@RokoMijic generate synthetic data + some way to discriminate higher quality from lower -> slowly inch towards better than was in the sources ... is easier for something like math proofs where checking much easier than generating