LIMEN reviewed-core: an evidence-tiered dataset of AI edge cases

A reviewer-safe, evidence-tiered catalogue of AI edge cases (failures, harms, regulator/court actions, security disclosures), and the data layer behind the LIMEN observatory (limen.eatf.eu) and the public Obscure AI atlas (obscure-ai.eatf.eu). It accompanies the methods/data preprint at 10.5281/zenodo.20700813. 294 cases. 248 evidence-grade, reported as separate denominator classes and never summed across the evidence-grade boundary: Regulator/court record 157; Contested/interim 80; Security disclosure (CVE) 11. Plus 46 media-documented incidents, excluded from the evidence-grade denominator. 32 national/supranational jurisdictions; evidence-grade 2016–2026, media back to 2010. Each case carries an evidence class, a claim ceiling (what the record does not prove) and, for ~93% of evidence-grade cases, a direct source link. The dataset is a catalogue of locatable, source-anchored cases — not a complete or representative sample of AI harm, and no count is a prevalence estimate. See README.md for the full field dictionary and reading guidance.