~/martinrussmann.com

The Case for a Normative Routing Layer Upstream of AI Agent Systems: Every major AI routing system optimizes the same question — which model handles this most efficiently — and none of them treat a different question as a first-class routing objective: whether autonomous processing is appropriate given the organizational stakes. GARA, the Governance-Aware Routing Architecture, specifies the structural pattern for a normative routing layer that sits upstream of any existing router or orchestration framework, evaluating governance-relevant properties of the information context — epistemic, distributional, normative, reliability, and learning properties — and enforcing non-compensatory constraints where certain conditions mandate human involvement regardless of how favorable all other conditions are. The architecture assigns each request to one of three processing regimes (AI-led, Hybrid, or Human-led) before the downstream router runs, producing an auditable governance state vector for every routing decision. The core architectural claim is that governance and performance optimization require separate layers with fundamentally different scoring logic, because compensatory scoring functions structurally cannot preserve non-compensatory safety properties — and this separation becomes more necessary as AI capability increases, not less.