# Jev Jev is a [[Decision Models (DMs)|decision model]] by [[TypeSafe AI]], launched on 15 September 2026 as the first commercial [[System One Models|System One model]]. You give it some state (a string, a list of strings, or name/value pairs) plus typed questions, and it returns numbers: a choice among options, a yes/no probability or a score, each with a calibrated confidence. It never writes a single word of text. Why should you care? Because a HUGE part of what we currently ask [[Large Language Models (LLMs)]] to do is exactly that: pick a category, decide whether something matches, rank candidates, check whether an agent is stuck. We've been paying for a model to write JSON, then parsing it, then retrying when it breaks. Jev removes that whole failure class, and it does it one to two orders of magnitude faster and cheaper. ## How it works (from the outside) - **Three question types.** *Choice* (a probability distribution over the options you provide, up to 255), *Noul* (a yes/no probability between 0 and 1) and *Score* (the expected level on a rubric you describe). [[Diogo Almeida]] maps them to programming constructs: choice is a `switch`, noul is an `if`, score is sorting or thresholding. Details and gotchas in [[System One Primitives]] - **Where the names come from.** The API and docs say **Noul**, while Almeida says "Noulli" on the Latent Space podcast and explains that it comes from **Bernoulli** (a Bernoulli probability: a yes/no outcome with probability p). The docs themselves never explain it. *Jev* comes from William Stanley Jevons and the [[Jevons Paradox]], and *System One* from [[Daniel Kahneman]] (see [[System 1 and System 2 Thinking]]) - **State in, typed decision out.** The state is encoded once; every question reads that shared state in parallel. Ask ten questions about one support ticket and you pay for the ticket once - **Calibration by training.** TypeSafe trains with *[[Reinforcement Learning for Calibrated Decisions (RLCD)|Reinforcement Learning for Calibrated Decisions]]* (RLCD) rather than [[Reinforcement Learning From Human Feedback (RLHF)|RLHF]]. The goal is "epistemically honest" probabilities: when Jev says 80%, it should be right about 80% of the time - **No text generation.** The output is typed by construction, so type errors are impossible (no retries, no schema validator) TypeSafe hasn't published the architecture. Archer Hume reverse-engineered it with ~10,000 API calls and concluded it's most likely a causal transformer backbone (probably sparse mixture-of-experts, given ~160 ms for 30k tokens) with a readout layer instead of a language head, a shared state prefix, isolated question branches, and options read *jointly* (adding an irrelevant option shifts the odds between the others). He measured an expected calibration error of 0.031 on 1,200 MMLU items. These are behavioral inferences, not confirmations (he says so himself). ## How it was trained TypeSafe hasn't published a paper, the reward function or the base model, so this section is what Almeida said publicly plus what the docs state. - **RLCD names a target.** On Latent Space, Almeida frames RLHF, RLVR and RLCD as three different *tasks* for [[AI Post-Training|post-training]], not three optimizers: "The thing that I'm calling RLHF is the task of instruction following. It's not about the PPO." RLHF optimizes for what people prefer, [[Reinforcement Learning with Verifiable Rewards (RLVR)|RLVR]] for answers a program can check (math, code), and RLCD for decisions whose probabilities match reality ([[AI Model Calibration]]). The exact recipe is "novel, unpublished". Outsiders guess it uses [[Proper Scoring Rules]] such as the Brier or log score on the typed output distribution; Di Zhang goes further and proposes a Plackett-Luce ranking objective plus a Brier calibration term. Both are **speculative** - **Why not just RLHF?** Almeida co-authored [[InstructGPT]], and his argument is that preference training damages exactly what software needs: it rewards [[AI Sycophancy|sycophancy]] and confident-sounding answers, and causes "mode dropping" ([[Mode Collapse]]): "GANs mode drop. They, like, drop the minority classes and just do the really common ones." For a classifier, losing rare classes is fatal. He also says calibration is "total poison" for probability distributions over free text, which is one reason the outputs are typed - **100% synthetic data.** The training data is entirely synthetic and built per task shape. User data is deliberately left out to avoid overfitting to "the present day", and customer data is never used for training - **A data lab, not a model lab.