I’ve recently been using Astra to help test and refine the point-deduction feature in one of my programs. As it turned out, Astra set an acceptance criterion for itself that was virtually impossible to meet. Consequently, it went through 30 rounds of self-correction without any intervention from me, consuming 1.5 times its weekly quota for the "20x" tier. Eventually, it was Opus 5.5 that identified the issue and corrected Astra’s objectives. This is a true story—I’m curious why this happened. Was my prompt not clear enough, or was Astra being overly cautious?