Defensible AI, not just accurate AI: why, in high-stakes domains, the model should read the evidence but never make the call.
