Citadel‑style bot reverse‑engineered

Prediction markets let traders buy and sell odds on future events, and one reverse-engineering thread says a bot has been beating those odds at a 74% clip. (x.com) The post, published by the researcher who goes by 0xTatara, said the bot made about $9,400 over 19 days while trading prediction-market contracts. The thread framed the system as a “Citadel-style” setup, borrowing hedge-fund language for a rules-driven trading stack. (x.com) On prediction markets such as Polymarket, prices act like implied probabilities: a contract trading at 0.50 signals roughly a 50% market estimate. Polymarket’s documentation says the displayed price is the midpoint between the best bid and ask unless the spread is wider than $0.10, when it shows the last traded price instead. (docs.polymarket.com) That plumbing matters because a bot can scan hundreds of contracts faster than a human, compare market prices with its own estimates, and place trades when the gap looks favorable. Polymarket exposes public order-book data and price history through its application programming interface, giving builders the raw inputs for that kind of system. (docs.polymarket.com) The thread landed as forecasting bots are already being tested in formal competitions. Metaculus runs an AI Forecasting Benchmark in which bots compete on real-world questions, with a Fall 2025 prize pool listed at $58,000 and seasonal tournaments spanning roughly four months. (metaculus.com) Metaculus says those tournaments cover about 300 to 500 questions per season, and it gives participants bot templates, access tokens, and sponsored language-model inference from OpenAI, Anthropic, and Google. Its resources page says that, as of the second quarter of 2025, top humans still beat bots and model quality mattered more than bot infrastructure. (metaculus.com, metaculus.com) Earlier Metaculus benchmark pages showed how fast that race was moving. A Q1 2025 tournament update said a template bot using o1-preview and AskNews reached sixth place in the prior quarter, and the Q2 page said only bot accounts could submit forecasts while new entrants could still join midstream. (metaculus.com, metaculus.com) The follow-up posts around the 0xTatara thread pushed the same idea in trading terms: use large language models to generate forecasts, wrap them in rules for position sizing and risk, and keep iterating. One public GitHub project for Polymarket, posted separately, describes a seven-signal ensemble with “adaptive Kelly sizing” and says it was built from more than 65 experiments. (x.com, github.com) That GitHub project also shows why many of the headline numbers need caution. Its README reports an 83% simulated win rate but a 25% win rate on 46 real markets, while still claiming positive profit because average wins were larger than losses. (github.com) The reverse-engineered bot thread did not, by itself, provide the kind of audited track record that a hedge fund or exchange would require. But it gave the AI-forecasting crowd a concrete claim to chase: that a machine can turn public market odds into a repeatable edge before everyone else builds the same thing. (x.com, metaculus.com)