Jev fills in forms.

1Hand it a form

Fernhill Cycles is a made-up bike shop. This message has just arrived. Four questions need answering before anyone can deal with it. Press the button and Jev answers all four.

That is the whole product. Text goes in with a list of questions. Typed answers come out, each with a number saying how sure it is. There is nowhere on the form for it to write anything else.

2Why a form and not a chat

Ask a chatbot the same four questions and you get paragraphs. A person can read paragraphs. A program cannot, so someone has to write more code to dig the answer out, and that code breaks when the wording changes.

Jev hands back values a program can use directly: refunds, 3.99 out of 4, 0.99. The rest of your code stays ordinary code. It reads the ticked box and gets on with it.

You can force a chatbot to reply in a fixed format too, and with some effort you can get rough probabilities out of one. The honest differences are three. Jev answers in a fraction of a second. It costs almost nothing, as section 14 shows. And the probability is the main thing it hands back, for every option, without you bolting anything on.

The makers say it cannot hallucinate. That is true in a narrow way: it cannot invent a sentence because it cannot write one. It can still tick the wrong box. Part three is about that.

3The three kinds of box

Every question on a Jev form is one of three kinds. These are the real questions and answers from the form above.

4What you staple to the form

Jev knows nothing about your business. It only sees what you attach. TypeSafe, the company behind it, calls the attachment the state. Same question each time here: is this customer entitled to a full refund under the shop's policy?

With only the angry message it shrugs: 0.56, close to a coin toss, because it has not been told the rules. Staple on the policy and the order date and it commits. Change one number in the order and the answer flips. The model did not get cleverer. You gave it what it needed.

5Many questions, one trip

All the questions on a form are answered at the same time, not one after another. Nine questions go out and come back in the same trip. Pick a message and Jev reads the room.

The questions cannot see each other's answers. If question two depends on the answer to question one, that is two trips.

6Acting on the number

The number is the useful part. You pick a line. Above it, the computer acts by itself. Below the opposite line, it moves on. Anything in between goes to a person.

Here are 24 made-up messages to the bike shop. Each was labelled by hand first, then sent to Jev alone with one question: Move the line.

At 90% there is one mistake, and it is a good one to stare at: "Do you refund the deposit if I change my mind?" is a question about the policy, not a request for money. Jev gave it 0.93. Raise the line to 95% and the mistake goes to a person instead, and so does one real request. That trade is yours to set, and you should set it using your own messages, not these.

7Noughts and crosses

You are X. Jev is O. Each turn it gets one question, "which empty cell should O play?", with the empty cells as the boxes. The percentages show how it spread its bet across the cells before it moved.

This game runs with no server. Jev's answer to every position it could ever face was recorded in advance: positions, 75 seconds of recording, in total.

8The Halberd Gallery job

A text game with one guard and one painting. You say what you try. Jev answers six questions about it: is that even possible, does the guard notice, how suspicious is he, what does he do, have you got the painting, are you outside. Ordinary code keeps score: each move adds to Desmond's alert level, and if he raises the alarm or the level reaches 100, you are caught. Pick an opening move.

9Jev paints

Jev has never seen a picture and cannot draw. So ask it about every pixel separately: "what colour is the pixel at column 4, row 7?", twelve colours to choose from. Every pixel goes on one form, up to 256 questions in a single trip.

It works better than it has any right to. It is also a useless way to make pictures. The point is the shape: hundreds of small judgments, all at once, for a fraction of a penny.

10It cannot think ahead

Jev reacts. It does not plan. Noughts and crosses makes that measurable, because the perfect move in every position is known. I marked all of its recorded moves against perfect play.

A child learns to block within an afternoon. Jev leaves the winning cell open nearly a third of the time, and misses its own win more often than that. You can beat it in three moves by going straight down the left side. Try it in section 7.

This counts every position that can legally happen, including silly ones no sensible game reaches. The lesson still stands. If the right answer needs two steps of reasoning, do the reasoning in code and ask Jev only for the judgment.

11The order of a list changes the answers

I took 17 issues of my newsletter. For each one I asked Jev, in a single form, whether each of the other 16 was essentially the same story. Then I asked again with the 16 listed in a different order. Nothing else changed.

Below is issue 11. Issue 12 really is the same story, written twice. Issue 7 is about something else. Shuffle the list and watch issue 7.

Put many items on one form and the scores smear on to their neighbours. The fix is dull and it works: one item per form. That is how the 24 messages in section 6 were sent.

12What other people found

These are other people's tests, not mine. I read each one and checked the figure quoted is in it.

• Sure is not the same as right. One tester found that on the rows where Jev said it was at least 90% sure, it was right 72.2% of the time. agentjournal.dev
• The numbers are rounded. Probabilities come back to two decimal places. In one ranking test of 360 rows, 53 tied at 0.99, so you cannot sort them. jev-orderby-bench
• The makers list the weak spots themselves. Counting, comparing dates, anything numeric, and forms stuffed with irrelevant detail. Their advice is the same as mine: do that part in code. TypeSafe's own list
• A fair complaint. The launch thread on Hacker News has people asking whether this is a classifier with good marketing, and pointing out there are no public benchmarks yet. The thread

13When it is the right tool

Three things need to be true. If one is missing, use something else.

1. Nothing clever is already making the decision. If a large model is already in the conversation, it will make this judgment itself and make it well. Adding Jev gains nothing. Jev belongs inside plain programs: a scheduled job, an inbox filter, a form handler.
2. There is a lot of it, or it has to be quick. Thousands of rows, or an answer needed while someone waits. For five decisions a day, speed and price do not matter.
3. A number you can draw a line on is useful. Act above 0.9, ask a person below. If you would act the same either way, you do not need the number.

14What it costs

You pay for what you send, at $0.042 per million tokens. A token is about three quarters of a word. Answers are free. These are the real bills for everything on this page.

15Calling it yourself

It is one web address. You post the text, the questions and a model name. This is the exact request behind the form in section 1, with the key swapped for a placeholder.

Access is by waiting list at the moment. Mine took a day. Useful links: the launch post, the documentation, prices and limits, the console where keys live. It is also listed on OpenRouter.

16Words

State

Whatever you attach to the form: a message, a record, a policy. Text or JSON. Jev sees this and nothing else.

Question

One box on the form. It has instructions and, for two of the three kinds, a list of allowed answers.

Choice

Pick one from a list. Up to 255 options. Comes back with a probability for every option.

Score

Place it on a ladder of two to ten described steps. Comes back as a position on the ladder, such as 3.99 out of 4.

Noul

A yes or no question answered as one number from 0 to 1. TypeSafe's own word. 0.5 means it cannot tell.

Criteria

The allowed answers and what each one means. Writing these clearly matters more than anything else you do.

Confidence

How bunched up the probabilities are. High confidence means it bet nearly everything on one box. It does not mean the box is right.

System One

TypeSafe's name for this kind of model, borrowed from Daniel Kahneman: fast gut reaction, as opposed to slow deliberate thought.