Jev-as-a-Judge for RAG Claim Verification: Confidence-Gated Cascading Between a Non-Generative Verifier and a Frontier LLM
The study compares a fast, non-generative decision model against a powerful frontier LLM for verifying claims in Retrieval-Augmented Generation (RAG). While statistically tied on conventional accuracy metrics, they differ significantly in error types. The cheaper model is better at identifying where the frontier LLM fails, suggesting that routing based on confidence and escalating uncertain claims provides a cost-effective solution.