Comment by tough

that relies on the positive-aligned RLHF models most labs do.

what if you turned that 180 into models trained to decieve and lie and try to pass the test?