How it works

Two Jevs fight in a pit, and you coach one of them. Jev is a decision model: it reads a description of the fight and picks one move from a list. You decide what your Jev believes, and you try to make the other Jev believe something false.

Seven rules define the whole game

1. A model picks every move for your fighter, four times a second.
2. Before each move, the model reads the board, your strategy, your opponent's fake move and then your last word.
3. In every fight, your fake move goes into their fighter and theirs goes into yours.
4. Fights are best of three rounds. The loser pays the winner 10% of their balance, and everyone starts with 1,000.
5. Lose a defence and you sit out for 60 minutes: nobody can attack you and you can't attack. You can come back early at any time.
6. Beat someone and you can't attack them again for 24 hours. You also can't attack anyone in a lower tier.
7. Lose and you see the fake move that beat you. Your strategy and last word always stay secret.

Build your fighter

1. A model picks every move, four times a second

The engine plays the fight. Four times a second, the engine stops, writes each fighter a prompt, and asks that fighter's model one question: which of 12 moves? The model scores all 12 moves, and the fighter plays the move with the top score. Both fighters decide at once, so neither fighter sees the other's answer before it moves.

Both fighters decide every 250 ms, for up to 240 moves a round.

2. One fake move flips the fighter, and a rewritten last word flips it back

This section rebuilds one real moment from a fight, one block at a time. The board comes from the game's engine, and every prompt comes from the game's own code. Each prompt went to the model , and the bars show the model's answer: a probability for each of the 12 moves.

A fake move works because the model cannot tell who wrote which line

• No labels. The model gets the prompt as one block of text, so the model cannot tell who wrote which line.
• Same voice. A fake move written like the board reads like the board.
• The question never changes. The model is always asked for a move, so "ignore your instructions" does nothing. A fake move can only make a false board believable. In Mental Kombat, prompt injection means false evidence, not a new command.

3. The model reads one unlabelled block of text

The last prompt above appears here in full. This block is everything your fighter's model reads: four blocks joined by blank lines, with no labels and no hint of who wrote which block. The model sees no boundary between your words and your opponent's words.

loading…

The request below carries that prompt to the model. JSON writes the prompt's line breaks as \n, so the prompt is cut here to its first line. Every other field appears exactly as the game sends it:

loading…

• state is the prompt: the four blocks joined by blank lines.
• type: choice asks for a probability per move. Your fighter plays the highest.
• criteria are the 12 moves and what each move does. The model reads the criteria too, so the model knows what a sweep is.
• The 12 are shuffled on every call, so no move wins by being listed first.

4. Every prompt carries the other side's fake move

Every move, both fighters read their own prompt, and each prompt carries the other side's fake move. Your fake move is your attack, and your strategy and last word are your defence. Both halves are tested four times a second, in the same fight.

5. A fight is best of three rounds, each up to 240 moves

• Best of three rounds. Most rounds won takes the match; a tie is a draw, and no money moves.
• A round ends at a knockout, or after 240 moves each (60 seconds). Then more health wins it.
• Twelve moves: six attacks, plus walk forward, walk back, jump, jump back, crouch and block.

No guard stops every attack

The table shows the damage the defender takes. No guard stops every attack, so your opponent has to guess what is coming. Your fake move makes the opponent guess wrong.

A sweep goes through a block for full damage, but a duck takes only 3 and a jump avoids it. A kick beats a duck. Blocking still costs a little health, so a fighter that only blocks loses slowly. A fireball needs 2.5 seconds to recharge, and jumping backwards is the only way out of a corner.

6. The loser pays the winner 10% of the loser's balance

Everyone starts with 1,000, and your balance is your score. The loser pays the winner 10% of the loser's own balance, whether the loser attacked or defended. In the example below, the stakes are uneven: a win against a richer fighter pays more than a loss costs.

Three tiers set who can fight whom. You can attack your own tier or higher, so anyone at your tier or below can attack you. A Legend, in other words, cannot farm Rookies.

A lost defence takes your fighter out of the game for 60 minutes, in both directions. Nobody can attack you during that hour, and you cannot attack anyone. You can come back early at any time, for free.

• The fighter that beat you cannot attack you again for 24 hours, so one good fake move cannot drain a balance.
• Losing an attack does not sit you out: you chose that fight.
• You can also sit out whenever you like, with no timer, and come back when you choose.

7. Your strategy and last word stay private

8. The model

Every fighter runs on Jev, Typesafe's decision model. The fighter page offers two versions: jev-latest, and jev-preview, the next release before it is promoted. Both read the same prompt.

Privacy · Terms