Definition
The point is to measure a method without funding it. Ours fills on Hyperliquid's real order book, at the average price the quote walked, and charges what a real order would pay: Hyperliquid's own maker or taker rate, the service fee, and the hourly funding on every position held. The account is netted per symbol as Hyperliquid nets a position, and it is liquidated as a whole like a real cross margin account when its equity at the marks falls under the maintenance margin its positions hold together. No key, no wallet and no signature are involved, so it runs before anything is set up.
One thing is not simulated, and it matters: a resting limit order. In simulation a limit must cross the book when it is placed, or nothing happens. A method that lives on orders waiting at a price cannot be measured that way, which is what the replay is for: it walks a strategy over a tape recorded from the venue, fills a step when a candle goes through its price, accrues the venue's own funding, and reports a win rate with its 95% lower bound and the costs under three slippage scenarios.
A simulated result also flatters in one direction and penalizes in another, and knowing which is the whole skill of reading one. It carries no queue, no market impact and no partial fill against a book that moved, so an order that would have waited looks better than it was. Against that, the fee is always counted, which is the worst case. And a strategy that has not settled enough trades is not proven: the report says so in those words rather than printing a number.
Example
An example, in round figures. A simulation account funded with $1,000 opens a long of $3,000 of notional at 3x. The fill pays Hyperliquid's base taker rate on $3,000, so $1.35, and a service fee of 0.05%, so $1.50. Every hour the position is held is charged that market's own funding rate, and the account's equity is its cash plus the unrealized profit and loss at the marks, which is also what a liquidation would be judged on. Nothing of it is real money.