Out-of-Distribution Detection

Explanation

Out-of-Distribution (OOD) Detection checks whether the model can recognise inputs that differ meaningfully from the data it was trained and validated on, rather than confidently producing answers for cases it was never designed to handle - critical in banking RAG and Summarization & Chat systems where an unfamiliar document type, language, or query pattern can otherwise yield silently wrong results. It is measured by the OOD Detection Rate, calculated as (True OOD Detection rates) / (Total OOD Cases); the control expects banking models to detect 5 out-of-distribution patterns. To implement it, define those OOD patterns, build a representative test set of in-distribution and out-of-distribution examples, add a detection mechanism (such as confidence scoring, novelty or distance checks, or a dedicated classifier) at the inference entry point, route flagged inputs to a fallback or human review, and log each detection decision as evidence. This is an Immediate-priority control. The threshold is a detection rate above 95%, meaning the system must correctly flag more than 95% of genuine OOD cases; falling below this should trigger an alert, investigation of the missed patterns, and retraining or recalibration of the detector. It maps to RBI Safety, Resilience and Sustainability, NIST MAP 2.1 and 4.1, and ISO Clauses 6.2.1 and 6.1.2.