Farkas-Certified Neural Safety Shield / 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": 0.00665
10 },
11 "sweep": [
12 {
13 "cfg": {
14 "lr": 0.001
15 },
16 "mean": 0.0009722943214001134
17 },
18 {
19 "cfg": {
20 "lr": 0.002
21 },
22 "mean": 0.0006509074009954929
23 },
24 {
25 "cfg": {
26 "lr": 0.003
27 },
28 "mean": 0.0004595852587954141
29 },
30 {
31 "cfg": {
32 "lr": 0.004
33 },
34 "mean": 0.0003482927131699398
35 },
36 {
37 "cfg": {
38 "lr": 0.005
39 },
40 "mean": 0.00024859366021701135
41 },
42 {
43 "cfg": {
44 "lr": 0.00067
45 },
46 "mean": 0.0011799605708802119
47 },
48 {
49 "cfg": {
50 "lr": 0.00134
51 },
52 "mean": 0.0008329209085786715
53 },
54 {
55 "cfg": {
56 "lr": 0.00268
57 },
58 "mean": 0.0005083113646833226
59 },
60 {
61 "cfg": {
62 "lr": 0.00402
63 },
64 "mean": 0.0003466853668214753
65 },
66 {
67 "cfg": {
68 "lr": 0.00536
69 },
70 "mean": 0.00022448933304985985
71 },
72 {
73 "cfg": {
74 "lr": 0.00665
75 },
76 "mean": 0.00013951877372164745
77 }
78 ],
79 "full": {
80 "mean": 0.00011474007897049887,
81 "std": 4.026902293849407e-05,
82 "per_seed": [
83 0.0001932993473019451,
84 0.0001438440231140703,
85 8.013794285943732e-05,
86 0.0001407937816111371,
87 0.00011875866039190441,
88 8.25471433927305e-05,
89 6.331295298878103e-05,
90 9.522678010398522e-05
91 ],
92 "n": 8
93 }
94 },
95 "idea": {
96 "mean": 0.05237803841009736,
97 "std": 0.008875649004641814,
98 "per_seed": [
99 0.04106063395738602,
100 0.058054666966199875,
101 0.0539846234023571,
102 0.06823215633630753,
103 0.05530509725213051,
104 0.03805254027247429,
105 0.05319170653820038,
106 0.05114288255572319
107 ],
108 "n": 8
109 },
110 "comparison": {
111 "delta_mean": 0.05226329833112686,
112 "idea_wins": 0,
113 "n_pairs": 8,
114 "per_seed_diffs": [
115 0.04086733461008407,
116 0.057910822943085805,
117 0.053904485459497664,
118 0.06809136255469639,
119 0.055186338591738604,
120 0.03796999312908156,
121 0.0531283935852116,
122 0.051047655775619205
123 ],
124 "p_value": 0.0081,
125 "mde": 0.007932781978214037,
126 "mde_rel_pct": 6913.697506042031,
127 "verdict": "idea worse (significant)",
128 "system_worked": false
129 },
130 "mechanism_signature": {
131 "farkas_shield": {
132 "constraint_count_full": 64,
133 "constraint_count_retained": 2,
134 "retained_fraction": 0.03125,
135 "certificate_max_residual": 0.0,
136 "certificate_max_offset_violation": 0.0,
137 "trained_test_samples": 400,
138 "full_reduced_max_output_difference": 0.0,
139 "full_reduced_decision_disagreement": 0.0,
140 "predicted_zero_disagreement_observed": true,
141 "confirmed": true
142 }
143 },
144 "idea_sweep": [
145 {
146 "cfg": {
147 "lr": 0.00665
148 },
149 "result": {
150 "mean": 0.05238204402849078,
151 "std": 0.008873981965744437,
152 "per_seed": [
153 0.04106547310948372,
154 0.0580538772046566,
155 0.0539894662797451,
156 0.06823457032442093,
157 0.055308111011981964,
158 0.03805993124842644,
159 0.053193967789411545,
160 0.05115095525979996
161 ],
162 "n": 8
163 }
164 },
165 {
166 "cfg": {
167 "lr": 0.0044555
168 },
169 "result": {
170 "mean": 0.05238870158791542,
171 "std": 0.008873361924843962,
172 "per_seed": [
173 0.04107039421796799,
174 0.0580560676753521,
175 0.053996019065380096,
176 0.06824300438165665,
177 0.055311888456344604,
178 0.03807036951184273,
179 0.053200963884592056,
180 0.05116090551018715
181 ],
182 "n": 8
183 }
184 },
185 {
186 "cfg": {
187 "lr": 0.0088445
188 },
189 "result": {
190 "mean": 0.05237803841009736,
191 "std": 0.008875649004641814,
192 "per_seed": [
193 0.04106063395738602,
194 0.058054666966199875,
195 0.0539846234023571,
196 0.06823215633630753,
197 0.05530509725213051,
198 0.03805254027247429,
199 0.05319170653820038,
200 0.05114288255572319
201 ],
202 "n": 8
203 }
204 }
205 ],
206 "protocol_notes": {
207 "task_match": "controlled pendulum dynamics; safety/control idea",
208 "epochs": 15,
209 "n_train": 1200,
210 "n_test": 400,
211 "idea_uses_same_base_architecture": true
212 }
213}