A call is priced at a historical quote at or after publish, not a fill. Crypto: first CoinGecko sample at or after tweet time (typically hourly; ~5m near now — not a 1-minute close). Equities: historical daily close (live only if the call is under 24h old). Missing history stays unpriced. Reference entry ≠ actual fill; system evaluation ≠ actual exit. Nothing here is real P&L.

Win rate uses decided calls only: priced, aged ≥3 days, and return ≠ 0 (flat/undecided calls are excluded). Fewer than 3 decided calls means no percentage and no composite score are shown.

A post scores only if the author identity is verified. Wrong author never creates a scorable call. Retrospective bragging never creates a successful entry.

AI classifies, extracts and summarizes — with every output bound to an evidence quote and human-reviewed. Returns, rankings, roles and scores come from deterministic code.

Deletions are synced; the text is taken down but the snapshot stays in the audit record and affected calls pause scoring into review.

Scores are versioned (score_version). Same inputs plus same versions always reproduce the same number.

Raw composite score is 0–100. A KOL needs ≥3 scorable calls (each with a return and age ≥3 days) or score is null. Points: Win rate 45 = (wins ÷ scorable) × 45; Sample size 15 = min(total calls, 20) ÷ 20 × 15; Avg return 10 = clamp(avg return to ±20%) ÷ 0.2 × 10; Drawdown quality 10 = (1 − clamp(avg max drawdown, 0–30%) ÷ 0.3) × 10 (missing → 5); Upvotes 20 = min(upvotes, 50) ÷ 50 × 20. Displayed on 0–5 via display = 5 × ln(1+100x)/ln(101), x = raw/100 — order-preserving rescale only. Ranking and settlement use raw.