A new testing protocol shows you can catch a chatbot's bad, made-up answers for a fraction of a cent in compute - as long as you keep retraining the filter for every new type of document it sees.
The paper introduces RAGScope, a protocol for evaluating "evidence gates": lightweight checks that scan a retrieval-augmented generation (RAG) system's retrieved sources and its answer text to flag likely hallucinations, without running a full second AI model to verify every response. The best version, RAGScope-E, scored 0.798 AUROC (Area Under the ROC Curve, a standard 0-to-1 score for how well a system separates good answers from bad ones, where 1.0 is perfect and 0.5 is a coin flip) across three RAGTruth benchmark tasks. It ran in 6.22 milliseconds per example on a plain CPU, compared with 145.75 ms and 223.07 ms for two heavier verifier models tested alongside it. Set to review just the riskiest 10% of answers, it flagged true hallucinations with 74.8% precision.
That speed matters because RAG systems - chatbots that look up documents before answering - need a cheap first pass to sort answers into auto-approve, send-to-human, and run-the-expensive-verifier buckets. But the paper's own stress test undercuts the pitch: on a 14,900-example set drawn from new sources, a gate calibrated in-domain hit 0.879 AUROC, then dropped to 0.466 - worse than a coin flip - when calibrated on different sources and tested cold. Retraining on 200 fresh labels per source only recovered it to about 0.685.
So this isn't a drop-in hallucination detector, it's a fast pre-filter that needs local retraining every time you point it at a new document type - a smaller claim than the headline metric suggests, but a more honest one.