"does Jev understand colour?" — an unofficial recreation

Type a word. A small, fast model weighs it against sixteen fixed colours, and the bar is that weighting drawn to scale — widest first.

Live: https://jev-colour.zammitpaul.com

Whose idea this is

His idea, his colours. Matt DesLauriers — @mattdesl on X, mattdesl on GitHub — posted this on 18 September 2026 and said how it worked: sixteen predefined colours, "just rendering their weighted probabilities". He linked no source, so I built a version that lives on a page you can type into. The sixteen colours are his too, from docs/palette.md in Bitframes (MIT) — copied into palette.js, sRGB and Display P3.

A test run

Sixteen football clubs and nations, typed straight into the live page (docs/teams.mp4):

phrase	Jev
man utd	red 100%
liverpool	red 91%, dark blue 5%, green 2%
juventus	black 57%, white 27%, gray 7%
borussia dortmund	yellow 95%, black 3%, orange 1%
barcelona	dark blue 44%, red 29%, light blue 16%
real madrid	white 99%, red 1%
inter	gray 33%, indigo 14%, green 13%
inter milan	dark blue 72%, black 18%, light blue 5%
inter miami	hot pink 60%, light pink 18%, dark blue 5%
la galaxy	dark blue 57%, black 29%, indigo 7%
seattle sounders	teal 31%, green 27%, dark blue 26%
brazil	green 89%, yellow 11%
argentina	light blue 91%, dark blue 4%, white 2%
netherlands	orange 85%, green 6%, light blue 4%
italy	green 59%, red 17%, yellow 10%
jamaica	green 72%, yellow 20%, teal 4%

inter alone is a prefix, not a club, and the answer says so; one more word and it is nerazzurri blue. italy comes back as the flag rather than the azzurri, which is the kind of disagreement the bar is good at showing.

How it works

Jev is TypeSafe AI's System One model: it answers typed questions with probabilities and cannot generate text. The whole page is one question (jev.js):

{
  colour: {
    type: 'choice',
    instructions: 'A person typed a word or phrase. If they had to paint it using only these sixteen
                   colours, which one would they reach for? Judge the phrase as a whole.',
    criteria: { beige: null, black: null, white: null, /* … all sixteen … */ },
  }
}

The answer comes back as a probability per colour, and that is the picture: a band is a colour's probability, the order is the ranking, and the only thing code does is drop anything below 0.05% and renormalise. The options are the palette's own names, except index 0, which the spec calls background (the beige of the paper) and Jev is shown as beige.

Typical answers, and what they cost (Jev's input tokens at $0.042 per million):

phrase	Jev	time	cost
tomato	red 100%	466 ms	$0.000017
london bus	red 99%, black 1%	848 ms	$0.000017
canada	red 78%, white 12%, green 6%, light blue 3%, dark blue 1%	292 ms	$0.000017
the matrix	green 59%, black 41%	356 ms	$0.000017
stormclouds	gray 92%, dark blue 5%, black 2%, indigo 1%	417 ms	$0.000017
retro	orange 33%, teal 13%, brown 11%, hot pink 11%, beige 7%, …	779 ms	$0.000017

Running it

npm install
TYPESAFE_API_KEY=… npm start   # http://localhost:8080

Without a key the server answers with weights derived from a hash of the phrase and says so on the page; the shape of everything else is the same.

GET /api/palette	the sixteen colours, who is answering, what has been spent today
POST /api/colour	{ "phrase": "london bus" } → the distribution
GET /up	healthcheck
/?q=london+bus	a link to an answer

Identical phrases are answered from memory, each visitor gets forty questions a minute, and the day stops when DAILY_USD_CAP is spent.

Deploying

config/deploy.example.yml is a Kamal config: copy it to config/deploy.yml, put TYPESAFE_API_KEY in .kamal/secrets, and kamal deploy.

Licence

MIT. The sixteen colour values come from Bitframes by Matt DesLauriers, also MIT.