loading
loading
Your eval scores or output quality slowly degrade over time even though your code didn't change.
The model provider silently updates the endpoint, real-world inputs shift away from what you tested on, or a dependency moves — and behavior that passed last month now fails. Because nothing in your repo changed, drift is invisible unless you run evals on a schedule and watch the trend, not just on commits. Example: your 'extract invoice total' prompt was 98% accurate in January and 91% in June with no code change, because the underlying model version moved.
Plainly
Think of Drift as the checklist that keeps the app open for real people. Your eval scores or output quality slowly degrade over time even though your code didn't change.
In practice
Use it when a change has to survive deploys, users, incidents, analytics, or billing reality. In practice, define the owner, input, output, and failure mode before you rely on it.