Failures / Standard term
Out-of-distribution (OOD)
An input or case outside the distribution used for a stated training or evaluation claim.
OOD is relative to a reference distribution. A legal question may be outside a retail support classifier's validated data even if a general-purpose model has seen legal text. Being OOD raises uncertainty about whether measured performance transfers; it does not prove that a particular answer will be wrong, and being in-distribution does not guarantee correctness.
Builder example
Products need an operational coverage claim they can observe, because the full pretraining distribution of a foundation model is rarely available. Input classifiers, retrieval coverage, uncertainty signals, and post-deployment outcomes can help route cases, but OOD detection itself has errors and needs evaluation.
Common confusion: OOD does not mean rare or unusual. A perfectly common request can be out-of-distribution if the model's training data never covered that domain. Frequency in the real world and coverage in the training set are separate things.