Comparison
DriftvsMetric drift
Drift
nothing in your code changed but the outputs are worse than they were in March, because the traffic or the model underneath moved.
Quality change over time without any deploy on your side: providers update models, user inputs shift, corpora age. Detected by scheduled eval runs and monitored online metrics; pinned model versions reduce it but expire when versions retire.
Full entry →Metric drift
the same metric name means something slightly different than it did in March, and the year-on-year chart is comparing two definitions.
A measure's definition changing over time without the change being versioned or announced. It is uniquely damaging because it corrupts comparisons rather than values: every number is correct on the day it was computed and the trend between them is meaningless. Versioning definitions and recording when each took effect is the only durable fix, since restating history is often impossible.
Full entry →