Certified ambiguity gating for LLM supervision / results.json

Mechanism failed

Raw ⬇ ZIP
 1{
 2  "audit": {
 3    "Dhat": 0.37833333333333335,
 4    "epsilon": 0.028846973043371422,
 5    "floor_LCB": 0.17474318014498097
 6  },
 7  "minimax_verification": {
 8    "D": 0.377,
 9    "D_over_2": 0.1885,
10    "minimax_risk_fair": 0.1885,
11    "random_predictor_worst_risk": 0.5981749879725199
12  },
13  "n_train": 1800,
14  "results": {
15    "ambiguity_gate": {
16      "ECE": 0.05871945929527283,
17      "accuracy": 0.8952
18    },
19    "confidence_matched": {
20      "ECE": 0.05871945929527283,
21      "accuracy": 0.8952
22    },
23    "hard_all": {
24      "ECE": 0.11987026920318603,
25      "accuracy": 0.8994
26    },
27    "soft_all": {
28      "ECE": 0.13260989146232607,
29      "accuracy": 0.9066
30    }
31  },
32  "retained_gate": 1119,
33  "retention": 0.6216666666666667,
34  "seed": 611,
35  "tau": 0.3
36}