benchgap
Calibration

DeepSearchQA → ApprenticeBench

ApprenticeBench is estimated from DeepSearchQA with a Hill curve fitted on 5 models measured on both: y = 0.0121 + (1.2000 − 0.0121)·x^6.00 / (1.18019^6.00 + x^6.00), R² = 0.66, cross-validated error 10.1 pp. It is used for 3 estimates.

Estimated modelDeepSearchQAApprenticeBenchSource
dots3-note Preview92.1%23.1%estimated ± 10.1 pp, low confidence
Mercury 2.534.0%1.3%estimated ± 10.1 pp, low confidence
Beam80.1%11.8%estimated ± 10.1 pp, low confidence