The global artificial intelligence (AI) evaluation system faces challenges of "measurement black boxes" and "strategic misjudgments." To systematically deconstruct its internal structure, this study integrates 22 mainstream AI evaluation indexes (637 original indicators), constructs a standardized library of 307 comparable indicators, and uses a dual-dimensional analysis (attention distribution and complex network modeling) to reveal its internal operational logic. The study, for the first time, identifies and names the dominant "national capacity-based, scale-oriented growth" evaluation paradigm in the field. This paradigm takes national capacity as its evaluation core, is driven by a "policy-technology" dual engine, exhibits a significant scale-oriented bias in its measurement standards, and consequently has structural blind spots in dimensions such as innovation quality and conversion efficiency. By providing a structural blueprint of this paradigm, this study offers a critical empirical pathway for avoiding strategic misjudgments and designing more balanced next-generation evaluation frameworks.