Comment by danbolt

I sometimes wonder if the verbosity gives the LLM more “margin for error” when generating text that passes requirements.