Obscure AI: a multi-source, evidence-tiered atlas of AI edge cases

Version 1.3.6 (2026-08-13) — documentation-and-validation release. The case records are unchanged: 1,142, as since 1.2.0. The correction. Section 4 of METHODOLOGY.md described github.com/tyche-institute/limen as the wider pipeline repository. That description was already inaccurate when version 1.3.5 was published: the repository had been rescoped to figures and supplementary tables on 4 August 2026 and holds no pipeline code. The security-lane pointer is github.com/tyche-institute/limen-tools; the remaining lane scripts are available on request. Three studies added to validation/. A caveat-withheld control on the blind tier rater, running the same seeded stratified sample in two arms that differ in exactly one input field and reporting the leakage delta. A caveat-content audit that refetches each sampled record's own cited source and checks the caveat against it, where previously only the presence of the field had been audited. And a model-sensitivity replay of the AI-materiality screening gate, which states plainly that the exact mid-corpus model switch is permanently unmeasurable — the second model now returns HTTP 410 Gone — and measures what can still be measured: whether the gate's decision depends on the model at all. An open, multi-source atlas of documented AI/algorithmic edge cases — security disclosures, regulator and court records, and media-reported incidents — spanning 1,142 cases across 64 jurisdictions and 38 source languages. Each record carries an honest evidence tier (security disclosure / regulator-court / contested-interim / media incident), a bounded claim, and a link to the primary or aggregator source. Built from public sources via a reproducible pipeline (multi-source ingest, materiality gating, verbatim-quote verification, human acceptance). Version 1.2.0 adds primary regulator/court records from Italy, Austria, Germany, Sweden, France, Canada (courts), the European Court of Human Rights, the CJEU, the EPO Boards of Appeal, the Netherlands and Kenya, plus a frozen-snapshot folder (2026-07-04, 1,044 cases) with the exact analysis inputs/outputs behind a companion analysis article. Version 1.3.6 corrects the METHODOLOGY description of the security-lane code pointer and adds three validation studies (a caveat-withheld control on the tier rater, a caveat-content audit against each record's own cited source, and a model-sensitivity replay of the admission screen); the case records are unchanged from 1.2.0 onward. CC BY-SA 4.0, from version 1.3.2, matching the AIAAIC material the corpus derives from.