gpt-4.1-nano as RAG judge barely beats chance: AUROC 0.603 on RAGBench

Giulio Zeloni · hf · 2026-10-07

Context: enterprises face recurring promote/revise/block decisions on RAG systems, but metrics come from fallible LLM judges. The authors present AGO, an evidence-first quality-gate framework deployed in industrial RAG assessment.
Design: a four-state decision model (missing data and judge errors are explicit outcomes), layered scoring (deterministic checks + local guardrails + structured LLM eval), a stratified beta-binomial gate quantifying regression risk, and mandatory meta-evaluation of the judge.
Findings (RAGBench, N=1200 per judge): gpt-4.1-nano detects non-adherent answers at AUROC 0.603 — barely above chance — despite flawless protocol output; gpt-4o reaches 0.783 but varies 0.62–0.88 across domains.
Gate study: under regression, the decision-grade profile cuts unsafe promotion to 22.2%–35.1% vs 29.3%–41.8% for a naive gate. Point estimates alone are not a release decision.

More from coding & agent