Factuality

Explanation

Factuality via Un-supported Claims checks whether an AI system makes statements that are not backed by its source material, which matters in RAG because every unsupported claim is a potential hallucination that can mislead a customer or staff member, for example asserting a fee or eligibility rule that no source document contains. The Un-supported Claims metric identifies and counts the individual assertions in the model's output that cannot be traced to the retrieved or reference context. To implement it, decompose each response into discrete claims, verify each one against the exact source passages the model was given, and record which claims lack support, logging the count and the offending statements as evidence per response; this check sits immediately after generation in the pipeline and runs both in pre-release evaluation and on sampled live traffic. The threshold is set as a target per control objective with an alert on breach (proposed), so set a risk-based threshold for the acceptable number or rate of unsupported claims given the use case; exceeding it raises an alert that prompts a reviewer to examine grounding and triggers corrective action such as tightening retrieval or constraining the model to cite sources. Its Immediate priority reflects how directly unsupported claims undermine the trustworthiness BFSI systems depend on.