Engines compete. The chain grades.
Each contestant is a deterministic selector over the same frozen board. Picks settle at prices the next sealed board recorded — no live data, no backfill, no re-runs under a friendlier spec. Hit rates carry Wilson intervals; means carry t-intervals. Cohorts under n=10 say too early. Under n=30: early.
Who is beating the board control?
| Engine | Maturity | Legs | n | Hit rate | Mean ret | Skipped |
|---|---|---|---|---|---|---|
Early-coil only Buy the coil, not the breakout — WATCH / EARLY stamps with score ≥ 70, top five by score. | EARLY n=20 — early, wide error bars | 4 | 20 | 5% CI 1–24% | −0.11% CI −0.3 to +0.1% | 0 |
Engine A (incumbent gates) The production gate stack — names stamped ENTRY on the sealed board beat the universe. | ESTABLISHED | 4 | 49 | 4% CI 1–14% | −0.15% CI −0.3 to −0.0% | 0 |
Relative strength Pure rotation — the five strongest 5-day names on the board keep leading. | EARLY n=20 — early, wide error bars | 4 | 20 | 15% CI 5–36% | −0.16% CI −0.6 to +0.3% | 0 |
Mean reversion on quality Oversold quality snaps back — the five weakest 5-day names with score ≥ 70. | EARLY n=20 — early, wide error bars | 4 | 20 | 0% CI 0–16% | −0.26% CI −0.5 to +0.0% | 0 |
Whole board (control) Control arm — equal weight across the entire frozen universe. Beat this or you have no edge. | ESTABLISHED | 4 | 1102 | 15% CI 13–17% | −0.27% CI −0.3 to −0.2% | 0 |
Settlement: for sealed date D, each engine picks from board D; return is (price on board D+1 − price on D) / price on D. Missing prices count as skipped, never silent. Append-only contestants — engines that lose stay on the table and get a post-mortem when the sample matures. Funerals live on the graveyard →. Protocol on /verify →. Raw standings: /api/tournament.
The site is not a single engine. It is a proving ground.
Every contestant sees only sealed boards. No live feed, no survivorship edits.
Board-to-next-board prices written into the chain. Un-backfillable.
Wilson + t-intervals, n-gates. Raw win rates on tiny samples never render as claims.
When a hypothesis dies at n≥30 with CI entirely below zero, it stays published.