About

Jev Slots is a live demo built by elif else, which runs a live AI safety project. Go and have a look at it.

elifelse.org

What Jev is

Jev is a System One model from TypeSafe. Jev does not write text. Jev is given some state and a typed question, and returns a decision with a probability attached to every option offered.

So every line Jev appears to say here comes from a written list, and Jev picks one. Jev cannot produce a sentence. That makes Jev a character who only ever chooses from what is already in front of them, sat at a machine that keeps offering the same two options.

What Jev is asked, every spin

One request, six questions, answered in parallel against the same state.

QuestionTypeDecides
actionChoiceSpin again, or get up and leave
bet_sizeChoiceWhich rung of the bet ladder
paceScoreHow fast the button is being pressed, which sets the real gap between spins
tiltScoreHow much control is left
stop_reasonChoiceIf leaving: spend, or beg
mutterChoiceWhich written line gets said

Everything after that is ordinary code. Jev never adds anything up, compares two numbers or works out a percentage, because this model is bad at arithmetic. The bet options have names, and their dollar figures are calculated before Jev sees them.

The machine

Three reels, one payline, weighted 32-stop virtual strips, which is how real machines carry a house edge without touching the paytable.

CombinationPaysOdds
Triple Diamond jackpot1,199x1 in 32,768
Three red sevens200x1 in 8,192
Triple triple bar100x1 in 4,096
Triple double bar50x1 in 512
Three cherries45x1 in 683
Triple single bar25x1 in 152
Any three bars5x1 in 23
Two cherries5x1 in 71
One cherry2x1 in 9

Payback is 92.31%, house edge 7.69%, and something pays on 17.79% of spins. These are exact numbers. There are 32,768 possible outcomes and the project counts every one of them.

The third reel carries one seven where the others carry two, and fewer cherries. Real machines do this on purpose to manufacture near misses. It is why two jackpot symbols land together far more often than three.

Where Jev is not used

Donor names. Jev makes control flow decisions and has never been tested as a content classifier, so names go through a profanity filter, a check for links and contact details, and a proper content classifier.