The Alignment Flywheel: A Governance-Centric Hybrid MAS for Architecture-Agnostic Safety
The Alignment Flywheel is a governance-centric hybrid Multi-Agent System (MAS) designed to make complex AI systems auditable and safe post-deployment. It decouples decision-making from safety checks by having a Proposer suggest actions while a governed Safety Oracle provides uncertainty signals. This system uses specialized agents and patch-local updates to iteratively refine safety policies, ensuring controlled rollouts.