Basin-Aware Hysteresis Guard / bench_report.json
Beats tuned baseline
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": 15,
11 "damping": 1.0
12 },
13 "sweep": [
14 {
15 "cfg": {
16 "lr": 0.001,
17 "epochs": 15,
18 "damping": 1.0
19 },
20 "mean": 0.016809451510198414
21 },
22 {
23 "cfg": {
24 "lr": 0.003,
25 "epochs": 15,
26 "damping": 1.0
27 },
28 "mean": 0.005665019620209932
29 },
30 {
31 "cfg": {
32 "lr": 0.01,
33 "epochs": 15,
34 "damping": 1.0
35 },
36 "mean": 0.0022826905187685043
37 }
38 ],
39 "full": {
40 "mean": 0.002359520163736306,
41 "std": 0.0007644030858149881,
42 "per_seed": [
43 0.001249853172339499,
44 0.002188401995226741,
45 0.002345386426895857,
46 0.0033471204806119204,
47 0.0033662922214716673,
48 0.0015826028538867831,
49 0.0030378259252756834,
50 0.0017586782341822982
51 ],
52 "n": 8
53 }
54 },
55 "idea": {
56 "mean": 0.0013422396659734659,
57 "std": 0.0003630924813985613,
58 "per_seed": [
59 0.001863177865743637,
60 0.001077942200936377,
61 0.001123852445743978,
62 0.0017142429715022445,
63 0.0011969851329922676,
64 0.0008323417860083282,
65 0.0017935478826984763,
66 0.0011358270421624184
67 ],
68 "n": 8,
69 "best_cfg": {
70 "lr": 0.01,
71 "epochs": 15,
72 "damping": 0.5,
73 "eps": 0.15,
74 "delta": 0.35
75 },
76 "sweep": [
77 {
78 "cfg": {
79 "lr": 0.001,
80 "epochs": 15,
81 "damping": 0.5,
82 "eps": 0.15,
83 "delta": 0.35
84 },
85 "mean": 0.01401310320943594
86 },
87 {
88 "cfg": {
89 "lr": 0.001,
90 "epochs": 15,
91 "damping": 0.7,
92 "eps": 0.25,
93 "delta": 0.35
94 },
95 "mean": 0.015718030161224306
96 },
97 {
98 "cfg": {
99 "lr": 0.001,
100 "epochs": 15,
101 "damping": 0.35,
102 "eps": 0.2,
103 "delta": 0.35
104 },
105 "mean": 0.013062152196653187
106 },
107 {
108 "cfg": {
109 "lr": 0.003,
110 "epochs": 15,
111 "damping": 0.5,
112 "eps": 0.15,
113 "delta": 0.35
114 },
115 "mean": 0.004635538120055571
116 },
117 {
118 "cfg": {
119 "lr": 0.003,
120 "epochs": 15,
121 "damping": 0.7,
122 "eps": 0.25,
123 "delta": 0.35
124 },
125 "mean": 0.004597917490173131
126 },
127 {
128 "cfg": {
129 "lr": 0.003,
130 "epochs": 15,
131 "damping": 0.35,
132 "eps": 0.2,
133 "delta": 0.35
134 },
135 "mean": 0.004916691017569974
136 },
137 {
138 "cfg": {
139 "lr": 0.01,
140 "epochs": 15,
141 "damping": 0.5,
142 "eps": 0.15,
143 "delta": 0.35
144 },
145 "mean": 0.0014448038709815592
146 },
147 {
148 "cfg": {
149 "lr": 0.01,
150 "epochs": 15,
151 "damping": 0.7,
152 "eps": 0.25,
153 "delta": 0.35
154 },
155 "mean": 0.0021654691954609007
156 },
157 {
158 "cfg": {
159 "lr": 0.01,
160 "epochs": 15,
161 "damping": 0.35,
162 "eps": 0.2,
163 "delta": 0.35
164 },
165 "mean": 0.0016093126323539764
166 }
167 ]
168 },
169 "comparison": {
170 "delta_mean": -0.0010172804977628402,
171 "idea_wins": 7,
172 "n_pairs": 8,
173 "per_seed_diffs": [
174 0.0006133246934041381,
175 -0.0011104597942903638,
176 -0.0012215339811518788,
177 -0.0016328775091096759,
178 -0.0021693070884793997,
179 -0.0007502610678784549,
180 -0.001244278042577207,
181 -0.0006228511920198798
182 ],
183 "p_value": 0.01485,
184 "mde": 0.0006846987059140442,
185 "mde_rel_pct": 29.018557096364123,
186 "verdict": "idea better (significant)",
187 "system_worked": true
188 },
189 "mechanism_signature": {
190 "prediction": "guard should increase perturbation recovery and reduce local recurrent gain",
191 "observed": [
192 {
193 "seed": 0,
194 "type": "baseline",
195 "recovery": 0.0,
196 "recurrent_norm": 2.6400585174560547
197 },
198 {
199 "seed": 0,
200 "type": "idea",
201 "recovery": 0.0,
202 "recurrent_norm": 2.653230905532837
203 },
204 {
205 "seed": 1,
206 "type": "baseline",
207 "recovery": 0.0,
208 "recurrent_norm": 1.7831625938415527
209 },
210 {
211 "seed": 1,
212 "type": "idea",
213 "recovery": 0.0,
214 "recurrent_norm": 1.9542343616485596
215 },
216 {
217 "seed": 2,
218 "type": "baseline",
219 "recovery": 0.0,
220 "recurrent_norm": 2.5040619373321533
221 },
222 {
223 "seed": 2,
224 "type": "idea",
225 "recovery": 0.0,
226 "recurrent_norm": 3.001326322555542
227 },
228 {
229 "seed": 3,
230 "type": "baseline",
231 "recovery": 0.0,
232 "recurrent_norm": 3.297189235687256
233 },
234 {
235 "seed": 3,
236 "type": "idea",
237 "recovery": 0.0,
238 "recurrent_norm": 2.7743632793426514
239 },
240 {
241 "seed": 4,
242 "type": "baseline",
243 "recovery": 0.0,
244 "recurrent_norm": 2.4658541679382324
245 },
246 {
247 "seed": 4,
248 "type": "idea",
249 "recovery": 0.0,
250 "recurrent_norm": 3.08595609664917
251 },
252 {
253 "seed": 5,
254 "type": "baseline",
255 "recovery": 0.0,
256 "recurrent_norm": 2.731605052947998
257 },
258 {
259 "seed": 5,
260 "type": "idea",
261 "recovery": 0.0,
262 "recurrent_norm": 3.1894450187683105
263 },
264 {
265 "seed": 6,
266 "type": "baseline",
267 "recovery": 0.0,
268 "recurrent_norm": 2.8795089721679688
269 },
270 {
271 "seed": 6,
272 "type": "idea",
273 "recovery": 0.0,
274 "recurrent_norm": 3.512533187866211
275 },
276 {
277 "seed": 7,
278 "type": "baseline",
279 "recovery": 0.0,
280 "recurrent_norm": 2.337895154953003
281 },
282 {
283 "seed": 7,
284 "type": "idea",
285 "recovery": 0.0,
286 "recurrent_norm": 3.0789902210235596
287 }
288 ],
289 "predicted_recovery_higher": true,
290 "observed_recovery_delta": 0.0,
291 "confirmed": false
292 },
293 "structural_match": "Dynamics track: actuated pendulum rollout directly tests recurrent stability/control."
294}