hmm - what I suspect is happening is that the AI is parsing the answer and will then occasionally format it however it wants to, rather than doing it in a precise way