Population-level model recovery for θ_E, γ_lens, e₁, and γ₁ using posteriors recovered from the fiducial test set. The black solid line is the ground truth of the test set, and the purple lines indicate the mean value of the converged chains. Note that we have removed 71 objects from our sample that had wider posteriors than the training distribution along one or more parameter axes, as this breaks the HBI framework (refer Section 3.3). We show results for e₁ and γ₁ only, as e₂ and γ₂ reproduce the same result. We note that the population model learns a relationship between ellipticity and shear encoded into the test. The population mean of γ_lens is well constrained by the fiducial test set, and we go on to demonstrate how this can change with different model assumptions or data quality in Figures 9 and 11.