The worker never writes its own test: how my agents are kept from grading their own homework

In my last post the rule was: no model says done, a command exits 0. The obvious hole came up in the...