Inside Our Risk Framework
A delta-neutral basis fund earns a thin, repeatable spread and survives on its ability to avoid the rare event that wipes out months of carry in an afternoon. The return is the easy part to describe. The discipline that keeps the return is the part an investment committee should interrogate, because a strategy with no directional view fails for structural reasons rather than for being wrong about the market. The Global Basis+ risk framework is built around that fact. It assumes the trade works and spends most of its effort bounding the ways it can break.
What follows is how we think about risk, organised around the failure modes that have historically taken down institutional basis books, and the controls we run against each one. The controls below are a representative selection of the parameters that govern the live book. All of them are indicative and subject to finalisation in the Offering Documents.
Principle one: constrain the system before you automate the response
There are two different questions in any risk framework. What can the system physically do, and what does the desk do when something goes wrong. We treat the first as an engineering problem with a hard answer and the second as a policy problem with a named owner, and we are precise about which is which.
On the first question the constraints are absolute. The onchain vault contract has exactly one permitted destination, the fund's Fireblocks vault. Venue keys can do two things, trade or return capital to that vault, with withdrawal destinations pinned at the protocol level. Every venue child order is an immediate-or-cancel limit placed inside a 25bps protection band from the live book mid, so nothing rests on the book, nothing walks it, and the engine never sends a market order. Each leg is pinned to 1x isolated before the first child order. Hyperliquid and Lighter orders are signed only inside an attested AWS Nitro enclave that independently re-checks leg, side, price, size, signer epoch and validity window against a signed parent authorisation. The engine host cannot forge a venue signature or decrypt a trading key, and it cannot create a Fireblocks whitelist entry. These are not promises about behaviour. They are limits on what the software is capable of attempting.
On the second question we are deliberately conservative about claiming automation we have not shipped. Delta rebalancing of the smaller leg, unwinding a filled first leg after a timing gap, funding-floor rejects at the order router, and NAV drawdown halts are monitored continuously and executed on operator approval under written policy bands. Today no drawdown threshold in the engine triggers an automatic unwind, and the monitor that flags adverse carry does not rebalance by itself. Those automations are in flight. We would rather an allocator diligence the framework we actually run than the one on the roadmap.
There is a reason to be careful here beyond honesty. An automatic unwind is itself a risk: it fires on bad data, on a stale oracle, on a venue print that reverses a minute later. Where a control is cheap and reversible we hardcode it. Where the action is a one-way door on a live book, a named human executes a written policy.
Principle two: zero leverage
The fund runs at zero net leverage, meaning 1x on every applicable perpetual position, pinned before the first order goes out. Leverage turns a manageable drawdown into a forced liquidation, and a basis trade does not need it to produce its target return. Removing it removes the single largest source of catastrophic failure before any other control is applied.
It also buys something specific in this market. Venues can auto-delever profitable shorts, and the ADL queue is ordered by profit and leverage. Running at 1x puts the fund near the back of that queue.
Directional and delta risk
The whole structure rests on the two legs cancelling each other's exposure to price. If that cancellation drifts, the book quietly takes on the directional risk it was built to avoid. Funding arrives as cash and changes the dollar size of one side relative to the other, the perp's mark price wanders from the index inside a window, and tokenised wrappers trade at small premiums and discounts.
The engine works to three bands. Legs are matched in USD notional with a target net delta of zero. The aim band is 25bps per child, and the engine sends correctives and then a reduce on the leader to get inside it. The accept band is 1% of requested gross: residual past that parks the position, new size stops, an approved reduce or close stays available, and an operator resolves it. On any reduce or close the engine sends a reduce-only IOC on the leader first and sizes the follower to what the leader actually filled, so an unwind cannot itself create the exposure it is meant to remove. A one-leg fill or a sustained gap between legs halts new size on that ticker until the pair is matched.
We also watch the carry itself. An internal adverse-carry signal flags capital sitting in a losing basis trade so it can rotate into a cleaner pair or into cash. The monitor flags and an operator approves the rebalance or close.
Market and liquidity risk
A position is only as good as the ability to exit it. Rather than publish a single liquidity number that would be gamed by the venue's own reporting, we run an eligibility screen on open interest and visible depth, with a standing rule that the fund should never be a material fraction of either for any ticker. If a proposed order would make it one, the order is capped or skipped. If the eligible universe shrinks, deployment is reduced rather than concentrated into what is left.
Execution quality is enforced by the same protection band that governs entries. If the book is not crossable inside 25bps of mid, or a realised fill lands far from the intended limit, the slice is not sent and the remaining children are aborted. In a stress unwind that logic inverts: closing the hedge outranks saving a few basis points.
The two legs also keep different hours. Cash equities at IBKR trade on exchange hours with T+1 settlement, perpetuals trade continuously, and Coinbase International can be closed or limit-only outside its session. No new cash-leg pair is opened outside market hours, and out-of-hours hedging is permitted only where it reduces residual delta or liquidation risk. Corporate actions get the same treatment. Dividends, splits and mergers may not be mirrored one-for-one on the perpetual, so the default on an approaching ex-date is caution, and the preference on messy structures is to exit ahead of them.
Pricing integrity is a named venue risk rather than a single divergence number. Mark price on HIP-3 markets is a deployer-controlled blend of cash-market sources, which means a bad oracle lets the mark be pushed, and when cash markets are shut the oracle can sit still while the perpetual keeps trading. Oracle latency, stale-window behaviour and funding caps pinned for several hours on a held name are all monitored and sized against.
