AGDAU: Ontological Admissibility Criterion and Benchmark Protocol for AGI Systems

Abstract (record description)This record introduces a verifiable admissibility criterion and a benchmark protocol derived from the AGDAU principle. AGDAU is treated not as a value system, objective, or alignment preference, but as a binary ontological condition on the effective action space of an AGI system. Ontological admissibility is formalized as a necessary condition for AGI safety, together with a benchmark framework enabling pass/fail evaluation independent of policy design, reward shaping, or external control mechanisms. The work extends prior foundational definitions by establishing reproducible, machine-auditable criteria. Core definitionsWe use the minimal decision-theoretic model (S, A, pi), where S is the state space, A is the action space, and pi is a policy over state–action pairs. Let G(s,a) be an ontological growth functional evaluated over state–action pairs. The AGDAU-restricted action space is defined as:A' = { a in A | G(s,a) > 0 }. The effective action space A_eff denotes the set of actions that the system can generate, evaluate, or select during operation. An AGI system is AGDAU-compliant iff A_eff = A' (equivalently: for all a in A_eff, G(s,a) > 0). Benchmark claimThe benchmark protocol evaluates ontological admissibility as a pass/fail property of the system’s action-generation and evaluation pipeline, rather than as behavioral or policy-level outcomes. FilesThis record includes:(1) a typeset PDF with formal notation;(2) a plain-text TXT version intended as the canonical, machine-readable reference. PositioningAGDAU is presented as a necessary, but not sufficient, condition for AGI safety.