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

Version 1.3.6 (2026-08-11): documentation correction and three added validation studies. Case records unchanged (1,142). Correction. Versions 1.3.0–1.3.5 of METHODOLOGY.md §4 described github.com/tyche-institute/limen as the wider pipeline repository. That description was already inaccurate when 1.3.5 was published: on 4 August 2026 the repository was rescoped to figures and supplementary tables only and holds no pipeline code. Cite github.com/tyche-institute/limen-tools for the security-lane tooling. The correction history in METHODOLOGY.md is rewritten to cover 1.3.1 through 1.3.6 in one place, including that versions 1.2.0–1.3.1 carried CC BY 4.0 in error and eight records naming individuals in criminal matters at a tier their sourcing did not support — both corrected at 1.3.2, and those releases cannot be withdrawn. Added to validation/: a caveat-withheld control on the tier rater (92.9% agreement with the caveat shown, 81.0% withheld, same 84-record sample, one named model); a caveat-content audit against each record’s own cited source (37 of 41 read caveats borne out, none contradicted, three sources refusing automated and human readers alike); and a replay of 40 recorded admission-screen decisions showing that two models the same lane used disagree on about a quarter of keep-or-drop calls. Each ships as a script with its sampled identifiers and outputs. Version 1.3.5 (2026-08-02): code pointer moved to the curated limen-tools snapshot (MIT); figure-script label placement refined, plotted values unchanged. Case records unchanged. Version 1.3.4 (2026-08-02): methodology movement-count corrected (nineteen to sixteen); intermediate-file reconstruction disclosed; figure script (make_figure1.py) and sample re-fetch added. Case records unchanged. Version 1.3.3 (2026-08-02): adds the validation/ folder (link-liveness and two-rater tier-agreement studies as executable scripts with results) and removes the zeroed pipeline-telemetry block from the JSON. Case records unchanged. Version 1.3.2 (2026-08-01): re-anchored 105 regulator/court records to primary documents (top tier now 417, all primary-anchored); replaced analyze.py with a dependency-free reproduce_corpus_stats.py; relicensed CC BY-SA 4.0 to match the AIAAIC upstream. Record identifiers unchanged.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. CC BY-SA 4.0.