As Fig. D.1, but comparing the ultranest and PyMultiNest samplers for the HD 19467 B default retrieval. Both samplers used 1000 live points and ∆ ln Z = 0.5; ultranest with a slice sampler (n_steps = n_dim = 30) and PyMultiNest with a constant sampling efficiency of 5%. This benchmark was run with an earlier setup than the final one, so the central values differ slightly from the adopted results quoted in the main text and Table C.1.