A learned correction layer applied on top of a task policy to improve safety without retraining the main policy.