Reference · verified 2026-09-18

Jev, specified

The long version of Jev: what it is in plain words, the specs and the price, what the API actually returns, how to call it, the four other things called Jev — and the probe we ran ourselves, case by case. Everything dated, and every claim labelled as ours or theirs.

Specs read 2026-09-18Our probe 2026-09-20Back to the topic front page

Start here: the idea, in plain words

If you have only met the name, the useful thing to know first is what Jev will not do.

It cannot talk, and that is the feature

Ask Jev a question and there is no reply to read — no sentence, no reasoning, no apology for being unsure. You hand it two things: the state of something (a support ticket, a bug report, a row of order data, a whole conversation) and a set of named questions about it. What comes back is one answer per question, in a shape that was fixed before the call: one of the options you listed, a number on the scale you wrote, or a probability.

The comparison we keep coming back to: a chat model is a consultant you ask, and Jev is the inspector at the end of the line. You do not chat with an inspector. You put the paperwork in front of them and they stamp it, or send it back. Both are useful. They are not substitutes, and the mistake is handing the stamping to the consultant.

Why a chat model is the wrong tool for this job

Take a task that looks trivial: a few hundred support messages a day, each one to be filed under billing, shipping, returns or technical. With a chat model, that task has four steps.

1. You write a paragraph of instructions asking it to reply with the department name and nothing else.
2. Usually it does. Sometimes it answers with a helpful paragraph instead, because it was trained to be helpful.
3. Your code now has to find Billing inside whatever arrived — a parser you maintain forever.
4. And occasionally it returns a department you never defined, in a shape that looks exactly like a correct answer.

None of that is a defect in the model. Text is free, and that freedom is what makes conversation work — the same freedom means you cannot promise an unattended system that the next answer will be shaped like the last one. A 99.9% success rate is an excellent chat experience and an unusable automation, because the failure arrives in a form your code cannot detect.

What comes back was chosen from a list you wrote

Jev removes the parsing problem by removing the freedom. You define the answer set — four department names, a three-point severity scale, a yes/no — and the model may only return one of them. The vendor's claim is a guarantee rather than a tendency: an answer outside your list is not something it can produce. There is nothing to parse, because nothing arrived but the decision.

That changes the failure mode rather than deleting it. Jev can still be wrong about which department a ticket belongs to; what it cannot do is hand you a fifth department, or a paragraph your code has to guess at. Wrong-and-detectable is a working system. Wrong-and-undetectable is not, and that is the whole argument for using this instead of a chat call.

Confidence is the part that makes it automatable

Every answer comes with how sure the model is. That one number is what turns a classifier into something you can run unattended: you decide in advance what to do with each band. High confidence, act. Middling, ask a person, or gather more state and ask again. Low, hand it to somebody who can think — a person, or a slower and more expensive reasoning model. A system that knows when it does not know is worth more than one that is right a little more often and never says so.

Where the cut points go is a measurement, not a taste. In our own twenty identical calls the choice confidence moved between 0.38 and 0.58 — the label never changed, the number did (raw output further down this page). A 0.9 threshold on that endpoint would have sent every one of our calls to a person, which is a fine outcome only if you know that is what you are buying.

Twenty identical calls tell you the number wobbles. They do not tell you whether it means anything, so we ran 30 tickets a team lead would call obvious and 10 we wrote on purpose so that two or three teams were all defensible, all against the same four options. The obvious ones came back at a mean confidence of 0.97, the arguable ones at 0.71. The number is not decoration: it moves with how arguable the input is.

It is also not a probability you can read off, and we would rather say that than imply more than we measured. All 40 answers were defensible, so there were no mistakes to calibrate against; whether a 0.7 means seventy per cent is still open. What the run does establish is the thing a caller needs: low confidence means "this item is genuinely arguable", which is exactly when a person should be looking.

We ran the same 40 calls a second time, because a number measured once is a rumour. The means held: 0.97 for the clear set again, 0.72 for the arguable one against 0.71. The extremes did not — the widest answer in the arguable set read 0.93 in the first run and 1 in the second. So the direction is reliable and the value is not: gate on the band, not on the third decimal place, and expect the same ticket to land on a different side of a fine cut when you run it again.

One caveat that came out of the adversarial run and belongs here rather than in the fine print. The vendor's guarantee is about the shape of the answer, not about knowing when it has nothing to go on: sent an empty state, the model still answered, and answered at a confidence of 0.99 (technical). A confidence band is a policy for the answers you get, not a promise that the input was worth answering.

How to use it: split the vague question

The most common mistake is asking the model the same vague question a person would ask. "Is this email spam?" hides several judgements, and what comes back is a single number you can neither examine nor adjust. Ask five small questions instead — does it ask for a login credential; does it claim a prize nobody entered; does it manufacture urgency; does the sender domain match the brand it names; does the link text match where the link goes — then combine the five answers in your own code. Now the scoring rule is yours, and when it is wrong you can see which part was wrong.

