Out-of-domain data

Data that falls outside the specific range, characteristics, or context (domain) a machine learning model was trained to handle.

Out-of-domain (OOD) data is any input that deviates from the model's defined operational scope, which is distinct from merely being out-of-distribution (OODist). For example, a model trained exclusively on financial news articles (the in-domain data) will encounter OOD data when processing medical journals: the vocabulary and semantic context are fundamentally different domains. The critical risk is 'silent failure': the model often generates a confident, yet incorrect, prediction because it cannot recognize the input's novelty. Effective OOD detection is mandatory for robust AI deployment (e.g., autonomous vehicles, medical diagnostics); it forces the system to flag the input for human review instead of making a high-stakes, low-quality inference.

https://medium.com/@shirou.matsumoto/two-oods-out-of-domain-and-out-of-distribution-8302195f03d5