Epistemic Production Is Not Organizational Persistence: A Falsifiable Architectural Hypothesis for Long-Horizon Artificial Intelligence

Problem. Contemporary evaluation of frontier artificial intelligence (AI) increasingly includes long-context, multi-session memory, tool-use, and stateful-workflow benchmarks. These instruments probe important persistence capacities, but they do not directly operationalize the joint trajectory-level construct examined here: organizational persistence (OP), the maintenance of claim/object identity, dependency structure, unresolved contradiction, status classification, and reconstructable lineage across interruption, revision, and representational change. Claim. Epistemic production (EP) and organizational persistence (OP) are distinct candidate dimensions of long-horizon artificial systems. We test whether high EP can coexist with low OP under perturbation and whether explicit mechanisms for preserving object identity, dependencies, status, and transition history improve OP beyond weaker persistence architectures. Epistemic Governance Architecture (EGA) is used here as a stipulative functional taxonomy for system-level grammars that organize tracked claims, evidence, relations, and transitions across time. Architectural necessity and functional minimality are separate hypotheses requiring different evidence and later adjudication. Method. We define OP behaviourally through recovery under perturbation, specify seven candidate organizational functions, and position them against mature and rapidly developing work on truth maintenance, provenance, cognitive architectures, organizational memory, governed agent memory, and long-horizon evaluation. We then present the frozen Premise Mutation–Rediscovery–Silent Status Transition (PM–R–ST) preregistration: six conditions ranging from bare probabilistic inference to persistent external structural state with executable transition control. The preregistered primary confirmatory contrast is F − E, which isolates executable enforcement from persistent state and recording. Evidentiary status. This manuscript reports no experimental results (N = 0). The empirical programme is preregistered and unexecuted. Internal observations from a multi-year AI-mediated research programme are reported only as motivating observations and are firewalled from validation. An unsuccessful attempt to maintain organizational state inside model context explains why an external realization is tested; it does not validate that realization. Falsification. The strongest programme-level null is commodity-sidecar sufficiency: generic logging, retrieval, workflow, version control, database state, and policy components can realize the relevant organizational functions without any performance advantage for a dedicated transition-control architecture. The preregistered primary confirmatory null is H0P (F ≤ E): executable enforcement adds no benefit over persistent external state alone. Retrieval, emergent model capability, ordinary software engineering, overconstraint, and measurement insensitivity remain admissible alternative explanations.