Tonn-AGI: A Falsifiable Architecture for Persistent Organizational General Intelligence

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Tonn-AGI proposes a falsifiable architecture for Artificial General Intelligence in which general intelligence is treated not as a capability that must reside inside a single model, but as a potentially emergent property of a persistent cognitive organization composed of replaceable models, agents, skills, tools, software, and verification mechanisms. Building directly on The AGI Category Error, the paper operationalizes the Organizational Deficit Hypothesis: the possibility that many of the cognitive primitives required for general intelligence already exist, while the remaining bottleneck lies in how those primitives are persistently composed, routed, verified, learned from, and reorganized over time. The proposed architecture combines a persistent epistemic and execution ledger, capability acquisition and recomposition, recursive work decomposition, adaptive cognitive routing, competitive generation, independent verification, organizational learning, component substitution, and Persistent Cognitive Hygiene — the maintenance of bounded active cognition over an expanding historical evidence base. A central distinction is introduced between ephemeral cognition and persistent organizational state. Individual workers may reason, fail, disappear, or be replaced without destroying the identity or accumulated knowledge of the cognitive system. Claims produced by models do not automatically become system truth; durable state transitions require evidence and verification. The paper further separates three learning timescales: within-task state adaptation, cross-task organizational learning from verified outcomes, and optional later parametric consolidation of repeatedly successful cognitive trajectories. This separation allows the organizational contribution itself to be tested while model capabilities remain fixed. Tonn-AGI is explicitly presented as a pre-implementation architecture and experimental pre-registration, not as an empirical declaration that AGI has already been achieved. The paper defines in advance the ablations, resource controls, component-substitution tests, unseen-task transfer requirements, No-Homunculus constraints, organizational-efficiency measures, and falsification conditions required to evaluate the hypothesis. Its central question is therefore empirical: Can a persistent organization of non-canonical, replaceable cognitive components develop transferable general capability that cannot be reduced to any single constituent operating alone? If the answer is negative, the architecture should fail under the specified tests. If the answer is positive, the implication is fundamental: the missing step toward AGI may not have been one final, sufficiently capable model, but the organizational structure required to make existing cognition persistent, composable, selective, verifiable, adaptive, and capable of surviving the replacement of its own components.