Designing the Output-Side Guardrails

You're building the output-filtering stage for a customer-facing LLM assistant that answers from a knowledge base and occasionally emits a structured {answer_md, citations[], suggested_action?} object for the frontend to render. Design the output pipeline:

1. What checks run, in what order, and which are cheap enough to run on every response versus which should be sampled or async?
2. What should happen when the schema check fails? When the groundedness check fails?
3. Why is a refusal template safer than letting the model generate its own refusal text, and where else in the pipeline does that same template get reused?