Design guardrails and fallback for LLM reliability

You operate a production application that uses an LLM to generate user-facing outputs (text actions, advice, summaries). The model is non-deterministic and sometimes produces unsafe, incorrect, or policy-violating content. Design the **safety and reliability layer** around the LLM. ## Requirements - Prevent unsafe or policy-violating outputs. - Handle model uncertainty and mistakes gracefully ("the model is wrong" scenarios). - Provide a **fallback / degrade strategy** during incidents, model regressions, or partial outages. - Keep latency overhead minimal and make decisions auditable. ## Discuss - Where guardrails should live in the system (before the model, after the model, or both). - Techniques: input validation, prompt constraints, output filtering, tool/action validation. - Monitoring and evaluation: what signals catch regressions quickly. - Incident response: rollout, rollback, and kill-switch mechanisms. - Tradeoffs between safety, user experience, and cost.