What's the difference between guardrails and evaluation? Aren't they the same?

"Guardrails are real-time gates — they block bad outputs before the user sees them. Evaluation is offline measurement — it tells you how good the system is over time. Guardrails: input sanitization, output content filtering, PII detection, token budget enforcement. They run on every request, add latency, and must be fast. Evaluation: golden set testing, LLM-as-judge scoring, user feedback analysis. Runs daily/weekly, can be slow, informs improvements. Think of guardrails as the seatbelt (prevents harm now) and evaluation as the crash test (improves safety for tomorrow)."