Jev can only choose. I made it write anyway.

Question
Try one
Jev says

The answer appears here, one word at a time.

Jev’s odds for this word, out of 1,000 choices

Jev is an evaluation model from TypeSafe. You give it a question and the possible answers, and it returns probabilities for those answers. You can try that in Ask-Jev. It cannot generate an answer of its own.

Gemini writes by choosing the next token, then choosing again with that token added to the answer. I wondered how far I could get by giving Jev words to choose from and putting that loop in my code.

Anyway, I made it write

The first problem was the vocabulary. Jev takes at most 255 options, and English has hundreds of thousands of words.

So I took a page whole book out of xkcd’s book. In Thing Explainer, Randall Munroe explains rockets, phones and the human body using only the thousand ten hundred most common English words (“thousand” isn’t one of them). Those words became Jev-GPT’s entire vocabulary.

It’s a tight list. “Boss” isn’t on it. Neither is “atmosphere,” so Jev has to explain the sky with “air” and “light.”

One word at a time

Say you ask “Why is the sky blue?” and Jev has written “The sky is” so far. To choose the next word, I give it options like:

The sky is blue
The sky is very
The sky is full
The sky is not

Each option contains the answer so far plus one candidate word. I ask Jev which one best begins a correct answer, append the winning word, and repeat.

A thousand words don’t fit

Jev takes 255 options and the list has a thousand. So the words play a tournament.

  1. Shuffle the vocabulary into four groups of 250 and send all four groups to Jev.
  2. Keep up to 25 candidates with nonzero probability from each group. These go to a final round with punctuation and an option to stop.
  3. The winner becomes the next word.

That is five calls for every word: four groups, then the final. Five calls to write “the.” A good word can lose because of the group it landed in.

Knowing when to stop

“Stop” is one option in the final round, next to a hundred words. It almost never won.

So when a sentence ends, Jev also gets a separate yes-or-no question: “Is the answer complete?” It runs alongside the final round, so there is a sixth call, but no third sequential round. The step still waits for both calls. At 50% or more, the answer ends. There are also limits of three sentences, 280 characters, and 80 generation steps.

In the demo above, click any word in Jev’s answer to see what else it weighed.

Jev against a language model

Under every answer, the demo shows Gemini 2.5 Flash-Lite. Same question, same brief, same ten hundred words. I picked it as a cheap, fast baseline.

In my runs Gemini took under a second for a whole answer. Jev took about half a minute, because every word waits for two rounds of calls: four groups at once, then the final. That measures my contraption, not Jev. Nobody should build this for real.

Gemini sometimes leaves the list, which the demo marks in red. Jev-GPT cannot: my code only offers words on the list. Gemini gets the vocabulary as an instruction. A structured-output schema could constrain its word choices too; this demo does not test that version.

What I learned

If you want prose, use a language model. Two things from this build apply to real software:

How Jev handles decisions, and what still needs testing against Gemini, is in Ask-Jev.