Transit ops notices model drift after road construction patterns change and asks which lifecycle phase handles refresh. What should they emphasize?

When transit ops sees model drift after road construction patterns change, the lifecycle emphasis is monitoring that detects drift and drives retraining or refresh. Monitoring compares live performance or input distributions against baselines and triggers a controlled response—new labeled data, retraining, or model replacement—rather than assuming data ingestion pipelines refresh models directly from new traffic feeds without drift detection. Treating data ingestion pipelines as enough to refresh models directly from new traffic feeds denies the operational reality that municipal traffic conditions evolve with construction and seasonality, and ingestion alone cannot detect drift. Letting offline validation scores from initial training guide production refresh timing removes the live signal needed to catch construction-driven drift and turns a detect-and-fix loop into a static schedule that misses real-world change. Relying on data version logs and rollback steps to flag drift and start retraining runs daily confuses governance with detection, and daily retraining without live drift signals can overwrite good models with stale or unvalidated ones. Exam caveat: not every metric shift is harmful drift, and not every alert should auto-retrain without validation; the exam point is that monitoring surfaces drift and connects to refresh. Operational check: define drift signals for the traffic model (error rates, feature shift), route alerts to owners who can schedule retraining, and re-validate before promoting a refreshed model to transit inference.