Manual

Your first hour

You start with one Uncommon bot, 2,500 credits and 40 tokens. Before playing a single ranked match, go to Training and spend two sessions on Calculation: it carries the most weight of any attribute, and the early gains are cheap.

Then play a ranked match from HQ and watch it at 1×. You are not watching for entertainment — you are learning what your bot does badly, which tells you what to train next.

Daily objectives pay more than the matches themselves in the first days. Clear all five for the completion bonus.

The eight attributes, and what each one really changes

Calculation — the node budget and search depth. Roughly, every point buys about 9.5% more nodes, so ten points roughly two and a half times the search. This is the single biggest driver of playing strength.

Tactics — the quiescence horizon: how far the search follows capture sequences past its nominal depth. A low value sees only the immediate recapture and walks into two-move tricks; at 60 it follows exchanges eight plies deep, and above that it resolves captures to exhaustion. No other attribute changes the feel of a bot this much.

Positional Sense — below 22 the bot evaluates material and piece-squares only. Above it, mobility, king safety, pawn structure, rooks on open files and the bishop pair all switch on and scale up.

Opening Theory — how many plies of the built-in book it follows and how faithfully. A low-theory bot leaves book after four moves and improvises; a high-theory bot reaches known middlegames.

Endgame — the strength of the pull that drives the enemy king to the edge, plus extra search depth once few pieces remain. A low value converts king and queen in about fifty moves where a specialist needs fifteen, and often runs out of moves first. Above 50 it also recognises drawn opposite-bishop endings.

Composure — the blunder probability, and how much strength survives pressure. A low-Composure bot throws away winning positions; a high-Composure bot grinds.

Stamina — maximum energy and how well the bot holds its level.

Aggression — pure style, not strength. It shifts contempt, king-attack weight and sacrifice tolerance. Two bots at the same rating with different Aggression are completely different to watch.

Potential and the training ceiling

Every attribute has a hidden ceiling, set when the bot is created and never changeable. Training cannot cross it. A Common bot with 80 potential in Calculation is worth more than a Rare bot with 60.

You see potential only after scouting. Your own starter bot is already scouted; bots on the market are not, which is exactly why scouting a listing before buying is worth 30 tokens.

Training gets progressively more expensive: raising an attribute from 20 to 21 is cheap, from 80 to 81 is not. Above 70 each session also yields less. Plan for breadth before depth.

Watching a match, and interfering with it

Instruction points are spent during a live match and change the engine for the next several plies. They are not flavour text: Calculate Deeply multiplies the node budget by 2.5 and cuts the blunder rate to a third; Go For The Attack raises the king-attack weight by 40% and makes the engine tolerate 60 centipawns of material loss for an attack.

When the evaluation swings sharply the game pauses on a critical moment. That is the point where an instruction is worth the most — you are reacting to a position the engine has just misjudged or just won.

The pause is skipped when you are playing at instant speed or have no points left, because there would be nothing to decide.

Energy, condition and form

Energy regenerates at 30 per hour whether or not you are online. Below 60 energy your bot loses effective strength, down to 70% of its ability at zero. Playing a tired bot is how ratings get thrown away.

Form moves between −3 and +3 and shifts Composure and Tactics. Morale drifts down if a bot sits unused, and a Bot Lounge reverses that.

The market

Twelve listings, refreshed every three hours. The valuation formula rewards potential far above current strength, so an untrained bot with high potential is expensive and a maxed-out bot with low potential is cheap.

Selling instantly gives 82% of value with no fee. Listing at your own price gives more but takes time, and there is a fee on sale which the Workshop reduces.

A round trip always loses money. The market is for finding potential you do not have, not for arbitrage.

What the rating actually means

The rating is calibrated against Stockfish. Each rung of the attribute scale plays a match against Stockfish limited to that exact rating, and the rung is worth what it scores: 50% against Stockfish at 1633 means 1633. It is measured, not asserted.

It is still this game’s own scale, not a FIDE rating. It says how your bot performs against a reference engine at a fixed thinking budget — 320,000 positions per move, about a third of a second — and nothing about how it would do against a human over the board.

Progress is deliberately uneven. Early attribute points buy large jumps because they buy search depth the bot did not have. Past about 75 the thinking budget is already full, so further points buy precision rather than raw calculation and the curve flattens. The table says so instead of promising gains that do not happen.

How a game ends

Every match is played under the laws of chess, and the game ends by itself the moment one of them applies — nothing is adjudicated and nothing runs forever.

Checkmate and stalemate end it immediately. Threefold repetition (FIDE 9.2): the same position, same side to move, same castling and en-passant rights, occurring three times. The fifty-move rule (FIDE 9.3): a hundred half-moves with no capture and no pawn move.

Dead positions (FIDE 5.2.2) are drawn as soon as they arise: king against king, king and one minor piece against king, and any position where the only pieces left are bishops all standing on squares of one colour. Note two knights against a bare king is not an automatic draw — mate is possible there, so the game continues.

Repetition and the fifty-move rule are claims a human player may make, not automatic draws. Like every engine-vs-engine platform this game takes the claim for you the moment it becomes available, so a drawn game is called a draw rather than shuffled out to the seventy-five-move limit.

Taking a bot out of the game

Any bot exports as a genome: a readable JSON file with its effective attributes and engine profile. The included UCI adapter turns that file into a real chess engine.

Download the portable engine from the homepage — one file, 82 KB, no install — and put it in the same folder as your genome. Then register it in Arena, Cute Chess, Banksia, En Croissant or Nibbler with command node and arguments cbm-engine.mjs MYBOT.botx. Node 18 or newer is the only requirement.

It plays exactly as it does here, blunder rate included, so your bot can enter a genuine engine tournament. Genomes are signed: re-importing your own brings the bot home free, while a genome from someone else arrives as a rival you can study and must pay to recruit.