Constructive Ontology: A Layered Constructive Verification Framework for AI “Confabulation”

LLM frequently produce “confabulations”, because the outputs lack corresponding objects of existence. Through a systematic examination of the transcendental philosophical traditions, the three major schools of mathematical philosophy (intuitionism, logicism, formalism), and contemporary AI engineering approaches, this paper concludes that the primary reason these approaches have not resolved AI confabulation is their failure to provide a unified, operational verification framework covering all cognitive domains (nature, morality, and aesthetics). To address this problem, this paper proposes “Constructive Ontology”—a transcendental philosophical framework grounded in the principle “to be is to be constructed”. Drawing on Kant’s three Critiques, it divides all cognitive domains into three regional ontologies—Theoretical Ontology, Practical Ontology, and Regulative Ontology—and further into nine objects of existence. Under the guidance of a unified categorical grammar of transcendental consciousness structure, corresponding construction rules and verification standards are established for each type of object. The concept of “Constructive Mismatch” is introduced to transform the AI confabulation problem into a measurable deviation between objects of existence and verification standards. This framework enables different objects to correspond to different constructive methods, transforming ontological objects of existence into effectively executable constructive procedures, thereby providing a philosophical foundation for verifiable and trustworthy AI.