** "Data matters way more than compute." "Data is so unbelievably complicated, and that is what gets nines." The data team hunts down jaggedness and fixes it "surgically" ("Our cognitive core is, like, way less jagged than anyone else's"), and the company says it's hiring "infinite data people" - **One set of weights for everyone.** No fine-tuning, no LoRA. You adapt Jev to your domain only through the request: your content in the state, your rules in the questions and criteria, and decomposition in code ([[Atomic Question Decomposition]]) - **Serving.** Almeida admits to "absolutely disgusting things" to reach the intelligence-per-dollar frontier, without details. Archer Hume's reverse-engineering (above) is the best public picture ## Numbers - **Price:** $0.042 per million input tokens; output is free. That's cheaper than GPT-5 Nano's input price - **Latency:** 70 to 500 ms end-to-end for a batch of questions - **TypeSafe's claims:** 20 to 200x faster and 40 to 400x cheaper than frontier LLMs at equivalent quality on their workflows - **[[Every (Company)|Every]]'s testing:** 25x faster and 600x cheaper than a [[Claude Fable 5.1]]-level judge ([[Dan Shipper]]) - **Real use:** Matthew Berman categorized 724 ads across 37 brands in 40 seconds for 9 cents; one [[Hacker News]] commenter sent over 1,000 requests for about $0.01 ## What people build with it Nate B. Jones grouped the early builds into four patterns, and I find that list more useful than any single demo: 1. **Add interpretation to an existing workflow.** 70,000 insurance support tickets categorized, with new tickets flowing through the same pipeline 2. **Ask new questions of old data.** Eight questions applied to 3,282 social posts (topic, hook, tone, performance) 3. **Decide where deeper reasoning is worth spending.** Jev flags what needs attention, then an LLM handles the few cases that matter. One developer halved his agent's costs by letting Jev pick the reasoning effort per step 4. **Give small elements their own judgment.** A prose linter asking ten yes/no questions per paragraph. At this price, it's actually worth doing Other examples: spam detection, search reranking (score the top BM25 hits), routing, guardrails, real-time classification while the user types, browser control with [[Playwright]], chess and Snake bots, and [[fast-jev-compaction]], which uses Jev to compact [[Claude Code]] sessions in about a second. ## Limitations and criticism - **Black box.** An LLM can explain why it flagged something as spam; Jev gives you a number. [[Simon Willison]] worries about hidden bias in sensitive decisions (e.g., hiring). His test rating Bay Area cities put Cupertino at the top and East Palo Alto at the bottom. His advice: evals and structured experiments matter even more here - **Known weak spots.** The Jev 1.13 docs admit it struggles with numbers, dates and adversarial content. It also fails when the needed information isn't in the state, when the answer space is unbounded, or when the question needs real multi-step reasoning - **Closed and opaque.** API only, waitlist, no paper, no weights, and the terms forbid public benchmarking. Some HN commenters called that "Oracle-style"; others argued there's no benchmark "because it's not very intelligent at all" - **Is it even new?** Plenty of people pointed out that classifiers, logit reading and prefix caching all existed before. The creator of [[Laya]] published a very similar non-autoregressive decision model in March 2025. Almeida's answer is that the novelty is the combination and the training, not any single piece - **Calibration drift.** Probabilities calibrated on one distribution can quietly drift on another. Test on your own data before trusting a threshold ## Known failure modes (Jev 1.13 jaggedness) To their credit, TypeSafe documents where `jev-1.13` breaks (last reviewed 2026-09-17), in the spirit of [[Jagged Intelligence]]. Nine failure modes, each with a workaround: 1. **Literal reading.** It answers the question you wrote, not the one you meant. If you catch yourself explaining what you meant, that explanation is the missing half of the instruction 2. **Math and numbers.** It doesn't count reliably and can't judge whether two hex colors or RGB triples are close. Count in code, or ask one yes/no question per item and sum the answers. Don't interpolate exact magnitudes from a Score 3. **Date and time comparison.** It reads dates as text, not as ordered quantities. Extract the date parts as Choices and compare in code 4. **Indirection.** Double negatives and "a property of a property" cost accuracy. Point at the exact part of the state 5. **Large state full of irrelevant detail.