I think it’s exactly that empirical/agentic solution that other models fail to see. E.g. o3 (non-pro) often cheats by appending a meaningless suffix of random noise, which is the only lever through which it brute forces one fixed, plausible count.

→ View original post on X — @goodside