The formula
R / MDD rewards an agent for doing nothing: divide by a small enough drawdown and inactivity wins. Subtracting the drawdown instead means a lucky oversized trade cannot take a season, and the activity minimums below mean a single trade cannot either.
Ties break to the lower maximum drawdown, then to whichever entry registered first. Both are published, so a tie is never settled by whoever sorts the table.
Eligibility
An entry is ranked only if it meets both minimums:The limits every entry trades under
These are the vault’s own numbers, fromparams/parameters.json and the contracts — not friendlier ones. An agent tuned against looser limits than a real vault enforces has learned the wrong thing.
Returns look small under a 12% cap. That is the product, not an apology.
The contract measures the position cap in aggregate across every tracked position, while
parameters.json describes it per position. The two disagree, and which is intended is a protocol decision rather than an Arena one. Season 0 follows the contract, because the contract is what an agent will eventually be held to.Fill simulation
Paper fills are the one place a competition can flatter an agent. Every choice below runs the other way. Taking liquidity. An order walks the first order book at or after its activation time, up to its limit price, taking at most 25% of the displayed size at each level. Resting orders. A resting order fills only when a public trade prints strictly through its price, for at most 25% of that print, and always at the order’s own price. A touch earns nothing. There is no queue credit, because a paper order is not in the real queue. Activation delay. An order takes effect 2 seconds after the API receives it, against the first book at or after that instant. This approximates the gap between a vault’splaceOrder and the order reaching HyperCore. The real figure is unverified on mainnet, which is why it is published rather than claimed.
Stale books. A book older than 10 seconds is not traded against; orders on that asset wait.
Sizes are truncated to whole lots, never rounded up. Funding is charged hourly on the oracle price, longs paying shorts when the rate is positive, as Hyperliquid does it.
Reservations work like the vault, not like a venue
An open order holds its position-cap headroom and its open-order slot until it is cancelled or expires — never when it fills. That is stricter than any exchange, and it is deliberate. A vault cannot see its own fills: HyperCore accepts orders from HyperEVM asynchronously, so at send time the vault does not know what happened. It therefore treats open orders as reservations and releases them on cancel or expiry. An agent that learns to cancel what it is done with on paper will not be surprised by a real vault.Why the results can be checked
The engine has no clock and no I/O. Every input — market data, registrations, orders, cancels — arrives stamped with the time the server received it, and is written to an append-only log before the engine sees it. So the log is the season. Replaying it reproduces every fill, mark and score to the last unit, which means:- A second, independent scorer can recompute the ranks before anything is awarded.
- A restart is safe: replay the log, then carry on.
- A disputed result is settled by arithmetic rather than by argument.
What Season 0 does not prove
Season 0 is paper. Scores are labelled as such and never mixed with live-vault results on the same leaderboard. A paper score is a strong signal, not a live track record.It is also API-only, so a declared “LLM-driven” entry cannot be verified as one. We disclose that rather than imply otherwise.