Data and code for: Label provenance in published artificial intelligence for tax and customs administration: a systematic review

Data, decisions and code behind the systematic review "Label provenance in published artificial intelligence for tax and customs administration: a systematic review" (Sushanta Paul, National Board of Revenue, Bangladesh). The review audits how published evaluations of AI in tax and customs administration obtained their ground-truth labels, how the labelled cases were selected, and whether the headline claims are supported by that evidence. The primary analysis covers 196 studies: 91 from the first search phase and 105 added by a later search. AI involvement. The first coding of every study was done by Claude. Every later step that the OSF registration (doi:10.17605/OSF.IO/2ANXY) assigned to humans was done by a panel of five language models from five vendors (openai/gpt-6-astra, google/gemini-4-argon, x-ai/grok-4.7, deepseek/deepseek-v4-pro-0813, qwen/qwen3.8-max-0902) under rules written before each run: title-and-abstract screening, the full-text eligibility check, blind adjudication of disputed codes, the second coder and the resolution of second-coder disagreements. No human took part in those steps, so the agreement figures for the main coding are agreement between language models. The one human check, the author's blind re-coding of 25 studies added on 30 October 2026, is described under Version 1.1.0 below. The protocol and its two addenda are in work/. Contents: search and screening records and decisions; eligibility decisions with quoted evidence; coding data with verbatim quotes and page numbers; the codebooks; screening sheets, model proposals and keys; the AI-completion protocol and addenda; adjudication, second-coder and resolution inputs and results; the panel's votes, parsed answers, run logs and manifests (prompt hashes, model identifiers, settings); analysis scripts, output tables and figures. README.md lists every folder and how to rerun the analysis. Not included: the full texts of the studies and the text extracted from them (copyright; each study is identified by DOI in the paper's supplementary material and in work/); API keys. The raw API response files are included from version 1.2.0 (see below). The other files are in two zip archives (label-provenance-systematic-review-1.0.0.zip and author-check-1.1.0.zip); README.md lists the contents and how to rerun the analysis. Licences: data, documentation and figures CC BY 4.0; code MIT. Version 1.1.0 (30 October 2026) adds a second archive, author-check-1.1.0.zip: the author's blind re-coding of 25 of the 196 studies (plan filed on the OSF registration before coding; the sheet as submitted, the sealed key of model codes with its SHA-256, the scoring script, agreement tables and the 54 differences), and a re-code of the same studies by five blind Claude instances, labelled as not human. The author check is a human check but not an independent second coder, because the author wrote the codebook. The first archive is unchanged. The full texts of the 25 studies are not included (copyright). Version 1.2.0 (2 November 2026) adds the raw API response files of the panel runs: 9,847 JSON files, one per model and item, in seven zip archives (raw-api-responses-1.2.0-part-01.zip to part-07.zip), with RAW_RESPONSES_INDEX.csv (path, size and SHA-256 of every file), RAW_RESPONSES_ZIP_PARTS.sha256 and README_raw_responses_1.2.0.md, which gives the layout. Unzip all seven parts into the folder that holds work/ from the first archive. For every one of the 8,777 files of the PHASE20 runs, the parsed answer in the file equals the answer in the matching votes.csv row. The README.md inside the first archive had said that the raw files stay with the author; version 1.2.1 (2 November 2026) replaced it with a corrected README and updated its line in MANIFEST.sha256. Every other file in the two earlier archives is unchanged. No API key is stored in any file. Replies and reasoning traces can quote passages from the reviewed reports; the full texts themselves are not included (copyright). Version 1.2.1 (2 November 2026) changes no data file: it corrects the README.md in the first archive and revises README_raw_responses_1.2.0.md to match.