Self-Improving AI Needs What Built Human Knowledge: Model Domains, Mappings, and Records

The applied knowledge that humans have accumulated over millennia exhibits a consistent underlying structure: knowledge is not a collection of propositions but a network of model domains, mappings, and records. Knowledge creation corresponds to three operations on this network — creating a domain, establishing a mapping, and generating a record. This paper argues that (1) this structure uniformly describes nearly all applied knowledge from arithmetic to engineering, and converges with several established theories in the philosophy of science and cognitive science (Hertz's principle of images, Rosen's modeling relation, Gentner's structure-mapping theory, dual-process theory, knowledge compilation); (2) the dominant AI paradigms — sequence prediction in large language models and end-to-end monolithic world models — each lack essential elements of this network, while the most successful reasoning systems of recent years (AlphaZero, AlphaGeometry, DreamCoder, program-synthesis ARC solvers) all contain recognizable counterparts of the three elements; and (3) on grounds of computational economics, any bounded-compute agent that must solve novel problems in an open world will almost inevitably evolve a structure of multiple models plus pruning plus caching. The framework thereby recasts "self-improving AI" as the autonomous, indefinite growth of an explicit knowledge network — new domains, new mappings, new records — rather than as updates to a fixed set of parameters. We close with a semi-formal framework, a module-level correspondence to existing technology, a set of falsifiable predictions, and a staged experimental program.