When to enter.

Same execution. $1,000 per account. Up to five $10 entries per side, including fees; at least ten seconds apart. Each pair receives the same forecast. Six percentage points of probability buffer plus five points of net buy edge. No pre-round entries, learned cells or resale overrides.

Entry windows only. Early DOWN enters only during seconds 0–120. Minute-3 UP enters only during seconds 120–180. Their controls may enter either side until second 270. Pending buys expire at the window boundary. Exits continue afterward under the same policy; stable positions wait for resolution in the final 30 seconds.

Actual fill constraints. Every order waits 500 ms. Fills use available Rust order-book depth at execution, at or better than the decision limit. Missing depth means no fill. Fees are estimated from verified market metadata and already included in P/L. Paper simulation cannot reproduce actual market impact or priority.

Prospective evidence. Rules were fixed before launch; historical rounds are not counted here. Signals, versions and availability are logged before outcomes arrive. Current champions may evolve; both members of each pair receive the same version. B held-move extrapolates a trained short-horizon forecast; it is not a separately trained five-minute model.

Comparing results. Candidate minus control P/L includes all common officially settled rounds after launch, even no-trade rounds and outages. Win rate uses traded settled rounds only. Charts and drawdown use the settled-round P/L curve, not intraround equity. Open positions and realized sales from unsettled rounds are shown separately.

No automatic winner. The first review is at 100 settled rounds. This is an exploratory test, not a promise of profit or statistical significance. No automatic promotion or real-money access. Existing portfolios continue independently.