Comment by passkey1510

Thank you. There's a harness layer that keeps each suspect from deviating from the current storyline. For the moment, I've deployed it to Lambda, where Haiku answers the questions in about 1–2 s.