Calibration

FrontierSWE v2 → DeepSWE

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

DeepSWE is estimated from FrontierSWE v2 with a Michaelis–Menten + offset curve fitted on 16 models measured on both: y = 0.0000 + 0.7732·x / (0.03667 + x), R² = 0.57, cross-validated error 4.2 pp. It is used for 1 estimate.

Estimated model	FrontierSWE v2	DeepSWE	Source
Inkling	4.1%	40.8%	estimated ± 4.2 pp, medium confidence