Delay-Budget Controller for Coupled Training / bench_report.json
Failed on benchmark
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 "epochs": 16
11 },
12 "sweep": [
13 {
14 "cfg": {
15 "lr": 0.001,
16 "epochs": 16
17 },
18 "mean": 0.003169315925333649
19 },
20 {
21 "cfg": {
22 "lr": 0.003,
23 "epochs": 16
24 },
25 "mean": 0.0013352147216210142
26 },
27 {
28 "cfg": {
29 "lr": 0.01,
30 "epochs": 16
31 },
32 "mean": 0.0008695064898347482
33 }
34 ],
35 "full": {
36 "mean": 0.0007919272829894908,
37 "std": 0.0001468373452544164,
38 "per_seed": [
39 0.0006091617397032678,
40 0.0008883575792424381,
41 0.0009651132277213037,
42 0.0010153934126719832,
43 0.000822118716314435,
44 0.0005955996457487345,
45 0.000717962218914181,
46 0.0007217117235995829
47 ],
48 "n": 8
49 }
50 },
51 "idea": {
52 "per_seed": [
53 0.0006370572373270988,
54 0.0005663432530127466,
55 0.0006719666416756809,
56 0.00119167345110327,
57 0.00044812753912992775,
58 0.00045454775681719184,
59 0.000672604946885258,
60 0.0006867195479571819
61 ],
62 "config": {
63 "lr": 0.01,
64 "delay": 3,
65 "beta": 0.9
66 },
67 "details": [
68 {
69 "metric": 0.0006370572373270988,
70 "history": [
71 0.47085492208600044,
72 0.0687079544365406,
73 0.05355181962251663,
74 0.013942306935787201,
75 0.011403620205819608,
76 0.01611782468855381,
77 0.005046065337955952,
78 0.006925272494554519,
79 0.003115472113713622,
80 0.0035618337569758296,
81 0.0017404001951217651,
82 0.001538055301643908,
83 0.0011132086324505508,
84 0.0007999823521822691,
85 0.0006484792870469392,
86 0.0005360299820313231
87 ],
88 "effective_delay": 0,
89 "effective_beta": 0.9,
90 "loss_spectral_proxy": 0.09964848842129591,
91 "gain_proxy": 0.22634743220793707
92 },
93 {
94 "metric": 0.0005663432530127466,
95 "history": [
96 0.5248998469114303,
97 0.07719431262463332,
98 0.060356385707855224,
99 0.020550017915666102,
100 0.02778526324778795,
101 0.006836556419730186,
102 0.01419046189635992,
103 0.0028096558526158333,
104 0.00680495067499578,
105 0.001783308652229607,
106 0.004530294346623122,
107 0.0010322351753711701,
108 0.0015835526073351503,
109 0.0008021262101829052,
110 0.0010320117487572133,
111 0.00069854547502473
112 ],
113 "effective_delay": 0,
114 "effective_beta": 0.9,
115 "loss_spectral_proxy": 0.11097327441321576,
116 "gain_proxy": 0.23964861900335657
117 },
118 {
119 "metric": 0.0006719666416756809,
120 "history": [
121 0.40221752375364306,
122 0.07661814518272876,
123 0.03775709956884384,
124 0.015351182222366333,
125 0.016443840116262435,
126 0.011262819766998292,
127 0.005225430428981781,
128 0.0064912005513906475,
129 0.0029075471498072147,
130 0.0035697654401883484,
131 0.00184590395539999,
132 0.0013333498127758503,
133 0.0012415370717644692,
134 0.0006822893698699772,
135 0.0008538498403504491,
136 0.0004581497609615326
137 ],
138 "effective_delay": 0,
139 "effective_beta": 0.9,
140 "loss_spectral_proxy": 0.08056035702057981,
141 "gain_proxy": 0.23553890194469956
142 },
143 {
144 "metric": 0.00119167345110327,
145 "history": [
146 0.5213569915294647,
147 0.08585628386586905,
148 0.05548876643180847,
149 0.01533776342868805,
150 0.026741317473351955,
151 0.0076001503318548205,
152 0.013853808455169201,
153 0.004096998870372772,
154 0.006211945214308798,
155 0.0028600253909826278,
156 0.0034357033739797773,
157 0.0013145605847239494,
158 0.0014876204589381813,
159 0.0012321077100932599,
160 0.0010768138803541661,
161 0.0009121003281325102
162 ],
163 "effective_delay": 0,
164 "effective_beta": 0.9,
165 "loss_spectral_proxy": 0.10794722299288356,
166 "gain_proxy": 0.2332885087540775
167 },
168 {
169 "metric": 0.00044812753912992775,
170 "history": [
171 0.4003379839658737,
172 0.0936926954984665,
173 0.06415947467088699,
174 0.011134160310029983,
175 0.019605498537421228,