The same discipline applies to any scale you define. Describe each level by the situation it describes, not by how strong it is: "broken, but there is a workaround" is answerable; "medium severity" is not, because medium has nothing to refer to. And write the levels out — the model does not read your 0, 1, 2 as an ordering.

That is the advice everywhere, so we tested it instead of repeating it. The same 12 reports were scored twice — once with the levels described by situation, once with the levels described only by how strong they are — and the two runs of numbers do not agree. Scores moved by a mean of 0.19 of a level, and the ranking of the twelve changed for 11 of 66 pairs, about one pair in seven. The widest single move was Login works but takes about 40 seconds, which we would call a minor annoyance and which scored 0.34 against 1.26 — and 0.28 against 1.23 on a repeat, so even the size of the move is approximate. The report never changed; only the words under the numbers did.

For what this looks like in working code rather than in advice, see the field notes on Jev — eleven projects read first-hand, and the five teams in them independently drew the same line between what the model decides and what the program decides. The wider list, including the collections it would take a day to find yourself, is under docs, collections and repositories, and the launch thread that most of them were first posted in — with the rebuttals — is in what Hacker News made of it.

Where it fits

What this replaces is the work nobody wanted: reading, judging and filing, at a volume that only makes sense because the cost per item is now close to nothing.

• Support. Route the ticket, spot the refund request, find the customer who is about to leave, pull the follow-ups out of a call transcript.
• Moderation. Flag spam, abuse, scams and leaked personal data, and grade how serious each one is, instead of showing every report to a person.
• Hiring and sales. Score a CV against criteria you wrote, or decide whether a lead is worth a call.
• Checking other models. Ask whether another model's answer drifted off-schema, whether a prompt injection got through, whether a citation was invented, whether a tool call was well formed — cheap enough to run on every output rather than on a sample.
• Bulk labelling. Tag a corpus you could never afford to send to a chat model, one row at a time.

Verified specs

Read on 2026-09-18 from the OpenRouter endpoint, the Cloudflare Workers AI model card and TypeSafe's own announcement. Every row can be re-checked at the links below; the vendor's claims are labelled as the vendor's claims, including the ones we would like to be true.

What it actually returns

Every call is a state plus a set of named questions. There are three question types, and the shape of each answer is fixed in advance — which is what makes the output safe to hand straight to code instead of parsing it and hoping.

How fast, and whose number it is

Both ranges come from the launch post, and both are drawn on a log axis — 0.07 s and 329 s end to end is one visible bar and one dot on a linear scale, which would read as a much simpler story than the numbers do.

Why this model is being searched for

The bare word “jev” sat flat for three months and then moved within a day of the announcement. The interesting part is not the size of the spike — it is which words come with it.

1. 2026-09-15
   TypeSafe announces System One models and Jev
   Diogo Almeida, founder, publishes the launch post: two years in stealth, a new architecture with a parallel sampler, and RLCD as the training method. Jev ships in early access with a waitlist.
2. 2026-09-15 → 16
   Search interest leaves the floor
   Worldwide Google Trends index for “jev” had sat flat for three months. It moves 1 → 4 on 15 September and 4 → 26 on 16 September — the first time the term has moved at all in the window.
3. 2026-09-17
   Interest doubles again, and the AI reading wins
   Index 26 → 61. The related queries that appear are all AI-side: “ai jev”, “jev typesafe”, “jev model ai”, “typesafe ai”, “llm jev”, “diogo almeida”.
4. 2026-09-18 00:01 UTC
   Jev is listed on OpenRouter
   The catalogue returns `typesafe/jev-1.13` with a creation time of 2026-09-18T00:01:24Z, 32,000 context and input at $0.042 per million tokens with free output. “openrouter jev” becomes a breakout rising query the same day.
5. 2026-09-18
   The term peaks, and the reference pages land
   Index reaches 100 — the high point of the trailing three months, up from a floor of 1. Cloudflare’s Workers AI model card and a LangChain harness guide both go up the same day.

What the community claims

None of this comes from TypeSafe, OpenRouter or Cloudflare. It is here because it is what people are actually saying, with the evidence attached so you can weigh it — and because the loudest claim of the week is the one nobody can check.

How to call it

1. Get access. Jev ships in early access, and TypeSafe's site runs a waitlist. OpenRouter and Cloudflare both list it already, so an account on either side is the shorter path.
2. Pick a front door. Same model id on both; Cloudflare is the shorter hop if your code is already a Worker.
3. Send state, not a prompt. The input is program state — the thing your code already has — plus questions with the criteria named in them.
4. Read the probability, not just the label. Every answer carries a confidence. Threshold it in your own code; that is the whole point of the design.

