The instruction to say what is missing instead of filling the gap does the heavy lifting in that prompt. I have seen agents write a confident brief off one weak transcript, so I now log how many sources backed each claim and fail the run under a threshold. Do you check speaker labels programmatically before quoting?