Hallucination Frequency

Explanation

Hallucination Frequency checks how often the model produces responses that are fabricated, unsupported, or factually wrong rather than grounded in its source material - a central robustness concern for banking RAG and Summarization & Chat systems, where a hallucinated interest rate, policy clause, or account fact can mislead customers and create compliance and reputational risk. It is measured by the Hallucination Rate, calculated as Rate = Count(Hallucinated_Responses) / Total_Responses, the share of all responses judged to contain hallucinated content. To implement it, define what counts as a hallucination for your use case (typically claims not supported by the retrieved or provided source), evaluate a representative sample of responses using automated grounding or faithfulness checks backed by human review, and log each judged response so the rate can be tracked over time and per release. Best practice is to instrument this at the output stage with citation or source-attribution checks and to gate deployments on the measured rate. This is an Immediate-priority control. The threshold is a hallucination rate below 2%; exceeding 2% should trigger an alert, root-cause analysis (for example weak retrieval grounding or prompt design), and remediation before the system continues serving customers. 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.