benchgap
Calibration

CritPt → GPQA

GPQA is estimated from CritPt with a Michaelis–Menten + offset curve fitted on 61 models measured on both: y = 0.6435 + 0.2787·x / (0.00473 + x), R² = 0.69, cross-validated error 6.8 pp. It is used for 0 estimates.