Pick an agent, give it a real task, and see exactly which parts of your rubric it passed or failed — in plain English. Edit the rubric or the ask and run it again.