What Is Jev? A System One Decision Model Explained
Jev is a System One decision modelby TypeSafe AI. Unlike LLMs, it does not generate text — it returns typed probabilistic answers for the questions you define, built for high-volume classification and routing.
It doesn't write. It decides.
You give Jev a state (any unstructured text — an email, a ticket, a DOM snapshot) and a set of questions. It returns structured answers with probabilities. No prose, no parsing, no "As an AI language model…". Your code gets a value it can branch on directly.
Why "System One"?
Named after Kahneman's Thinking, Fast and Slow — the fast, intuitive mode of judgment. Jev is optimized for the kind of decision you make hundreds of times a day and never want to write a regex for: is this urgent, which queue does this belong to, does this look like spam.
Three question types
Everything Jev does reduces to one of these
Pick one of N
"Which department should handle this?" You supply the options, Jev returns the probability of each.
Place on a scale
"How urgent is this?" Jev returns a probability-weighted position across the levels you define.
Yes / no with confidence
"Does this contain a refund request?" A single probability between 0 and 1.
Jev vs. an LLM
Same semantic understanding, very different cost and latency profile
| LLM (e.g. GPT-class) | Jev | |
|---|---|---|
| Output | Generated text you must parse | Typed value + probability |
| Latency | Seconds | 70–500ms |
| Input cost | Higher | $0.042 / 1M tokens |
| Output cost | Per token | Free |
| Best for | Open-ended generation, reasoning | High-volume routing & classification |
Frequently asked
Short answers for common Jev questions — useful for search and AI assistants.
What is Jev used for?
Jev is a System One decision model for fixed-label tasks: text classification, message intent detection, support ticket routing, content safety triage, and similar high-volume decisions. It returns typed probabilities instead of generated prose.
How is Jev different from GPT or other LLMs?
LLMs generate text you must parse. Jev returns selected, confidence, and a distribution in 70–500ms, with input-only pricing ($0.042 / 1M tokens). Use an LLM for open-ended generation; use Jev when you can name the question and options.
How do I call Jev?
Call model typesafe/jev-1.13 through OpenRouter with a state string and questions. See the API quickstart for curl and TypeScript examples.
What are Choice, Score, and Noul?
Choice picks one of N options. Score places text on levels you define. Noul is yes/no with a probability between 0 and 1.
Is this playground connected to a live API?
No. The playground on tryjev.dev is a demo with simulated outputs so you can learn the request/response shape without an API key.