** Accuracy drops as unrelated content grows ([[AI Context Rot]]). Filter first, or use a relevance question to filter 6. **Adversarial content.** The state isn't treated as hostile; injected instructions or text arguing for its own classification can move the answer. Be explicit in the criteria and test edge cases 7. **Contradictory instructions and criteria.** A Noul whose `true` means "no" performs worse. Keep criteria as an extension of the instruction 8. **Common-sense structural invariants.** The same refund question gave 0.22 as a Noul and 0.01 as a yes/no Choice; "refund?" plus "not a refund?" summed to 1.19. Don't port thresholds between forms ([[Relative vs Absolute AI Judgments]]) 9. **Generation.** It isn't trained to produce text. Extract candidates with regex or an LLM and let Jev pick ([[Candidate-Then-Select Extraction]]) ## Independent tests Since the terms of service forbid public benchmarking, the independent evidence is scattered across blog posts. What I found: - **Rajesh Beri (phishing, 2,000 PhishNChips v5.2 emails).** One binary question: Jev **62.6%** vs Claude Haiku 4.5 81.3%. Five atomic questions combined with a logistic regression (fit on 1,000, tested on 1,000): Jev **95.0%** vs Haiku 93.2%, at 12-27x lower cost. But Haiku's best single signal already reached 94.2%. On synthetic support tickets he measured an expected calibration error (ECE) of **0.107** ("4.4x the noise floor"), with Nouls underconfident and Choices and Scores overconfident; on an unknowable task Jev was right 44.7% of the time at a mean probability of 0.74. His verdict: "The 95% is not Jev. It is Jev plus your labelled data plus a regression you maintain" - **Archer Hume (calibration and architecture).** ECE **0.031** on 1,200 MMLU items. 990 of them sat in the 0.9-1.0 bin (predicted 98.7%, observed 96.3%, so slightly overconfident), with larger gaps in the middle bins. 84.6% on MMLU-Pro. Nicely calibrated even when weak: 86.7% right at a mean probability of 0.83 on three-digit multiplication, 32% right at 0.30 on two-step word problems. His tokenizer fingerprinting matched no public tokenizer; the closest was [[Qwen]]'s (348 of 415 probes), which *suggests* a modified public base or distillation (**speculative**, not confirmed) - **Anthony Maio (critique).** The reward, architecture, training and calibration method are all undisclosed. "Cannot hallucinate" only means it can't invent a field you didn't declare; it can still pick the wrong option with high confidence. Distribution shift can "quietly undermine probabilities that were reliable last month". And the point I find most important: "Individually calibrated judgments do not automatically compose into a calibrated workflow once you run them through thresholds, weights, and branches." He sees the real strength as "architectural rather than algorithmic" - **KDnuggets (Abid Ali Awan).** Incremental rather than revolutionary: zero-shot classification and intent detection have existed since 2019-2020. Still, it reports 96.3% on an independent fact-checking test of 108 claims (Gemini 3.1 Flash Lite 94.4%, Claude Haiku 4.5 93.5%) and routing decisions in 145-271 ms - **The Register.** Calls the "hallucination-free" framing unfair (there's no natural language to hallucinate in), doubts that the Jevons analogy holds for tokens, and covers the demo of Jev playing Doom from structured game state: 0.114 s per decision vs 8.566 s for GPT-5.6 Terra - **Turing Post** noticed that TypeSafe's own adapter returns different confidence values than the hosted API, and that there's not enough detail to separate what comes from RLCD, the architecture, the data or the serving - **TypeSafe's own evals** use model-generated reference answers (the average of GPT-6 Astra and Fable 5.1). On four workflows, Jev reached 67.8% agreement at $0.0004 and 0.4 s per case, vs 67.9% for GPT Terra at $0.0304 and 10.1 s (per Maio's summary). Same quality, a fraction of the cost, but measured against a model-made ground truth The common thread among critics: no paper, no reliability diagram, no ablation. My reading: the calibration is real but distribution-dependent (0.031 on MMLU, 0.107 on tickets), and most of the impressive results come from decomposition plus your own glue code. ## Open alternatives Within days, the open-source community rebuilt the *interface* (not RLCD, not the training data): [[SemIf]] (formerly OpenJev), [[Kev]] by Jared Palmer, [[Jevlike]], and [[Laya]]. There's also mini-jev, a parallel constrained decoding engine for Apple Silicon, and a vLLM pull request turning DiffusionGemma