On 14 September 2026, Diogo Almeida posted that TypeSafe is releasing Jev, its first System One model. A System One model evaluates a state and returns typed answers and probabilities. Jev understands natural language. It does not write replies, produce code, or explain its reasoning.
Think of Jev as a function call, he wrote: unstructured state in, typed probabilistic decisions out. Code keeps the workflow. The model supplies a judgment where ordinary rules cannot read messy text.
You pack the facts the model is allowed to see into one bundle they call state: a support message, the charges, the refund rule. You then ask small questions. Choice picks one option from a list you wrote, such as billing, technical, or account. Score places the state on ordered levels you wrote, such as how angry the customer sounds. Noul is the probability that a yes-or-no statement is true, from 0 to 1. All three can share one request. They see the same state and are scored independently, so adding questions barely changes how long you wait.
A Noul near 0.5 means yes and no look about equally likely. It is not a medium amount of the thing you asked about. Noul has no separate confidence field. Choice and Score do: that number collapses how peaked the probability distribution is, so code can decide whether to act, to ask a person, or to stop. Calibration is measured across groups of predictions. It does not guarantee that one answer is correct.
They trained this with a method they call Reinforcement Learning for Calibrated Decisions, or RLCD. The aim, on their account, is that a 0.8 should come true about 80 percent of the time in a batch of similar calls. Higher confidence means higher accuracy, they say. That is their claim to test, not a result from this note.
Language models generate strings one token at a time. Software has to parse them. There is always some risk they go off the rails, Almeida wrote. A trained classifier maps new inputs onto a label set you froze when you trained it. Change the question or the labels and you collect data and train again. Jev's questions are written in language for this call. The possible answers are still fixed for that call. TypeSafe's contrast is with language models. The classifier is the other common tool for the same job when the label set never moves.
Their building guide says keep control flow, deterministic rules, and side effects in code. Insert System One where the program needs a common-sense judgment over unstructured data. In an agent that already uses a language model, the outer loop still calls that model to write and to use tools. Jev sits on the judgments that loop used to bury inside a prompt: which tool this turn, whether the last reply is good enough to send. The language model keeps talking. Jev only returns the typed answers the harness branches on.
They list Jev at 70 to 500 milliseconds end to end, and $0.042 per million input tokens, with output free. They cannot prove the price is not subsidized, Almeida wrote. The homepage figures of 193.6 times faster and 444.6 times cheaper come from their workflow evals. He wrote that those sit on the high end of real-world gains, and that the workflows were made by people on the model capabilities team, so some bias could exist.
Almeida wrote that Jev cannot hallucinate and never makes type errors. The 0 percent on their plot is not an empirical count. The output always fits the shape you defined, so they put 0 percent on the chart. Jev can still pick the wrong option among the ones you listed. Typed output guarantees the shape of the answer. It does not guarantee that the answer is right.