Humans see it instantly — LLM doesn’t

Part of it is just how the problem is framed. Models usually jump straight into trying to produce a path because that’s what the prompt suggests, so they skip that initial check entirely. If you force them to first answer if its even solvable they normally do a lot better.