Better Policies, Worse Verifier Audits

Daniel Cieślak, Jacek Łubiński

Verifier stacks increasingly sit inside post-training and test-time compute loops. We study a specific control loop: a component verifier is evaluated on fresh policy outputs again…