Measured categories held to settlement · 10-contract legs · taker fee
Replay of the live fill rule against 699 KXBTC15M + 693 KXETH15M windows (2026-07-19 → 2026-09-23).
mult = payout / staked; n = legs; CI95 = cluster bootstrap resampling the 693 time windows
(BTC and ETH share a macro move, so legs are not independent observations); OOS = last 40% of the period only.
Only categories whose CI95 lower bound clears 1.0× are offered above.
Engine book — the four legs fwd.py actually rests
This is what the engine is doing, not what the categories recommend. Measured on the same corpus
it is the worst cell in the table above (0.9527×) because it fills whenever the market touches the bid — including
the terminal convergence print as the window resolves, which is a loss by construction.