Your production LLM application serves 1M requests per day. Over time, output quality degrades — some due to model drift, some due to data distribution shift, some due to hallucinations. Design a monitoring system. Cover: what signals indicate quality degradation without human evaluation on every output, how to detect hallucinations at scale (factuality checks, citation validation, consistency checks), how to set up alerting thresholds that catch real regressions without alert fatigue, and how to distinguish model degradation from prompt drift from upstream data changes.