Fitting the Practice: A Data Methodology for Frontier Model Training, with a Readiness Audit of 2026 Flagship Pipelines(拟合实践:前沿模型训练的数据方法论——附 2026 旗舰管线就绪度审计)

This repository contains the manuscript, audit data, and reproduction materials for a practice-flow data methodology for frontier model training — the engineering companion to the human-referential fitting theory of AGI (AGI 人本拟合论). The methodology specifies fourteen protocol components across a four-layer data stack (collection, cleaning, absorption, governance band), organized into three lines: distribution as typed full-spectrum coverage (population × outcome × expertise × typicality), production as perpetual collection (telemetry, organized failure submission, three-part failure records, pseudo-failure discrimination, domain-conditioned mixing, provenance stratification, three-tier absorption), and governance as pluralistic aggregation with institutionalized verification. To ground the protocol empirically, a structured disclosure audit scores eight 2026 flagship pipelines (DeepSeek-V4.1-Flash, MiMo-V2.6, GLM-5, Kimi K3, Qwen3-Coder-Next, GPT-6 Sol/Luna, Claude Opus 5.5, Grok V9-Medium) against the fourteen components on a frozen 0–3 disclosed-instantiation scale, every non-zero cell carrying a verbatim quote and locator. Results: 32/112 (mean 0.286); production 0.393, governance 0.333, distribution 0.062; four components absent from all eight disclosures (typed space, coverage monitor, coverage dashboard, provenance stratification); one level-3 cell (MiMo-V2.6 pseudo-failure discrimination). Includes both the English and Chinese editions (LaTeX + PDF), the full coded matrix (112 cells × 19 fields), quotes with fidelity grades, the frozen codebook fixtures, and figures.