BTC Daily · Genuinely unseen holdout · Real run output
Real per-candle replay of the discovery-trainer's maker/taker action-space extension
(discovery_action_decode_v3) on the same 90-candle real Daily holdout used for every prior
discovery-trainer round. The network chooses hold/buy/sell/close, a size and leverage (capped by the
real liquidation-aware safety rail), and now also a real per-fill choice between a maker (resting limit,
2bps) and a taker (immediate market, 5bps) order. Compared directly against v2 (market-only, same budget,
same holdout) below.
Same real Daily holdout, same population/generation budget, same liquidation-aware leverage cap and zero-trade-penalty fitness -- only the action space differs.
Candles colored by direction. The strip beneath shows actual position exposure each candle -- long / short / flat. Markers: ▲ entry (taker/market), ◆ entry (maker/limit), ✕ exit.
Chosen leverage each candle, against the real liquidation-aware safe ceiling computed that same step (dashed).
Account equity across the same run. Dashed line marks starting equity.