Jev returns typed Choice / Score / Boolean answers with calibrated confidence. This hub maps a quick start, capability boundaries, ecosystem tools, and starter use cases. We are not affiliated with TypeSafe.
What Jev is → access paths → one evaluate call → confidence thresholds → hard boundaries.
01
TypeSafe’s first System One model: state + typed questions in → decisions + probabilities out — no chat text.
02
Two public paths — TypeSafe console key, or call typesafe-ai/jev via Vercel AI Gateway.
03
Send shared state + a map of named questions; read typed answers and confidence in code.
04
Every answer includes calibrated confidence — automate high confidence, escalate the rest.
05
Jev does not generate prose, code, or free-form tool strings — surrounding code owns loops and safety.
Thin placeholders you can adapt — not production recipes, and not invented benchmarks.
Placeholder guardrail. Launch theme: verify / guardrail / jailbreak detection. Outline only.
Get the prompt →Placeholder: mirror the Vercel Gateway sample shape (shared state, multiple typed questions). Not a production recipe.
Get the prompt →Placeholder agent-loop control. Schema-bound Choice keeps tool names from being invented. Outline only.
Get the prompt →Placeholder based on typesafe-computer-use. Mark community demo. Outline only.
Get the prompt →Placeholder based on kitze/unclutter. Mark extension demo. Outline only.
Get the prompt →Official TypeSafe / Gateway paths, plus independent community projects (OpenJev→SemIf, computer-use, Unclutter). SimpleJev held until a public URL is locked.
Official TypeSafe Jev — early access / hosted System One decisions. Independent site; we are not TypeSafe.
Documented Gateway path for Jev. Check current Vercel billing — do not invent free credits.
Independent community project (OpenJev lineage → SemIf). Reproduces an interface pattern on open models; not affiliated with TypeSafe.
System One vs LLMs, question composition, confidence failure modes, and computer-use patterns.
ecosystem
Official TypeSafe Jev, Vercel AI Gateway access, and community OpenJev experiments — keep the labels straight.
foundations
Jev selects typed decisions with probabilities. LLMs generate text. Use both — do not force one job onto the other.
Wrong model id, OpenJev≠Jev confusion, over-automating low confidence, expecting chat prose.
Local OpenJev probabilities look decisive, but TypeSafe Jev (or humans) disagree on the same rows.
OpenJev documents uncalibrated option softmax. Do not treat it as TypeSafe calibrated confidence.