Research on a Next‑Generation AI Technology Pathway Based on Hierarchical Verification and Public Co‑Construction: An Engineering Contingency Plan for Eliminating Hallucinations in Large Language Models in the Post‑Scaling Law Era

Large language models built on the Transformer architecture have now fallen into an impasse where the marginal returns of parameter scaling continue to decline. Beyond GPT-8, the physical constraints of the Scaling Law are approaching their limit, while the reliability requirements for deploying AI in high-risk vertical domains—healthcare, judiciary, finance, and public policy—are growing exponentially. To date, humanity has not yet deciphered the fundamental mechanism of general intelligence from first principles; relying blindly on the spontaneous emergence of intelligence to drive industrial iteration lacks both engineering rigor and practical controllability. This paper proposes a systematic solution that does not depend on unknown breakthroughs in intelligence and can be implemented entirely with existing mature engineering technologies: a dual-layer decoupled architecture composed of a "bottom-layer world common-sense foundational model" and an "upper-layer language generation and expression module." Using a three-tier hierarchical structured knowledge base—comprising an Absolute Truth Layer, an Expert Verification Layer, and a Public Common-Sense Layer—the framework establishes a fully closed-loop collaborative verification mechanism of "AI candidate generation → quantitative content-deviation detection → lightweight mass verification → expert final adjudication on disagreements." This enables full-process common-sense constraints and factual fidelity for model outputs. The scheme thoroughly decouples objective world knowledge from language generation, adopts publicly available authoritative encyclopedias and peer-reviewed academic findings as the bottom-layer absolute truth baseline, and distributes the massive cost of fact verification through society-wide public collaboration. Compared with continuously scaling model parameters, this pathway exhibits long-term diminishing marginal cost, with each generation delivering quantifiable and traceable hallucination-suppression metrics, thus clearly charting the generational evolution of large models in the Post-Scaling Law era. At the same time, a standardized and structured world public knowledge base will lay a universal public infrastructure foundation for the next paradigm shift in artificial intelligence. The paper adheres to a core engineering philosophy: before the underlying scientific mechanism of intelligence is clarified, the primary priority of the AI industry is to achieve ultimate reliability in information automation across all scenarios.