Jev-as-a-Judge for RAG Claim Verification: Confidence-Gated Cascading Between a Non-Generative Verifier and a Frontier LLM

A non-generative decision model called Jev was compared against a frontier LLM for verifying whether evidence supports claims, using 495 human-labeled claims from LLM-AggreFact. The two systems showed a statistically indistinguishable difference of -0.56 points on the conventional metric, but disagreed on 15.2% of claims, with Jev admitting 23.2% of unsupported claims versus the frontier model's 14.1%. The cheaper system costs 187 times less to run than the frontier LLM.