Projective Boundary Certificates for Neural Selective Prediction / results.json
Failed on benchmark
1{
2 "seed": 3002,
3 "n": 20,
4 "delta": 0.1,
5 "reps": 200000,
6 "moment_check": {
7 "1": {
8 "empirical": 0.9523219823374796,
9 "theory": 0.9523809523809523,
10 "abs_error": 5.897004347277246e-05
11 },
12 "2": {
13 "empirical": 0.9089856354477753,
14 "theory": 0.9090909090909091,
15 "abs_error": 0.00010527364313372978
16 },
17 "5": {
18 "empirical": 0.7998104126740451,
19 "theory": 0.8,
20 "abs_error": 0.0001895873259549763
21 },
22 "10": {
23 "empirical": 0.6664358861689239,
24 "theory": 0.6666666666666667,
25 "abs_error": 0.00023078049774283294
26 }
27 },
28 "beta_upper_quantile": 0.10874906186625444,
29 "scalar_beta_coverage": 0.899365,
30 "deletion_checks": {
31 "scalar_max": {
32 "boundary_size": 1.0,
33 "equivalence_rate": 0.9499999999999997,
34 "projectivity_rate": 1.0
35 },
36 "adaptive_rank_control": {
37 "boundary_size": 1.0,
38 "equivalence_rate": 0.9071749999999997,
39 "projectivity_rate": 0.9546605263157896
40 }
41 },
42 "mini_experiment": {
43 "batches": 1000,
44 "test_points_per_batch": 10000,
45 "baseline_split_conformal_coverage": 0.889,
46 "projective_boundary_beta_coverage": 0.889,
47 "unstable_control_false_beta_coverage": 0.404,
48 "mean_scalar_true_risk": 0.047448855636694875,
49 "mean_adaptive_control_risk": 0.22837482515636362
50 },
51 "interpretation": "Max order-statistic has K=1, exact Beta(1,n), and passes deletion tests; adaptive rank is not projective and its naive K-only beta certificate fails."
52}