The artificial intelligence industry defines intelligence implicitly through what it measures: reasoning ability, knowledge recall, code generation, language fluency. This paper argues that this definition is incomplete in a specific and consequential way. It captures the generative dimension of intelligence while omitting the stabilising dimension — the structural capacity to determine whether outputs correspond to reality. The paper proposes a reframing: intelligence is most usefully understood as the capacity to reduce uncertainty about reality. Under this framing, reasoning, memory, prediction, and language are not intelligence itself — they serve intelligence, as mechanisms by which uncertainty can be reduced. Verification — the process of testing outputs against reality — provides the stabilising function that maintains intelligence's relationship with the world it operates in. The scientific method is the existence proof. It converts unreliable individual cognition into reliable collective knowledge not through the quality of individual reasoning but through verification architecture. The cybernetics tradition provides the theoretical framework: Ashby's Law of Requisite Variety establishes that a regulatory mechanism must match the complexity of the system it regulates, and current AI verification mechanisms fall far short of the generative capabilities they are supposed to govern. Five independent intellectual traditions — Western neural network research, Chinese cognitive philosophy, Indian logical epistemology, cryptographic formal methods, and category-theoretic provenance — have converged on the same structural diagnosis from different starting points: the current AI paradigm is structurally incomplete because it implements generation without verification. A companion paper (Paper 5: The Verification Substrate) specifies the architectural principles that follow from this reframing. ---