01 / HOW JEV WORKS
State and typed questions become decisions
State is the text or structured text available for a decision: a customer message, incident report, or application record. A typed question says what you want to know and constrains the answer space.
State
“The customer says they were charged twice.”
Typed questions
Which team should handle this?
BillingHighest-probability option in this illustration.
The application still owns the action. Selecting Billing can route a ticket, but it does not prove a duplicate payment happened or authorize a refund. Code and policy decide what happens next.
Current TypeSafe documentation supports strings, JSON objects, and arrays of text values as state. Jev does not accept audio or video input. Image input is an experimental, opt-in feature available since the Jev 1.14 release (also on 1.13), not recommended for production; see the Jev 1.14 page.
Jev question types
02 / NATIVE PRIMITIVES
Noul, Choice, and Score
TypeSafe Direct exposes three native Jev question types. Each keeps the answer inside a shape your software knows in advance.
A yes/no judgment as a probability
Ask a bounded yes/no question such as “Does this ticket convey urgency?” Noul returns one number from 0 to 1: the probability that the answer is yes. A value near 1 favors yes; a value near 0 favors no.
One selection from bounded options
Supply named options with distinct meanings. Jev selects the highest-probability option and returns the full probability distribution plus a confidence value under the current TypeSafe API contract.
A probability-weighted value across ordered levels
Define an ordered rubric, from “No urgency” through “Critical,” for example. Jev returns a probability-weighted score, a distribution over the levels, and confidence. The result can land between levels.
- 0No urgency
- 1Low
- 2Moderate
- 3High
- 4Critical
03 / WHERE JEV FITS
What is Jev good at?
Jev fits decisions where the input may be nuanced but the allowed result is known. These are workflow patterns, not claims that Jev Playground has benchmarked every example.
- Routing
- Choose which known queue, team, or workflow should receive an item.
- Classification
- Select among named categories with operationally distinct meanings.
- Relevance and filtering
- Judge whether supplied text is relevant before downstream processing.
- Workflow gating
- Assess whether evidence meets a defined condition before code proceeds.
- Semantic verification
- Check whether a claim is supported by supplied evidence or a rule.
- Urgency and qualitative scoring
- Place an incident or message on an ordered, described rubric.
- Agent next-action selection
- Choose among actions the application has already permitted.
04 / WHERE JEV STRUGGLES
What is Jev not good at?
TypeSafe documents these jagged edges for Jev 1.13. Design around them instead of treating every task as a model question.
- Open-ended text generation
- Use a generative model for replies, articles, code, and other created text.
- Arithmetic and exact counting
- Count, calculate, and enforce numeric identities in code.
- Date and time comparison
- Extract bounded components when needed, then parse and compare dates in code.
- Multiple layers of indirection
- Reduce the hops, point directly to relevant state, or decompose the task.
- Large states full of irrelevant context
- Retrieve and filter first; send only what each question needs.
- Adversarial input
- Write precise criteria, test edge cases, and keep application guards around actions.
Jev can also read criteria literally. Align instructions and criteria, make boundary cases explicit, and test on representative inputs before deployment.
05 / DIFFERENT JOBS
Jev vs generative models
Generative model
- Output
- Text, code, explanations, or other generated content.
- Typical task
- Draft a reply, explain a concept, write code, or summarize a document.
- Who consumes it
- People or software that parses and validates the generated result.
Jev
- Output
- A bounded typed decision with probabilities.
- Typical task
- Choose a route, assess a proposition, or score against a rubric.
- Who consumes it
- Application code that branches, applies policy, or asks for review.
The two can work together. The practical distinction is what the workflow needs next: a constrained decision for code, or newly generated content. TypeSafe positions Jev as a System One Model rather than a generative chat model; this page does not infer an undisclosed implementation from that positioning.
How to read probabilities
06 / USING PROBABILITIES
The top answer is only part of the decision
Both examples below select Billing. Their distributions suggest different application policies.
A clear lead
Billing dominates the alternatives. An application might route directly if its tested policy and other checks permit it.
A close call
Billing and Technical are almost tied. Asking for clarification or human review may be more useful than acting immediately.
Illustrative distributions, not live model results. Each totals 100%.
