What is a decision model like Jev?

Choosing instead of writing

A chat model writes an answer word by word, and the answer can be anything. A decision model works more like a multiple-choice test. You give it some text, a question and the allowed answers, and it picks one.

TypeSafe AI launched Jev in September 2026 and calls it a “System One” model, after the fast, intuitive thinking that the psychologist Daniel Kahneman described. Jev can choose from up to 255 options, rate something on a scale, or give the chance that a statement is true.

What is inside is not public

TypeSafe says Jev uses a new design that returns all the answers to one request together. It has not published the model’s size, how it was built or a technical report, so nobody outside the company can check how it works inside.

Its rivals are more open. Amazon’s Strands Decider 2B and Cloudflare’s Clef are free to download. Amazon built its model on an existing small language model and swapped the part that writes text for a part that scores each option.

Can you trust the percentage?

A decision model might say “Billing: 94%”. That number is only useful if it is honest: of all the times it says about 90%, it should be right about 9 times in 10. Experts call this calibration.

Jev also gives a confidence: how sure it says it is. TypeSafe itself warns that high confidence does not mean the answer is right. Its documents also list weak spots, such as counting, dates and a habit of picking the first option too often.

So test it on your own examples first. Then set a threshold: let it act alone only above the confidence where your tests show it is usually right, and send the rest to a person. The diagram above lets you try this.

An old idea, a new product

Picking a label is classification, one of the oldest jobs in machine learning. BERT (2018) showed that a pre-trained language model can learn it with just one extra output layer.

Decision models package this for AI agents and apps. You send text and questions of a set type, and the model returns probabilities, never free text. Jev takes up to 64,000 tokens per request and charges $0.042 per million input tokens, with output free.

Three models, three approaches

TypeSafe says Jev uses a new architecture and a “parallel sampler” that returns all answers to one request together. It has published no size, base model, technical report or calibration metric.

A model’s head is its last part, which turns its numbers into an output. Amazon’s Strands Decider 2B is built on Qwen3.5-2B, with the text-writing head swapped for a small “pointer head” that scores each option.

Cloudflare’s Clef (27 billion parameters) and Clef-flash (9 billion) start from Qwen base models. They add a small trained head and adapters, small add-on layers, and also read images.

Both rivals publish their weights under the Apache 2.0 licence. Their speed and accuracy figures are the makers’ own.

Calibration and thresholds

Calibration means the stated confidence matches reality: of 100 answers given at 0.8, about 80 should be right. Guo and others (2017) found that modern deep networks tend to be overconfident, and that a simple fix, temperature scaling, repairs much of it.

TypeSafe’s documents say Jev’s confidence shows how spread out its probabilities are, not whether the answer is correct. Use the confidence, once tested on your own data, to set a threshold: act alone above it, and send the rest to a person or a larger model. To set it, run a few hundred of your own labelled examples and find the confidence above which the model was right often enough.

What outside tests show

On Jevals, a site that says it is unaffiliated, Jev scored 69.0 on yes-or-no medical questions, against 73.0 for Gemini 3.8 Flash. A September 2026 paper found Jev within three points of GPT-6 when the answer is in the text, but behind on maths, code and logic.

The same paper found that accepting Jev’s confident verdicts and passing the rest to a bigger model matched or slightly beat GPT-6 as a judge, at about 41% of the cost. That paper is a preprint, not yet reviewed by other experts. We found no independent test that compares Strands Decider or Clef directly with Jev.

Interfaces and internals

Jev exposes three primitives: Choice, one of up to 255 options with a probability for each plus a confidence; Score, 2 to 10 ordered levels; and Noul, the chance of yes. TypeSafe claims a new architecture, a parallel sampler and a training method it calls RLCD, but publishes no size, base model or report.

The open rivals document their designs. Strands Decider 2B replaces Qwen3.5-2B’s language-model head with a pointer head of just over one million parameters that scores each option.

Clef and Clef-flash train a small head and adapters on Qwen3.8-27B and Qwen3.5-9B, with a Brier loss and a reinforcement learning step that Cloudflare also calls RLCD. Neither source makes clear whether it is the same method as TypeSafe’s.

Calibration and scoring rules

Guo and others (2017) showed that modern deep networks tend to be overconfident, and that temperature scaling fixes much of this without changing accuracy. The Brier score (1950) grades probability forecasts by squared error, so 0 is perfect. It mixes calibration with accuracy, so check calibration on its own too, for example with a reliability chart.

AWS reports its own JevBench run for Strands Decider: 167 of 231 right, or 72.3%, with a calibration error of 0.050. TypeSafe reports no calibration metric for Jev, and its documents say that confidence measures spread, not correctness.

Generative models can decide too

A large model can be made to choose with structured outputs, which force the answer into a JSON schema, including fixed lists. Constrained decoding allows, at each step, only tokens that keep the output valid.

That fixes the shape, not the truth: OpenAI warns that structured results can still contain mistakes. Token probabilities can serve as confidence, but they are not correctness either.

Evidence and cascades

Cloudflare’s own tables show the spread. On BANKING77, a test of sorting banking questions into 77 types, Clef scores 94.2 and Jev 79.7; on GPQA Diamond, a reasoning test, Jev 78.3 and Clef 48.0. Cloudflare also claims a median of 524.1 ms for Jev against 38.8 ms for Clef-flash, without describing its method.

Li and others (2026) found Jev within three points of GPT-6 when the answer is in the text, but behind on maths, code and logic. In this preprint, accepting confident verdicts and escalating the rest matched or slightly beat GPT-6 as a judge, at about 41% of the cost. RouteLLM (2024) cut costs by over 2 times in some tests by routing between a strong and a weak model.

TypeSafe lists Jev’s own weak spots: literal reading, counting and dates, double negatives, long irrelevant text, injected instructions and a habit of picking the first option too often.