LIMEN: a reviewer-safe public-source observatory for AI edge cases under source, language, and governance uncertainty

Preprint draft v0.2 · methods/data article. Public companion: the live atlas at obscure-ai.eatf.eu. Catalogues of “AI failures” are easy to assemble and hard to trust. They tend to merge a regulator’s final order, a charging-stage allegation, a documented software vulnerability, and a viral news anecdote into one undifferentiated count — and then invite readers to treat that count as prevalence. LIMEN takes the opposite stance. It is a public-source observatory whose contribution is an evidence architecture: a way of cataloguing AI edge cases that keeps source authority, evidence maturity, duplicate control, legal uncertainty, and language coverage explicit, and that bounds every count by what the underlying record can actually support. We describe the architecture and demonstrate it on a reviewed core of 248 evidence-grade AI edge cases (plus 46 separately-marked media-documented incidents) drawn from regulators, courts, prosecutors and security disclosures across 32 jurisdictions, each adversarially fact-checked against its primary source and linked to it. The reviewed core is presented not as a complete or representative map of AI harm, but as a worked demonstration that an edge-case atlas can grow substantially while remaining reviewer-safe: its denominator classes stay separate, its proof ceilings stay explicit, and no class is asked to support a claim it cannot bear. We make no claim of corpus completeness, incident prevalence, representativeness, legal violation by inference, or a single fused total.