Calibration

C-Eval → CritPt

Name: benchgap: LLM benchmark scores, measured and estimated
Creator: benchgap

CritPt is estimated from C-Eval with a inverse Michaelis–Menten curve fitted on 5 models measured on both: y = 0.00006·(x − 0.0000) / (0.0000 + 0.9350 − x), R² = 0.83, cross-validated error 1.1 pp. It is used for 1 estimate.

Estimated model	C-Eval	CritPt	Source
LongCat-Flash-Lite-Sparse	85.8%	0.1%	estimated ± 1.1 pp, low confidence