SOS Backup Shield for Learned Policies / bench_report.json
Mechanism confirmed, baseline not beaten
1{
2 "bench_version": 1,
3 "track": "robust_cbf_pendulum_policy",
4 "model": "local_mlp_tiny",
5 "metric_direction": "lower is better",
6 "n_seeds": 8,
7 "baseline": {
8 "best_cfg": {
9 "lr": 0.006,
10 "eps": 0.0,
11 "epochs": 15
12 },
13 "sweep": [
14 {
15 "cfg": {
16 "lr": 0.001,
17 "eps": 0.0,
18 "epochs": 15
19 },
20 "mean": 0.021794200874865055
21 },
22 {
23 "cfg": {
24 "lr": 0.003,
25 "eps": 0.0,
26 "epochs": 15
27 },
28 "mean": 0.002334954886464402
29 },
30 {
31 "cfg": {
32 "lr": 0.006,
33 "eps": 0.0,
34 "epochs": 15
35 },
36 "mean": 0.0018331437022425234
37 }
38 ],
39 "full": {
40 "mean": 0.0019978795608039945,
41 "std": 0.0005495084055684182,
42 "per_seed": [
43 0.0015628907131031156,
44 0.0015461408765986562,
45 0.0021467802580446005,
46 0.0020767629612237215,
47 0.0030877823010087013,
48 0.0011752876453101635,
49 0.0020312685519456863,
50 0.0023561231791973114
51 ],
52 "n": 8
53 }
54 },
55 "idea": {
56 "mean": 0.05749530019238591,
57 "std": 0.00835564815799482,
58 "per_seed": [
59 0.06218996271491051,
60 0.06785300374031067,
61 0.04927119240164757,
62 0.05358738824725151,
63 0.042678408324718475,
64 0.06384517252445221,
65 0.05430787801742554,
66 0.06622939556837082
67 ],
68 "n": 8,
69 "selection_sweep": [
70 {
71 "cfg": {
72 "lr": 0.001,
73 "eps": 0.15,
74 "epochs": 15
75 },
76 "mean": 0.05939904507249594
77 },
78 {
79 "cfg": {
80 "lr": 0.003,
81 "eps": 0.25,
82 "epochs": 15
83 },
84 "mean": 0.058225386776030064
85 },
86 {
87 "cfg": {
88 "lr": 0.006,
89 "eps": 0.35,
90 "epochs": 15
91 },
92 "mean": 0.06890939734876156
93 }
94 ]
95 },
96 "comparison": {
97 "delta_mean": 0.05549742063158192,
98 "idea_wins": 0,
99 "n_pairs": 8,
100 "per_seed_diffs": [
101 0.06062707200180739,
102 0.06630686286371201,
103 0.04712441214360297,
104 0.05151062528602779,
105 0.039590626023709774,
106 0.06266988487914205,
107 0.05227660946547985,
108 0.06387327238917351
109 ],
110 "p_value": 0.0081,
111 "mde": 0.007832081629823283,
112 "mde_rel_pct": 392.0197084688862,
113 "verdict": "idea worse (significant)",
114 "system_worked": false
115 },
116 "mechanism_signature": {
117 "prediction": {
118 "margin_slope": -2.0,
119 "lower_bound_shift_per_epsilon": 1.0
120 },
121 "observed_from_trained_models": {
122 "margin_slope_mean": -2.0,
123 "lower_shift_mean": 1.0,
124 "mean_intervention": 1.705237858260225,
125 "feasible_fraction_mean": 1.0,
126 "n_models": 8
127 },
128 "confirmed": true
129 },
130 "custom_track": {
131 "name": "robust_cbf_pendulum_policy",
132 "file": "robust_cbf_bench.py",
133 "domain": "dynamics"
134 }
135}