Two measured studies, their labelled data, and two running indexes of what is built on Jev, TypeSafe AI's typed-decision model.
A typed-decision model wired into five places in a Dungeons & Dragons co-DM, measured on four benchmarks (table rules, routing, contradictions, asset kinds) against the code it replaced, a frontier model called as a classifier, Laya, the open-weight model marketed as its replacement, and Qwen 3.8 read as a System 1 sampler.
8,916 skill-selection decisions per arm (743 skills across 12 job postings), rerun with every arm on equal terms against 720 pairs graded by Claude, GPT and Gemini: Jev, GLM-5.3-flash, Haiku 4.5 and Qwen 3.8 asked 60 skills per call and one per call, Qwen 3.8 read as a System 1 sampler, Laya, embeddings, a substring matcher, and the same postings as page images.
Every labelled item from both studies, each with the exact state and questions Jev was sent and the answer it is scored against: 720 skill–posting pairs with three graders' votes and every arm's score, and the four co-DM sets (94 table-rule paragraphs, 60 contradictions, 90 routing messages, 177 art-library rows). JSON Lines, with a README.
Tools, demos and findings from the TypeSafe Discord, GitHub and X, each scored 1 to 5 and tagged; searchable by text, tag, source, score and repository stars.
Named products and companies running Jev in production, each with a professional-usefulness score and its source, and a weekly highlight.