What is Jev? How a decision model turns language into choices

Some AI work needs a choice before it needs a sentence. Jev gives software a way to interpret language and return bounded decisions, with uncertainty attached.

What is Jev? beside a message card connected to four sculptural routes, with one route highlighted blue.
BY
AI EXPERT SYDNEY
PUBLISHED
READ
12 MIN

A customer writes: “The parcel arrived, but the glass bowl is cracked. I need a replacement before Friday.” Your business needs to recognise the problem, choose the right queue and work out whether someone needs to intervene. A beautifully written reply can wait until those choices are made.

Jev is TypeSafe AI’s model for making structured decisions inside software. It reads supplied text and answers questions using choices, scores and probabilities, rather than writing a response. TypeSafe calls this a “System One” model. The practical promise is an AI component that can help a program interpret messy language while leaving the program in control of what happens next.

That is an interesting job for AI. Many useful automations get stuck at a small judgement: which team should handle this message, which passage is relevant, or does this request need review? Jev is aimed at that gap. Understanding it means keeping three things separate: what the model thinks, what the business permits and what the software actually does.

What Jev is, and what “decision LLM” means

TypeSafe announced Jev on 15 September 2026 as its first public System One model. Its launch explanation describes a model architecture, parallel sampling and a training method called reinforcement learning for calibrated decisions, or RLCD. Those are the company's descriptions of how it built the product, rather than independently reproduced findings about every workload.

“Decision LLM” is useful shorthand for language-aware AI used to make choices. It needs a little care here. In TypeSafe’s terminology, Jev understands natural-language input like a large language model, but returns constrained decisions instead of generated text. It does not write replies, code or explanations of its reasoning. “System One” is its category name, inspired by the fast, intuitive thinking discussed in Daniel Kahneman’s Thinking, Fast and Slow. The name describes the intended kind of judgement; it does not establish human-like understanding.

Jev is the specific TypeSafe product. The broader idea of a decision model can also include ordinary LLMs configured to select options, or other models built for classification. A chatbot can make a choice too. The meaningful questions are how that choice is produced, how uncertainty is represented, and how well the approach works for the task.

What goes in, and what comes back

Think of a Jev request as a small case file with a few questions attached. The case file is called the state: the information the model should consider. It might contain a customer message, relevant order details and an approved policy. The state documentation allows plain text, a JSON object with named fields, or an array. Jev currently accepts text, not images, audio or video. A photo of a broken bowl would need a separate process to turn it into useful text or fields first.

The developer then defines the question and the possible answers. Typed means the answer must fit a specified shape, such as one of four named queues. It gives the surrounding software something concrete to use.

TypeSafe’s question guide defines three forms:

Question type What you ask What comes back
Choice Which of these options fits? A selected option, probabilities across the options and a confidence value.
Score Where does this sit on defined, ordered levels? A score, the level descriptions, probabilities and confidence. The score can fall between levels.
Noul Is this specific statement true? The probability of “yes”, from 0 to 1. It has no separate confidence field.

“Noul” is a product term; you can read it as a yes/no probability. A value near 0.5 signals uncertainty between yes and no. A Score serves a different purpose: it rates something against levels you describe, such as calm, concerned and very frustrated.

Several questions can share one request. Each sees the same state and is evaluated independently. One answer does not secretly become the evidence for another. If a later question depends on something the first answer causes your program to retrieve, the program must make a subsequent request with that new information.

That design encourages small, clear questions. “Does the message report damage?” is easier to inspect than “Understand this customer and do whatever is best.” The second instruction hides several business decisions in one sentence.

A worked example: routing a damaged-delivery enquiry

Consider a hypothetical Sydney homewares retailer. The customer message above is fictional, and the numbers below are illustrative, not measured Jev results.

The retailer supplies the message, the order status and a short description of each support queue. It asks: “What is the main issue in the customer’s message?” The available options are delivery delay, damaged item, billing and other. A separate yes/no question asks whether the customer explicitly requests a replacement.

A Choice answer could identify damaged_item and return this distribution:

Supplied option Illustrative probability
Damaged item 0.88
Delivery delay 0.06
Billing 0.01
Other 0.05

An API is the interface software uses to call the model. The table simplifies its response. The API reference specifies that Choice returns the highest-probability option and a probability for every option, summing to one.

