Research Data for “From Extraction Accuracy to Analytical Admissibility: A Human–AI Division of Labor Protocol for LLM-Based Supply-Chain Risk Intelligence”

Abstract

This dataset contains the reproducibility materials supporting the study “From Extraction Accuracy to Analytical Admissibility: A Human–AI Division of Labor Protocol for LLM-Based Supply-Chain Risk Intelligence.” The package supports an offline benchmark of four open-weight large language models—DeepSeek-V3.2, Qwen3.8-Max, Kimi K2.5, and GLM-5.3—for supply-chain risk intelligence under a text-bounded evaluation protocol. It includes two datasets: an internal development set of 58 English-language nickel-supply articles and a held-out general-news set of 97 articles, including 11 risk-positive cases. Model inputs are restricted to four fields: article identifier, headline, publication date, and excerpt/summary. The archived materials include frozen labels, prompt architectures, public model outputs, scoring scripts, result tables, data dictionaries, and documentation. The evaluated tasks cover risk-relevance screening, Tier-1 and Tier-2 mechanism classification, temporal judgment using TRUE/FALSE/UNKNOWN labels, and an operational human-review triage analysis. The H032 case is retained as an illustrative definitional-boundary case. The package reproduces the manuscript’s main screening, mechanism-classification, temporal, and triage results from archived model outputs without rerunning any model API. Full copyrighted news articles, private reasoning traces, credentials, personal information, and local machine paths are not included. The dataset is intended to support transparent verification of the reported human–AI division-of-labor analysis.