APR-Lite Phase 5 Technical Note Routing Brain v2: Governed Intent Classification, Agent Selection, Topology Construction, and Pruning in Regulated Multi-Agent AI Systems

This technical note describes the Routing Brain v2 implemented in APR-Agentic v5.0.0 as Phase 5 of the APR-Lite governance platform. The Routing Brain v2 replaces the Phase 3 routing stub with a full governed routing resolution engine. Every routing request undergoes: drift-state evaluation against the current session, intent classification via the APR-Lite substrate (enforcing sector rules, privilege ceilings, and envelope constraints on the input itself), candidate agent selection from the agent_registry with sector defaults fallback, deterministic pruning of unsafe, frozen, and policy-forbidden agents, topology construction (sequential, parallel, or hierarchical), routing plan persistence, and governance evidence emission. The implementation is validated against a 16-test smoke suite. A key architectural invariant established in Phase 5: intent classification is content evaluation and must route through the substrate, not a local classifier.