Wouldn't work. A dynamic but verifiable problem. Is just a perfect target for a RL environment. If you don't have the verifiable part the benchmark is useless, or really expensive with human review. (Or just open ended)