Adversarial Decision-Equivalent Training / results.json
Failed on benchmark
1{
2 "seed": 2028,
3 "checks": {
4 "predictions": {
5 "onset_boundary": "regret starts when delay exceeds true margin 1.0",
6 "linear_scaling": "regret = delay - 1 in the disagreement interval (slope 1)",
7 "scale_law": "scaling costs and delays by s scales regret by s and thresholds by s",
8 "zero_perturbation": "d=0 has zero extra regret for a nominally equivalent predictor"
9 },
10 "observed": {
11 "onset_grid": 1.05,
12 "last_positive_grid": 2.45,
13 "fitted_regret_vs_delay_slope": 1.0000000000000002,
14 "max_linear_absolute_error": 2.220446049250313e-16,
15 "scale_rows_s_regret_expected": [
16 [
17 0.5,
18 0.25,
19 0.25
20 ],
21 [
22 1.0,
23 0.5,
24 0.5
25 ],
26 [
27 2.0,
28 1.0,
29 1.0
30 ],
31 [
32 3.0,
33 1.5,
34 1.5
35 ]
36 ],
37 "max_scale_absolute_error": 0.0,
38 "zero_delay_worst_regret": 0.0
39 },
40 "theory": {
41 "onset": 1.0,
42 "end": 2.5,
43 "slope": 1.0
44 }
45 },
46 "results": {
47 "mse": {
48 "nominal_regret_mean": 0.0,
49 "worst_regret_mean": 0.0,
50 "worst_regret_max": 0.0,
51 "adversarial_flip_rate": 0.0,
52 "cost_mse": 0.0005127362417326554
53 },
54 "nominal": {
55 "nominal_regret_mean": 0.0,
56 "worst_regret_mean": 0.09968487394957982,
57 "worst_regret_max": 0.4384453781512603,
58 "adversarial_flip_rate": 0.325,
59 "cost_mse": 0.041750840342126685
60 },
61 "adversarial": {
62 "nominal_regret_mean": 0.0,
63 "worst_regret_mean": 0.0,
64 "worst_regret_max": 0.0,
65 "adversarial_flip_rate": 0.0,
66 "cost_mse": 0.05633179241077497
67 }
68 },
69 "note": "Exact hard paths are used for evaluation; training uses differentiable cross-entropy over the exact two-route oracle."
70}