Same, I think you both have great points. Idk how you can debug effectively (the model itself) without reasoning traces