Adjoint-Weak Fractional Residuals / bench_report.json

Mechanism confirmed, baseline not beaten

Raw ⬇ ZIP
  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_hyperparams": {
  9      "lr": 0.01,
 10      "epochs": 16
 11    },
 12    "sweep": [
 13      {
 14        "lr": 0.001,
 15        "seed": 0,
 16        "mse": 0.0016373801045119762
 17      },
 18      {
 19        "lr": 0.001,
 20        "seed": 1,
 21        "mse": 0.0020086602307856083
 22      },
 23      {
 24        "lr": 0.001,
 25        "seed": 2,
 26        "mse": 0.001998279709368944
 27      },
 28      {
 29        "lr": 0.001,
 30        "seed": 3,
 31        "mse": 0.0037652531173080206
 32      },
 33      {
 34        "lr": 0.001,
 35        "seed": 4,
 36        "mse": 0.002077344572171569
 37      },
 38      {
 39        "lr": 0.001,
 40        "seed": 5,
 41        "mse": 0.002141711302101612
 42      },
 43      {
 44        "lr": 0.001,
 45        "seed": 6,
 46        "mse": 0.0035790614783763885
 47      },
 48      {
 49        "lr": 0.001,
 50        "seed": 7,
 51        "mse": 0.00521521782502532
 52      },
 53      {
 54        "lr": 0.003,
 55        "seed": 0,
 56        "mse": 0.0012488373322412372
 57      },
 58      {
 59        "lr": 0.003,
 60        "seed": 1,
 61        "mse": 0.0012570498511195183
 62      },
 63      {
 64        "lr": 0.003,
 65        "seed": 2,
 66        "mse": 0.0009830091148614883
 67      },
 68      {
 69        "lr": 0.003,
 70        "seed": 3,
 71        "mse": 0.0018271016888320446
 72      },
 73      {
 74        "lr": 0.003,
 75        "seed": 4,
 76        "mse": 0.0011851947056129575
 77      },
 78      {
 79        "lr": 0.003,
 80        "seed": 5,
 81        "mse": 0.0011170521611347795
 82      },
 83      {
 84        "lr": 0.003,
 85        "seed": 6,
 86        "mse": 0.0019040972692891955
 87      },
 88      {
 89        "lr": 0.003,
 90        "seed": 7,
 91        "mse": 0.0017286647344008088
 92      },
 93      {
 94        "lr": 0.01,
 95        "seed": 0,
 96        "mse": 0.0005306488601490855
 97      },
 98      {
 99        "lr": 0.01,
100        "seed": 1,
101        "mse": 0.0006936120917089283
102      },
103      {
104        "lr": 0.01,
105        "seed": 2,
106        "mse": 0.0006226884433999658
107      },
108      {
109        "lr": 0.01,
110        "seed": 3,
111        "mse": 0.0011054250644519925
112      },
113      {
114        "lr": 0.01,
115        "seed": 4,
116        "mse": 0.0005887902225367725
117      },
118      {
119        "lr": 0.01,
120        "seed": 5,
121        "mse": 0.0005850467132404447
122      },
123      {
124        "lr": 0.01,
125        "seed": 6,
126        "mse": 0.000752387335523963
127      },
128      {
129        "lr": 0.01,
130        "seed": 7,
131        "mse": 0.0007097375928424299
132      }
133    ],
134    "full": {
135      "per_seed": [
136        0.0005306488601490855,
137        0.0006936120917089283,
138        0.0006226884433999658,
139        0.0011054250644519925,
140        0.0005887902225367725,
141        0.0005850467132404447,
142        0.000752387335523963,
143        0.0007097375928424299
144      ],
145      "mean": 0.0006985420404816978,
146      "std": 0.00016862358035770679
147    }
148  },
149  "idea": {
150    "best_hyperparams": {
151      "lr": 0.01,
152      "alpha": 0.8,
153      "epochs": 16
154    },
155    "sweep": [
156      {
157        "lr": 0.001,
158        "seed": 0,
159        "mse": 0.0016854512505233288
160      },
161      {
162        "lr": 0.001,
163        "seed": 1,
164        "mse": 0.0019817491993308067
165      },
166      {
167        "lr": 0.001,
