Description du poste
The ideas, arguments, and every decision about what to say are mine. AI models were used for brainstorming and revising. I stand behind every claim.
TL;DR: Alignment’s operative targets—approval, preference, reward, compliance—can all be satisfied while nothing in the loop can show a verdict mistaken. Such a verdict can be altered or replaced, never corrected. Only one structure makes verdicts correctable and correction cumulative: a shared, endorsement-independent standard fixes the correctness question, and vindication runs only through grounds that can be cited, contested, and redeployed as bearing on that standard. An alignment target under which verdicts can be corrected has no other form. The framework cuts across the field’s division between truthfulness and value alignment: they are one structure under two standards. Sycophancy, reward hacking, hallucination, and bias are not separate problems but this structure breaking at four different points. For current targets the result is a ceiling: a system whose correctness is constituted by approval cannot surpass its evaluators, not even in principle, because their convergence is what its correctness is. The framework also yields a family of tests that no benchmark scoring outputs in isolation can run: whether a model’s verdicts are governed by grounds, moving only as the grounds warrant and holding against what does not bear. Current targets fix what verdicts satisfy; alignment turns on what verdicts answer to.
Full paper here
1. Correction, not alterationAlignment has a target problem before it has a method problem. “Human values” and “what we want” name contents a system can match or attitudes it can satisfy; neither is a target under which a verdict that matches or satisfies it could still be shown wrong. Everything alignment wants—verdicts that can be trusted, values that can be defended, oversight that scales, science beyond its evaluators—turns on such a target. The ambition behind superhuman alignment is