A field guide to AI that answers in types, not prose

What it does

This is a community-curated field guide and browsable directory for TypeSafe’s Jev model, which the project calls a “System One” judgment engine. Instead of generating prose, Jev returns typed primitives—Choice, Score, and Noul—so application code can route tickets, score frustration, or flag urgency without parsing chat output. The repository compiles SDKs, integration paths, live demos, and third-party evaluations into a site with a downloadable resources.json directory.

The interesting bit

The README is unusually honest about what Jev is not: it explicitly warns that code still owns policy and side effects, and it labels community inclusions as “not a claim that TypeSafe has reviewed or approved them.” That transparency is rare for an awesome list. The site also doubles as a coding-agent skill, offering a structured context file so agents can browse the directory without hallucinating links.

Key highlights

• Documents three answer shapes: Choice (classification with per-option probabilities), Score (ordered rubric), and Noul (yes/no probability).
• Maps five deployment routes—TypeSafe direct, Cloudflare Workers AI, Netlify AI Gateway, Vercel AI Gateway, and OpenRouter—with explicit caveats that they are not equivalent SDKs.
• Features live builds such as Jev Chess and Jevtown, plus independent benchmarks like Jevals.com and JevBench.
• Ships a machine-readable resources.json and an agent skill definition for automated browsing.
• 102 contributors, but explicitly disclaims affiliation with TypeSafe AI.

Caveats

• The repository is a directory, not a framework; most value comes from following outbound links to external repos and docs.
• Several access routes and featured projects are third-party or experimental, and the README warns that availability and terms change.

Verdict

Worth bookmarking if you are evaluating structured-output models or routing LLM calls through typed decisions. Skip it if you are looking for a single installable library—this is a curated map, not a codebase.