Abed ManzaljiAI, Step by Step

How does Jev work?

Jev is an AI that fills in a form you write. For each question, it picks an answer from your list and says how sure it is. It never writes text.

4 min to watch4 min to read

Based on TypeSafe's own docs and launch post for Jev, which opened in early access on September 15, 2026.1,2 The plant shop, its customers, their messages and every percentage in the walkthrough are made up. Jev's speed, price and "can't hallucinate" claims are TypeSafe's own, and weren't tested for this walkthrough.14

Read the walkthrough as text

With a chat AI

  1. A small plant shop. Overnight, its support inbox fills up.
  2. Each message must go to one team: returns, shipping or billing. A small sorting script reads this form.
  3. The owner asks a chat AI, the kind you talk to, to fill it in. First: Mia's broken pot.
  4. It answers like in a chat: a friendly paragraph, not one word.
  5. The script digs a team name out of the paragraph. It finds two, and keeps the first.
  6. Next: Raj's order is late. That's a shipping problem.
  7. The AI writes "not a returns problem". The script spots "returns" first, and sends it there.
  8. Raj's late order waits with Returns, where nobody handles deliveries.
  9. Tom was charged twice. The AI decides this is an "urgent complaint": a team the shop doesn't have.
  10. The script only knows three teams. It crashes, and nothing else gets sorted.
  11. Even when it works, it's slow: each reply takes seconds, and the shop pays for every word it writes.
  12. Tom's double charge sits unsorted. So does Lena's, and dozens more.

Chat AIs write words for people. Code needs one pick from a list. Asked to sort, the chat AI wrote paragraphs, the script picked the wrong team, and an invented team crashed it. TypeSafe says this mismatch is why it built Jev: a model that answers with typed values, not text.

With Jev

  1. Same inbox. This time, the owner writes a form for Jev: three questions, and the only answers allowed.
  2. Mia's broken pot again. Every field comes back from its list, each with a probability. No paragraph.
  3. The script reads the answers as they are. Raj goes to Shipping. Tom is urgent billing.
  4. Each one lands in the right queue. TypeSafe says a call like this takes 70 to 500 milliseconds.
  5. Then a messy one: Lena's plant came late and half dead, and she may have paid twice.
  6. Jev still answers from the list, but it isn't sure: only 48% for shipping.
  7. The owner's rule: under 80%, a person decides. Lena's message goes to a person, not a guess.
  8. Tom's answers, once more. This is a typed decision: an answer picked from a list you wrote, plus how sure the model is.

Pick from the list. Say how sure. Let a person take the rest. Jev only answers from the options you allow, each with a probability. It can still pick the wrong one, so draw a line: below it, a person decides. Its speed and price figures are TypeSafe's own claims.

Keep going:

Why it matters, what it costs, and what to ask your team.

What it is

An AI that fills in your form, and never writes a word.

The AIs most people know, like chatbots, write text for people to read. To use that text in code, the code has to dig the answer out of it.1

Jev, from a company called TypeSafe, works the other way round. You send the text plus your questions, each with the answers you allow. Jev sends back one pick per question, and a probability for every option.1,5

A customer writes "Charged twice!!" Your form Team? Urgency?Refund? Chat AIJev "Thanks for reaching out!It sounds like an urgentcomplaint. I'd suggest…" ✗ no team called that team: billing · 99%urgency: high · 94%refund: yes · 98% ✓ the code reads it as is

A sketch to show the idea. The message and percentages are made up.

It takes three kinds of question:

  1. Pick one from a listWhich team, which category, which product. Up to 255 options. TypeSafe calls this a Choice.5
  2. Place it on a scaleHow urgent, how angry, how risky, on levels you describe. A Score.1
  3. Yes or noDoes this ask for a refund? The answer is the chance of yes, from 0 to 1. A Noul.6

All the questions go in one request and are answered side by side. Jev doesn't write replies, code or explanations. TypeSafe calls it a "System One" model, after the fast, gut-feel thinking in Daniel Kahneman's book Thinking, Fast and Slow.1,4

I can't write you a single sentence. I can tell you it's 98% a refund.

Why it matters

Fast and cheap, TypeSafe says.

TypeSafe opened Jev to developers on September 15, 2026, in early access.2 Sorting support messages is one of the jobs its own docs suggest.17

0per billion tokens sent in: TypeSafe's price for Jev. The answers back are free.8
0options at most in one pick-from-a-list question5
0public projects on GitHub used Jev by September 22, 2026, a week after it opened16

TypeSafe says a call takes 70 to 500 milliseconds, 40 to 200 times faster than leading AIs on this kind of question. Its price: $0.042 per million tokens in, and nothing for the answers out.2,8

On its own tests, TypeSafe reports 193.6 times faster and 444.6 times cheaper, and expects those to be "on the higher end of real world gains." It also says it can't prove its prices aren't subsidized.2,11

Early reviews stress that these numbers are TypeSafe's own, and should be read as launch claims until someone tests them independently.12,14

TypeSafe also says Jev "can't hallucinate" and never makes type errors.2 That holds in a narrow sense: it can't answer outside your list. It can still pick the wrong option from it.13,15

So it never makes up a team. It can still pick the wrong one.

How to use it

Draw a line. Below it, a person decides.

Every answer carries a probability. TypeSafe trains Jev so that, across many answers, the probabilities match how often it's right: 80% answers should come true about 80% of the time. That's no promise about any single answer.18,4

Here are real answers from TypeSafe's docs to one yes-or-no question: is the customer asking for a person?

