Failures / Standard term
Distribution shift
A difference between the data distribution used to build or evaluate a model and the distribution encountered after deployment.
The shift may affect inputs, labels, relationships between them, user behavior, time, language, or operating conditions. Performance may decline, remain stable, or even improve, so the shift and its consequence must be measured separately. Foundation-model training data is usually only partly known, which makes product-specific evaluation especially important.
Builder example
A polished demo can miss production formats, customer segments, changing policies, or tool failures. Versioned inputs, slice-level outcome metrics, and cases drawn from live failures reveal where the deployed workflow no longer matches its tested conditions.
Common confusion: Distribution shift can happen across customer segments within the same product. A model that performs well for one industry vertical may fail in another, even when the task looks identical, because the underlying data patterns differ.