Automated Geographic Risk Revision for Financial-Economic Crime: A Deliberative Mixture-of-Experts Framework

Geographic risk is a dimension of the essence in anti-money laundering (AML) and counter-terrorist financing (CFT) frameworks, yet most existing models treat country profiles as static and retrospective. Recent literature explores event-driven risk assessment and adverse media monitoring but lacks scalable, explainable systems tailored to compliance. This paper proposes a mixture-of-experts framework using large language models to assess ingested news content that is calibrated against structured geographic indices. The research explores whether deliberative multi-agent systems improve accuracy in multi-class geographic risk classification. Experimental results show that the framework outperforms zero-shot NLI baselines, achieving over 90% precision and recall post-calibration, especially in high-severity risk tiers. These findings support the integration of structured and unstructured data into dynamic compliance systems. The proposed architecture advances regulatory technology by introducing a modular, auditable, high-performance solution for real-time geographic risk monitoring, thereby bridging the gap between static indices and event-sensitive financial crime detection.