I think it's already out of date with verifiable reward based RL, e.g. on maths domain. When "correctness" arguments fall, the argument will probably just shift to whether it's just "intelligent brute force".