A Multi-Agent Audit Framework for High-Stakes Reasoning: Evaluation and Interpretability in Clinical Mental Health Screening

The article introduces a Multi-Agent Audit Framework designed for high-stakes reasoning tasks, specifically in clinical mental health screening. This framework employs a modular LangChain workflow comprising a Perception Agent, Knowledge Retrieval-Augmented Generation (RAG), Chain-of-Thought (CoT) inference, and an Audit verification stage, demonstrating a reduction in Mean Absolute Error (MAE) for PHQ-8 depression severity prediction from 5.35 to 5.02 when evaluated on the DAIC-WOZ dataset. This approach enhances interpretability and reduces hallucination in AI-assisted decision support, offering a scalable solution for practitioners aiming to improve the reliability of AI systems in sensitive domains.