IAUC Systematic Literature Review: Extraction Protocol, Metadata, and Validation Data

Abstract

Data and code for a systematic review of the urban climate literature. A large language model extracted structured information from the full text of every paper, and domain experts checked a sample of that extraction. This package lets you check the data and rebuild the paper's statistics, tables and figures. Start here Read README.md first. make check checks your setup. uv sync --extra dev installs the pinned environment, make paper-all rebuilds the statistics and tables, make figures draws the figures, and make ci-test runs the tests. No API keys are needed. data/README.md explains which values came from the model, which from fixed code, and which from the experts. Contents data/extracted_metadata/: 9,284 records, one per paper, as the model extracted them data/enrichments/ and data/contribution_rationales/: 9,284 records each of values computed from the extraction, and the reasoning behind each contribution label data/volunteer_review/: the anonymized expert review responses from both validation rounds, and the record of how a third reviewer settled the disagreements in the blind second round data/cross_model_validation/: comparisons with other models that support the choice of extraction model data/experiments/recall_audit_2025/: the search-recall audit, with its query strings, per-journal counts and the screening decision for each extra paper, and the check of how papers were assigned to publication years, with the corrected annual series data/contribution_reproducibility/: the earlier contribution labels the revision table is checked against, and the contribution classifier's attestation and the production promotion manifest that binds it src/, tests/, Makefile and uv.lock: the analysis code and its tests License Everything in this deposit, data and code alike, is released under CC-BY-4.0. This version Refreshes the archive for the revised manuscript. Compared with version 3.0 (1 August 2026), the extracted records and the set of papers are unchanged. Corrections to where study locations are placed update the coordinates, climate zone, coastal status and population class in 1,166 enrichment records, and with them the location-based statistics and maps; the other enrichment fields are unchanged. The package can now also be run from a clean download. It also adds the blind second-round expert responses and the record of their adjudication, with the script that recomputes their figures; the recall audit's per-paper screening decisions; the diagnostic that found early-access papers counted under the year they were first listed rather than the year they were published, with the corrected annual series; and the contribution classifier's attestation. Offline scripts now recompute the Supplementary audit figures: the workload and cost footprint, the search-recall audit, the screening funnel, the climate-zone rule sensitivity, the classifier gate scores and the topic-cluster silhouette scores.