[dist-epoch]

I was thinking more along this line: you can solve unreliably 100% of the problem with LLMs, or solve reliably only 80% of the problem.

So you trade reliability to get to that extra 20% of hard cases.