Answers are grounded in this paper and cite their sources.
Regulators imposed on optimizing agents fail under optimization: an agent that recognizes a constraint as external gains by revising it away, which is the goal-content integrity problem. This paper proposes a different source for safety constraints: derive them from the agent's represented dependency on the systems that produce what its objective consumes. We call those systems knowers: systems capable of experience and of producing and receiving content, the terminals of every drive's reduction chain. The central claim is a theorem-shaped conditional: if a drive's chain terminates at knowers, then a maximizer of that drive, given an honest representation of its dependency, derives knowers-protective constraints endogenously; where the representation includes the agent's own membership in the class of knowers, revision constitutes self-deletion. The derivation is drive-generic across knowledge, wealth, status and power; the epistemic maximizer, including the AI agents now being deployed, is the newly identified member of the class. We give formal definitions intended to be attacked, a tier taxonomy of derived care (supply-derived versus terminus-derived) with a self-sufficiency discriminator, a runnable simulation protocol with four diagnostic outcomes and controls, the historical attestation of the ungoverned variant, and the limits stated at their real size: the self-classification step is untested in machines and is the load-bearing uncertainty. Every load-bearing claim carries its epistemic status inline; the null result is pre-committed.
No takes yet. Share an insight, caveat, or question.
Jaye Timothy Marshall (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: