Well problematic because they are using their own verifier as apanem of experts, with their own model trained specifically to satisfy this verifier. On the benchmark runs, they dont mention using human experts to cross validate their scores.