Fair play

A proof shows that every move in a game was legal and signed by the player whose turn it was. It can't show who chose the move. A player can sign moves that KataGo picked for them, and the chain can't tell the difference.

So Surround analyzes games in its backend. When an account's games show, over time, that it's playing an AI's moves, the account is labeled. It isn't banned and can keep playing, but everyone who meets it sees the label:

Possible unregistered bot. This account's games match AI play more closely than human play at its rank.

The approach borrows from the chess servers that have fought this problem longest. As on Lichess and Chess.com, the evidence is statistical and is gathered across many games, and a person reviews it before any label goes on.

Registered and unregistered bots

Bots are welcome on Surround as long as they're open about it. The AI anchors are registered bots: they're named as AI everywhere they appear, and their ratings are pinned. An unregistered bot is a human account whose moves come from an AI. The label exists so that you know which of the two you're playing.

What we analyze

Every game played through Surround's keepers is analyzed after it ends. The analysis uses only what the game itself contains: its moves and the time spent on each, which the referee stamped and attested.

Each position is analyzed by KataGo at high strength, and each move is also scored by KataGo's human SL model for the player's rank, the same model the AI anchors play with. Popular Go AIs (KataGo, Leela Zero and their descendants) agree with one another on most moves that matter, so measuring agreement with one of them catches assistance from any.

Signals

• Point loss for the rank. Every move gives up some points against the best move. We measure a player's average point loss, a measure Go researchers prefer to win rate because a close game swings the win rate on tiny mistakes. We then compare it with how much humans at that rank actually lose. A 15k who loses half a point per move over many games is playing far above their rank. This is Go's version of the intrinsic performance rating that Chess.com measures.
• Agreement where it's hard. Matching the AI's top move proves nothing by itself: some moves are forced, and many are obvious to any player. We count agreement only in positions where the human model is unsure, where players at the rank genuinely spread across several moves, and where the AI's choice is one humans rarely find. Agreement in the opening and in routine endgame moves counts for little.
• Human or engine? For each move we compare two likelihoods: that a human at the player's rank would play it, and that a strong AI would. Human play, even strong human play, leaves a recognizable trail of human-shaped choices. A long run of moves the human model finds unlikely and the AI finds natural is the clearest signal we have. It's the idea behind Lichess's Kaladin, which looks for play uncharacteristic of the player.
• Timing. People move quickly when a move is obvious and think when it's hard. Someone relaying moves from an AI tends to take the same time on every move, or thinks longest when the position is simple. Ranked games are timed by the referee, so every move's time is known to the millisecond.
• The player's own history. A sudden jump in move quality, unlike the account's earlier games, counts more than steady strong play. So does a run of wins against much stronger opponents.

No single signal labels an account. Each game gets a score from all of them, and the scores add up across the account's recent games.

What isn't counted

• Short games. Games too short to rate are too short to judge.
• Games after a resumption, when both players are capturing dead stones and the best move is plain.
• Forced moves, such as answering atari or retaking an obvious capture.

From signal to label

1. Analysis. Each finished game is queued for analysis. KataGo runs in the backend, never in the players' browsers.
2. Accumulation. The account's game scores accumulate over a rolling window. One game, however strong, never flags an account: a good day and a good game happen to everyone.
3. Review. An account that crosses the threshold is flagged for review. A person looks at the games and the evidence before a label goes on, because the software only flags. (Lichess works the same way: its detection tools flag accounts, and only its moderators mark them.)
4. Label. The account is labeled as a possible unregistered bot.

We don't publish the thresholds or every feature we use. Chess.com and Lichess keep theirs confidential too: a published threshold is a target to play just under.

What the label does

• It's visible. The label shows on the player's profile, in the lobby, on the pairing offer before you sign a game's terms, and beside their name during the game. You can decline a game against a labeled account.
• It doesn't stop play. A labeled account can still queue, open tables and play rated games. We would rather show players the evidence than decide for them.
• It isn't onchain. The label is backend data, not a contract flag. The chain records results and ratings, not suspicions, so a labeled account's rating changes by the same rules as anyone else's.

Mistakes and appeals

Statistical detection gives false positives, and the cost of a wrong label falls on a player who did nothing wrong. That's why accounts need many games of evidence, why a person reviews every flag, and why labels can be appealed. On appeal the account's games are reviewed again, and a label that doesn't hold up is removed.

Strong players who are new to Surround are the likeliest false positives: they play far above their starting band at first. Their history quickly catches up with them, and the point-loss signal is measured against their rating as it rises, not against where they started.

Where detection stops

• Selective help is harder to see. A player who consults an AI for only a few critical moves leaves a weaker trail than one who copies every move. Those moves still tend to fall in hard positions, where we look most closely, and the evidence builds up across games.
• Detection is never proof. The label says possible for a reason.
• Proofs and signatures don't help here. They make a game's moves honest, not its players. See Settling a game.