CHOOSEAgents & automation
Jev: AI for Decisions
Nikhil Mourya@GonnabeNikhil
Jev is an AI model designed to make specific, bounded decisions within software applications, outputting typed decisions and probabilities rather than conversational text.
TakeIt has three main primitives: Choice picks one option from a list and returns probabilities for each option.
Pattern↑ Jev AI decision models
JEV is a new AI model built for decisions, not conversations.
Released by TypeSafe AI on September 15, 2026, Jev does not chat, write code, summarize documents, or generate paragraphs.
That is the point.
Jev is designed to make the small decisions that software constantly needs:
→ Is this ticket urgent?
→ Which team should handle it?
→ Should this request use a cheap or powerful model?
→ Is this shell command dangerous?
→ Does this retrieved passage answer the question?
TypeSafe calls Jev a System One model.
The simple idea is:
State goes in. A typed decision and probabilities come out.
You give Jev text or JSON describing the situation, then define the questions you want answered.
It has three main primitives:
Choice picks one option from a list and returns probabilities for each option.
Score puts something on an ordered scale, such as low, medium, or high.
Noul handles yes or no style questions by returning the probability that the answer is true.
So instead of asking an LLM to generate a paragraph saying which team should handle a ticket, your application can get:
Engineering: 0.91
Billing: 0.06
Sales: 0.03
Your normal code then decides what happens.
That is why Jev is often described as a smart switch statement. The model provides the judgment. Your code controls the branches.
This matters because modern agents can make many small decisions during one task. Which tool should I use? Is the result correct? Is this action risky? Should I continue? Should I escalate?
Today, these decisions are often handled by general purpose LLMs. That means token generation, parsing, validation, and sometimes retries.
Jev is built specifically for bounded decisions.
TypeSafe reports 70 to 500 ms latency, pricing of $0.042 per million input tokens, with output free, and headline comparisons of around 200x faster and 400x cheaper than comparable LLM workflows. Those large comparisons come from TypeSafe's own evaluations, so they should not be treated as universal guarantees.
Another important feature is confidence.
If Jev says:
Billing: 52%
Technical: 46%
Billing technically wins, but the decision is uncertain.
You can automate high confidence cases, send medium confidence cases to a stronger model, and send low confidence cases to a human.
TypeSafe says Jev uses Reinforcement Learning for Calibrated Decisions, or RLCD, to make its confidence better reflect accuracy.
But Jev can still be wrong.
It may not invent an option outside your schema, but it can confidently choose the wrong valid option. Type safety prevents invalid outputs, not incorrect judgment.
Jev can fit into AI systems for model routing, tool risk checks, agent verification, support classification, retrieval filtering, safety checks, and high volume classification.
It is not the right tool for writing, summarization, code generation, exact arithmetic, counting, date comparison, or tasks where the possible answers are unknown.
The simplest mental model is:
LLMs generate. Jev judges. Code decides what happens next.
And if normal code can solve the problem reliably, use normal code.
Jev is most interesting when the values are known, but their meaning requires judgment.
Here are the full reference links:
typesafe.ai/blog/introducing-system-one-mo
typesafe.ai
langchain.com/blog/building-a-harness-with
flaviocopes.com/jev/