01 / THE INTERFACE
How state becomes a decision
State is the information available for a decision: a customer message, an incident report, or a record from your application. A typed question defines what you want to know and the answer space you will accept.
State
“The customer says they were charged twice.”
Typed question
Which team should handle this?
BillingHighest-probability option in this illustration.
The application still owns the action. Selecting Billing can assign a ticket to a queue; it does not prove a duplicate payment happened or authorize a refund. Those are separate questions and checks.
The Gateway evaluation interface also accepts structured state and multiple questions. This playground keeps the interaction focused on text state and one question at a time.
02 / THE ANSWER SPACE
Boolean, Choice, and Score
Choose the type that matches the decision. These names and fields follow the AI SDK integration used by Jev Playground.
How likely is a proposition to be true?
For “Was a refund issued?”, the returned probability is P(true), between 0 and 1. It is not a native true verdict. With P(true) = 0.99, our interface derives Yes ≈ 99% and No ≈ 1% using 1 − P(true).
Naming matters: TypeSafe’s direct API calls this Noul and returns a noul field. Vercel’s evaluation API uses boolean and probability.
Which option fits this state?
Supply one question and a bounded set of named options. Our Support Routing preset offers Billing, Technical, Sales, and Other. The answer contains a selected choice; its distribution shows how probability is spread across the options.
Choose labels that lead to distinct actions. If two options describe the same thing, a split distribution may reflect an ambiguous question rather than an unhelpful model.
Where does it sit on an ordered rubric?
Score uses levels in order, from lowest to highest. Our Incident Urgency preset uses this scale:
- 0No urgency
- 1Low
- 2Moderate
- 3High
- 4Critical
The result contains a score and can include a probability for each level. Scores can be fractional; 4/4 is not the only possible result. Rubric labels define the meaning of the scale, not a universal urgency standard.
Vercel documents these three question types. The installed SDK permits Choice/Score distributions to be absent; the real calls below returned them. A missing probability should not be displayed as zero.
03 / FIRST-HAND EVIDENCE
What Jev actually returned in our tests
On September 18, 2026, we ran each existing preset once on the public Jev Playground. These were real calls through Vercel AI Gateway to typesafe-ai/jev. Gateway logs identified TypeSafe AI as the provider.
The supplied state said the support agent had issued a full refund. The UI showed Yes 99% and No 1%.
Observed server + network time: 905 msA duplicate subscription charge was reported. Billing received 100%; Technical, Sales, and Other received 0%.
Observed server + network time: 399 msThe state described a complete production outage affecting all customers. Level 4 received 100%.
Observed server + network time: 693 msThese results demonstrate the response shape. They do not establish accuracy, calibration, or expected latency. Different inputs can produce different decisions. The responses reported two-decimal probability rounding, so a displayed 100% is not a guarantee of correctness.
Timing includes application, Gateway, and network overhead. An earlier local validation’s first request took about 6.4 seconds; these three production calls were quicker. Neither small sample predicts your workload. For current cost, check the Jev pricing calculator; vendor workflow speedups are not a promise for every request.
04 / FROM AN ANSWER TO AN ACTION
The top answer is only part of the decision
Both examples below select Billing. They should prompt different questions about whether to act automatically.
A clear lead
Billing dominates the alternatives. An application might route directly if its policy and other checks permit it.
A close call
Billing and Technical are almost tied. Asking one clarifying question may be more useful than choosing immediately.
Illustrative distributions, not live model results. Each totals 100%.
An example application policy
- High confidenceAutomate a permitted, reversible step.
- Medium confidenceGet more evidence or run an additional check.
- Low confidenceRoute to human review.
This is a policy you design, not a Jev requirement. Choose thresholds using representative data and the cost of being wrong. Check the gap between competing options as well as the highest probability. A separate Choice/Score confidence field, when supplied, is not interchangeable with the selected option’s probability.
Calibration is a property of many predictions: among comparable events assigned 80%, roughly eight in ten should occur in a well-calibrated system. It does not certify any individual answer. TypeSafe describes calibration as a training objective; validate it for your own task.
05 / DIFFERENT JOBS
Jev and general-purpose LLMs
ChatGPT / Claude-style LLMs
- Output
- Generated tokens forming text, code, or structured content.
- Typical task
- Write a reply, explain a concept, produce code, or summarize a document.
- Software integration
- Extract or validate the decision you need. Many APIs also support schema-constrained output.
Jev
- Output
- A typed decision with probabilities within a defined answer space.
- Typical task
- Choose a route, assess a proposition, or score against a rubric.
- Software integration
- Use the answer in a branch or policy; still handle failures and evaluate decision quality.
Jev is not a replacement for ChatGPT or Claude for open-ended generation. They solve different tasks. A workflow could use Jev to select a support queue and a generative model to draft the customer reply.
Structured output alone is not the distinction. An LLM can also return a valid category. TypeSafe positions Jev around training and serving decision probabilities as the model’s main output, rather than adapting a conversational response into a decision.
