Guardrails

Runtime filters, classifiers, or policy checks that sit around an LLM to block harmful inputs/outputs or enforce topical scope.