LLM Context Window Management in Production: A Governance Checklist

Your model advertises a 200K token window. Your agent starts producing worse answers somewhere around 90K, and nobody notices for three weeks, because nothing throws. That gap between advertised capac