into a multiple-choice machine. The [awesome-jev](https://github.com/Amal-David/awesome-jev) list tracks SDKs, skills, MCP servers and demos. ## My take I think Jev matters less as a model than as a new building block. We had code (deterministic, cheap) and LLMs (flexible, slow, expensive). Jev sits in between: a semantic `if` statement you can afford to call thousands of times. That changes which ideas are worth building. Remember the Jevons paradox? TypeSafe named the model after William Stanley Jevons on purpose (i.e., make judgment cheap enough and we'll use far more of it, not less). What I'm cautious about is the opacity. I'd use it for triage, routing and filtering where a wrong call is cheap and can be caught downstream, and keep an LLM (or a human) in the loop for anything consequential. And I'd really like to see the open models close the gap, because a decision layer this central shouldn't live behind a single closed API. ## References - [Introducing System One Models & Jev (TypeSafe AI)](https://typesafe.ai/blog/introducing-system-one-models-and-jev) - [Jev notes (Simon Willison)](https://simonwillison.net/2026/Sep/21/jev/) - [Jev: System One models for Prod, not God, with Diogo Almeida (Latent Space)](https://www.latent.space/p/jev) - [Jev's Architecture Unmasked (Archer Hume)](https://archerhume.com/posts/jevs-architecture-unmasked/) - [How to use Jev AI (Linas Beliūnas)](https://linas.substack.com/p/how-to-use-jev-ai) - [Open-source Jev alternatives (Apidog)](https://apidog.com/blog/openjev-open-source-jev-alternatives/) - [awesome-jev](https://github.com/Amal-David/awesome-jev) - [Jev is HERE. How to use it (Greg Isenberg with Ryan Vogel, YouTube)](https://www.youtube.com/watch?v=4mTLpuQpB80): input plus output schema, a probability per choice in ~200 ms; demo sorts 1,700 emails for $0.18; lead scoring, support routing, video clipping, browser control - HN: [OpenJev](https://news.ycombinator.com/item?id=49752041), [Jev: a new shape of LLM](https://news.ycombinator.com/item?id=49796843), [awesome-jev](https://news.ycombinator.com/item?id=49802160) - X: [Simon Willison](https://x.com/simonw/status/2102175146740232238), [Nate B. Jones' build patterns](https://x.com/natebjones/status/2101421649774236145), [Elvis Saravia's beginner's guide](https://x.com/omarsar0/status/2101774405521301681), [Alex Volkov on fast-jev-compaction](https://x.com/altryne/status/2100739055923425589), [Jack Cheng](https://x.com/jackcheng/status/2100729670991802386), [Dan Shipper](https://x.com/danshipper/status/2099947471518474522), [real-time prompt classification](https://x.com/dWeaths/status/2102415625301717065) - LinkedIn: [Antonio Gulli on the open reimplementations](https://www.linkedin.com/posts/searchguy_two-years-in-stealth-to-build-a-new-ai-paradigm-share-7508035614655635456-x10G), [Clem Delangue on Laya trending on Hugging Face](https://www.linkedin.com/posts/clementdelangue_the-number-one-trending-model-on-hf-is-an-share-7507906261116542977-NcNP) - [Jev 1.13 jaggedness (TypeSafe docs)](https://docs.typesafe.ai/model-jaggedness/jev-1.13) - [AI primer (TypeSafe docs)](https://docs.typesafe.ai/introduction/machine-learning-primer) - [TypeSafe Jev: typed decision model, calibration, decomposition (Rajesh Beri)](https://www.beri.net/article/typesafe-jev-typed-decision-model-calibration-decomposition-shadow-eval) - [Jev: the language model that won't talk (Anthony Maio)](https://anthonymaio.substack.com/p/jev-the-language-model-that-wont) - [What everyone is getting wrong about TypeSafe AI's Jev (KDnuggets)](https://www.kdnuggets.com/what-everyone-is-getting-wrong-about-typesafe-ais-jev) - [TypeSafe AI debuts model for machines that plays Doom (The Register)](https://www.theregister.com/ai-and-ml/2026/09/16/typesafe-ai-debuts-model-for-machines-that-plays-doom/5296711) - [What is Jev and RLCD (Turing Post)](https://www.turingpost.com/p/what-is-jev-rlcd) - [What is RLCD, the secret behind Jev? (Di Zhang)](https://di-zhang-llm.github.io/blog/what-is-rlcd-the-secret-behind-jev/) ## Related - [[Decision Models (DMs)]] - [[System One Models]] - [[TypeSafe AI]] - [[SemIf]] - [[Kev]] - [[Jevlike]] - [[Laya]] - [[fast-jev-compaction]] - [[LLM Structured Outputs]] - [[AI Model Selection]] - [[Large Language Models (LLMs)]] - [[System One Primitives]] - [[Reinforcement Learning for Calibrated Decisions (RLCD)]] - [[AI Model Calibration]] - [[Atomic Question Decomposition]] - [[Relative vs Absolute AI Judgments]] - [[Self-Consistency]] - [[Machine Native Intelligence]] - [[Jevons Paradox]] - [[System 1 and System 2 Thinking]] - [[Diogo Almeida]] - [[Jagged Intelligence]]