This repository provides a reproducible framework for auditing domain alignment in latent representations learned from heterogeneous materials datasets. The code implements unsupervised latent encoder training, representation-level validation, and quantitative alignment diagnostics independent of predictive accuracy.