How a verdict is made

1. Principles

Three rules govern every measurement:

• Chain data only. We decode raw transactions. We never trust a platform's self-reported trade feed or P&L.
• The follower's ledger, not the leader's. Every number answers one question: what would a copier — entering late, paying costs, holding every bag — have realized?
• Addresses, never persons. Reports describe the measured activity of a wallet address. They attribute nothing to any individual or entity.

2. Data sources

Trade histories are decoded from Solana transaction data (token balance diffs and parsed swap transactions). Prices come from public minute-resolution OHLCV candles across three independent providers, cross-checked. Current prices for mark-to-market use the deepest live pool per token; ghost pools under $100 liquidity are ignored. Leaderboard appearances are our own periodic captures of public leaderboards, timestamped.

3. Decoding trades

A swap event is recognized when a wallet's token balance moves opposite to its cash leg (SOL/wSOL/USDC/USDT) in one transaction — which filters airdrops, transfers, and LP operations. The decoder is DEX-agnostic (routing and multi-hop included). Where swap activity exists but no events decode as the address's own retail trades, the address is ruled NOT A STANDARD WALLET rather than scored.

4. The follower simulation (FSP)

Every completed round trip (FIFO-paired buy→sell) is re-executed as a follower would experience it:

• Delays: entry and exit each shifted +5s, +30s, +120s, +300s behind the wallet. The gating delay for verdicts is +30s.
• Position size: $500 per trade, flat — what a retail copier actually risks.
• Costs: DEX fees, priority fees, and depth-based slippage scaled to the pool's measured liquidity.
• Fills: the minute-candle close at trade-time + delay, empirically validated against real observed executions (see §9).
• Exclusions: holds over 30 days, unresolvable pools, and trades with no candle data are excluded and disclosed with counts on the report.

5. Abandoned-position accounting (survivorship correction)

Round trips only cover trades a wallet chose to complete — the exact survivorship bias leaderboards monetize. So every FIFO-leftover buy older than a 3-day grace period is marked to current price at the same $500 sizing, with entry costs applied. Dead and unpriceable tokens mark toward zero: a follower holds those bags too. Entries with sub-$1 recorded cost (dust and airdrop-shaped events a follower could not mirror) are excluded. Mark-to-market gains are clamped at +2000% — anything past that is degenerate price data, not edge (§8).

The report's blended expectancy per mirrored buy combines completed trips and abandoned positions — the single number closest to "what does copying this wallet pay?"

6. All-positions win rate

Closed-trade win rates ignore the never-sold tail. Ours doesn't: closed winners ÷ (closed trades + never-sold positions). A wallet that sprays 2,000 buys, closes 15 winners, and abandons the rest shows ~1% here — regardless of what its realized-only win rate claims.

7. Statistical gating

A positive average is not evidence; variance is enormous. Per-trade expectancy at the gating delay is bootstrapped (1,000 resamples, 95% interval). COPYABLE requires the entire interval above zero, plus profit factor ≥ 1.3 and a minimum trade sample. The bootstrap is seeded deterministically from the wallet address itself, so identical inputs reproduce the interval — and the whole report — bit-for-bit.

8. Sanity clamps (we assume our own data can be wrong)

• Single-trip returns outside [−99%, +2000%] are rejected as bad data, never counted as edge.
• Fills below a minimum plausible price (ghost candles) are rejected.
• Abandoned-position marks are clamped identically (§5).
• Scores abort — writing nothing — when a data provider fails mid-pull. Provider downtime is never allowed to masquerade as "this wallet has no data."

9. Validation — how we check ourselves

• Fill-model validation: our modeled fills are compared weekly against the median of real observed executions in the same pool within ±90s of the simulated time. Current measured error: 0.06–0.10% median, <1% at P90 — far below the expectancy differences verdicts rest on.
• Forward calibration: weekly, we test whether verdicts predicted follower P&L on trades made strictly after the verdict. Published as accrued.
• Verdict stability: re-scored wallets are monitored for verdict flips under unchanged methodology — our public drift alarm.
• Reverse verification: every report carries a SHA-256 hash over the exact decoded event set (signature | token | side | block-time, sorted). Anyone can re-pull those transactions from any Solana RPC, re-run this methodology, and confirm both the hash and the verdict.

10. Verdict definitions

• COPYABLE — positive blended follower expectancy; PF ≥ 1.3; full 95% CI above zero; adequate sample.
• MARGINAL — positive point estimate but CI includes zero, or sample too thin to distinguish edge from luck.
• UNCOPYABLE — negative follower expectancy after costs.
• UNCOPYABLE — ALL-IN — completed trades positive, but mirroring every buy including never-sold positions is net negative.
• UNMEASURABLE AT MINUTE RESOLUTION — median hold under 30 minutes; any edge lives inside a single candle where minute data cannot resolve it. We decline to guess.
• NOT A STANDARD WALLET — swap activity present, none decodable as the address's own retail trades (program/vault/aggregator pattern).
• INSUFFICIENT DATA — fewer than 3 completed, priceable round trips.

11. Versioning

Any change that can alter a verdict or a printed figure — thresholds, cost model, clamps, decoder behavior, statistical method — bumps the methodology version. The version appears on every report and is stored with every verdict in our history, so any past report can be audited against the exact rules that produced it.

• 2026.07.21 — current. Added evidence hashing and report provenance; abandoned-mark dust exclusion and +2000% clamp formalized.