Recovery of population-level model parameters using two different test sets. In purple we show our fiducial test set. In orange we take the same test set and change the mass–light ellipticity relation to following q_m ≥ q_l − 0.1 (taken from SLACS lenses; A. S. Bolton et al. 2008; C. Y. Tan et al. 2024). Inset panel: we show the two different mass–light ellipticity relations against the cloud of gray scatter points representing the mass–light ellipticity relationship in the training data. We demonstrate how sensitive the population-level parameter recovery is to such a shift, and we suggest that we need a more informative cpdf in order to work with such distribution shifts seen in real data.