OSL-QIB Contractive State Observer / bench_report.json
Mechanism confirmed, baseline not beaten
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 "gain": 0.0
11 },
12 "sweep": [
13 {
14 "cfg": {
15 "lr": 0.001,
16 "gain": 0.0
17 },
18 "mean": 0.001116242419811897
19 },
20 {
21 "cfg": {
22 "lr": 0.003,
23 "gain": 0.0
24 },
25 "mean": 0.0005947774989181198
26 },
27 {
28 "cfg": {
29 "lr": 0.01,
30 "gain": 0.0
31 },
32 "mean": 0.00012917288222524803
33 }
34 ],
35 "full": {
36 "mean": 0.00010188088299400988,
37 "std": 6.96920114960574e-05,
38 "per_seed": [
39 0.00021521856251638383,
40 0.00019868022354785353,
41 5.157031409908086e-05,
42 5.122242873767391e-05,
43 2.9590093618026003e-05,
44 7.646851008757949e-05,
45 4.320807784097269e-05,
46 0.00014908885350450873
47 ],
48 "n": 8
49 }
50 },
51 "idea": {
52 "mean": 5.228218060437939e-05,
53 "std": 4.3368261054941446e-05,
54 "per_seed": [
55 0.0001497500343248248,
56 9.526080248178914e-05,
57 2.5922774511855096e-05,
58 3.981342524639331e-05,
59 1.754087134031579e-05,
60 2.408532782283146e-05,
61 2.590940312074963e-05,
62 3.9974805986275896e-05
63 ],
64 "n": 8
65 },
66 "comparison": {
67 "delta_mean": -4.959870238963049e-05,
68 "idea_wins": 8,
69 "n_pairs": 8,
70 "per_seed_diffs": [
71 -6.546852819155902e-05,
72 -0.00010341942106606439,
73 -2.5647539587225765e-05,
74 -1.14090034912806e-05,
75 -1.2049222277710214e-05,
76 -5.238318226474803e-05,
77 -1.7298674720223062e-05,
78 -0.00010911404751823284
79 ],
80 "p_value": 0.0081,
81 "mde": 3.3421829705425835e-05,
82 "mde_rel_pct": 32.804809620064724,
83 "verdict": "idea better (significant)",
84 "system_worked": true
85 },
86 "mechanism_signature": {
87 "description": "trained GRUCell hidden-state observer correction on pendulum windows",
88 "prediction": "local trained recurrent map should contract latent perturbations in V=||e||^2",
89 "trained_model_measurement": {
90 "gain": 1.0,
91 "lr": 0.01,
92 "predicted_V_ratio_bound": 0.9832114113428112,
93 "observed_mean_V_ratio": 0.3183068037033081,
94 "observed_max_V_ratio": 0.39757975935935974,
95 "certificate_contraction_predicted": true,
96 "observed_contraction": true,
97 "confirmed": true
98 },
99 "idea_sweep": [
100 {
101 "cfg": {
102 "lr": 0.01,
103 "gain": 0.25
104 },
105 "result": {
106 "mean": 7.185172580648214e-05,
107 "std": 6.221648707698256e-05,
108 "per_seed": [
109 0.00020611767831724137,
110 0.00011917055235244334,
111 3.0685481760883704e-05,
112 2.7863250579684973e-05,
113 1.7419906725990586e-05,
114 4.1515966586302966e-05,
115 2.652378134371247e-05,
116 0.00010551718878559768
117 ],
118 "n": 8
119 }
120 },
121 {
122 "cfg": {
123 "lr": 0.01,
124 "gain": 0.5
125 },
126 "result": {
127 "mean": 6.361333316817763e-05,
128 "std": 5.613212133729689e-05,
129 "per_seed": [
130 0.00017784775991458446,
131 0.00013154245971236378,
132 2.7753183530876413e-05,
133 3.123701753793284e-05,
134 1.546565181342885e-05,
135 2.7983382096863352e-05,
136 2.439024865452666e-05,
137 7.268696208484471e-05
138 ],
139 "n": 8
140 }
141 },
142 {
143 "cfg": {
144 "lr": 0.01,
145 "gain": 1.0
146 },
147 "result": {
148 "mean": 5.228218060437939e-05,
149 "std": 4.3368261054941446e-05,
150 "per_seed": [
151 0.0001497500343248248,
152 9.526080248178914e-05,
153 2.5922774511855096e-05,
154 3.981342524639331e-05,
155 1.754087134031579e-05,
156 2.408532782283146e-05,
157 2.590940312074963e-05,
158 3.9974805986275896e-05
159 ],
160 "n": 8
161 }
162 }
163 ]
164 },
165 "idea_best_cfg": {
166 "lr": 0.01,
167 "gain": 1.0
168 },
169 "protocol_notes": {
170 "track_match": "dynamics: controlled damped pendulum",
171 "epochs": 18,
172 "n_train": 800,
173 "n_test": 300,
174 "lr_union": [
175 0.001,
176 0.003,
177 0.01
178 ],
179 "paired_seeds": 8,
180 "baseline_sweep_seeds": 4
181 }
182}