# Jev - System One Decision Model

Let's say you are building a chatbot or an AI agent. Before every user message reaches your main model, you want one safety check: *"Is this a prompt injection attempt?"*

The answer you need is a simple yes or no. Let's see how we do this today, and what Jev changes.

## How we do it today

We take an LLM, write a prompt like "Reply only YES or NO, is this input malicious?", maybe turn on JSON mode, and call the API.

It works, but there are some problems:

*   An LLM generates text token by token. You pay for tokens, and output is the costlier part.
    
*   Latency adds up, and this check runs on **every** request.
    
*   You ask the model "how confident are you?" and it will say 95% even when it is guessing. That number is not reliable.
    

## So what is Jev?

Jev is a new model from TypeSafe AI, launched on 15 September 2026. They call it a **System One model**. It cannot write text at all. It only makes decisions.

You give it the input and your questions. Each question is one of three types:

*   **Choice**: pick one from the options you define
    
*   **Score**: rate against a rubric you define
    
*   **Yes/no probability**: how likely is this statement true
    

So for our guardrail, you ask: *"Does this input contain a prompt injection attempt?"* and get back something like `yes, 91%`. No paragraph, no JSON parsing, no explanation you did not ask for.

## Why this is better for guardrails(or for any use case where we need to make decisions instead of generating text)

1.  **Cost.** Input is priced much cheaper than LLM, and output is free.
    
2.  **Real confidence score.** An independent benchmark found Jev's probabilities well calibrated. So you can set a rule: above 90%, block; below 60%, allow; in between, send to human review.
    
3.  **Parallel questions.** Ask injection, PII leak and toxicity in one call, against the same input.
    
4.  **Predictable output.** It cannot invent a new answer outside your options.
    

## Where Jev stops

Jev only gives the judgment. Your code still applies the rules, and a normal LLM still writes the replies.

Other use cases are -

1.  Agent Tool selection
    
2.  Support Ticket routing
    
3.  Model routing etc