Experiment 001 · model jev-1.13.0

A city of a thousand people. Nobody scripted them.

One small AI model, Jev by TypeSafe, decides what every one of them does next. All thousand, about once a second, for a cent and a half. This page is a recording of it thinking. The city itself is live.

1,000
minds a tick
0.98 s
median, all of them
$0.015
per tick, measured
REC Day 3 · 00:00 tick 000
minds asked
Jev answered in
that is
for
Every dot is a person; its colour is what Jev decided they would do. Replayed from the logs of a real live run () without calling the model again. The timings are the ones measured that night; the green bar fills in the time Jev actually took. So far in this recording: 0 decisions for $0.00.

What one second buys

1,000people, each asked what they will do next, each given a probability for every choice they have
0.98 sfor the whole city: ten requests of a hundred people, in parallel. 95th percentile 1.32 s
$0.015when every single person is asked. $0.0073 on an average tick, because people are only asked again when their situation changes

Code runs the world: time, hunger, fire, gossip, injury, death. Jev is a “System One” model: it supplies only judgment. It never writes a word and never sees a number, only plain phrases that code wrote: “starving”, “was robbed by Cora”, “wolves, right here”.

One mind, in full

This is everything Jev knew. This is what it said.

A real decision from the recording above, chosen by a rule, not by taste: the first person with danger on top of them about whom Jev was unsure.

state.villagers.what Jev was sent

        
probabilitieswhat Jev answered

Code rolls dice against those odds. Try it.

Twelve minds, one verdict

A thief is caught. The city decides.

Code seats twelve citizens, friends and rivals first. Jev judges each one once: would this person vote guilty? Visitors call the verdict before the votes turn over. This one really happened.

0 guilty0 not guilty

The number on each juror is the probability Jev gave a guilty vote; the vote is drawn from it. Gold: tied to the accused. Blue: tied to the victim. Tested before it was built, with pass marks committed in advance: what held and what failed.

Citizens of X

A thousand citizens. One owner each.

Sign in with X inside the city and the citizen you were given becomes yours alone. Owners score by calling what citizens do before it happens. This board is live.

of 1,000 citizens claimed

Claim yours
  1. Loading the live board…

Failures first

None of these raised an error. All of them were wrong.

Confident, well-formed, wrong behaviour, visible only in what a whole population did. Each was found by replaying the exact moment and changing one field.

  1. 80 of 80

    The frozen village

    We told Jev what each villager was already doing. Every one of them chose it again, at 0.92 confidence.

  2. day 3

    The chapel

    We described what the crowd nearby was doing. 73 of 78 joined in. By day three the whole village was praying, starving.

  3. 34.6

    The famine

    Starving farmers answered “eat” when there was no food, never “farm”. That needs a “therefore”, and Jev takes one step. 34.6 of 80 died, per run, until code closed the gap.

  4. −0.03

    The alibi

    All that a witness for the defence moved a jury. Jev under-weights evidence of innocence, so the game does not rely on it.

The expensive one: villagers[37]

Jev cannot count, and an array index is a count. Past about the twentieth item in a list it answered, confidently, about nobody in particular. We had run a whole pre-registered evaluation that way. We re-ran it keyed by name and published both.

gap between greedy and loyal villagers, by position in the list

0–190.51 · 0.44
20–390.45 · 0.02
40–790.40 · 0.00

keyed by name by array index

Eighteen findings, the method, the limits →

Every dollar

$14.20 for all of it.

Seven rounds of evaluation against baselines, twenty benchmarks, and the live city: 17,713 requests, 338 million tokens, each one a line in a ledger. Our own waste is in it, on purpose.

Evaluations against baselines, 5 seeds per condition$9.21
The live city, tick by tick$2.75
Development runs of the village of 80$0.95
Twenty benchmarks: every finding on this page$0.68
Daily character sheets, first city run, juries, screening$0.62

Live, the city costs about $3.50 an hour whether one person is watching or ten thousand, and nothing when nobody is. Ledger as of 19 September 2026; the full table is in the study.

It is running right now.

You will be given one citizen the moment you arrive. Follow them. Call what they do. Claim them with X, and nobody else can: there are only a thousand.

Enter Jevtown

Free. No account needed. Works on a phone.