Symmetry-Resolved Fourier Bifurcation Monitor / bench_report.json

Failed on benchmark

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_cfg": {
  9      "lr": 0.01,
 10      "clip": 1.0
 11    },
 12    "sweep": [
 13      {
 14        "cfg": {
 15          "lr": 0.001,
 16          "clip": 0.5
 17        },
 18        "mean": 0.25846289470791817
 19      },
 20      {
 21        "cfg": {
 22          "lr": 0.001,
 23          "clip": 1.0
 24        },
 25        "mean": 0.2586120665073395
 26      },
 27      {
 28        "cfg": {
 29          "lr": 0.001,
 30          "clip": 5.0
 31        },
 32        "mean": 0.264630239456892
 33      },
 34      {
 35        "cfg": {
 36          "lr": 0.003,
 37          "clip": 0.5
 38        },
 39        "mean": 0.010474085924215615
 40      },
 41      {
 42        "cfg": {
 43          "lr": 0.003,
 44          "clip": 1.0
 45        },
 46        "mean": 0.009937021299265325
 47      },
 48      {
 49        "cfg": {
 50          "lr": 0.003,
 51          "clip": 5.0
 52        },
 53        "mean": 0.012804438127204776
 54      },
 55      {
 56        "cfg": {
 57          "lr": 0.01,
 58          "clip": 0.5
 59        },
 60        "mean": 0.011152858496643603
 61      },
 62      {
 63        "cfg": {
 64          "lr": 0.01,
 65          "clip": 1.0
 66        },
 67        "mean": 0.00783492624759674
 68      },
 69      {
 70        "cfg": {
 71          "lr": 0.01,
 72          "clip": 5.0
 73        },
 74        "mean": 0.008992245304398239
 75      }
 76    ],
 77    "full": {
 78      "mean": 0.009677478694356978,
 79      "std": 0.0029150067189096305,
 80      "per_seed": [
 81        0.010251475498080254,
 82        0.005092725623399019,
 83        0.009792184457182884,
 84        0.006203319411724806,
 85        0.008501410484313965,
 86        0.010244788601994514,
 87        0.014377173036336899,
 88        0.012956752441823483
 89      ],
 90      "n": 8
 91    }
 92  },
 93  "idea": {
 94    "mean": 0.009327424864750355,
 95    "std": 0.004120870426856388,
 96    "per_seed": [
 97      0.01298802811652422,
 98      0.0055983662605285645,
 99      0.01386063452810049,
100      0.00352195231243968,
101      0.00725489342585206,
102      0.0069480095990002155,
103      0.00851170253008604,
104      0.015935812145471573
105    ],
106    "n": 8
107  },
108  "comparison": {
109    "delta_mean": -0.00035005382960662246,
110    "idea_wins": 4,
111    "n_pairs": 8,
112    "per_seed_diffs": [
113      0.002736552618443966,
114      0.0005056406371295452,
115      0.004068450070917606,
116      -0.0026813670992851257,
117      -0.0012465170584619045,
118      -0.003296779002994299,
119      -0.005865470506250858,
120      0.0029790597036480904
121    ],
122    "p_value": 0.78655,
123    "mde": 0.0029343297181013495,
124    "mde_rel_pct": 30.321221164892698,
125    "verdict": "no significant win",
126    "system_worked": false
127  },
128  "mechanism_signature": {
129    "prediction": "smallest Fourier recurrent margin predicts dominant hidden-channel Fourier sector",
130    "predicted_vs_observed": [
131      {
132        "seed": 0,
133        "metric": 0.01298802811652422,
134        "predicted_sector": 5,
135        "observed_sector": 29,
136        "predicted_margin": 0.9399786254849777
137      },
138      {
139        "seed": 1,
140        "metric": 0.0055983662605285645,
141        "predicted_sector": 56,
142        "observed_sector": 20,
143        "predicted_margin": 0.9352336144834422
144      },
145      {
146        "seed": 2,
147        "metric": 0.01386063452810049,
148        "predicted_sector": 62,
149        "observed_sector": 14,
150        "predicted_margin": 0.9405557241088474
151      },
152      {
153        "seed": 3,
154        "metric": 0.00352195231243968,
155        "predicted_sector": 22,
156        "observed_sector": 26,
157        "predicted_margin": 0.9286565991151029
