This chapter presents Bayesian model selection—traditionally performed through evidence evaluation—as inference of an indicator variable that labels the model that generated the data (from among a range of possibilities over which prior probabilities are, naturally, assumed). It thus introduces the technique of simulation-based model selection through classification and quantifies the reliability and refinedness of the results via bespoke diagrams. Finally, it discusses the implications of amortisation to visualising and overcoming Occam’s razor: the preference for a parsimonious description of the data-generating process.