Frequently asked questions
Everything people ask about Jev before they build on it, answered plainly and linked to where each answer came from.
What is Jev?
Jev is TypeSafe AI's first System One model: a decision engine that takes a state plus typed questions and returns typed answers with probabilities and confidence, instead of generating text. It runs in 70 to 500 milliseconds, and output tokens are free. It is built for judgments your code can act on directly: classify, route, score, verify, branch.
What is a System One model?
A class of model built for fast, structured decisions inside software rather than conversation. TypeSafe introduced the term, inspired by Kahneman's fast System 1 thinking. The design choices follow from that goal: bounded outputs instead of strings, parallel sampling instead of token generation, and calibrated probabilities instead of human-preference optimization. See the glossary.
How fast is Jev, and how much does it cost?
TypeSafe reports 70 to 500 milliseconds end to end per decision, at $0.042 per million input tokens with output tokens free. The community's measured examples line up: 178 millisecond median decisions in the browser agent jev-ultrafast, 81 milliseconds per trade block in jev-trader, and about $0.04 per 1,000 alerts triaged in jev-oncall. Treat vendor multiples as directional and measure on your own workload.
Can I use Jev without the TypeSafe waitlist?
Yes, two documented routes. The Vercel AI Gateway serves typesafe-ai/jev through the AI SDK, including an experimental evaluate path and an HTTP endpoint at /v1/evaluate. OpenRouter serves typesafe/jev-1.13 through its decisions API. Neither requires the TypeSafe waitlist. Note that the OpenRouter route is not a chat-completions endpoint: the request shape carries state and typed questions. Details are on the getting started page.
What are Noul, Choice, and Score?
The three question types, mixable in a single parallel request. Noul asks a yes/no statement and returns a probability from 0 to 1. Choice picks one option from a list of up to 255 and returns the pick, every option's probability, and a confidence value. Score rates the state on an ordered rubric and returns a weighted score, the distribution, and confidence. Ten questions cost roughly the time of one.
Can Jev hallucinate?
It cannot return a value outside your schema: the possible answers are defined before the call, so invalid categories, malformed objects, and invented fields are impossible by construction. It can still pick the wrong valid option. The honest framing: no type errors, plenty of room for wrong answers, which is why confidence gating and measurement exist. See what independent tests found.
How accurate is Jev?
There is no single number, and any vendor who quotes one is selling something. On TypeSafe's own four-workflow evaluation, Jev agreed with a frontier-model consensus on 67.8 percent of cases at about $0.0004 per case, roughly tied with GPT-5.6 Terra at a fraction of the cost. Independent studies measured specific behaviors: 96.5 percent on a prompt-injection benchmark, ties with Cohere on reranking, and the failure modes documented on the rigor page. Accuracy is a per-task property: label your own cases and measure before automating.
What is Jev good at?
The fuzzy middle between brittle rules and expensive chat models: smart if-statements inside ordinary software. Concretely: classification and routing, rubric scoring and ranking, verification and guardrails, real-time decisions in games and agents, and bulk judgment over large corpora. The usecases page has more than a thousand projects sorted by category, and the patterns page covers what transfers.
What is Jev bad at?
TypeSafe publishes a per-version jaggedness page, which saves everyone a week of debugging. The summary: it reads literally, it is not a calculator, dates are text to it, accuracy falls as state fills with irrelevant detail, adversarial text can steer answers, it cannot abstain unless you give it an out, and option order can matter. The remedies are in the glossary and the failure modes section.
How is Jev different from an LLM?
An LLM generates strings token by token and you parse structure back out of them. Jev generates the structure directly in one parallel pass: no parsing layer, no invented fields, no refusals, two orders of magnitude faster, and free outputs. The trade is flexibility: Jev cannot write, summarize, or reason in the open-ended way an LLM can. The strongest systems in this ecosystem use both, with Jev making the bounded decisions and an LLM handling the generation around them.
Is Jev free?
Output tokens are free, inputs are billed at $0.042 per million tokens, and that is the entire pricing story. Whether your workload is cheap depends on state size: the community runs show 1,018 papers classified for about $0.08, 1,000 emails triaged for 3 cents, and 1.7 million posts scored for $0.65. The counterexample to respect: large states with many scoring calls can push cost up, so filter state in code before sending it.
Can I run Jev locally or on-device?
Not the hosted model itself, but the interface travels. Open reimplementations like SemIf, jevlike, and Laya reproduce typed decisions by reading option logits from small local models, and PocketJev runs decisions on an iPhone with MLX. The published comparisons generally favor the hosted model on decision quality, and the open ones win on latency and independence. Start with the hosted API and keep the local path as your hedge. See open models.
How do I choose confidence thresholds?
From measured curves on your own labeled data, never from a default. Label a few hundred real cases, compute precision, recall, and coverage at candidate thresholds, price both error directions, then pick the operating point and pin the model version you tuned it against. The full checklist is on the rigor page. This is also the highest-leverage consulting work: it is where production failures are prevented.
How do I get my project listed on JevOps.wiki?
Use the submit form. No account and no email: it goes to the maintainer for review, and approved entries appear in the community section of the usecases page. The wiki is curated manually for signal, so dead links, thin demos, and duplicates are skipped silently.
Last updated 30 September 2026.