Jev is a new AI model from Type safe designed for fast, low-cost decision-making rather than text generation. Instead of writing responses, it chooses from predefined options and returns calibrated probabilities, making it useful for quick classification tasks inside software workflows.
What makes Jev different
Most AI models, especially chatbots, generate text one token at a time. Reasoning models go further by producing longer chains of thought, which makes them closer to “system two” thinking: slow, deliberate, and suited to complex problems.
Jev is positioned as a “system one” model, borrowing from Daniel Kahneman’s Thinking, Fast and Slow. System one thinking is fast and automatic, like instantly knowing that 2 × 2 = 4. Jev is built for similar quick judgment calls in software, not for writing prose or solving complex reasoning tasks.
Example: handling a support email
For an email like:
“I was charged twice this month. Please fix it.”
Software may need to answer several quick questions:
- Is this a refund request? Yes or no.
- Which team should handle it? Billing, technical support, or sales.
- How urgent is it? Low through critical.
A traditional LLM could be asked to return structured JSON, but its confidence score may not reflect its true uncertainty. Jev instead returns probabilities directly.
How Jev works
Jev takes two kinds of input:
- State: the data being evaluated, such as the support email and recent customer charges.
- Questions: the specific decisions to make.
The questions can be different types:
- A yes/no question, such as whether it is a refund request.
- A multiple-choice question, such as which team should process it.
- A score or scale, such as urgency from low to critical.
Jev does not generate text. It returns probability values for the available answers all at once. For example:
- Refund request:
0.9, meaning a 90% probability that it is a refund request. - Team: billing might receive
0.85, with other options receiving lower probabilities. - Urgency: a score near high or critical.
Because the answer must come from the provided choices, Jev is constrained to the options given by the developer.
Why calibrated probabilities matter
Jev is trained to produce calibrated decisions. A calibrated model’s confidence should match reality: when it says something has an 80% probability, it should be correct about 80% of the time.
This allows developers to set thresholds in code. For example:
- If refund probability is above
0.9, send it directly to the refund queue. - If it falls into an uncertain middle range, route it to a human.
- If it is below
0.1, treat it as not being a refund request.
The higher the cost of a mistake, the higher the threshold can be set.
Training approach
Traditional language models are usually pre-trained to predict the next token, then post-trained to behave more usefully as chatbots.
Common post-training methods include:
- RLHF: reinforcement learning from human feedback, where human preferences shape model behavior. This can make models sound confident even when wrong.
- RLVR: reinforcement learning with verifiable rewards, where answers are judged automatically, such as whether code passes tests or a math answer is correct.
Jev uses a different method called RLCD, or reinforcement learning for calibrated decisions. It is rewarded not just for picking the right option, but for assigning probabilities that accurately reflect how likely each answer is to be correct.
Use cases and limitations
Jev could be useful for:
- Routing support tickets.
- Classifying customer messages.
- Acting as a guardrail around chatbots.
- Detecting jailbreak attempts.
- Making fast decisions across database rows or log lines.
However, it has limits:
- It currently only takes text input.
- It is not strong at math or counting.
- Like other AI systems, it can be tricked by malicious or hidden instructions in the data it reads.
How Jev fits with LLMs
Jev is not meant to replace large language models. Instead, it can work alongside them.
A workflow might use:
- Jev to quickly classify an incoming support email.
- An LLM to write a thoughtful customer response.
- Jev again to classify the customer’s reply.
In this setup, Jev handles fast “system one” judgments, while an LLM handles slower “system two” reasoning and writing.
Name origin
Jev is named after William Stanley Jevons, the economist associated with Jevons’ paradox. In 1865, Jevons observed that as steam engines became more efficient, Britain used more coal, not less.
The implication is that if fast judgment models become cheap enough, they may be used far more widely than LLMs are today.