Design guardrails and fallbacks for a reliable LLM service

A product depends on calls to a large language model. Design the guardrails and fallback paths that keep it reliable: what you check on the way in and on the way out, what the system does when the model times out, errors, is overloaded or returns something unusable or unsafe, and how it falls back (retries, another model, a degraded answer) without making things worse. Say how you would measure whether the guardrails work.