06 / TYPESAFE’S TERMINOLOGY
What does “System One Model” mean?
TypeSafe uses the name for models designed for quick, focused judgments in software. It references the fast, intuitive System 1 / slower, deliberate System 2 distinction popularized by Daniel Kahneman. It is a design direction, not a claim that a model thinks like a person or never makes mistakes. See TypeSafe’s explanation in the sources below.
RLCD: training for decisions
Reinforcement Learning for Calibrated Decisions is TypeSafe’s name for its training approach. Its AI primer describes optimizing decisions and probability estimates, rather than free-form responses that people prefer. The launch article also describes a purpose-built architecture and parallel output sampling. These are vendor descriptions; they do not disclose a complete recipe for reproducing the model.
What about “zero hallucinations”?
TypeSafe uses that phrase on its website. Its launch article ties the claim to guaranteed schema matching. Staying inside an allowed answer space does not mean choosing the right answer: Billing is a valid label even when Technical would be correct. Treat type safety and semantic correctness as separate properties.
07 / FIT AND LIMITS
Where a bounded decision helps
These are possible workflow designs, not claims that we have evaluated Jev on each task.
- Support routing & lead classification
- Choose a support queue from a message, or classify an inquiry as a sales lead, existing customer, or unrelated request.
- Agent next action & workflow gating
- Choose among permitted next steps, or check whether the supplied evidence is sufficient before allowing a workflow to continue.
- Verification & moderation review
- Ask whether a transcript records a refund, or whether a comment needs review under a stated policy. Keep enforcement in application code.
- Urgency & risk classification
- Score an incident against operational severity levels, or flag a suspicious account event for investigation. A score is a review signal, not proof.
What Jev is not good for
Writing articles, long-form conversation, code generation, document summaries, creative writing, and detailed explanatory answers all need generative output. Those are not the job of this interface. Jev can assess a summary against a rubric; it does not write the summary for you.
For multi-step reasoning, separate the task into focused questions and combine answers in code. Supply the evidence and criteria the decision needs. Current TypeSafe documentation supports text-based state, not direct image, audio, or video input.
08 / A SMALL TECHNICAL EXAMPLE
The same pattern in code
This server-side example matches the evaluation API in our installed AI SDK 7.0.105. Keep AI_GATEWAY_API_KEY on the server. An evaluation can fail; your application still needs validation, timeouts, and error handling, as our playground does.
import { createGateway, experimental_evaluate as evaluate } from 'ai';
const gateway = createGateway({
apiKey: process.env.AI_GATEWAY_API_KEY,
});
const result = await evaluate({
model: gateway.evaluationModel('typesafe-ai/jev'),
state: 'The support agent issued a full refund.',
questions: {
refunded: {
type: 'boolean',
instructions: 'Was a refund issued?',
},
},
maxRetries: 0,
providerOptions: { gateway: { zeroDataRetention: true } },
});
console.log(result.answers.refunded.probability);The value is produced by the model, not hard-coded. Vercel’s evaluation docs require AI SDK 7 or later; this is not a chat-completions endpoint.
Read the Jev API guide for setup, copyable examples, response contracts, and current pricing.
Try the same pattern in Jev Playground09 / COMMON QUESTIONS
A few useful distinctions
Is Jev an LLM?
It understands natural-language input, but TypeSafe presents Jev as a System One decision model. It exposes typed answers and probabilities, not a general-purpose text-generation interface. Calling it a small chatbot misses what the product does.
Can Jev generate text?
Not free-form replies, articles, code, or reasoning explanations. A Choice answer such as Billing is a label from the options you supplied, not a newly written response.
Can Jev be wrong?
Yes. A valid answer type does not make the underlying judgment correct. A mistaken interpretation of the state can still produce a perfectly valid label or score.
Does 99% mean 99% accuracy?
No. It is a probability assigned to one answer in one context. To assess calibration or accuracy, you need many representative examples and known outcomes. Our three test calls do not establish either.
Does Jev issue refunds or route tickets itself?
No. In these examples it evaluates the supplied state. Your application decides whether to call a billing tool, assign a ticket, ask for more information, or do nothing.
How can I try it?
Open Jev Playground, choose Support Routing, Refund Verification, or Incident Urgency, and run the example. Then change the state or criteria to see how the returned decision changes. No signup is required; usage limits apply.
10 / SOURCES & VERIFICATION
What this guide is based on
Information checked against current TypeSafe and Vercel documentation on . Model behavior, interfaces, and pricing can change.
- TypeSafe AI homepage — product positioning.
- Introducing System One Models & Jev — launch claims and their qualifications.
- System One, AI primer, and Noul — terminology and native API distinctions.
- Vercel evaluation documentation and Jev model listing — the Gateway interface used here.
- Our September 18 production validation — the three observed results in this guide, with provider and ZDR confirmed in Gateway logs.
Jev Playground is an independent tool, not TypeSafe’s official website. The explanatory policy examples are ours; vendor claims are attributed, and the three test calls are not a model-quality evaluation.