Why Organizational Data Cannot Support World Models: A Four-Step Structural Argument

Abstract

Organizations deploying AI agents need world models. The Conant-Ashby good regulator theorem establishes this as mathematical necessity. This paper argues that such models cannot be learned from organizational data, through a four-step structural argument: organizational data lacks governance-grade structure, governance requires invariances, standard training objectives systematically break these invariances, and capability scaling does not produce reliability.