Model Drift

A change in the observed behavior or performance of a model or system over time. It may result from shifts in input data, a silent update to a hosted model, dependency changes, product evolution, or degradation in the relationship between signals and expected outcomes.

Diagnosis requires versions and continuous evals. An aggregate decline does not identify the cause, and stable input distributions do not guarantee stable system behavior. Task- and segment-level monitoring, canary tests, and rollback help distinguish model drift, data drift, and application regressions.