Jev replace answers with decisions

TypeSafe unveils Jev, a model that replaces text generation with structured decisions, priced at $0.042 per million input tokens.

An artificial intelligence model that writes no response but instead selects an option, assigns a score, or estimates the probability that a statement is true. That is the trade-off offered by Jev, TypeSafe’s first publicly released model, designed to make decisions that software can use directly.

Introduced on September 15, 2026, Jev inaugurates a category the company calls “System One Models.” The name refers to the “System 1” described by Daniel Kahneman: a fast, intuitive way of thinking, as opposed to the slow, deliberate reasoning associated with System 2.

TypeSafe is therefore not trying to replace a conversational assistant. Jev cannot write an email, generate code, explain its answer, or carry on a conversation. It receives a state describing a situation and a series of predefined questions, then returns structured values accompanied by probabilities.

Giving up text generation is central to the project. A large language model typically produces its response one token at a time. Each element depends on those that came before it, allowing the model to write content of varying lengths and formats but requiring sequential computation.

Jev processes multiple questions in parallel. Because it does not have to construct a sentence, it can calculate the requested answers directly and return them in a format the application already understands. TypeSafe claims latency ranging from 70 to 500 milliseconds, depending on the request.

The documentation describes three question types. `Choice` selects an option from a list and returns the probability associated with each one. `Score` evaluates the situation across several ordered levels. `Noul` answers a binary question with a value between zero and one.

A support application could, for example, submit a customer’s message as the state. In a single request, Jev could identify the relevant department, assess the customer’s level of frustration, and estimate the urgency of the request. The application’s code would then decide whether to display a response, trigger an action, or transfer the case to a person.

All three formats can be combined in a single call. Each question is evaluated independently from the same state. According to TypeSafe, adding several questions has little effect on response time because their results are calculated simultaneously.

That independence also comes with a drawback. A question does not automatically see the answers given to the others. When a decision depends on several factors, the development team must separate them and then write the rules that combine their results.

Instead of asking directly whether an invoice should be paid, a developer could have Jev evaluate whether it matches the purchase order, whether supporting documents are present, whether it may be a duplicate, and whether the banking details are consistent. The application would then apply its payment or escalation rules.

TypeSafe describes this approach as a series of “intelligent conditions.” The model handles judgments that are difficult to encode precisely, while calculations, thresholds, and actions remain deterministic. The software therefore retains control over the procedure.

This structure makes Jev resemble a specialized function more than an autonomous agent. Its value appears in systems that must classify, filter, score, or route large volumes of data, including invoice processing, moderation, agent monitoring, incident detection, request prioritization, and compliance reviews.

The model can also be used to evaluate another AI system’s work. An application could ask whether a response complies with a policy, whether a tool call appears dangerous, or whether an agent trace should be reviewed. Jev does not correct the content itself. It produces the signal the application uses to decide what happens next.

The response type is defined before execution. A field expected to contain a choice therefore cannot suddenly become a paragraph or contain a value missing from the list. TypeSafe presents this property as a mathematical guarantee against formatting errors.

The company goes further in its marketing by claiming that Jev “cannot hallucinate.” That statement, however, relies on a very narrow definition of the term. Jev cannot invent a new output type or produce an option outside the schema because it does not freely generate text.

It can still choose the wrong category, assign an inappropriate score, or estimate a probability incorrectly. A decision can be wrong even when its structure is valid. The guarantee therefore applies to the form of the response, not its truth.

The absence of generated text reduces some of the risks associated with general-purpose models. Jev cannot add a fabricated explanation, call a tool using an unexpected format, or disregard the schema to pursue a different task. It does not, however, eliminate errors in judgment or problems in the data it receives.

TypeSafe accompanies its decisions with probabilities and, for `Choice` and `Score`, a confidence indicator between zero and one. This indicator summarizes how concentrated the distribution is: one clearly dominant option produces high confidence, while several similarly likely possibilities result in a lower value.

The confidence documentation recommends defining different behaviors based on this result. A low-risk decision might be applied automatically at a moderate threshold. A financial or irreversible operation could require a higher value, an additional confirmation, or human approval.

This feature makes it easier to build systems capable of recognizing some ambiguous situations. It does not, by itself, prove that the probabilities correspond precisely to the model’s actual success rate. A confidence measure derived from a distribution can remain high even when the model is wrong.

Demonstrating genuine calibration would require verifying across many cases that decisions reported with 90% confidence are actually correct roughly nine times out of ten. TypeSafe says its new