PolyLeviathanPolymarket Insider
Activity Detection.
PolyLeviathan scores every trade on Polymarket for insider-like behaviour using a machine learning model, 30 seconds after each trade hits the order book. Suspicious wallets surface in the Scanner as they appear.
Context
Insider trading on Polymarket is a documented problem.
Prediction markets have a structural vulnerability: anyone with advance knowledge of an event outcome can profit by trading before that information reaches public odds. On Polymarket, traders are pseudonymous and the blockchain is the only record. That makes insider trading easier to do and harder to prosecute than in traditional financial markets.
In late 2025 and early 2026, major news outlets documented multiple confirmed or suspected cases. An OpenAI employee lost their job over it. Twelve coordinated wallets made over $1M on a market specifically built to detect insider activity. Polymarket filed formal CFTC cases and hired Palantir to police the problem.
PolyLeviathan's model gives retail traders the same visibility that institutional risk desks pay for. You see the flagged trades as they happen, not after the market has already moved.
WIRED · Feb 2026
OpenAI fires employee for Polymarket insider trading
OpenAI terminated the employee after investigating their activity on Polymarket and other prediction platforms.
CoinDesk · Feb 2026
Twelve wallets made $1M on a market built to catch insider traders
Coordinated wallets bet heavily on a market specifically designed to detect insider activity, then won.
The Atlantic · Mar 2026
Insider Trading Is Going to Get People Killed
A major feature on how prediction market insider trading creates dangerous real-world incentives around geopolitical events.
Yahoo Finance · Mar 2026
Polymarket brings Palantir and TWG AI in to police suspicious bets
Polymarket filed formal CFTC cases and announced partnerships to detect and report suspicious trading.
How it works
Every trade comes in as it executes
PolyLeviathan processes every trade on the Polymarket CLOB as it happens. Each trade gets attributed to its wallet address and parsed for size, direction, market, and timing.
The model extracts features per trade
For each trade, the classifier looks at: position size relative to market liquidity, time to resolution, wallet age and historical win rate, price impact of the trade, and whether the wallet has been flagged before.
XGBoost outputs a score from 0 to 1
An XGBoost gradient boosting classifier, trained on historical Polymarket data with known resolution outcomes, assigns each trade an insider probability. A score near 1.0 means the trade closely matches trades that preceded surprising resolutions in the training set.
Two thresholds trigger alerts
Scores above 0.1 get logged as Suspicious. Scores above 0.5 get flagged as Potential Insider. Both appear in the Scanner live. The thresholds are calibrated so high-confidence signals are rare by design — currently 57 wallets above 0.5 at any given time.
Score reference
What each score level means.
No anomaly. Trade pattern is within normal range for this market and wallet type.
Some features are atypical. Large size, unusual timing, or a wallet with little prior history. Worth watching, but low confidence.
Multiple features align with the known insider pattern. Strong signal. Not a verdict.
How we differ
Retail gets the same feed Polymarket's own team monitors.
In early 2026, Polymarket brought in Palantir and TWG AI to monitor suspicious trading. That monitoring is internal. It works backward from flagged outcomes, and it's invisible to retail traders until a case gets filed months later.
PolyLeviathan's model runs forward. Every trade scores 30 seconds after execution, before the market has time to move on the information. You see the flag when it fires. Not in a press release after the fact.
The formula is public. The thresholds are calibrated and documented. The feed is free. Every flagged wallet, every suspicious score, and every high-confidence signal surfaces in the Scanner in real time.
Caveats
Signal, not verdict.
A score of 0.9 means a trade strongly resembles the pattern of historical insider trades. It does not mean the trader had insider knowledge. Large, well-timed bets by sophisticated traders or market makers produce high scores without any wrongdoing.
False positives cluster in two places: low-liquidity markets where any moderately large trade moves the price significantly, and markets that resolve quickly after opening. Both look like insider patterns but often aren't.
The thresholds are conservative on purpose. Lowering the cutoff would increase volume but degrade signal quality. 57 wallets sit above 0.5 at any given time. That rarity is a feature.
PolyLeviathan flags anomalies. Conclusions are yours to draw.
Watch the feed live.
Every flagged trade, clarification alert, and regime signal across active Polymarket markets, as they appear.