MarCompBench v1 — restricted companion record (consultancy reference verdicts)

This restricted companion record holds the two files of the MarCompBench v1 release that derive from a third-party controlled professional gap-analysis report and are therefore not openly republished: rating-matrix-full.csv (all three reference verdict columns per item) and overrides-full.csv (per-item panel modal verdict, all three references' verdicts, and the Author endorsement flag), plus a README stating the terms. Access is limited to peer review and to verification of the statistics published in the accompanying paper and the open record (DOI 10.5281/zenodo.21357226, latest version), granted case by case under confidentiality. No license is granted over these files; do not redistribute them or publish their contents at item level. To verify: place rating-matrix-full.csv in the open record's directory and run python3 compute_agreement.py; the script then recomputes every statistic in the paper, including all consultancy pairs, Krippendorff's alpha over three raters, and the three-way consensus split. Accompanies: Torgersen (2026), Machine reproducibility and human disagreement in maritime compliance assessment: evidence from a triangulated benchmark on a DP operations manual.