Set thresholds using representative data and the cost of being wrong. Consider the selected option’s probability and its lead over alternatives. Provider confidence, when returned, is not the same field as the selected option’s probability. The API guide shows an example policy.
Calibration is a population-level property: among comparable events assigned 80%, roughly eight in ten should occur in a well-calibrated system. It does not certify one answer. TypeSafe describes calibration as a training objective; validate it for your own workload.
Model terminology
07 / SYSTEM ONE & RLCD
TypeSafe’s model terminology
TypeSafe defines System One Models as AI models built to make fast, structured decisions that software can use directly. The name draws on the fast, intuitive System 1 concept popularized by Daniel Kahneman. It describes a design direction; it does not mean the model thinks like a person or cannot make mistakes.
RLCD: training for decisions
Reinforcement Learning for Calibrated Decisions is TypeSafe’s name for its training approach. Its materials describe optimizing decisions and probability estimates rather than free-form responses people prefer. The launch article also describes a purpose-built architecture and parallel sampling; these are vendor descriptions, not a complete public recipe.
What about “zero hallucinations”?
TypeSafe uses that phrase in the schema and output-space sense. Staying inside an allowed answer space does not guarantee semantic correctness: Billing is a valid Choice even when Technical would be correct. Type safety is not the same as being right.
Recorded response examples
08 / FIRST-HAND RESPONSE EXAMPLES
What Jev returned in three real calls
On September 18, 2026, we ran the three original single-question presets once through 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 state said the support agent had issued a full refund. The UI showed Yes 99% and No 1%.
A duplicate subscription charge was reported. Billing received 100%; Technical, Sales, and Other received 0%.
The state described a complete production outage affecting all customers. Level 4 received 100%.
These results demonstrate response shape. They do not establish accuracy, calibration, or expected performance. Different inputs can produce different decisions, and a displayed 100% is not a guarantee of correctness.
09 / TRY OR INTEGRATE
See a decision, then choose your next step
Run a real Jev example in the Playground without visitor signup, or use the beginner guide to choose among the Playground, Direct API, official SDKs, and Agent Skill.
Ready to integrate? The Jev API guide owns authentication, provider contracts, request examples, responses, and errors.
Questions and answers
10 / QUICK ANSWERS
Jev FAQ
What is Jev?
Jev is TypeSafe AI’s flagship and first System One Model. It evaluates text-based state against typed questions and returns bounded decisions with probabilities for software workflows.
Who makes Jev?
Jev is made by TypeSafe AI.
What does Jev return?
TypeSafe Direct returns Noul, Choice, or Score answers with probabilities. Under the current API contract, Choice and Score also expose confidence.
Is Jev a chatbot?
No. Current Jev is designed for bounded decisions inside software, not an open-ended conversational response.
Can Jev generate text or code?
No. Jev 1.13 is not trained to generate text. Use a generative model when the task is to create prose, code, or another open-ended output.
What is a System One Model?
TypeSafe uses the term for a class of AI models built to make fast, structured decisions that software can use directly. It does not imply human-like thought.
What is Noul?
Noul is TypeSafe Direct’s native yes/no primitive. Its value from 0 to 1 represents the probability that the answer is yes.
Can Jev be wrong?
Yes. A type-safe answer stays inside the allowed schema, but it can still be the wrong semantic answer. Test representative cases and keep safeguards around consequential actions.
How can I try Jev?
Try Jev online in this site’s real Playground without visitor signup.
How do I integrate Jev?
Use the Jev API guide for current provider contracts, authentication, examples, and error handling.
Sources
11 / SOURCES & VERIFICATION
Primary sources for this guide
Information checked against current TypeSafe documentation on . Model behavior and interfaces can change.
- TypeSafe AI — model definition and FAQ (opens in a new tab)
- Jev launch article — System One, RLCD, and vendor claims (opens in a new tab)
- System One — model class and practical workflow (opens in a new tab)
- API reference — Noul, Choice, Score, probabilities, and confidence (opens in a new tab)
- Models — Jev 1.13, supported inputs, and current alias (opens in a new tab)
- Jev 1.13 jaggedness — documented limitations (opens in a new tab)
- AI primer — RLCD, decisions, and calibration (opens in a new tab)
Jev Playground is an independent tool, not TypeSafe’s official website. Vendor claims are attributed, and the three historical calls above are response examples rather than a model-quality evaluation.