Constrained adaptation of language models on a manifold of valid adapters, studied as a safety mechanism: poisoning resistance, out-of-distribution detection, and few-shot regularization, by construction rather than by heuristics.