Architecting Optimized Legal Corrections: A Neuro-Symbolic Pipeline of Agentic AI Compliance

Agentic AI systems capable of autonomous reasoning and tool use pose new challenges for legal compliance and accountability. This paper presents a neuro-symbolic compliance pipeline that integrates large language model (LLM) reasoning with satisfiability modulo theories (SMT) solving to perform optimized legal corrections—systematic adjustments that restore compliance while preserving statutory and operational constraints. Through a multi-agent architecture, the system automates legal text parsing, constraint synthesis, and solver-guided self-repair, incorporating SMT feedback into LLM prompts for iterative correction.

Evaluated on 479 financial enforcement cases from Taiwan's Financial Supervisory Commission, the framework achieves a 99.79% formal verification rate and demonstrates the complementary strengths of neural interpretation and symbolic validation. While standalone LLMs reach 72.0% feasibility accuracy and 15.14% compliance-generation success, the hybrid approach provides a scalable, verifiable foundation for explainable legal reasoning.