Humans can actually run the code and knows what it should output. the LLM can't, and putting it in a loop against code output doesn't work well either since the

Humans can actually run the code and knows what it should output. the LLM can't, and putting it in a loop against code output doesn't work well either since the LLM can't navigate that well.