Current AI systems, regardless of scale, are static pattern matchers. They do not learn from mistakes, do not actively experiment, and cannot abstract concepts from physical interaction. Humans are fundamentally different: we extract key tokens through positive and negative feedback from interaction, name them, build constraint systems, and continuously iterate. This paper argues that achieving genuine AGI requires a dual-architecture system: a frozen fast reflex layer for pattern completion and prediction, an open slow learning layer for physical interaction and verification, connected by an abstraction engine that drives token discovery and evolution through physical feedback — not text labels. We provide a complete architectural design, component specifications, training procedures, and a three-phase implementation roadmap. The contribution of this paper lies not in citing existing research, but in proposing an engineerable AGI system design from first principles.