@mjd Do you have a reason to expect whether replacing conversation with directly amending the language model's output (in a similar way to one that people use to guarantee that the output is e.g. valid json) and using that amended output for context would be better/worse?