About

How Jev works

Jevilize looks like a toy: you describe something, and the things that fit float up. Underneath is a different kind of language model. It never writes a word. It answers typed questions with probabilities, and it answers 149 of them in one go. This page walks through how, with real numbers from this site.

A tour in debug mode: the real UI, its numbers, and what each one means Β· 3:41
  1. A model that doesn't write
  2. Anatomy of a typed question
  3. What comes back
  4. Inside: one pass, one slot per question
  5. 149 questions, one call
  6. From probability to physics
  7. Wording is part of the question
  8. Limits, plainly

1 A model that doesn't write

A chat model answers by generating text, one token at a time. Ask it to sort 149 things and it writes a list: that takes seconds, the list has to be parsed, and asking twice can give a different list. It might also invent an item or a category you never offered.

Jev, from typesafe.ai, is a System 1 model: the fast, intuitive kind of judgement, as opposed to the slow, deliberate reasoning a chat model does. You give it a state (what the question is about) and typed questions, each with a fixed set of answers. It returns a probability for every answer. There is no text to read back, and nothing outside the options you wrote can come out.

A chat model

  • writes an answer, token by token
  • you parse what it wrote
  • can answer outside your options
  • seconds for a long answer

Jev

  • picks among the answers you defined
  • returns numbers: a probability for each
  • cannot answer outside the options
  • 149 answers in about 0.4 s

2 Anatomy of a typed question

This is exactly what Jevilize sends for one item when you type things you can wear:

"state": "Sorting objects into the category: things you can wear",
"questions": {
  "i0": {
    "type": "choice",
    "instructions": "Object: scarf. Does it match \"things you can wear\"?",
    "criteria": {
      "yes": "clearly and typically matches \"things you can wear\"",
      "no":  "does not match \"things you can wear\", or only in unusual edge cases"
    }
  }
}
state
The context every question shares. Here it's the category you typed.
questions
A map of named questions. Jevilize names them after the items (i0 is the scarf) and sends all 149 at once.
type: "choice"
The answer is one of the keys in criteria. Here that is yes or no, but a choice can have many options: an intent, a language, a period.
instructions
The question itself. Jev only ever sees the item's name, never its picture.
criteria
Each option, with a description of when it applies. The descriptions steer the answer (see wording).

3 What comes back

For every question, a choice, a probability for each option (they add up to 1), and a confidence:

"i0": { "choice": "yes", "confidence": 1.0, "probabilities": { "yes": 1.0, "no": 0.0 } }

Confidence is the gap between the two most likely options. In every answer we checked, it matched the top probability minus the runner-up. Real answers for things you can wear in winter:

itemyesnoconfidencereading
🧣 scarf1.000.001.00certain
πŸ•ΆοΈ sunglasses0.660.340.32leans yes, not sure
🧒 baseball cap0.490.510.02a coin flip
πŸŽ’ backpack0.150.850.69fairly sure it's no

Low confidence is information, not an error. It tells you where the question itself is borderline, so you can treat those answers differently: ask again with better wording, or hand them to a slower model.

4 Inside: one pass, one slot per question

TypeSafe hasn't published Jev's internals. The model class is described in the open by decider, an independent reproduction (not affiliated with TypeSafe AI). There, a request becomes a single prompt with one answer slot per question:

  1. Render. The state and every question with its lettered options (A = yes, B = no) become one prompt, with an empty answer slot after each question.
  2. One forward pass. The model reads the whole prompt once. Nothing is generated, so all slots are filled by the same pass.
  3. Read the slots. At each slot it takes the scores for the option letters, and only the letters that are valid for that question.
  4. Softmax at a fitted temperature. Those scores become probabilities. The temperature is tuned on labelled data so that β€œ70%” means right about 70% of the time. This is called calibration.

Because the model only scores the options you gave it, it cannot answer outside them, and there is no output to parse. Because nothing is generated, one question costs almost the same as many.

5 149 questions, one call

Measured from here against the Jev API, three runs each:

β‰ˆ 0.32 s1 question, round trip
β‰ˆ 0.39 s149 questions, round trip
β‰ˆ 70 mswhat the other 148 cost
~12k / 4.5ktokens in / out for 149

Most of the round trip is the network and the request itself. The questions share one state and one forward pass, so adding 148 more barely moves the clock. That is what makes it practical to ask about every item on every keystroke pause.

6 From probability to physics

Jevilize turns each yes probability into physics with three rules:

  • At or above 75%, the magnet pulls. The item gets a slot in a row under the text box. Every physics step its velocity is nudged toward that slot and gravity is cancelled. A higher probability pulls harder (a gain of 0.06 at 75%, up to 0.16 at 100%).
  • Below 75%, gravity wins. When an item loses its slot, it gets a small shove and falls. Items stranded up top get extra gravity that grows until they reach the pile.
  • Near its slot, an item stops colliding with loose items, so nothing can ride up on the row and sit there.

The threshold is a choice, not a fact. Move it and watch what rises for things you can wear in winter:

🧣 scarf 100%🧀 gloves 100%🧦 socks 100%πŸ§₯ winter coat 100%πŸ₯Ύ hiking boot 99%⌚ wristwatch 89%πŸ₯½ goggles 89%πŸ‘Ÿ running shoe 84%⛸️ ice skate 84%
πŸ•ΆοΈ sunglasses 66%πŸ’ diamond ring 64%πŸ‘“ glasses 55%🧒 baseball cap 49%🎩 top hat 45%πŸͺ– helmet 45%🎧 headphones 42%πŸ‘• t-shirt 41%πŸ’Ž gem stone 25%πŸŽ’ backpack 15%

At 75% the sunglasses (66%) fall. At 60% they rise, with the diamond ring. Lower still, and the coin flips come up too. Too low lets in the wrong things; too high drops the right ones.

7 Wording is part of the question

The criteria text changes the answer. With loose criteria (yes: "matches"), sunglasses scored about 0.95 for things you can wear in winter. With Jevilize's stricter wording ("clearly and typically matches" against "… or only in unusual edge cases"), they drop to about 0.66–0.71. Nothing else changed.

Write the options so they describe what you mean, including the edge cases, and keep them looking like buckets. Long rules inside a question tend not to be followed by models this size.

8 Limits, plainly

  • Close to, not exactly, repeatable. Three identical calls gave sunglasses 0.67, 0.73 and 0.70. Clear cases don't move; borderline ones wobble. Jevilize caches each answer, so the same words give the same pile while it is cached.
  • Words are ambiguous. For Instruments, a compass scores 84%: it is a scientific instrument. Jev reads words, not your intent.
  • It only knows the name. β€œred apple” is red to Jev because of the word, not the picture.
  • One pass, no working out. Multi-step reasoning or arithmetic needs to be split into several questions, or left to a slower model.

Try it

Describe something and turn on Debug: every item shows its probability and hitbox, and clicking one opens its full decision. The items page lets you ask Jev about all 149 and read the exact question for each.

How the video was made

  • It replays this app's own interface: its styles and its debug drawing, fed with real Jev answers.
  • The pile is a Matter.js simulation with this app's exact sizes and magnet rules.
  • The clock pauses inside one call and slows for the replays, so each number can be explained while it is on screen.
  • Rendered frame by frame in headless Chromium at 1920Γ—1080 and encoded with ffmpeg.
  • The voice is a neural text-to-speech model. Whisper word timings put each beat on its spoken word. Every sound is synthesized in code.