Independent playground · no signup
Jev AI Playground — Try a Typed Decision
Jev is TypeSafe AI’s System One decision model. Hand it a state and a typed question — it hands your code a decision with probabilities. No prose. No parsing.
state + question + choices → decision + probabilities
Try a decision, right here
Interactive DemoIllustrative values, not real API calls — the flow and output shape mirror the real thing. For live calls, get a key and follow the API guide.
Pick a preset (or fill in your own state and question), then run.
How to read the results
Choice answers come with a probability bar per option — they always sum to 100%. The winner is the tallest bar, but the number to act on is confidence: a high confidence means the model is sure of its read of the state, while a low one is your signal to route to a human instead of trusting the pick. In production that single field is the difference between “auto-handle it” and “flag it for review”.
Score answers sit on the scale you defined — a 1.4 on a 0–2 rubric means “between upset and very upset, leaning upset”. The distribution over your levels shows how firm that placement is, and the legend is echoed back in the answer so the result stays auditable after the fact.
Noul answers are a single 0–1 probability with no separate confidence field — the probability is the answer. The practical pattern is to threshold it in code: 0.7 and above treat as yes, 0.3 and below treat as no, and everything in between is a gray zone that goes to a human or a second check. Where you set the cutoffs should follow the cost of being wrong in each direction.
The four presets, in real terms
Route a support ticket. This is the classic two-queues-touch-it case: a double charge (billing) locked behind a broken login (account). Watch which team wins and with what confidence — ambiguous tickets are exactly what the confidence field exists for. Edit the ticket text toward one problem and the answer moves with it.
Flag an urgent message. An ops inbox triage: a failing webhook plus a rising error rate should push the urgency reading up. Try rewriting the message vague and hedged — “things might be a bit slow lately” — and watch the same judgment calm down. Direct phrasing is what these models read best.
Score customer frustration. A retention play: score inbound email before a human replies, and let a 2.0 on “very upset” jump the queue. The rubric is yours to change in the Levels editor — rename “very upset” to “escalate” and you have a prioritization signal.
Confirm a meeting time. A scheduling assistant’s smallest unit of work: does the thread support this slot? Alex can only do Thursday afternoon, the calendar is free — the reading should lean yes. Add a conflict to the state text and watch it shift. Every preset is editable, and the demo responds to whatever you change.
When a decision model beats a chat model
Anywhere a judgment has a small, known answer range, a typed decision is cheaper, faster, and safer to consume than generated text.
Ticket routing
"Which team should own this?" is a small answer space. A decision model picks one option and tells you how sure it is — no prompt engineering, no JSON repair.
Intent classification
Label a message against your fixed set of intents and route it downstream in a single call, with probabilities you can log and monitor.
Grading against a rubric
Score a review, an email, or a lead against your own levels — and get a distribution over the rubric, not vibes.
Relevance filtering
"Is this urgent?" "Is this a refund request?" Yes/no with a probability you can threshold in code and route to automation or a human.
What is Jev?
Jev is the first publicly released System One Model, built by TypeSafe AI in San Francisco and launched in late September 2026. It is not a chat model. You send a state — any material worth judging — plus typed questions, and it returns decisions your code can branch on directly: a choice with probabilities, a score against your rubric, or a yes/no probability.
There is no text to parse. Decisions come back typed and calibrated, fast and cheap enough to sit inside product loops — $0.042 per million input tokens, output free (see the full pricing math). It speaks plain HTTP, with official Python and JavaScript SDKs (get started in five minutes).
Frequently asked questions
What is Jev AI?
Jev is a decision model from TypeSafe AI, and the first publicly released "System One Model". You send it a state — any material worth judging — plus typed questions, and it returns decisions with probabilities instead of generated text.
Who makes Jev?
TypeSafe AI, a San Francisco company founded in 2024, released Jev in late September 2026. This site is an independent resource and is not the vendor.
What is a System One Model?
The name comes from Kahneman’s "System 1" — fast, focused judgment. A System One Model is built to return typed answers your code can branch on directly: a choice with probabilities, a score against your rubric, or a yes/no probability. No text generation, no parsing.
Is Jev an LLM?
No. Chat LLMs generate text for humans; Jev returns typed decisions for software. It does not write replies, generate code, or explain its reasoning — it decides, with calibrated probabilities.
What can Jev be used for?
Classification, routing, relevance filtering, grading against a rubric — any product step where a semantic judgment has a small, well-defined answer range. Typical jobs: "which team handles this ticket", "how frustrated is this customer", "is this message urgent".
Is there a Jev API?
Yes. You POST a state plus typed questions to api.typesafe.ai/v1/systemone with a Bearer key from the TypeSafe console. Our API guide walks through it with copy-paste examples.
Can I try Jev online for free?
Yes — the playground on this page runs the same input/output flow in your browser, no signup. The values shown are illustrative demos; for live calls you need your own API key from the TypeSafe console.
Is this site affiliated with TypeSafe AI?
No. JevAI Online is an independent resource, not affiliated with, endorsed by, or sponsored by TypeSafe AI. Values shown in the playground demo are illustrative examples, not real API calls.