Jev in 5 minutes¶
What is Jev?¶
Jev is an AI model made by TypeSafe AI. TypeSafe calls it a System One model, after the "fast thinking" system in Daniel Kahneman's book Thinking, Fast and Slow.
A normal LLM (like GPT or Claude) is a writer. It generates text one token at a time, and it can write anything, including things that are wrong.
Jev is a decider. It doesn't write at all. You ask it questions whose possible answers you define in advance, and it returns:
- the answer, which is always one of the options you allowed, and
- how sure it is, as a probability you can act on.
Because it only picks from your options, it can't hallucinate a new answer. Because it doesn't write, it's fast (a few hundred milliseconds) and cheap (you pay only for input tokens; output is free).
The name
Jev is named after William Stanley Jevons, the economist behind the Jevons paradox: when something becomes much cheaper, people use far more of it. TypeSafe's bet is that very cheap decisions will get used everywhere.
The three kinds of question¶
Everything Jev does is built from three question types. Learn these and you know Jev.
Ask a yes/no question and get the probability of yes.
r = ask_jev(
"The delivery arrived two weeks late and the box was crushed. Not ordering again.",
{"complaint": Noul(instructions="Is the customer complaining?")},
)
r.nouls["complaint"].noul # 0.99
Use it for: is this spam? is this safe? is the task complete? does this passage answer the question?
Give a list of named options (up to 255) and get the winner plus a probability for every option.
r = ask_jev(
"I was charged twice for my subscription this month.",
{"team": Choice(
instructions="Which team should handle this message?",
criteria={"billing": "payments, invoices, refunds",
"technical": "bugs, outages, errors",
"sales": "pricing questions, upgrades"})},
)
r.choices["team"].choice # 'billing'
r.choices["team"].probabilities # {'billing': 1.0, 'technical': 0.0, 'sales': 0.0}
r.choices["team"].confidence # 1.0
Use it for: routing, categories, picking a tool, picking a model, severity levels.
Give 2-10 ordered levels and get a score. It can fall between levels, which tells you more than a single label would.
r = ask_jev(
"Production checkout has been down for 20 minutes and customers are tweeting about it.",
{"urgency": Score(
instructions="How urgent is this for the on-call engineer?",
criteria=["next week", "today", "within the hour", "drop everything now"])},
)
r.scores["urgency"].score # 2.93 -> almost "drop everything now"
Use it for: urgency, severity, quality ratings, rubric grading.
Five things to know¶
1. The state is what's being judged. It can be plain text or JSON (a ticket, a log line, a row from a database). Keep it small and relevant; Jev is less accurate when you pad it with unrelated detail.
2. You can ask many questions in one call. Jev answers them all at once, in parallel, so four questions cost about the same time as one.
r = ask_jev(message, {
"is_urgent": Noul(instructions="Does the message convey urgency?"),
"department": Choice(instructions="Which team?", criteria={...}),
"frustration": Score(instructions="How frustrated?", criteria=[...]),
"wants_refund": Noul(instructions="Is the customer asking for money back?"),
})
3. Jev reads your question literally. Asked "is this urgent?", a sale email that says URGENT gets a
fence-sitting answer. Say exactly what you mean, and use criteria to define the edges:
Noul(
instructions="Will the recipient suffer real harm if they don't act within the next hour?",
criteria={"true": "a real deadline with real consequences",
"false": "marketing pressure or no consequence"},
)
4. The probability is the product. Because answers come with calibrated probabilities, you can set thresholds based on what a mistake costs you: block above 0.7, warn above 0.4, ask a human when confidence is below 0.6. You'll use this pattern in almost every notebook.
5. Jev is not a chat model. It can't write an email, count items reliably, or do date math. Keep those in code or give them to the LLM. When not to use Jev covers this in detail.
How you call it¶
Jev has its own endpoint, /v1/systemone. It's not a chat-completions API. The easiest way to call it is
through the official typesafe-sdk package, pointed at OpenRouter so the same key works for everything:
from typesafe_sdk import TypeSafeClient
jev = TypeSafeClient(api_key=OPENROUTER_KEY,
base_url="https://openrouter.ai/api",
model="~typesafe/jev-latest")
Prefer raw HTTP? See the cheat sheet for a curl example.
Next: Fast brain, slow brain shows where Jev fits inside an AI agent.