For years, the alignment community has focused almost entirely on the model’s output — making sure the final tokens are safe, helpful, and honest. RLHF, DPO, constitutional AI, output filters — all of it operates at the surface level. But what if the m…