FinRegAgents: A Multi-Agent RAG Framework for AI-Assisted Financial Regulatory Audits with Confidence-Aware Validation

Abstract

We present FinRegAgents, an open-source multi-agent framework that simulates supervisory-style financial regulatory audits using retrieval-augmented generation (RAG). The system evaluates institutional compliance against four major European regulatory frameworks—GwG (Anti-Money Laundering), DORA (Digital Operational Resilience Act), MaRisk (Minimum Requirements for Risk Management), and WpHG/MaComp (Securities Trading Act)—across 94 audit fields organized in declarative JSON catalogs. Our key contribution is a confidence-aware validation architecture that addresses the critical gap between RAG retrieval and regulatory assertion: a retrieval quality gate prevents hallucination on thin evidence, a composite confidence score (retrieval relevance, evidence coverage, type matching, LLM self-assessment) quantifies assessment reliability, and structural validation detects phantom citations, placeholder artifacts, and logical inconsistencies before findings enter the audit report. We formalize the coverage–verification gap—the systematic discrepancy between a system's ability to produce assessments and its ability to verify them—drawing on parallels to ad-tech fraud detection. We describe the architecture, the confidence scoring framework, the validation pipeline, and report generation across three output formats. FinRegAgents is released under Apache 2.0.

Keywords