How the numbers are made.

Long sleeves: BTC from Binance USDT spot, SPY / GLD / SHY from Yahoo Finance daily, macro features from FRED. Intraday: Binance USDT-perp 5-minute bars over the curated universe.

Long sleeves: 2018-01-01 through the current week — covering the 2018-19 bear, the March-2020 liquidity event, the 2021 mania, the 2022 bottom, and the 2024-25 bull. Intraday: the past 90 days, replayed against historical signal state.

90d R, win-rate, and average R-per-trade come from a rolling 90-day window of simulated runs against live Binance bars. Live operator strategies on real venues run continuously but enter the published ranges only with the operator's explicit consent.

Long sleeves: 5 bps round-trip per rebalance, applied to every band-cross. Intraday: maker fees on entry (−0.5 bps net), taker fees on stop (+0.5 bps), funding overlay re-applied every 8 hours. Slippage assumed constant: 1 bp on majors, 2 bps on alts.

The long-sleeve universe is fixed (BTC / SPY / GLD / SHY) — no asset is added or removed mid-period. The intraday universe is a curated B-only confluence set; pairs that later fail liquidity or spread filters are dropped forward only, never retroactively.

Paper means orders are matched by an internal simulation venue against the same live bars the engine sees, with the cost model above. Live means orders hit a real exchange API. Where ranges combine both, the strip is labelled CALIBRATED · 2018-2026 BACKTEST + 90D LIVE — both inputs are disclosed.

The max-drawdown target is a model assumption under the cost / regime distribution observed in-sample. Out-of-sample regime shifts — correlation breaks, liquidity vacuums, venue outages, prolonged grind-downs — can push drawdown past the target before the throttle takes effect.

Failed deployments, unreconciled partial fills, manual interventions through the dashboard, and returns generated while the kill switch was engaged are excluded from the published numbers. The audit log records them; they are not folded into target ranges.