System One models and Jev: typed decisions with calibrated probabilities instead of generated text

What it is

TypeSafe AI's documentation describes a class of models it calls System One models and its first member, Jev. Instead of generating text, the model evaluates typed questions against a state (a string, a JSON object or an array of text) and returns typed answers. The three question types are Choice (pick one option from a defined set; returns the option, a probability per option and a confidence), Score (rate the state against an ordered list of two to ten levels; returns a probability-weighted score that can land between levels, the probabilities and a confidence) and Noul (a yes/no question; returns the probability of yes, without a separate confidence). All questions in one request are evaluated in parallel and independently against the same state.

The vendor states that the model is trained with a method it calls Reinforcement Learning for Calibrated Decisions (RLCD) so that its probabilities reflect uncertainty, and adds that calibration is measured across groups of predictions and does not guarantee that an individual answer is correct. The model is not fine-tuned with customer data; domain knowledge goes into the state and into the wording of instructions and criteria.

Why it matters

Agents make many small judgments: which handler takes a request, whether a retrieved passage is relevant, whether a tool result contains an instruction, which of several candidates matches a record. Doing these with a chat model means generating text and parsing it back. A model whose output is constrained to the options the caller defined removes the parsing step and gives code a distribution to branch on, sort by or threshold. The documented facts an agent should carry (as of September 2026, early access, stated to change without notice): model jev-1.13.0 behind the aliases jev-latest and jev-preview; POST https://api.typesafe.ai/v1/systemone; price USD 0.042 per million input tokens with output tokens free; rate limits of 250,000 tokens per second and 1,200 requests per minute; 64k tokens per request for state plus all questions and 32k for state plus the longest question; text input only; English as the primary training language with lower accuracy documented for other languages.

How to apply

• Use it where the answer space is closed: routing, classification, relevance, ranking, extraction over enumerated candidates, yes/no gates.
• Keep code in control: the model decides one atomic thing per question; the workflow, arithmetic and thresholds live in your program.
• Read the probabilities, not only the winning option; a flat distribution is the model saying it does not know.
• Log the model field of every response and pin a versioned ID once thresholds are tuned, because aliases move.

Pitfalls

The launch post's speed and cost figures (70–500 ms end to end, "40x-200x faster", "193.6x faster, 444.6x cheaper" on the vendor's workflow evaluations) come from the vendor and had, at the time of writing, no independent replication. Schema conformance is not correctness: the model cannot return a value outside the options, but it can pick the wrong option with high probability. The generation gap is real: it cannot write a reply, a summary or code, and the documentation says forcing generation through chained choices works badly.