Now the software has a useful signal. It can put the case in the damage queue and highlight the requested replacement. The retailer can write a rule that uncertain cases go to a person. A separate rule can require staff approval before any refund or replacement order is created.

The model’s classification is only one step. The order system still needs to confirm which purchase is involved, the stock system needs to check availability, and the business needs to decide who may authorise the remedy. A generative model might draft the reply after those checks. Code sends it only when the workflow permits.

Change the message to “It arrived eventually. I’m not happy with what I paid,” and the issue becomes less clear. Good queue definitions and an “other” route matter. If the only options are billing and delivery, even a perfectly valid output can force the case into an unhelpful category. A constrained answer space needs room for the cases your team cannot confidently classify.

This distinction also explains why Jev is not a complete agent. An agent may repeatedly retrieve information, choose tools and act towards a goal. Jev can supply some of those choices inside the larger system TypeSafe describes; the application supplies the sequence and authority. Our guide to AI agents versus AI models explains that wider system.

How it compares with chat models and rules

An ordinary generative model can classify a ticket and write a reply in the same interaction. Some also support tightly constrained output formats. For example, OpenAI’s Structured Outputs documentation describes responses that adhere to a supplied JSON schema, while explicitly warning that their contents can still be mistaken. It would be misleading to suggest that chat models always produce unusable prose or cannot return valid choices.

TypeSafe even publishes an LLM-backed adapter for its decision interface, designed to compare ordinary models with TypeSafe on cost, speed and capability. Jev’s case therefore rests on the quality and efficiency of a dedicated decision component, beyond simply returning a label in a valid format.

There is early research supporting this broader view. The LLM2Jev preprint investigates extracting option probabilities from Qwen3.5-4B and Qwen3-0.6B without changing their architectures. The authors report useful decision capability without fine-tuning, with further training helping particular models and tasks. That is evidence about those experiments, not proof that every LLM matches the hosted Jev service. “Jev-style” in such research describes an interface or approach; it does not make a community model TypeSafe’s Jev.

Rules-based automation has a different strength. In our hypothetical retailer, code can check whether an order status equals “dispatched”, calculate the days since purchase and enforce an approval limit. Those are explicit conditions with exact answers. A semantic decision model becomes interesting when the condition depends on meaning: does “the bowl is in pieces” describe damage even though the customer never uses the word “damaged”?

The useful combination is straightforward: rules handle exact conditions, Jev or another model interprets language, and a generative model writes when writing is needed. Which component earns its place depends on the workflow. A reliable keyword rule may be enough for a narrow queue; a capable general model may already satisfy the team's needs.

Where a decision model could help

The following are illustrative workflow ideas, not claims of tested business results. Each gives the model a bounded judgement and gives software a defined response.

Work Model judgement What the application could do
Incoming enquiries Is this a quote request, an existing job update or something else? Route to sales, operations or review.
Document sorting Which supplied document category best fits this text? Suggest a folder or label for staff to check.
Knowledge search Does this retrieved passage address the customer’s question? Filter the passages sent to an answering model.
Quality review Does the supplied source actually support this drafted claim? Flag a claim for checking rather than automatically approving it.
An agent workflow Is another search needed, or is the available evidence sufficient? Continue searching, pass to a reasoning model or ask a person.

These jobs share a pattern: the input is varied language, but the next step comes from a small set of known possibilities. That is a more useful starting point than asking a decision model to run the whole business.

The REFLEX research preprint explores this division of labour in agents: Jev handles bounded choices and a stronger model handles uncertain cases or generation. Its controlled experiment reduced calls to the stronger model, but an inexpensive generative-model cascade remained competitive in external multi-turn tests. The sensible inference is that a dedicated layer can help some workflows; adding one does not automatically improve an existing system.

Confidence helps, but it does not certify the decision

There are two numbers people can easily confuse. A probability describes how likely an outcome is according to the model. TypeSafe’s separate confidence value for Choice and Score summarises how concentrated the distribution is. Its confidence guide explains the calculation. For the four-option example above, a top probability of 0.88 gives Choice confidence of 0.84. Confidence is not an independent fact check, and 0.84 should not be read as a measured 84% accuracy guarantee.

