Position paper on AI oversight quality as a training signal — written with Claude, by an actuary. Feedback welcome on whether the mechanism holds.
The Anthropic/Pentagon situation in February prompted me to think about a governance problem I haven't seen framed quite this way: not whether AI companies should refuse certain uses, but what happens to AI models when human oversight of their outp…