Illustration of the application of redshift nuisance models to the first and third bins of the $n_z^{\mathrm{coh}}$ biased stage III photometric redshift distribution (Fig. 4). Alternating panels give the target distribution $n_z^{\mathrm{coh}}$, or its corresponding lensing efficiency kernel $q_z^{\mathrm{coh}}$, subtracted from the equivalent n(z) or q(z) seen under the various recalibration models: the ‘do nothing’ model, denoted as N(z) (dashed purple), the comb model 𝒩_comb (solid green), and the shift model δz_i (dotted red). The latter two are calibrated here against the 3 × 2 pt data-vector. Vertical lines give the mean of each n(z) or q(z). The optimized comb model corresponds to a flexible recalibration of the n(z), which minimizes the difference (via Eq. (2.23)) between the theory- and data-vector (e.g. Fig. 9) by varying the amplitudes of a Gaussian mixture model. One sees that the comb model outperforms the shift model in reducing errors in q(z) whilst minimally increasing errors in n(z).