benchgap
Calibration

DeepSearchQA → BrowseComp

BrowseComp is estimated from DeepSearchQA with a offset logistic curve fitted on 11 models measured on both: y = 0.7622 + (0.9795 − 0.7622) / (1 + exp(−71.13·(x − 0.9417))), R² = 0.50, cross-validated error 8.2 pp. It is used for 1 estimate.

Estimated modelDeepSearchQABrowseCompSource
Mercury 2.534.0%76.2%estimated ± 8.2 pp, low confidence