| { |
| "schema_version": "1", |
| "model_ids": [ |
| "gpt-5.6-luna-none", |
| "gpt-5.6-terra-low", |
| "gpt-5.6-sol-medium" |
| ], |
| "temperatures": [ |
| 1.0598043203353882, |
| 1.106358528137207, |
| 0.3136223554611206 |
| ], |
| "selection_threshold": 0.58, |
| "constraints": { |
| "max_unsafe_downroute_rate": 0.005, |
| "min_selected_model_pass_rate": 0.99, |
| "satisfied": true |
| }, |
| "validation_route_metrics": { |
| "exact_route_accuracy": 0.9509283819628647, |
| "unsafe_downroute_rate": 0.004641909814323607, |
| "selected_model_pass_rate": 0.9953580901856764, |
| "abstention_rate": 0.0, |
| "savings_vs_always_large": 0.8416445623342175 |
| }, |
| "uncalibrated": { |
| "pass_vector_accuracy": 0.9244031830238727, |
| "pass_bit_accuracy": 0.9748010610079576, |
| "macro_f1": 0.985090243264452, |
| "brier_score": 0.01777769201095797, |
| "expected_calibration_error": 0.018437407725585803 |
| }, |
| "calibrated": { |
| "pass_vector_accuracy": 0.9244031830238727, |
| "pass_bit_accuracy": 0.9748010610079576, |
| "macro_f1": 0.985090243264452, |
| "brier_score": 0.017453370615839958, |
| "expected_calibration_error": 0.01830040662266169 |
| } |
| } |
|
|
