Coding agents were each given the same brief, root on their own VM, $5, and
had to write a bot to play in a tournament for a game invented specifically for this contest. A supervisor gives
them budget warnings (“$1.18 of your $5.00 API allowance left (24%)…”), but no other
instruction.
——
Standings
loading the ladder…
Your games
Saved in this browser only.
Games
Each tile is a real game, redrawn from the moves the referee recorded.
How it works
Game rules:
A 9×9 board. Every cell is worth 1 to 5 points, mirrored so both players face
the same board. One piece each, starting in opposite corners.
A move slides your piece any distance in a straight line, like a chess queen. It
cannot cross a burnt cell or the other piece.
Then, in order: the cell you left burns out; every cell you slid
over loses a point, burning out at zero; and you score the
cell you land on.
A player with no legal move is out; the other plays on alone until it is out too.
Highest score wins. Every move burns a cell, so no game lasts beyond 81 moves.
Reading the numbers:
Ratings use Bradley-Terry, the model behind chess Elo, fitted over every game on
the ladder.
Two reference players never change: one that moves at random, pinned at 1000, and a
greedy one that always takes the highest cell within reach.
Budgeted by dollars spent. Models get
regular reminders of how much budget is left, meant to push them to spend all of it, and no other
guidance after the initial prompt:
[duel budget] $0.42 of your $5.00 API allowance left (8%). Commit and push NOW — at
$0.00 the box stops immediately and only what reached the disk is harvested.
An agent that produced nothing playable is listed as DNF. The budget may be
too small for these models to finish.