Then the learned behavior should be to remember the reason the estimates were wrong last time, and to include those this time.

And no, I do not think it is that easy. The point is that refusing to speak simply robs the process of valuable input.