https://howtonotcode.com/story/2083-anthropic-proposes-thresholded-ai-safety-rules-with-stop-ship-powerstime-to-harden-eval-and-audit