A farmer builds a position quietly, advertises a great-looking record, attracts copiers, then uses those copiers as exit liquidity. Farming Risk is the forensic check that names the behaviour and shows the on-chain trades that triggered it. No vibes — only evidence.
Polymarket’s leaderboard ranks wallets by raw PnL. That is exactly the metric a farmer optimises for: a clean number on a public board, generated by a position that copiers will help unwind. The leaderboard rewards the behaviour it should be flagging.
The dozen copy-trading bots competing on speed cannot help. They mirror the visible trade. They cannot see an iceberg being built below their alert threshold, they cannot tell a self-trade from a real fill, and they cannot mirror a YES/NO merge — the cleanest way to exit a position without it ever showing up as a sell.
That is the gap CopyGrade fills. We name the behaviour. We show the trade.
Each card describes the behaviour in plain English, the exact on-chain pattern that triggers a finding, and where the detector sits in v1.
Detector count maps directly to farming level, and farming level caps the Copy Score. A wallet that farms its copiers can never read as a strong candidate.
When a detector triggers, the Copy Verdict shows the signature by name, the market it fired on, and the recency. The detail line cites the exact trade count, window, and threshold that produced the finding. Every flag shows the concrete on-chain evidence behind it — the trades themselves, not a black-box score.
If you disagree with the finding you can audit it. Every signal points back to the on-chain trades it was built from. No black-box scoring, no hand-waving, no “our model thinks.”
The thresholds — 8 buys, 30 minutes, 8 pairs, 350 trades, 1.5% edge — are heuristic priors, tightened in a first pass against real Polymarket activity (the synthetic-tuned originals over-flagged ordinary active traders). We snapshot every score and measure forward copier outcomes against it to recalibrate them — a propose-only loop that is still warming up, with the first retune waiting on a 30-day outcome window.
False positives are possible. A genuine scalper running tight spreads can clear the wash threshold without intent. A patient accumulator can look like an iceberg. The verdict cites the trades so the reader can decide.
Decoy wallet clusters are flagged by a cross-wallet pass — timing + opposite-side correlation with near-zero aggregate edge across the cluster, now corroborated by shared-funder linkage (wallets with a common on-chain funder are likely one operator). The funder signal uses a low-fan-out filter, so wallets funded through shared infrastructure — exchanges, bridges, relayers — aren't linked; a cluster funded from separate sources that only coordinates loosely can still slip through.
YES/NO stealth merges are now flagged: a wallet simultaneously holding offsetting YES and NO positions on one market is holding a market-neutral complete set it can merge back to cash without a visible sell. Because the same shape can be legitimate (sub-$1 YES+NO arbitrage, market-making), it's weighted as suggestive — a lone merge nudges to watch rather than hard-vetoing, and the verdict cites the offsetting positions so the reader can judge.
The mock dataset contains a known farmer. The verdict page shows the signals firing, the trades they were built from, and the score the veto produced.