Ten Is Not a Hundred

August 18, 2026 | Source: Towards Data Science

Tags: RAG, hallucination detection, MiniCheck, LLM judges, numerical reasoning, LettuceDetect

A controlled study of five hallucination detectors shows all major tools fail to catch simple numerical errors in RAG outputs — four of five score near-random AUROC (0.51–0.59), with only MiniCheck reaching 0.75 — exposing a critical blind spot in production RAG pipelines.

Details

Researcher Javier Marin built a test set of 200 RAG responses from the RAGTruth corpus — 100 accurate and 100 with a single digit swapped (e.g., $1,000 instead of $10,000) — and measured how five hallucination detection approaches handled the cases. The results are sobering: Ragas-style LLM judges scored AUROC 0.57, embedding similarity 0.51, entailment models 0.59, and LettuceDetect (published state-of-the-art at its size) 0.58. Only MiniCheck reached 0.75. The core reason: embedding methods compress the full response into a vector where a one-digit swap produces an almost identical representation — the semantic distance between 1,000 and 10,000 is nearly zero. The author extends the analysis to practical operating points, showing that at real false-positive thresholds, none of the detectors provide actionable signal. This matters most for financial, medical, and legal RAG applications where numerical precision is high-stakes. The article includes a reproducible methodology using publicly available RAGTruth data. Anyone running RAG over structured numerical data should treat this as a confirmed blind spot in current tooling — no off-the-shelf solution reliably catches this failure mode.