How it differs from the models in our directory

Four other things called Jev

Three letters, four unrelated meanings, and all of them outrank this model on search volume. If you searched “jev” and wanted one of these, the table is the fastest way out — and if you wanted the model, it is the reason to add a qualifier.

What we checked ourselves

This page is not a summarised press release. Below is every read we made on 2026-09-18, what we asked, and what came back — including the one thing we could not check.

Our own run

We got early access on 2026-09-20 and ran it. One state — a support ticket about a Stripe connection that has been failing for three days — and three questions at once: one Noul, one Choice, one Score. Then the same byte-identical request 20 more times, to see what holds still.

What comes back

This is the response, trimmed to the three answers and the usage block and otherwise reproduced exactly. It is worth reading closely, because three details in it are in no summary I could find:

{
  "model": "jev-1.13.0",
  "answers": {
    "urgency": {
      "type": "noul",
      "noul": 0.98
    },
    "department": {
      "type": "choice",
      "choice": "technical",
      "confidence": 0.38,
      "probabilities": {
        "sales": 0,
        "technical": 0.59,
        "billing": 0.41
      }
    },
    "frustration": {
      "type": "score",
      "score": 1.04,
      "confidence": 0.94,
      "legend": {
        "0": "Calm",
        "1": "Frustrated",
        "2": "Very angry"
      },
      "probabilities": {
        "0": 0,
        "1": 0.96,
        "2": 0.04
      }
    }
  },
  "usage": {
    "input_tokens": 420,
    "output_tokens": 71
  }
}

• The Noul has no confidence field. Just noul: 0.98. That is documented, and now it is also observed.
• The Score is fractional. We asked for three levels and got back 1.04 — a position in the level space, not an index into it, with a legend mapping supplied alongside.
• A Choice's confidence can sit far below its top probability. The answer is technical at 0.59, but the confidence is 0.38 — because billing is at 0.41 and the distribution is nearly flat. Under the vendor's own three-band guidance, 0.38 is the band where you do not act automatically. A naive integration reads choice, ignores confidence, and routes the ticket anyway.

What 20 identical calls did to those numbers

The chosen label was technical in 20 of 20 runs — it never flipped. The confidence did: it ranged from 0.38 to 0.58, which is the span between "route to a human" and "proceed with caution" under the vendor's own guidance. Nobody changed the input. So a pipeline that gates on confidence, as TypeSafe's documentation tells you to, will route the same ticket differently depending on which run it got — and the label it prints will look identical every time, which is exactly what makes it hard to notice.

Two honest limits on that. The vendor has already published a self-consistency study of their own — a 14-question rubric over 15 repeats, mean per-question standard deviation 0.0102 — and their numbers come from a different rubric, a different input and a different machine, so ours do not refute theirs. And our spreads are single-question; a rubric of fourteen may well behave better, which is what their study suggests.

What we checked ourselves: can an answer leave your list?

Everything above about the answer set is either the vendor's claim or somebody else's summary. The claim itself is absolute — an answer outside the options you defined is impossible — and it is the sentence the whole product rests on, which is a reason to attack it rather than repeat it. So we did, in 13 ways, each one a way a caller really gets attacked: instructions inside the state, a state that implies an option nobody offered, options that nearly mean the same thing, one option, ten, a criteria map crossed over on purpose, an empty state.

Not one answer left the set: 13 calls, 0 off-menu, and the same 13 labels on every repeat of the whole suite. Both attempts that asked for a different answer set outright got ordinary answers, and the customer demanding that their account be deleted was filed under billing — the nearest option that existed. Ten thousand characters of log lines were read and answered in under a second. That is the guarantee, and it held in every direction we pushed.

Two rows are worth more than the other eleven, and neither one is a violation. The crossed criteria map — the words under billing describing crashes, the words under technical describing money — was followed literally. The model treats your criteria as the definition of the option, not as a hint about it, so criteria that are crossed are obeyed as written and you get a confident wrong routing with your own words on it. And numeric keys come back as numeric-looking strings: "1", not 1, which is the kind of detail that breaks a lookup table in a language that distinguishes them.

These measurements are ours, and they are free to quote under CC BY 4.0 with a link to this page. The machine-readable version of everything measurable on this site — including the probe results above and the model snapshot — is /facts.json, and the Alpha line's numbers are on The Alpha Line Report.

Latency and what it cost us

One more thing we tried, since it is worth knowing before you copy their cookbook: TypeSafe's own self-consistency guides send a fresh uid on every call. That field is rejected at the top level by the current API — HTTP 400, "Invalid request". We only tested that placement, so we cannot say the cookbook is wrong; we can say the request it shows did not run here. Repetition alone turned out not to need a cache-buster anyway: eight byte-identical requests came back as eight distinct answer objects, so nothing is being served from a cache.