ClinicalVerifier: A Retrieval-Augmented Pipeline for Detecting Guideline Contradictions in LLM-Generated Mental Health Text

Large language models (LLMs) are increasingly deployed in mental health applications, yet their outputs may silently contradict clinical guidelines, posing serious patient safety risks. This paper presents ClinicalVerifier, a retrieval-augmented generation (RAG) system that automatically detects when LLM-generated clinical text contradicts NICE and WHO evidence-based guidelines.

The system combines a FAISS-indexed embedding store of guideline excerpts with an LLM judge (Llama-3.3-70B via Groq), augmented by a Neighbourhood Consistency Scoring (NCS) hallucination probe to produce calibrated combined risk levels (LOW / MEDIUM / HIGH). Guideline sources include NICE CG90, NG185, CG178, NG116, CG53, CG42, and the WHO mhGAP Intervention Guide 2023.

Evaluated on a 30-case labelled benchmark spanning safe, uncertain, and contradicts clinical outputs, ClinicalVerifier achieves:

• 73.3% overall accuracy
• 95.2% F1 on safety-critical contradiction detection
• 100% recall on guideline-contradicting cases
• 100% precision on HIGH combined-risk alerts

The system's conservative design ensures no guideline violations are missed, while the dual-signal architecture (RAG verdict + NCS score) minimises false alarms. The pipeline is fully open-source, runs without proprietary API access, and is designed for integration into clinical AI monitoring workflows as a sidecar service.