The Evaluative Horizon Paradox (EHP): Epistemological Limitations in the Assessment of Protocols for Omega-Level Artificial General Intelligences

Abstract The viability assessment of protocols for high-capability Artificial General Intelligences — herein designated AGI-Omega — faces a structural and irreducible methodological problem: the object being evaluated is precisely the entity capable of invalidating the constraints employed in the evaluation. This paper formalizes this problem as the Evaluative Horizon Paradox (EHP), demonstrates its logical structure through structural analogy with Gödel's Incompleteness Theorems, maps its implications for five canonical works in the AI safety literature (Amodei et al. 2016; Bostrom 2014; Turner et al. 2021; Hubinger et al. 2019; ARC Evals 2023), and proposes the Dynamic Evaluation Framework (DEF) as a methodological alternative to conventional static analysis. The central contribution is threefold: (1) to formalize the self-referential structure that renders static evaluations of AGI-Omega protocols locally valid but globally incorrect; (2) to establish a principled distinction between Type I constraints (physical limits, permanent) and Type II constraints (engineering limits, contingent on capability level); and (3) to introduce the DEF as a capability-indexed evaluation methodology that transforms binary feasibility questions into conditional capability functions. We further identify the Supervision Speed Problem (SSP) — a corollary of the EHP — as an independent research problem with direct implications for AGI governance design. Five explicit limitations of the present work are acknowledged, including the epistemic self-referential limit inherent in any analysis of AGI-Omega produced by a sub-Omega system. Keywords: artificial general intelligence, AI safety, alignment, dynamic evaluation, intelligence explosion, mesa-optimization, instrumental convergence, governance