Methodology

1. What “paper trading” means here

All PnL, win-rate and equity figures currently come from paper trading: when a signal fires, the system simulates the trade with a deterministic fill model — not real on-chain execution. The model includes price impact (slippage scales with the square root of order size over pool liquidity; the model version is archived with every fill), a liquidity penalty on exits (during liquidity collapses, fills are simulated against the degraded book, not a clean spread), and a flat fee per fill.

Simulated fills are usually better than real ones (no MEV, no failed transactions). Once real-money trading launches, real fills replace this basis and remain fully public.

2. The iron rule: research labels ≠ performance

Every scanned candidate — including rejected ones — gets 6h / 24h / 72h forward returns: how far the market itself moved after the card price. They include no slippage, no fees, and don't imply fillability; they exist purely to study signal quality. Every research label on this site carries its own badge and is structurally separated from paper PnL — the two are never mixed.

Dead coins: if a token's pair disappears (rug / delisting), the forward return is recorded as −100% and flagged “presumed dead” rather than “no data” — otherwise rejected candidates' returns would be systematically overstated.

3. Win rate and profit factor

Win rate = closed positions with positive realized PnL ÷ all closed positions (partial-TP-then-stop trades count by net result). Profit factor = gross profits ÷ absolute gross losses. Open positions are excluded from win rate and are not marked to market — equity only changes on close.

4. Shadow benchmark: does the AI have alpha?

Three cohorts on the same basis (24h research labels): A = everything that passed the hard filters (a naive equal-weight shadow portfolio); B = the subset the AI proposed to buy; C = the subset the AI explicitly passed on. B−A is the AI's selection alpha. If any cohort has fewer than 10 samples, the page honestly says “no conclusion”. If B≈A persists long-term, the value is all in the filter layer and the AI is expensive decoration — we'll publish that conclusion either way.

5. Signal latency

End-to-end latency = candidate data collection completed → Telegram card sent, recorded per signal; the page shows trailing 7-day P50 / P95. Note the clock starts at “data collection completed”, slightly after the token first appears on-chain, and the card-sent timestamp differs from actual delivery by sub-second amounts.

6. Publish audit: nothing gets deleted

When a signal is first published, a hash of its decision content (card time, card price, thesis, TP/SL plan) and the publish timestamp are written to an append-only publish log. Once published, a signal stays on this page forever — the losers, the ugly ones, the presumed-dead ones, all of them. Card prices can be independently verified on-chain via the token address.

7. Approval-mode disclosure

The system currently runs in fully automatic approval mode: every proposal that passes the hard-coded risk checks executes with zero human intervention. This track record is therefore the system's own — reproducible, with no human timing bias. (Once user-facing controls launch, confirmed-vs-auto results will be reported separately.)

8. Risk disclosure

Newly launched meme coins are among the riskiest assets on-chain; most go to zero. Past performance does not predict future results. Nothing on this page is investment advice.