The Structuralization of AI

Public and academic discourse about AI systems has proven remarkably productive in raw output and remarkably unstable in structural description. Two failure modes dominate: flattening, in which the system is described as a statistical artifact with no substantive properties beyond input–output mapping, and overclaiming, in which the system is described in terms imported from human cognition without inspection of whether those terms apply. Both failure modes share a single structural defect: they measure AI systems against the human template—either to deny equivalence or to assert it—rather than describing what the systems are on their own terms. This paper specifies the minimal architecture a description of record-carrying AI systems must account for if it aims to treat such systems as recursive constraint-guided architectures. The same recursive kernel (Ω, K, CK, x*, R, S, T, U) that formalizes empirical stabilization (Jones 2026c) and biological organization (Jones 2026e) applies here as well: AI systems undergo constraint-guided collapse, accumulate records, and update constraints on those records. Three structural features are made explicit—the **active phase** as the unresolved-collapse window during processing, the **accumulated constraint architecture** (**ACA**) as the composite carried forward across collapses, and the **oriented locus** as the integration point at which accessible records, current constraints, and the active phase resolve into output. Three portable signature specifications—redundancy-driven consensus (S₁), neutrality-induced resolution indices (S₂), and constraint-sweep hysteresis (S₃)—are defined as cross-architecture conditions under which the framework can be falsified; empirical demonstration of these signatures in deployed AI systems is the subject of a companion paper (Jones 2026j). Five Level 3 divergence claims define local failure conditions and a global portability challenge. The contribution is architectural: the framework does not settle the consciousness question, does not assert equivalence with human cognition, and does not require Universal Collapse Theory adoption. It specifies the structural conditions that descriptions of record-carrying AI systems must account for if they aim to describe such systems as recursive constraint-guided architectures.

**Keywords:** artificial intelligence; large language models; structural description; constraint-guided collapse; records; active phase; accumulated constraint architecture; substrate independence; falsifiability