Outcomes of resuming end-to-end discriminative models by refitting in CLAP illustrated for the SDSS, CFHTLS, and KiDS target samples. The SDSS target sample is split into a reference subsample (used to refit the model) and an inference subsample (held out for inference), while the CFHTLS and KiDS target samples are only used as reference samples. The results obtained by No-NIR CLAP for the KiDS dataset are not shown in order to avoid redundancy. For each dataset, only one of the ten CLAP models in the ensemble is selected for demonstration. First row: mean redshift residuals as a function of photometric redshift z_photo computed using the reference samples with and without refitting. The results for the SDSS inference subsample are also shown. As in Fig. 6, the results for the CFHTLS and KiDS datasets are shown in the low redshift ranges in addition to the full ranges. Second row: σ_MAD as a function of r-band magnitude. Third row: cumulative offsets $\text{Δ}{F}_{p_{\text{Σ}}-h}$ between each of the stacked recentred probability densities p_Σ (z − z_photo) and the corresponding z_spec − z_photo histogram. The vertical dashed line in each panel indicates the zero point where z and z_photo coincide. Fourth row: PIT distributions. The y-axis in each panel is truncated at 0.5 for clearer illustration.