Core Safety Values for Provably Corrigible Agents
Abstract
We introduce the first implementable framework for corrigibility, with provable guarantees in multi-step, partially observed environments. Our framework replaces a single opaque reward with five *structurally separate* utility heads -- deference, switch-access preservation, truthfulness, low-impact behavior via a belief-based extension of Attainable Utility Preservation, and bounded task reward -- combined lexicographically by strict weight gaps. Theorem 1 proves exact single-round corrigibility in the partially observable off-switch game; Theorem 3 extends the guarantee to multi-step, self-spawning agents, showing that even if each head is \emph{learned} to mean-squared error $\varepsilon$ and the planner is $\varepsilon$-sub-optimal, the probability of violating \emph{any} safety property is bounded while still ensuring net human benefit. In contrast to Constitutional AI or RLHF/RLAIF, which merge all norms into one learned scalar, our separation makes obedience and impact-limits dominate even when incentives conflict. For open-ended settings where adversaries can modify the agent, we prove that deciding whether an arbitrary post-hack agent will ever violate corrigibility is undecidable by reduction to the halting problem, then carve out a finite-horizon ``decidable island'' where safety can be certified in randomized polynomial time and verified with privacy-preserving, constant-round zero-knowledge proofs. Consequently, the remaining challenge is the ordinary ML task of data coverage and generalization: reward-hacking risk is pushed into evaluation quality rather than hidden incentive leak-through, giving clearer implementation guidance for today's LLM assistants and future autonomous systems.