168        "seed": 2,
169        "mse": 0.002013311255723238
170      },
171      {
172        "lr": 0.001,
173        "seed": 3,
174        "mse": 0.0036906730383634567
175      },
176      {
177        "lr": 0.001,
178        "seed": 4,
179        "mse": 0.002067329129204154
180      },
181      {
182        "lr": 0.001,
183        "seed": 5,
184        "mse": 0.00208784150891006
185      },
186      {
187        "lr": 0.001,
188        "seed": 6,
189        "mse": 0.003548926906660199
190      },
191      {
192        "lr": 0.001,
193        "seed": 7,
194        "mse": 0.0051675597205758095
195      },
196      {
197        "lr": 0.003,
198        "seed": 0,
199        "mse": 0.0012303238036110997
200      },
201      {
202        "lr": 0.003,
203        "seed": 1,
204        "mse": 0.0012324448907747865
205      },
206      {
207        "lr": 0.003,
208        "seed": 2,
209        "mse": 0.000958611664827913
210      },
211      {
212        "lr": 0.003,
213        "seed": 3,
214        "mse": 0.001847315113991499
215      },
216      {
217        "lr": 0.003,
218        "seed": 4,
219        "mse": 0.0011463016271591187
220      },
221      {
222        "lr": 0.003,
223        "seed": 5,
224        "mse": 0.0010947113623842597
225      },
226      {
227        "lr": 0.003,
228        "seed": 6,
229        "mse": 0.001916371053084731
230      },
231      {
232        "lr": 0.003,
233        "seed": 7,
234        "mse": 0.0017388231353834271
235      },
236      {
237        "lr": 0.01,
238        "seed": 0,
239        "mse": 0.0005385181866586208
240      },
241      {
242        "lr": 0.01,
243        "seed": 1,
244        "mse": 0.0007211403572000563
245      },
246      {
247        "lr": 0.01,
248        "seed": 2,
249        "mse": 0.000598620914388448
250      },
251      {
252        "lr": 0.01,
253        "seed": 3,
254        "mse": 0.001104937051422894
255      },
256      {
257        "lr": 0.01,
258        "seed": 4,
259        "mse": 0.000581878877710551
260      },
261      {
262        "lr": 0.01,
263        "seed": 5,
264        "mse": 0.0006115676951594651
265      },
266      {
267        "lr": 0.01,
268        "seed": 6,
269        "mse": 0.0007518851780332625
270      },
271      {
272        "lr": 0.01,
273        "seed": 7,
274        "mse": 0.000709107902366668
275      }
276    ],
277    "per_seed": [
278      0.0005385181866586208,
279      0.0007211403572000563,
280      0.000598620914388448,
281      0.001104937051422894,
282      0.000581878877710551,
283      0.0006115676951594651,
284      0.0007518851780332625,
285      0.000709107902366668
286    ],
287    "mean": 0.0007022070203674957,
288    "std": 0.00016781088891265367
289  },
290  "comparison": {
291    "delta_mean": 3.6649798857979476e-06,
292    "idea_wins": 5,
293    "n_pairs": 8,
294    "per_seed_diffs": [
295      7.869326509535313e-06,
296      2.7528265491127968e-05,
297      -2.4067529011517763e-05,
298      -4.880130290985107e-07,
299      -6.911344826221466e-06,
300      2.6520981919020414e-05,
301      -5.021574907004833e-07,
302      -6.296904757618904e-07
303    ],
304    "p_value": 0.52365,
305    "mde": 1.4286583623107655e-05,
306    "mde_rel_pct": 2.0452002592794516,
307    "verdict": "no measurable effect",
308    "system_worked": false
309  },
310  "mechanism_signature": {
311    "trained_model_behavior": {
312      "mean_weak_prediction_correlation": 0.9371926834185919,
313      "mean_strong_prediction_correlation": 0.7029504552483559
314    },
315    "prediction": "weak residual should couple to trajectory while reducing derivative noise amplification",
316    "confirmed": true
317  }
318}