Jev

How it differs from a typical LLM

Large language models (LLMs) such as ChatGPT and Claude build sentences by adding one token at a time. Jev does not generate sentences at all. When it receives a situation to judge (text or JSON) and a 'typed question' defined in code, it returns only the answer to that question in a fixed format. It does not produce explanations, summaries, supporting sentences or code. That is why it is often likened to a 'smart if statement': instead of a person writing out every rule for a program's branches (if), the model reads the situation and decides which way to go. The 'System One' in the name comes from 'System 1', the fast, intuitive mode of thinking described in Daniel Kahneman's Thinking, Fast and Slow.

Three output types (primitives)

Choice picks one of several options (up to 255) and also returns the probability distribution over all options and a confidence value. Example: whether to route an inquiry to 'refunds, shipping or technical support'. Score assigns a score on an ordered scale of 2 to 10 levels and returns a probability-weighted average score and the distribution. Example: answer quality rated 1 to 5. Noul returns the probability that a statement is true as a single value between 0 and 1. Example: a 0.87 probability that 'this email is spam' is true. Because the probability itself is the signal, Noul does not return a separate confidence value.

Why it is fast and cheap

Jev is non-autoregressive: it does not generate tokens one at a time. It reads the situation once and evaluates all questions simultaneously to produce answers in a single pass, so adding more questions barely increases response time. The published response time is about 70–500 milliseconds per request. TypeSafe states that for these kinds of judgment tasks it is 193.6 times faster and 444.6 times cheaper than LLMs, but these are the provider's own measurements, and the company describes them as upper-end figures. It has published a price of $0.042 per million input tokens, with no charge for output.

Confidence and training method

The key feature of Jev is that it also gives a number for 'how confident it is'. TypeSafe says it trained the model with a method called RLCD (Reinforcement Learning for Calibrated Decisions), so that its probabilities match actual outcomes (calibration) rather than the answers people prefer. This makes it easier to design systems that split work between automation and human review using a threshold, such as 'process automatically if confidence is 0.9 or higher, otherwise hand off to a person'.

Things to know

Because Jev cannot output values outside the defined options, it does not produce broken formats or make up irrelevant labels. However, it can be confidently wrong. It accepts only text input, is weak at date/time calculations and math, and its accuracy drops when a lot of irrelevant information is mixed in. Tasks that require writing still need an LLM, so it is often used not as a replacement for LLMs but as a component that handles fast decisions alongside them. As of late September 2026, it is in early access.

Example

In a customer support agent, Jev's Choice classifies incoming inquiries into one of 'refunds, shipping, technical support or other', and if the confidence is below a threshold, the inquiry is handed off to a human agent. In the same request, Noul returns the probability that 'it contains profanity' and Score returns an urgency rating from 1 to 5 at the same time. It is also used for guardrails that check whether answers written by another LLM followed policy, for reranking search results, and for choosing which tool an agent should use next (routing).

Related terms

Decision model LLM AI agent Guardrails RLHF Hallucination

References

• TypeSafe AI official site (Jev introduction, 2026)
• TypeSafe docs: Introduction (Choice, Score, Noul)
• What Is Jev? Inside TypeSafe AI's System One Model (Maxim AI)
• What is Jev (2026)? TypeSafe AI's System One model (DigitalOcean)