LLM-Compiled Wikis for Cross-Institutional Policy Retrieval: Benchmark, Compiled Wiki, Evaluation Results, and Audit Log

This deposit contains the full research artifacts supporting themanuscript "LLM-Compiled Wikis for Cross-Institutional PolicyRetrieval: A Benchmark on Chinese University Grant Policies"(B. Zhang, 2026, Journal of Future Artificial Intelligence andTechnologies, forthcoming). Contents:- A T3a benchmark of 28 cross-institutional policy-conflict queries over 22 Chinese university grant-management policy documents, with article-level gold citations.- Six compiled wiki pages produced by an LLM-based wiki-compilation pipeline (one per policy dimension).- Full evaluation results from the canonical run reported in the manuscript: 84 raw predictions, aggregated metrics, retrieval-only recall analysis, failure-mode breakdown, and statistical- significance tests.- A structured human-audit log of all 172 LLM-extracted gold citations across 6 representative queries.- All prompt templates, evidence-window keyword lists, and configuration files needed to re-run the pipeline.- The Python source code of the wiki compiler, baseline systems, evaluation harness, and analysis scripts.- The final manuscript and the point-by-point response to reviewers. The 22 source policy Markdown files are derived from publiclyavailable administrative regulations on the official websites of 19Chinese universities; original copyright remains with each issuinginstitution and the Markdown files are redistributed under fair-useresearch-purpose terms. All other artifacts are released underCC-BY 4.0 (data) and MIT (source code). See LICENSE.md in the archivefor details. Source URLs for the 22 institutional documents will be added tocorpus_metadata.csv in a forthcoming v1.1 release; original documentsare publicly available via search on each university's officialwebsite by document name.