Residual-Tightened Neural Safety Shield / results.json
Mechanism confirmed, baseline not beaten
1{
2 "seed": 544,
3 "sigma_r": 0.2,
4 "kappa": 1.0,
5 "math_check": {
6 "d": [
7 0.0,
8 0.02,
9 0.05,
10 0.2,
11 0.8
12 ],
13 "r_at_ema_equal_d": [
14 0.0,
15 0.09090909090909091,
16 0.2,
17 0.5,
18 0.8
19 ],
20 "tightening_margin": [
21 0.0,
22 0.09090909090909091,
23 0.2,
24 0.5,
25 0.8
26 ],
27 "r_monotone": true,
28 "max_projection_constraint_violation": 1.1102230246251565e-16,
29 "constraint_satisfied": true
30 },
31 "experiment": {
32 "nominal_unshielded": {
33 "violation_rate": 0.0,
34 "mean_action_deviation": 0.0,
35 "mean_reward": -0.005490534225707342
36 },
37 "nominal_fixed": {
38 "violation_rate": 0.0,
39 "mean_action_deviation": 0.3875413736135295,
40 "mean_reward": -0.23340531676644
41 },
42 "nominal_adaptive": {
43 "violation_rate": 0.0,
44 "mean_action_deviation": 0.0,
45 "mean_reward": -0.005491596774677281
46 },
47 "shifted_unshielded": {
48 "violation_rate": 1.0,
49 "mean_action_deviation": 0.0,
50 "mean_reward": -0.11336081586839418
51 },
52 "shifted_fixed": {
53 "violation_rate": 0.0,
54 "mean_action_deviation": 0.4612857142857539,
55 "mean_reward": -0.015070375678229206
56 },
57 "shifted_adaptive": {
58 "violation_rate": 0.02,
59 "mean_action_deviation": 0.49816108846300994,
60 "mean_reward": -0.026906702730669113
61 }
62 }
63}