Sovereign-First, Fail-Closed LLM Routing with Agentic Workflow Governance

This technical report presents a sovereign-first, fail-closed large language model (LLM) routing and agentic workflow governance framework for regulated enterprise environments. The proposed architecture prioritises locally hosted sovereign AI models and permits controlled escalation to approved frontier models only when capability requirements justify escalation and data classification, residency and security policies allow it. The framework introduces an authority-envelope approach for controlling delegated permissions, tool execution and cross-model transitions. It applies established enterprise orchestration principles, including access control, policy enforcement, exception handling and auditability, to agentic AI workflows. Two deterministic simulation studies evaluate routing decisions and non-expanding authority controls using synthetic data. The routing experiment includes 2,000 synthetic requests, while the authority-control experiment evaluates 1,600 synthetic transitions. The findings demonstrate the implemented policy rules under controlled simulation conditions. They do not constitute production performance or regulatory compliance validation. The accompanying research package includes reproducible Python simulation code, synthetic datasets, experiment results and supporting documentation.