Atria Dawn Preview: an AI model trained end-to-end on verifiable research

training on verifiable research seems like the right constraint. what kind of claim still trips it up most?