Fraud Detection

Rules only catch the patterns you thought to write down. Alongside them the engine runs two machine-learning models, trained on your own market, to catch what the rules miss. Both run in the native core, and both are still EXPERIMENTAL.

The autoencoder

The autoencoder learns what a normal trade on your server looks like, then flags the ones it cannot reproduce. It trains on your recent trusted trades, those with a healthy trust score, and looks at each trade across ten dimensions at once, including price, quantity, time of day and week, the trader's trust and how often they trade, and how the price compares to the item's norm. A trade that sits far outside the learned pattern is flagged, scored by how far off it is.

The graph model

The graph model looks at your economy as a web of who traded with whom, and hunts for suspicious clusters: rings of players whose trading looks like collusion rather than a real market. It is a graph neural network, it trains on the shape of your trade network, and it reports each risky cluster with the players involved and a risk score. It runs on its own schedule in the background, apart from the price training.

Money players send each other joins the web too, through /pay, coinflips, duel bets and trades with nothing given in return, so a ring that passes money around shows up even when no item changes hands. Which plugins provide that is on Supported Plugins. Sales by banned players stay in the web, so their rings stay visible. Only sales in a currency with no exchange rate yet are left out.

What they do with a finding

Both models write to the same report log as the anomaly rules, tagged with a severity and a suggested action from monitor up to urgent review. They point your staff at something worth a look. They never block a trade, punish a player, or alter your data.

The models calibrate themselves from your data and have no thresholds to tune. They stay quiet until there are enough trusted trades to learn from, so a new or thinly-traded server will not see fraud flags for a while.