Method of building confidence in our ML estimator for data outside the training domain. Left: Train ensemble of ten regression NNs on the fiducial hydro BAHAMAS simulations. We show the combined PDF of log(σ_DM/m) for the known BAHAMAS simulations (blue shades) and blind uniform random noise (hatched black). We find that the regressor consistently estimates a significant, positive cross-section of ∼0.6 cm²g⁻¹ with no regard to its confidence, presenting the issue with direct regression estimators. Right: First and third latent dimensions from our compact clustering algorithm trained with the same fiducial BAHAMAS as known (blue shades) and the random noise dataset as unknown (hatched black). The first latent dimension corresponds to log σ_DM/m, which we would naively assume that the noise dataset has a cross-section of ∼0.1 cm² g⁻¹; however, from the third dimension we see that the noise dataset shares no similarities with the known simulations and therefore, cannot be trusted.