Look into splitting policy reasoning apart from koreagrafeye

[termontwouter]

This is more modular, allowing us to exchange the policy reasoner separate from the orchestrator. Whatever policy reasoner then feeds its conclusions to whatever orchestrator.