benchgap
Calibration

HealthBench Hard → CritPt

CritPt is estimated from HealthBench Hard with a Hill curve fitted on 11 models measured on both: y = 0.0000 + (0.2636 − 0.0000)·x^6.00 / (0.18557^6.00 + x^6.00), R² = 0.51, cross-validated error 7.4 pp. It is used for 0 estimates.