158      },
159      {
160        "seed": 4,
161        "metric": 0.00725489342585206,
162        "predicted_sector": 16,
163        "observed_sector": 2,
164        "predicted_margin": 0.9472155426115653
165      },
166      {
167        "seed": 5,
168        "metric": 0.0069480095990002155,
169        "predicted_sector": 59,
170        "observed_sector": 18,
171        "predicted_margin": 0.9560543716567116
172      },
173      {
174        "seed": 6,
175        "metric": 0.00851170253008604,
176        "predicted_sector": 37,
177        "observed_sector": 30,
178        "predicted_margin": 0.9501085228512387
179      },
180      {
181        "seed": 7,
182        "metric": 0.015935812145471573,
183        "predicted_sector": 8,
184        "observed_sector": 0,
185        "predicted_margin": 0.9243731275838336
186      }
187    ],
188    "agreement_fraction": 0.0,
189    "confirmed": false
190  },
191  "idea_sweep": [
192    {
193      "cfg": {
194        "lr": 0.001,
195        "margin0": 0.2
196      },
197      "full": {
198        "mean": 0.2551296819001436,
199        "std": 0.06966527729488384,
200        "per_seed": [
201          0.1645231992006302,
202          0.286499947309494,
203          0.23894044756889343,
204          0.3685573637485504,
205          0.13928723335266113,
206          0.2740314304828644,
207          0.2623410224914551,
208          0.30685681104660034
209        ],
210        "n": 8
211      }
212    },
213    {
214      "cfg": {
215        "lr": 0.003,
216        "margin0": 0.2
217      },
218      "full": {
219        "mean": 0.013181458110921085,
220        "std": 0.006411194717789758,
221        "per_seed": [
222          0.01743449829518795,
223          0.00642412668094039,
224          0.02289336547255516,
225          0.004465762060135603,
226          0.010089418850839138,
227          0.00904592964798212,
228          0.013677984476089478,
229          0.02142057940363884
230        ],
231        "n": 8
232      }
233    },
234    {
235      "cfg": {
236        "lr": 0.01,
237        "margin0": 0.2
238      },
239      "full": {
240        "mean": 0.009327424864750355,
241        "std": 0.004120870426856388,
242        "per_seed": [
243          0.01298802811652422,
244          0.0055983662605285645,
245          0.01386063452810049,
246          0.00352195231243968,
247          0.00725489342585206,
248          0.0069480095990002155,
249          0.00851170253008604,
250          0.015935812145471573
251        ],
252        "n": 8
253      }
254    }
255  ],
256  "track_justification": "Dynamics is structurally matched: the task is an actuated pendulum rollout and the intervention monitors recurrent stability sectors.",
257  "protocol": {
258    "paired_seeds": [
259      0,
260      1,
261      2,
262      3,
263      4,
264      5,
265      6,
266      7
267    ],
268    "baseline_sweep_seeds": [
269      0,
270      1,
271      2,
272      3
273    ],
274    "epochs": 5,
275    "batch": 128,
276    "baseline_grid": [
277      {
278        "lr": 0.001,
279        "clip": 0.5
280      },
281      {
282        "lr": 0.001,
283        "clip": 1.0
284      },
285      {
286        "lr": 0.001,
287        "clip": 5.0
288      },
289      {
290        "lr": 0.003,
291        "clip": 0.5
292      },
293      {
294        "lr": 0.003,
295        "clip": 1.0
296      },
297      {
298        "lr": 0.003,
299        "clip": 5.0
300      },
301      {
302        "lr": 0.01,
303        "clip": 0.5
304      },
305      {
306        "lr": 0.01,
307        "clip": 1.0
308      },
309      {
310        "lr": 0.01,
311        "clip": 5.0
312      }
313    ],
314    "idea_grid": [
315      {
316        "lr": 0.001,
317        "margin0": 0.2
318      },
319      {
320        "lr": 0.003,
321        "margin0": 0.2
322      },
323      {
324        "lr": 0.01,
325        "margin0": 0.2
326      }
327    ]
328  }
329}