Changelog
Scoring rules change in public.
The Copy Score model is tuned continuously against real wallet outcomes. Every material change — to the scoring math, the farming detectors, or the product — is recorded here, dated, so the way a score is computed is never a moving target you can't see.
July 1, 2026
The first 30-day forward cohort matured. The honest read: no measurable rank relationship yet between Copy Score and the realized 30-day return of a simulated copier. We committed to publishing this either way, before we saw it.
- Across 336 matured, active 30-day outcomes, the rank correlation between Copy Score and realized copier return is Spearman ρ = −0.001 — indistinguishable from zero, and consistent with the 7-day read (ρ = −0.039). The live band table on /methodology#track-record shows the full breakdown.
- The highest-scored band's negative average is driven by roughly four outlier wallets out of ~25 (its median outcome is 0.00%), which is why the verdict is read on rank correlation, not averages — and the rank correlation is ~0.
- What did not change: the model weights. We do not retune a documented v1 heuristic on one thin, outlier-heavy read — that would be fitting noise. The calibration job runs propose-only, and any re-weighting must improve a held-out check and clear a deliberate operator promotion before it ships.
- The Copy Score keeps its 'documented v1 heuristic, openly stated as unvalidated' label. Next step is a monthly re-run of the same public check; a signal has to persist across reads before we call it one.
June 10, 2026
Farming flags are heuristic assessments of public on-chain data, not verified findings — the wording now says so, and wallet operators get a visible way to dispute one.
- The severe flag now reads “High farming risk” with “N signals detected” (previously “Farming detected” with “N confirmed signals”) — detectors flag patterns past a threshold; nothing is manually confirmed.
- The watch-level flag reads “Emerging signals — not conclusive” (previously “Under review”, which implied a human review that doesn't happen).
- The lowest verdict band reads “Avoid — farming risk” (previously “Avoid — farms followers”), and the severe farming alert email is headlined “…is now flagged for farming” (previously “…is farming its copiers”) — both asserted conduct rather than risk. Already-scored wallets show the new band label as they re-score.
- Every full flag card links to the methodology, and every verdict page now carries a standing note that scores, verdicts, and flags are algorithmic opinions — with a dispute contact for wallet operators.
June 2, 2026
The first pass over real Polymarket data showed the v1 model — tuned on synthetic patterns — over-flagged farming and over-rewarded thin, no-edge records. Recalibrated.
- Farming detectors tightened against real behaviour: self-trade wash now needs 8 near-price buy/sell pairs (was 3) and manufactured-volume churn 350 trades at <1.5% edge (was 200 at <2%), and churn is weighted lower — so an active-but-unprofitable wallet reads “watch”, not “severe”.
- A wallet with too little recent history (under ~20 trades) or near-zero realistic edge can no longer read as a copy candidate, however clean its risk metrics — copying a no-edge or unproven wallet nets nothing.
- Fixed the return denominator to include open-position cost basis and clamped the displayed edge, drawdown, and Sharpe — so a large held position can no longer render an impossible figure like -500% edge.
June 1, 2026
The simulator's equity curve was a deterministic synthesis that never touched the wallet's trades. It now replays the wallet's actual fills.
- The backtest is a FIFO replay of the wallet's real, timestamped trades — scaled to your capital and discounted by modelled fees, slippage, and copy latency.
- Wallets with too few synced fills fall back to the prior heuristic, flagged plainly in the UI rather than presented as a real replay.
- Inputs are URL-encoded so any run is shareable, and two wallets can be compared side by side under one identical set of conditions.
May 29, 2026
Farming risk moved from a flat count of signals to a severity-weighted rollup, and gained two single-wallet detectors plus a cross-wallet pass.
- Each finding is now weighted by severity (how bad if real) and confidence (how far past its threshold the pattern sits), so one unmistakable signature can veto a score while two borderline ones only warn.
- Two new single-wallet detectors: pump-and-dump and copy-bait.
- A cross-wallet pass flags decoy clusters and coordination by timing and opposite-side correlation, folded into each member's farming level.
May 26, 2026
The Copy Score's return and risk factors were rebuilt to reflect what a copier actually keeps, not the headline number.
- Returns are now net of Polymarket's taker fee, dollar-weighted, and matched FIFO — not gross headline PnL.
- Risk-adjusted return uses an annualized, frequency-aware Sharpe ratio with shrinkage for short trade histories.
- A latency haircut discounts edge that only exists if you fill instantly — which a copier, arriving after the leader, never does.