What? Can they just look if they fetched certain documents/conversations and feeded them into the training loop?