Fitting the Practice, Not the Works: A Human-Referential Theory of AGI(拟合实践,而非作品:人本参照的AGI理论)

This repository contains the manuscript, codebook, coded corpus, and reproducible analysis for a theory of artificial general intelligence in three connected claims. Conceptually, AGI is a philosophical concept: as long as "generality" is referenced to the open set of human practice, it cannot be reduced to a technical specification. In implementation, the means of AGI is the fitting of human practice — not the fitting of humanity's accumulated works; the theory's core is a temporal clause: "The object of fitting is not humanity's works but its practice — the full spectrum and the full course of human experience. The stock has a wall; the flow does not." Normatively, AGI should be positioned as an augmentative tool, not a new subject. To make the theory auditable, we built a structured evidence map: a purposefully sampled corpus of N = 80 works coded on two stance axes (descriptive and normative) and eight fitting-object dimensions with explicit/implicit modality. Key quantified results: the experience-type coverage axis (success/failure, expert/novice, typical/marginal) is explicitly treated in 0/80 works (exact binomial 95% CI [0.0%, 4.5%]) while implicitly touched in 9/80; practice-as-flow is explicit in only 2/80 works, both from 2026; a registered absence-search (8 query families across arXiv, ACL Anthology, and Chinese-language web search) confirms zero D8-explicit hits. Four termination boundaries for flow-side fitting and explicit falsification conditions are stated. This paper builds on a prior empirical audit of failure data in LLM training (DOI 10.5281/zenodo.23355929); engineering protocols for practice-flow data are future work. Includes the English manuscript (LaTeX + PDF), the codebook (dual stance axes A1/A2 + dimensions D1–D8 + modality rules R1–R5, CR1–CR6), the full coded corpus (80 works × 24 fields), and a single-file reproduction script.