One Swapped Digit Beats Every Major Hallucination Detector
Change one number in a grounded RAG answer (10,000 for 1,000) and every mainstream hallucination detector scores near chance: AUROC 0.51-0.59 for embedding similarity, entailment models and LLM judges, with MiniCheck barely better at 0.75. A warning for any production pipeline that trusts faithfulness checks on numeric answers.