[ltononro]

Using LLMs to judge this could be very hackable, shouldnt it be? Is that the best practice we expect? Would be very interested in see some ablations about failure modes where the LLMs tried to hack it somehow. Or failures from the llmaaj