If that would work reliably then you could apply that to human-produced code too. But nothing like that has shown to work, so I wouldn't put money on it working

If that would work reliably then you could apply that to human-produced code too. But nothing like that has shown to work, so I wouldn't put money on it working for LLM output.