On the other hand, if it represents scratch space in English, it's a lot easier to see how it justifies its answer and to tell where it's gone wrong. Debuggability seems pretty important?

Maybe it just needs more training at "thinking out loud" so it does it without prompting?