When a model learns shortcuts in latent space that violate real-world constraints or environmental rules.