Abstract

We introduce a unified Bayesian anomaly-detection framework for Cosmology, applied to the REACH global 21cm probe and also Type Ia supernovae. This approach embeds data-integrity beliefs directly into the inference process. Rather than excising contaminated or anomalous data points, the method employs a piecewise likelihood constrained by a Bernoulli prior and an Occam penalty, allowing anomalies to be down-weighted automatically while performing numerical sampling for parameter inference. When applied to supernova light curves, the framework yields precise estimates of brightness scaling, stretch, and colour, while also automating supernova sample and band selection. In the context of global 21 cm cosmology, it offers a principled way to mitigate radio-frequency interference (RFI), particularly within the band of interest. We also discuss additional potential applications of this methodology.