What is a System One model and why we need it?

TypeSafe AI announced Jev on 15 September 2026, a "System One model" that takes a piece of state and a list of typed questions and returns typed answers with calibrated probabilities in a claimed 70 to 500 milliseconds without generating text, available in early access behind a waitlist. TypeSafe, founded by former OpenAI researcher Diogo Almeida and backed by $40 million, trained Jev with a method it calls RLCD (Reinforcement Learning for Calibrated Decisions), which optimizes for calibrated probabilities rather than human preference; the launch demo showed Jev deciding in 0.114 seconds against 8.566 seconds for GPT-5.6 Terra. The model exposes three primitives — Choice, Score, and Noul — and is positioned as a new slot beside an LLM for small, frequent, structured decisions rather than a replacement for it.

What is a System One model, and where does it go in your stack? Sergei GordeichukThe master of Stackness https://stackness.dev/stck/gordeychuk s On 15 September 2026, TypeSafe AI announced Jev https://typesafe.ai/blog/introducing-system-one-models-and-jev , which it calls a System One model: a model that takes a piece of state and a list of typed questions and returns typed answers with calibrated probabilities, in a claimed 70 to 500 milliseconds, without generating a single word of text. It is in early access behind a waitlist. As of 18 September 2026, exactly one Stackness user lists Jev https://stackness.dev/tools/typesafe-jev in their stack, and that user is the founder of Stackness, so we will not be reading adoption into it. This is not another "what is Jev" explainer. Within three days of the launch that ground was thoroughly covered, and a fourth day of it would help nobody. The question for a site about stacks is what kind of thing a System One model is, and where it goes. Our answer: a new slot, not a replacement. It sits beside your LLM and takes the small, frequent, structured decisions the LLM is overqualified and overpaid for. That fits what the previous article https://stackness.dev/blog/which-ai-coding-tools-are-developers-switching-to found: Stackness profiles add AI tools beside each other far more than they switch between them. What is a System One model? A System One model answers typed questions about a piece of state and returns typed answers, each with a probability, instead of generating text. The name is borrowed from Daniel Kahneman's Thinking, Fast and Slow , where System 1 is fast and intuitive and System 2 is slow and deliberate. An LLM working through chain-of-thought reasoning is System 2. A System One model is meant to be the fast half, on its own. The interesting part is not the speed, although the speed is what the launch leads with. It is the narrowing. A large language model can be asked anything and answer anything, which is why so much engineering goes into forcing its output into a shape a program can read. Jev starts from the shape, according to the docs https://docs.typesafe.ai/introduction : state and a schema of questions in, values that fit the schema out. TypeSafe was founded by Diogo Almeida, formerly of OpenAI, whom The Register describes https://www.theregister.com/ai-and-ml/2026/09/16/typesafe-ai-debuts-model-for-machines-that-plays-doom/5296711 as a co-inventor of RLHF and ChatGPT, and has raised $40 million. Jev was trained with a method TypeSafe calls RLCD, Reinforcement Learning for Calibrated Decisions, which optimises for epistemically honest probabilities rather than human preference. If that holds, it is a real shift in objective: a model trained to be liked says 95 percent, and a model trained to be calibrated is supposed to say 60 when it means 60. The launch demo has Jev playing Doom, deciding in 0.114 seconds against 8.566 seconds for GPT-5.6 Terra. Doom is not a workflow, although some weeks in production argue otherwise. What do the three primitives do, and why does decomposition matter? The docs https://docs.typesafe.ai/introduction define three primitives. Choice picks one option from a list. Score rates the state against a rubric. Noul says whether a statement is true and returns a value from 0 to 1. Every answer carries its own confidence. Questions are evaluated in parallel and in isolation, so adding more barely changes response time. The documented pattern is to break a judgement into atomic questions and compose the result in code. That last sentence is the real design idea. The usual way to get a decision out of a language model is to describe the whole problem in a prompt, ask for a reasoned answer, and hope the reasoning stays on the rails. The System One pattern inverts that. Instead of asking "should this ticket go to billing?", you ask ten small questions: does the message mention a charge, is a refund requested, is the tone escalating. The rule that combines them lives in your code, where it can be read, tested and versioned. A prompt that drifts is hard to notice. A function that composes ten booleans and a score is boring, which is the highest compliment available to a routing function. What has TypeSafe demonstrated, and what has it only asserted? TypeSafe has published latency, price and correctness claims, a Doom demo, a live evals dashboard and an open adapter for re-running its comparison on other models. It has not published an independent evaluation, a production SLA, or any way to audit why an individual answer came out as it did. Claim by claim, separated because the launch material does not always separate them: - 70 to 500 ms end to end , framed as 40 to 200x faster than frontier models on System One tasks. What backs it: TypeSafe's own workflow evaluations, the Doom demo at 0.114 s against 8.566 s for GPT-5.6 Terra, and an MIT-licensed adapter for re-running the comparison. Still open: TypeSafe itself expects the figures to be "on the higher end of real world gains", and no production SLAs are published. - $0.042 per million input tokens , with output tokens listed as "FREE too cheap to meter ". What backs it: the published price. Still open: early access behind a waitlist, so nobody outside TypeSafe has a production bill. - Type errors and schema violations are "mathematically impossible" . What backs it: it follows from the design, typed values rather than text to parse. Still open: a valid schema says nothing about a correct choice, as Anthony Maio