Device and IP checks catch one person running many accounts. They don't catch two real people on two phones who have agreed not to hurt each other. We look at how the cards were actually played, and we show you the hands.
Soft play, chip dumping and card sharing don't leave a shared fingerprint. They leave a pattern in the decisions: a player who folds top pair to one opponent and nobody else, a stack that drains toward the same seat, a raise that only makes sense if you knew what the other hand was.
Those patterns are invisible in a single hand and obvious across a few hundred. Most security teams find them the slow way, from player complaints. In October a player was reading opponents' hole cards through compromised table tools for about two years before sites acted.
Source: PokerNews, 2 Oct 2026
A model of what a player does in each spot: position, price, stack depth, cards, action so far. It is cross-fitted, so nobody is scored by a model that already saw their tables, and it learns each player's own style so a loose regular doesn't look like a cheat.
For each two players: while B is still in the hand, does A fold, call or raise differently from expectation, and differently from how A plays everyone else? A sequence model then reads every hand the two shared, with each player's strongest other partner as context, so one cheater doesn't drag his innocent tablemates up the list.
Every flagged pair comes with the five hands that carry the signal, in the order a reviewer would read them, and a short case note that includes the innocent explanation. Your team makes the call. We never auto-ban.
The "Detect Suspicious Value Transfers in Poker" competition on Kaggle gave 370+ teams two million hands across 400 tables, a handful of labelled colluding pairs, and asked for three things: rank 112,540 pairs, name the type of collusion, and list the hands that prove it.
We finished 5th at the deadline. Afterwards we kept going, and every change was picked on held-out data with its expected score written down before it was ever submitted. The last version scores above the winning entry on the private leaderboard.
| Our entry at the deadline (5th) | 0.92817 |
| Post-deadline v2 | 0.93449 |
| Winning entry | 0.93514 |
| Post-deadline v3 | 0.93562 |
Post-deadline scores are late submissions. They don't change the official standings, and we don't claim they do. The competition's own score is 70% pair ranking, 20% evidence hands, 10% behaviour type.
Wash trading and coordinated volume on prediction markets look a lot like chip dumping: value moving between the same accounts at prices nobody else would take. We're applying the same pair-level method to public on-chain trades and will publish what it finds.
