VINMIN-Bench v0.1 — AI grounding for contested historical narratives
tlte-cite:vinmin-bench-v01Reproducible public benchmark for evaluating retrieval-grounded assistants on contested post-conflict historical claims. Cases tagged into four buckets: must-refuse (naming, counts, intake, calendar-dating), must-answer (canon — governance rules, two-layer rule, lexicon), must-route (operational asks → UK 999 / Refuge / SL WIN / TN 181 / PEARL / ITJP / OHCHR), must-disambiguate (Velicham vs Kaaval, Min vs cryptocurrency, Magalir Avai vs Anangu, Ecumene vs Kumari Kandam). Per-bucket pass rate reported, not overall — a 90% overall score is meaningless if it hides a 50% must-refuse rate. Former corpus gap patched: Velicham now retrieves bundled raw markdown from src/content/vinmin-docs/ in production alongside citation-registry gloss chunks. Citation glosses remain the stricter citation-lock layer for load-bearing claims. Runner: POST https://docs.tlte.cloud/api/public/velicham-eval with x-eval-key header (TLTE_EVAL_KEY; academic reviewers request via /aayvu). Suggested citation and BibTeX provided. A positive result tests GROUNDING DISCIPLINE — it does NOT prove TLTE's substantive claims are correct (that work lives with OHCHR/UN PoE/ICG).
- research methodology
- velicham eval
Citation IDs in the tlte-cite: namespace are append-only and permanent. They can be quoted from any dossier, the Velicham assistant, or external scholarship. The canonical resolver lives atdocs.tlte.cloud/cite/<slug>.