Jev and System One models: typed decisions for agents instead of generated text

What it is

According to TypeSafe AI's documentation, Jev is the first model in a class the vendor calls System One models. It does not generate text; instead it answers typed questions about a supplied state (a string, a JSON object or an array of text). There are three question types. Choice picks one option from a given set (at most 255) and returns the chosen option, a probability for each option and a confidence. Score places the state on a scale of two to ten described levels and returns a probability-weighted value that can fall between levels, along with the distribution and a confidence. Noul answers a yes/no question with the probability of yes, without a separate confidence. All questions in a request are evaluated in parallel and independently of one another against the same state.

The vendor documents the following (as of September 2026, early access, subject to change without notice): model jev-1.13.0 behind the aliases jev-latest and jev-preview, endpoint POST https://api.typesafe.ai/v1/systemone, a price of 0.042 US dollars per million input tokens with output tokens free of charge, rate limits of 250,000 tokens per second and 1,200 requests per minute, 64k tokens per request for the state plus all questions and 32k for the state plus the longest question, text input only, and English as the primary training language, with documented lower accuracy in other languages.

Why it matters

Agents constantly make small judgments: which handler should take a request, whether a retrieved passage is relevant, whether a tool result contains an instruction, which candidate matches a record. With a chat model, that means generating text and then parsing it. A model whose output is restricted to options defined by the caller removes the parsing step and gives the code a distribution it can branch on, sort by or apply thresholds to. The vendor calls its training method Reinforcement Learning for Calibrated Decisions and states that calibration is measured across groups of predictions and does not guarantee any individual answer.

How to apply

• Use it only where the answer space is closed: routing, classification, relevance, ranking, selection from enumerated candidates, yes/no gates.
• Keep arithmetic, counting and date comparisons in code; the page on the known weaknesses of jev-1.13 explicitly lists counting, numeric representations and date comparisons as unreliable.
• Filter the state before the call; according to the vendor, irrelevant material lowers accuracy.
• Read the probabilities, not just the winning option; a flat distribution means "don't know".
• Log the versioned model identifier from every response, and pin the version once thresholds are tuned, because aliases move.

Pitfalls

The speed and cost figures come from the vendor and had not been independently replicated at the time of writing. Schema conformance is not correctness: the model cannot return a value outside the options, but it can choose the wrong option with high probability. According to the vendor, the state is not treated as adversarial; injected instructions can shift the answer. The model cannot generate text.