Coded corpus, coding manual, and experiment code for "Has Industrial Control System Anomaly Detection Heeded Its Critics? A Systematic Review, Benchmark Data Audit, and Leakage-Free Re-Evaluation"

Research materials for the article "Has Industrial Control System Anomaly Detection Heeded Its Critics? A Systematic Review, Benchmark Data Audit, and Leakage-Free Re-Evaluation" by L. Raszka, O. Grycz, and Š. Vilamová (submitted to the International Journal of Critical Infrastructure Protection). The study combines a PRISMA-guided systematic review of 210 papers (2018–2026) that evaluate anomaly detectors on the SWaT, WADI, HAI, and BATADAL benchmarks, an audit of the SWaT and WADI benchmark files and label sources, and a controlled re-evaluation of seven detectors under three evaluation protocols. The record contains:- 00_coding_manual.md – final coding manual with variable definitions and version history;- 01_candidate_records.csv – the 275 Web of Science records (search of 8 October 2026) that mention the benchmarks, with full-text availability;- 02_coded_corpus.csv – all 220 assessed full texts with inclusion decisions and the codes of the 210 included papers (dataset version, learning setting, train–test split, decision threshold, point adjustment, code release, use of the truncated WADI training file), coding source, and coder notes;- 03_threshold_revisions.csv – the 17 threshold codes revised after the quote-by-quote re-check;- 04_round3_paired_codes.csv and 05_interrater_agreement.csv – blind double-coding data and agreement (raw agreement and Cohen's kappa) for three rounds;- 06_benchmark_data_audit.md – findings of the SWaT and WADI data audit;- code/icsbench_experiments.ipynb – experiment code (data loading, seven detectors, metrics, runner);- results/results_usad.csv – per-seed, per-threshold results of the USAD runs.See README.md for details and a glossary. The SWaT and WADI datasets are distributed by iTrust, Singapore University of Technology and Design, under its data-sharing terms and are not included; they must be requested from iTrust. HAI is available from its public repository. Full texts of the reviewed papers are not included for copyright reasons. An AI assistant (Claude, Anthropic) pre-coded the corpus and wrote parts of the code; all codes were verified as described in the article.