176 0.011598305366933346,
177 0.005959209538996219,
178 0.0064916802942752835,
179 0.0022103236243128776,
180 0.004928581081330776,
181 0.0015571040101349353,
182 0.0025297450833022595,
183 0.0013386529870331287,
184 0.000932006910443306,
185 0.0012357496097683907,
186 0.000636223186738789
187 ],
188 "effective_delay": 0,
189 "effective_beta": 0.9,
190 "loss_spectral_proxy": 0.07627690377854396,
191 "gain_proxy": 0.22375892249812535
192 },
193 {
194 "metric": 0.00045454775681719184,
195 "history": [
196 0.5347963106632233,
197 0.08657035626471042,
198 0.06304869443178177,
199 0.011364500075578689,
200 0.01554770177230239,
201 0.00924489352852106,
202 0.0039030701480805875,
203 0.004997533429414034,
204 0.0027877656742930413,
205 0.0023218371625989675,
206 0.002174343951046467,
207 0.001435691132210195,
208 0.0014137867919635028,
209 0.0010988898482173681,
210 0.0007053272798657418,
211 0.0005926695314701646
212 ],
213 "effective_delay": 0,
214 "effective_beta": 0.9,
215 "loss_spectral_proxy": 0.11111526176250495,
216 "gain_proxy": 0.20464232012118197
217 },
218 {
219 "metric": 0.000672604946885258,
220 "history": [
221 0.5318101173639298,
222 0.07845896065235138,
223 0.0610912823677063,
224 0.019874949678778647,
225 0.01623648837208748,
226 0.01566230658441782,
227 0.005753589915111661,
228 0.007185849454253912,
229 0.003104981817305088,
230 0.002994525693356991,
231 0.0018818091508001089,
232 0.0013551599718630313,
233 0.0013808830687776208,
234 0.0009053897112607956,
235 0.0008721279329620302,
236 0.0006727166124619544
237 ],
238 "effective_delay": 0,
239 "effective_beta": 0.9,
240 "loss_spectral_proxy": 0.11220600557712056,
241 "gain_proxy": 0.25874861419365247
242 },
243 {
244 "metric": 0.0006867195479571819,
245 "history": [
246 0.5368606805801391,
247 0.07741906464099885,
248 0.03744138062000275,
249 0.02687046892940998,
250 0.010084275305271149,
251 0.013893421962857247,
252 0.004783628340810538,
253 0.005570829231292009,
254 0.0029055914096534252,
255 0.002362888865172863,
256 0.0016249716910533606,
257 0.0010325243696570397,
258 0.0009152249037288129,
259 0.0007940076058730483,
260 0.0005743569927290082,
261 0.0005937276943586766
262 ],
263 "effective_delay": 0,
264 "effective_beta": 0.9,
265 "loss_spectral_proxy": 0.11372852844335014,
266 "gain_proxy": 0.23705882672768308
267 }
268 ]
269 },
270 "comparison": {
271 "delta_mean": -0.0001257972362509463,
272 "idea_wins": 6,
273 "n_pairs": 8,
274 "per_seed_diffs": [
275 2.7895497623831034e-05,
276 -0.0003220143262296915,
277 -0.0002931465860456228,
278 0.0001762800384312868,
279 -0.00037399117718450725,
280 -0.00014105188893154263,
281 -4.5357272028923035e-05,
282 -3.499217564240098e-05
283 ],
284 "p_value": 0.11805,
285 "mde": 0.00016050424424104764,
286 "mde_rel_pct": 20.26754825710148,
287 "verdict": "no significant win",
288 "system_worked": false
289 },
290 "math_check": {
291 "omega": 0.3089419282163646,
292 "tau_crit": 6.289312566882576,
293 "relative_magnitude_error": 5.234701561107613e-13,
294 "phase_error": 0.0,
295 "margin": 1.0
296 },
297 "structural_match": "dynamics control/stability track; same rnn_small system and task",
298 "custom_track": null,
299 "mechanism_signature": {
300 "predicted": "delay reduction when estimated tau exceeds 0.8 tau_crit; beta reduction in filter branch",
301 "observed_effective_delays": [
302 0,
303 0,
304 0,
305 0,
306 0,
307 0,
308 0,
309 0
310 ],
311 "observed_effective_betas": [
312 0.9,
313 0.9,
314 0.9,
315 0.9,
316 0.9,
317 0.9,
318 0.9,
319 0.9
320 ],
321 "observed_loss_spectral_proxy": [
322 0.09964848842129591,
323 0.11097327441321576,
324 0.08056035702057981,
325 0.10794722299288356,
326 0.07627690377854396,
327 0.11111526176250495,
328 0.11220600557712056,
329 0.11372852844335014
330 ],
331 "confirmed": true
332 }
333}