Signal LLM-Assisted Review Prioritization for German Statutory Health Insurance Websites: A Multi-Stage Corpus Audit

Summary

Martin Möller published a paper on arXiv (cs.CY) on August 4, 2026, presenting a multi-stage workflow for prioritizing substantive review needs across German statutory health insurance (SHI) fund websites, whose content can shape health and benefit expectations but exceeds continuous specialist review capacity. The study analyzed 56,198 pages from 84 SHI websites and sub-sites, combining deterministic screening, model-assisted triage and in-depth review, minimum evidence checks, temporal-validity safeguards, and paired-model comparison, explicitly framed as reproducibility-bounded rather than a validated detector. Every page received a review state, with the workflow generating 35,998 review records and routing 21,452 to case review, with workload concentrated in transparency, legal framing, medical content, contradictions, and AI-related failure-mode signals. A routing stress test on a 300-page lower-priority sample surfaced a signal on 100 pages (33.3%), and across 182 matched cases, two models agreed 75.8% of the time (kappa=0.532). The author emphasizes the workflow produces a prioritized workload rather than error prevalence or final legal, medical, or insurer-level findings, and that public claims require human adjudication.