https://anthonymaio.substack.com/p/jev-the-language-model-that-wont points out, and The Register notes the "hallucination-free" comparison with text models is not like for like. - Probabilities are calibrated, via RLCD . What backs it: a live evals dashboard with per-case disagreement walkthroughs. Still open: no independent calibration study, and a wrong probability comes with no reasoning to inspect. - Adding questions barely changes response time . What backs it: the documented parallel, isolated evaluation. Still open: not yet measured outside TypeSafe. Two of those deserve restating. A schema guarantee stops malformed output and does nothing to stop the model picking the wrong option; the result is a perfectly well-formed mistake. And there is no explanation available at all. A model that cannot write a paragraph also cannot write a paragraph about why it was wrong. Credit where it is due, and it is due. TypeSafe shipped the announcement with a "we love skeptics" section arguing the case against itself, an MIT-licensed adapter so anyone can re-run the comparison on competitor models, and a live evals dashboard. That is unusually good behaviour for a launch. It does not close the open cells, but it means someone other than TypeSafe can close them. Where does a System One model go in a stack? Beside the LLM, not in place of it. The LLM keeps the work that needs text or extended reasoning: writing, explaining, planning, conversation. The System One model takes the decisions that are small, structured and frequent: which queue this goes in, whether an alert is worth a human, whether a row matches the filter, whether an agent's run finished acceptably. In a pipeline that means one call per row, per event or per tool call. That is the shape of the Stackness move Use a fast small decision model instead of an LLM call for structured filtering and scoring https://stackness.dev/moves/use-a-fast-small-decision-model-instead-of-an-llm-call-for-structured-filtering-and-scoring-2 , by Sergei Gordeichuk, which tags Claude Code https://stackness.dev/tools/claude-code and Jev https://stackness.dev/tools/typesafe-jev : the agent does the open-ended work, and the decision model sits at the checkpoints asking typed questions about the state. TypeSafe's own use cases read the same way: customer service routing, invoice handling, security alerts, reviewing completed AI agent runs, guardrailing LLM output, and map-reduce over large datasets. It is the co-use pattern from the previous article in its purest form, because a System One model could not replace the LLM even if you wanted it to. On the Stackness data, the honest sentence is short. As of 18 September 2026, Jev has one user on Stackness, and that user is the founder. That is the sort of adoption curve a person can produce by logging in. When there are enough profiles to say something, this is where we will say it. What would have to be true before it earns a place in yours? Three things. You need a decision in your pipeline that is high-frequency, structured and currently handled by a regex you are ashamed of or an LLM call you are paying too much for. You need the latency and cost figures to hold on your own data, not on TypeSafe's evaluations. And you need to live with a wrong answer that arrives with a probability and no explanation, by tolerating it or routing low-confidence cases elsewhere. The first condition is the one most teams fail. Plenty of pipelines have exactly one LLM call, and it is doing real work that needs text. There is nothing to move. The pattern pays off where an LLM is being used as a very expensive classifier, or a classifier is being used because an LLM was too slow. The second is now cheap to test: run your own state through your current model and through Jev, and see whether the gap looks like 40x or 4x. TypeSafe has already told you to expect less than its headline. Take them up on it. The third is a design decision, not a benchmark. Calibrated probabilities let you set a threshold and send everything under it to a human or a slower model. If every decision must be explained after the fact, a model that produces no reasoning is the wrong tool regardless of speed. That is not a criticism of the category. It is the boundary of it. Key numbers - 15 September 2026 is when TypeSafe AI announced Jev https://typesafe.ai/blog/introducing-system-one-models-and-jev , in early access behind a waitlist. - 70 to 500 ms is the claimed end-to-end latency, framed as 40 to 200x faster than frontier models; TypeSafe says the figures are likely on the higher end of real-world gains. - $0.042 per million input tokens is the published price, with output tokens listed as free. - 0.114 seconds against 8.566 for GPT-5.6 Terra is the per-decision gap in the Doom demo, a demo rather than a workflow evaluation. - 1 Stackness user lists Jev as of 18 September 2026, the founder, so the number carries no signal yet Stackness, September 2026, see data sources https://stackness.dev/about/data-sources . Quick answers What is a System One model? A model that takes a piece of state and a list of typed questions and returns typed answers with calibrated probabilities, without generating text. The name comes from Kahneman's fast, intuitive System 1, rather than the slow, deliberate System 2 of an LLM reasoning step by step. Is Jev a replacement for an LLM? No. It cannot generate text, so it cannot write, explain, plan or converse. It sits beside an LLM and takes the frequent, structured decisions, which is why the Stackness move that uses it tags Claude Code alongside it. What are Choice, Score and Noul? The three primitives in the Jev docs. Choice picks an option from a list, Score rates the state against a rubric, and Noul returns a 0 to 1 value for whether a statement is true. Each answer carries its own confidence. Does "schema violations are mathematically impossible" mean the answers are correct? No. It means the output always fits the type you asked for. It can still be the wrong option, and there is no reasoning to inspect when it is. Are the speed claims independently verified? Not as of 18 September 2026. They come from TypeSafe's own workflow evaluations and a Doom demo. An MIT-licensed adapter lets others re-run the comparison. How many developers use Jev on Stackness? One, as of 18 September 2026, and it is the founder of Stackness. That is far too early to read anything into, and we are not going to.