--- title: "Jev and the Case for AI That Only Decides" description: "TypeSafe AI's Jev cannot write a sentence. It returns typed, probability-weighted decisions instead. What System One models are, how they differ from LLMs, and where they fit." date: 2026-10-09 tag: ai author: Lloyd Rowat canonical: https://grayforgelabs.com/posts/jev-system-one-models --- # Jev and the Case for AI That Only Decides *2026-10-09 · ai · by Lloyd Rowat · [Grayforge Labs](https://grayforgelabs.com/)* A surprising amount of the AI inside AI products is a frontier model being asked a multiple-choice question. Is this support ticket about billing or a bug? Does this page need a lawyer's eyes? Should this request go to the cheap model or the expensive one? Each of those calls pays for a model that could write a sonnet, then waits several seconds while it produces a sentence that the surrounding code immediately parses back down to one word. Jev is built for exactly that question. TypeSafe AI, a San Francisco lab founded by former OpenAI researcher Diogo Almeida along with Erik Gafni and Sasha Sheng, came out of stealth on September 15 with a $40 million seed round led by DCVC and a model that cannot write anything at all. It only decides. ## A Model That Only Answers the Question TypeSafe calls Jev a "System One" model, borrowing the term from Daniel Kahneman's *Thinking, Fast and Slow*, where System 1 is the fast, automatic judgment you use to read a face or swerve around a pothole, and System 2 is the slow, deliberate reasoning you use for a tax return. The name itself nods to William Stanley Jevons, the economist who noticed that making coal cheaper to use led people to burn more of it. The bet is the same: make a judgment cost a fraction of a cent and software will start making far more of them. Using it looks different from prompting a chatbot. You send two things: the state, which is whatever the model should look at (an email, a document page, a JSON blob of game state), and a set of questions you define ahead of time. Each question has a fixed type. A choice picks from a list of options, a score rates something on an ordered scale, and a yes/no question returns a probability. Jev answers all of them in one call and returns a probability for every possible answer plus a confidence estimate, rather than a paragraph. ## How It Differs From an LLM An LLM generates text one token at a time, and everything it does, from writing code to classifying a ticket, comes out of that loop. If you want a structured answer, you ask nicely, maybe enforce a JSON schema, and parse the result. Jev skips generation entirely. Coverage of the launch describes it as non-autoregressive, meaning it produces its answers together in a single pass instead of word by word, though TypeSafe has not published the architecture in detail. That design changes a few things that matter in practice: - **Output is typed.** The answer is always one of the values you allowed. TypeSafe markets this as a model that "can't hallucinate," which is true only in the narrow sense that it cannot return something outside your schema. It can still pick the wrong option. - **You get probabilities, not just answers.** An LLM saying "billing" tells you nothing about how sure it was. Jev saying billing at 0.93 and bug at 0.05 lets code set thresholds and escalate the uncertain cases. - **Speed and cost.** TypeSafe claims 20 to 200 times faster and 40 to 400 times cheaper than frontier models on decision tasks, with latency between 70 and 500 milliseconds and pricing of $0.042 per million input tokens. - **No explanation.** It cannot tell you why it chose something, write a reply, or reason through a problem step by step. The headline numbers deserve some salt, since they are TypeSafe's own best cases. Independent testing has been more modest but still points the same direction: LangChain found Jev 5 to 6 times faster than Claude Sonnet at the classification step of a document review pipeline. A 5x speedup on the most frequently called step of a pipeline is still a large improvement, even if it is not 200x. ## Where You Would Use It **Routing and triage.** This is the most natural fit. LangChain's example was legal discovery, where every page of a document dump has to be reviewed before it goes to the other side, and a single case can run to hundreds of thousands of pages. Jev answers three questions per page, and each answer decides the route: most pages pass straight through, some go to an LLM for redaction, and a few go to an attorney for a privilege call. The same pattern works for support queues, content moderation, and choosing which model should handle a request based on its difficulty. **Evals and LLM-as-judge.** Grading agent outputs is a huge volume of small yes/no and score judgments, which is exactly Jev's shape. LangChain tested it against GPT-5.6 and Claude Sonnet 4.6 as an agent-eval judge and reported that it agreed with a human reviewer on every call, at 0.44 seconds per call against more than two seconds for the LLMs. The full run cost $0.34 with Jev versus $28.17 with Sonnet. **Guardrails inside agents.** Agents make dozens of small decisions per task: is this tool output an error, is the task finished, does this action need approval. Handing those to a fast decision model leaves the expensive LLM for the parts that need language or reasoning, which is the same division of labor I ended up with in [Songtrail](https://grayforgelabs.com/posts/everything-the-model-cant-see), where each component answers only the question it is equipped for. **Real-time loops.** TypeSafe pitches Jev for games, simulations, and robots, where a decision has to land inside a frame budget. A sub-second call is fast enough for an NPC choosing between a handful of behaviors. So far the demos run in simulators rather than on physical hardware, so the robotics case remains unproven. ## When It Is the Wrong Tool Anything open-ended still belongs to an LLM. Jev cannot write, summarize, or explain, and its own documentation warns that it is unreliable at arithmetic and date comparisons, a weakness it shares with the LLMs behind [the viral strawberry and carwash failures](https://grayforgelabs.com/posts/llm-reasoning-viral-failures). It also reads your questions literally and answers what you wrote rather than what you meant, accuracy drops when you stuff the state with irrelevant material, and it does not treat that state as hostile. If users can control the text going in, a cleverly written input can argue its way to the classification it wants. The interesting shift is less about Jev specifically and more about the idea that an AI system can be built from parts with different jobs, instead of sending every question to the biggest model available. Whether TypeSafe ends up owning that layer or simply proves it should exist is an open question. Either way, there are a lot of support tickets out there still waiting several seconds for a frontier model to say "billing."