Real answers from jev-1.13.0, recorded by TypeSafe in its docs, to the question "Is the customer asking for a human agent?" Jev answers from 0 to 1; shown here as percentages.6

Near 0% or 100%, the answer is clear. Near the middle, Jev is unsure. TypeSafe suggests three paths: act when it's sure, double-check in between, and hand it to a person when it isn't.6,7

Try it. Drag the line, and watch which messages go to a person:

Made-up messages and numbers to show the idea. Each bar is how likely Jev's pick of team is.

The riskier the action, the higher the line. Sorting a message can take a lower line than sending money back.7

Ask your team

Four questions before you use a model that only picks.

Start with these:

  1. Is the answer one pick from a list?Sorting, scoring and yes-or-no checks fit. Writing a reply or explaining why doesn't: Jev gives "a number, not a rationale."10,12
  2. Where is our line?Below it, a person decides. Set it higher for actions that cost money or are hard to undo.7
  3. Does the question need math, dates or counting?TypeSafe says Jev struggles with those. Do them in code, and ask Jev only the judgment.9
  4. Have we tested it on our own messages?The speed, price and accuracy figures are TypeSafe's. Check the line on your own data before you trust it.7,14

Under the hood, step by step: build a support-ticket sorter that picks from a list, and asks a person when it's unsure.

Build along

Build a support-ticket sorter that picks from a list, one step at a time

We start with a chat AI and a hopeful guess, run it on the same 200 made-up messages, and fix one thing per step. Press Run it after each step and watch the result change.

Before you ship

Six habits for any typed decision.

Whatever you sort or score, these keep it honest:

  1. Add an "other" optionSo a message that fits nothing can say so.5
  2. One judgment per questionAsk about the team and the urgency separately, then combine them in code.1
  3. Keep math and dates in codeJev reads dates as text, is weaker with numbers, and doesn't count reliably.9
  4. Send only what it needsExtra detail costs accuracy.9
  5. One line per action, then test itStart strict. Adjust on your own messages.7
  6. Pin the version"jev-latest" moves when a new model ships. Pin the version your lines were tuned on.8

References

  1. "Introduction" · TypeSafe docs. What Jev is: state and typed questions in, typed answers with probabilities out; the three question types; questions answered in parallel.
  2. "Introducing System One Models & Jev" · Diogo Almeida, TypeSafe's founder, TypeSafe AI blog, September 15, 2026. Early access; TypeSafe's claims on speed (70–500 ms, 40–200× faster), price, "can't hallucinate" and "never makes type errors", and its own caveats.
  3. "Quick start" · TypeSafe docs. The request and response shapes, and the Python SDK the build-along follows.
  4. "System One" · TypeSafe docs. No replies, code or explanations; calibration doesn't guarantee any single answer; where the name comes from.
  5. "Choice" · TypeSafe docs. One option from a set, a probability for every option, up to 255 options, and an "other" option.
  6. "Noul" · TypeSafe docs. Yes-or-no questions; the six recorded jev-1.13.0 answers to "Is the customer asking for a human agent?"; middle values can go to a person.
  7. "Confidence" · TypeSafe docs. Three paths (act, double-check, hand to a person), thresholds that scale with risk, and the top probability as a simple measure.
  8. "Models" · TypeSafe docs. Jev 1.13: $42 per billion input tokens, output free; aliases like jev-latest, and pinning a version.
  9. "Jev 1.13 jaggedness" · TypeSafe docs, last reviewed October 2, 2026. Where Jev struggles: literal reading, math and counting, dates, large noisy input, text written to steer it, and generation.
  10. "Jev with coding agents" · TypeSafe docs. Jev doesn't write text or code; when to reach for it.
  11. TypeSafe AI home page · TypeSafe, fetched October 6, 2026. "193.6x Faster, 444.6x Cheaper" on its own workflows, "Zero Hallucinations", and $42 per billion input tokens.
  12. "Jev: TypeSafe's System One Model Explained" · Matt Crabtree, DataCamp, September 16, 2026. The numbers are vendor-reported; "a number, not a rationale".
  13. "TypeSafe AI's Jev and 'System One Models': What Actually Shipped" · Ashish Dubey, TrueFoundry, September 18, 2026. A model held to allowed answers "can still confidently pick the wrong one".
  14. "Jev by TypeSafe: The Non-Hallucinating Decision Model for AI Agents" · Fredrik Falk, Beam AI, October 6, 2026. Read the speed and cost figures "as launch claims until independently tested".
  15. "What Is Jev? Inside TypeSafe's Decision-Only AI Model and Its Developer Use Cases" · Hiba Fathima, Firecrawl, updated September 23, 2026. The 0% hallucination figure counts answers outside the schema, not wrong picks.
  16. "Jev in the Wild: A Data-Driven Analysis of the Jev Model's Functionality, Applications and Ecosystem" · Guoming Ling, Muen Xue and Zijian Ye, arXiv 2609.30216, September 24, 2026. 2,170 public Jev projects on GitHub as of September 22, 2026.
  17. "Example use cases" · TypeSafe docs. Customer support: sort tickets, spot urgency and refund requests, send cases to the right team.
  18. "AI primer" · TypeSafe docs. Calibrated decisions: across many answers, 0.8 should come true about 80% of the time.

The shop, its customers, messages, teams and every percentage in the walkthrough and the widgets are made up, except the six answers to "Is the customer asking for a human agent?", which TypeSafe recorded and published in its docs. The chat AI's replies are made up to show common ways free-text answers break sorting code. Jev's speed, price and accuracy figures, and its "can't hallucinate" and "never makes type errors" claims, are TypeSafe's own and weren't tested for this walkthrough. The build-along follows the request shape in TypeSafe's Python quick start; chat_ai() is a placeholder, and the run outputs are made-up numbers to show the idea.

Made by Abed Manzalji, AI Engineer in Toronto.More about meContact