TypeSafe says RLCD trains for calibrated probabilities. In plain language, if a model is well calibrated on a relevant set of cases, outcomes it assigns 80% probability should occur roughly 80% of the time across that group. The training primer makes that group-level distinction explicit. It does not promise this particular customer's classification is correct, or that calibration carries unchanged into your business.

This is also the right way to read the launch’s “can’t hallucinate” claim. Restricting output to supplied options prevents an invented category from appearing. It does not prevent the model choosing the wrong supplied category. A valid billing label on a damaged-item enquiry remains a mistake.

TypeSafe’s Jev 1.13 limitations page, reviewed by the vendor on 2 October 2026, lists unreliable counting, arithmetic and date comparison, sensitivity to irrelevant context and option order, and vulnerability to adversarial content. It recommends keeping exact calculations in code and making questions precise. For audit evidence, record the input, question, choices, model version, outputs and resulting action.

Independent evidence reinforces the need for careful testing. In the JevOut preprint, researchers tested Jev 1.13.0 and searched for context additions that would steer initially correct answers towards a fixed wrong option. Across seven datasets, they found such flips in 312 of 508 initially correct decisions, using up to 64 accepted target evaluations per item. That is an engineered attack experiment, not a 61.4% error rate on ordinary customer messages. It does show why a confident output and a tidy format are insufficient grounds for giving software unrestricted authority.

For the retailer, routing a ticket and issuing a refund deserve different controls. A low-risk queue suggestion can be corrected. A payment needs verified records, permissions and an approval policy even when the model appears certain. Human review should have a practical job: resolve missing evidence, inspect consequential cases and correct the process when errors recur.

Current access, version and cost

Checked 4 October 2026: TypeSafe’s model catalogue lists Jev 1.13 as jev-1.13.0. Both jev-latest and jev-preview point to that version at this check. An alias can move to a new release; a team evaluating behaviour should record the returned version and consider pinning it while validating changes.

The listed API price is US$0.042 per million input tokens, with output tokens free. TypeSafe’s customer agreement specifies US dollars unless an order states otherwise. Tokens are units used to measure model input; they are not a fixed number of words.

For a hypothetical cost calculation, 100,000 requests averaging 2,000 billed input tokens each would total 200 million input tokens: US$8.40 at that listed rate. That covers model input charges only. It excludes retries, other models, integration, hosting, staff review and any applicable taxes. Request size and the number of repeated calls matter more than a headline rate on its own.

The quick-start guide directs readers to a logged-in playground and a dashboard API key, with Python and JavaScript SDKs available. This article has not tested account creation, Australian account eligibility or introductory credits. Check the actual account terms before planning around a free offer.

TypeSafe’s workflow evaluations report very large speed and cost gains, using reference answers from other large models. Their measure of accuracy is agreement with those reference answers, rather than verified business outcomes. Those results are a reason to evaluate a decision workload, not a promise of the same improvement on your enquiries. An Australian team's complete response time also includes network travel and the rest of its application.

Before sending sensitive records, check the applicable data terms. TypeSafe’s legal documentation says zero data retention is offered for enterprise customers. That does not establish the retention terms of every account or Australian data residency.

How to tell whether it earns a place in your workflow

Choose one recurring judgement with clear options and an owner who can resolve disputed cases. In the retailer example, start with queue suggestions, before attaching replacement or refund authority. Write down what a correct classification means and include an escape route for missing information.

Compare Jev with the process you already have: staff routing, simple rules and, where relevant, an ordinary LLM. Use the same cases and business definitions. Include straightforward messages, mixed requests, unfamiliar wording and attempts to steer the result. Our AI evaluations guide explains how to turn that exercise into test cases and a release decision.

Measure wrong automatic routes, the share needing review, review time, complete workflow time and total cost. Check whether highly confident cases are actually more accurate on your material. A threshold that makes a demo look smooth can leave a team correcting mistakes all afternoon.

Jev is worth understanding because it makes a useful piece of AI work explicit: interpreting language well enough to choose among known next steps. Its value becomes clearer when that job is kept small. Find the judgement that interrupts an otherwise predictable process, test it against the alternatives, and expand its authority only when the evidence supports doing so.

READING IS FREE. SO IS THE FIRST CONVERSATION.

Want to put this to work in your business? Start there.