711843639612198
120      },
121      {
122        "cfg": {
123          "lr": 0.001,
124          "weight_decay": 0.0,
125          "alpha": 1.0
126        },
127        "mean": 0.00134452705970034
128      },
129      {
130        "cfg": {
131          "lr": 0.001,
132          "weight_decay": 0.0,
133          "alpha": 2.0
134        },
135        "mean": 0.0013239112158771604
136      },
137      {
138        "cfg": {
139          "lr": 0.001,
140          "weight_decay": 0.0001,
141          "alpha": 0.5
142        },
143        "mean": 0.001358222885755822
144      },
145      {
146        "cfg": {
147          "lr": 0.001,
148          "weight_decay": 0.0001,
149          "alpha": 1.0
150        },
151        "mean": 0.0013323448947630823
152      },
153      {
154        "cfg": {
155          "lr": 0.001,
156          "weight_decay": 0.0001,
157          "alpha": 2.0
158        },
159        "mean": 0.001312095031607896
160      },
161      {
162        "cfg": {
163          "lr": 0.003,
164          "weight_decay": 0.0,
165          "alpha": 0.5
166        },
167        "mean": 0.0008442715043202043
168      },
169      {
170        "cfg": {
171          "lr": 0.003,
172          "weight_decay": 0.0,
173          "alpha": 1.0
174        },
175        "mean": 0.0008238502341555431
176      },
177      {
178        "cfg": {
179          "lr": 0.003,
180          "weight_decay": 0.0,
181          "alpha": 2.0
182        },
183        "mean": 0.0008038862433750182
184      },
185      {
186        "cfg": {
187          "lr": 0.003,
188          "weight_decay": 0.0001,
189          "alpha": 0.5
190        },
191        "mean": 0.0008234149281634018
192      },
193      {
194        "cfg": {
195          "lr": 0.003,
196          "weight_decay": 0.0001,
197          "alpha": 1.0
198        },
199        "mean": 0.0008033880876610056
200      },
201      {
202        "cfg": {
203          "lr": 0.003,
204          "weight_decay": 0.0001,
205          "alpha": 2.0
206        },
207        "mean": 0.000783455150667578
208      },
209      {
210        "cfg": {
211          "lr": 0.01,
212          "weight_decay": 0.0,
213          "alpha": 0.5
214        },
215        "mean": 0.00044341940520098433
216      },
217      {
218        "cfg": {
219          "lr": 0.01,
220          "weight_decay": 0.0,
221          "alpha": 1.0
222        },
223        "mean": 0.0004603777633747086
224      },
225      {
226        "cfg": {
227          "lr": 0.01,
228          "weight_decay": 0.0,
229          "alpha": 2.0
230        },
231        "mean": 0.0004929038841510192
232      },
233      {
234        "cfg": {
235          "lr": 0.01,
236          "weight_decay": 0.0001,
237          "alpha": 0.5
238        },
239        "mean": 0.00046618811757070944
240      },
241      {
242        "cfg": {
243          "lr": 0.01,
244          "weight_decay": 0.0001,
245          "alpha": 1.0
246        },
247        "mean": 0.0004956323173246346
248      },
249      {
250        "cfg": {
251          "lr": 0.01,
252          "weight_decay": 0.0001,
253          "alpha": 2.0
254        },
255        "mean": 0.000535586521436926
256      }
257    ]
258  }
259}