The cash rail carries its own risk. USDC over Circle CCTP funds movement between Interactive Brokers and the perpetual venues. A sustained move off peg, a delayed mint or redeem, a partner outage on the fiat ramp, or regulatory action against the issuer pauses the rail, with an exit to cash or a swap into another stablecoin under risk assessment.
Venue and counterparty risk
Both legs depend on venues that can pay out at scale, and a venue that halts withdrawals or fails outright leaves one leg hedging nothing. Venue weight follows how battle-tested a venue is, its available liquidity, and internal due diligence. Offchain, Interactive Brokers is the most robust name in the mix, a licensed broker-dealer with SIPC coverage and excess insurance, sized against published per-customer caps. Onchain, Hyperliquid and Coinbase carry the most weight, with Lighter held at a second-rank weighting and sized accordingly.
Hyperliquid's equity and commodity markets add a layer that a generic venue policy would miss. They are listed under HIP-3 by third-party deployers, principally Trade.xyz, and the deployer controls the oracle blend, the liquidity and the fee design. Deployer stake, oracle latency, and sudden changes in fee or rebate design are monitored, and new Hyperliquid exposure is capped if the deployer looks stressed. Cross-chain movement uses canonical platform bridges, whitelisted and fixed in code, and the engine cannot transfer to an arbitrary contract.
Concentration is bounded at the trade level. No single trade may exceed 35% of the book. There is no numeric per-venue cap; venue exposure is managed through weighting and through the systemic-risk view below.
Drawdown governance
Above every position-level control sits a portfolio-level policy with two bands. A 3% drawdown from the high-water mark opens the risk-off band: new risk halts and the weakest pairs are cut. An 8% drawdown is a hard stop and a full unwind, with investors informed immediately. Redeployment after either requires CIO and risk review.
Both bands are executed by an operator under written policy. We have described above why the fund-level unwind is not wired to a threshold today, and the automation is in flight.
Operational risk
The failures that ended institutional crypto books in 2022 and 2023 were rarely about the trade. They were about custody, key management, and where the assets actually sat.
All capital sits in a Fireblocks MPC vault with Ascent GFS, the fund administrator, holding independent signing authority on every account. A Transaction Authorisation Policy governs destination whitelists, per-asset velocity caps and transaction vetoes, with higher quorums required for fund-level actions such as processing deposits and servicing withdrawals than for systematic trade execution. Trading keys for the onchain venues are generated and used only inside the Nitro enclave, with the decryption key released by AWS KMS solely to an enclave whose measured image matches the approved value. Keys for Coinbase, Binance and IBKR cannot be enclave-protected because they must be generated on those platforms, so they are held encrypted on the engine host and constrained by the same trade-or-return-to-vault permissions. Emergency functions carry 48-hour timelocks. An independent watchdog agent continuously reconciles onchain and venue balances against the signer's own log, and treats anything it cannot account for as a highest-severity alarm.
On the contract side, Pashov Audit Group reviewed Jetstream in April 2026 and returned 25 findings: two critical, five medium and eighteen low. The criticals covered a redemption-funding accounting gap and an incorrect yield-fee calculation. All 25 were resolved and the fixes re-audited. Neither critical would at any point have enabled user funds to leave Jetstream or resulted in a loss of funds. Rankin Berkower serves as independent fund auditor. The separation of execution, custody, administration and audit is deliberate, and it is the part of the framework that exists entirely outside the trade.
The three risks position-level neutrality does not cancel
Most of the framework above works ticker by ticker. Three risks hit many names at once, and delta neutrality does nothing about them. We name them rather than distribute them across a dozen mitigants.
Hyperliquid-ecosystem concentration. Bridge design, a relatively small validator set, a closed-source core binary and HIP-3 deployer concentration mean one Hyperliquid failure could impair the short leg across a large part of the book simultaneously. Hyperliquid remains the deepest decentralised book for TradFi perpetuals, so the mitigants are structural: cap single trades, keep Lighter and Coinbase in the mix, and hold the cash leg at IBKR so spot custody never sits on the same stack as a Hyperliquid short.
Funding-regime compression. The strategy needs structurally positive funding, or a positive cross-venue gap, after fees. As professional shorts arrive that premium can shrink below the cost of running the book. Crypto-native basis went through exactly this cycle. Our read is that the skew on TradFi perpetuals is a multi-year phenomenon driven by global demand migrating onto 24/7 markets, and that even a temporary compression leaves this opportunity set well above crypto basis and private credit. That is a view, and it is the view the fund is exposed to.
Forced intervention and ADL. Venues can delist, settle at a chosen mark, or auto-delever profitable shorts. No control fully removes this. Running 1x, spreading across names and venues, and flattening quickly after an ADL print are the available mitigants.
What the framework is for
None of these controls improve the return in a calm market. They exist for the days that decide whether a basis fund compounds for years or ends in a single headline. A 16% net target is only worth pursuing if the structure underneath it is built to still be standing after the market does something unexpected, and that is the standard we hold this framework to.
For professional and accredited investors only. Access is subject to KYC, accreditation and jurisdictional eligibility determined by the Fund and its administrator; subscriptions from OFAC-sanctioned and other restricted jurisdictions are prohibited. In Hong Kong this material is directed only at professional investors within Schedule 1 of the SFO (Cap. 571). Not SFC-authorised. Strictly private and confidential. This is not an offer or solicitation. All terms and risk controls are indicative and subject to finalisation in the Offering Documents, which prevail over anything stated here. Risk controls reduce but do not eliminate the risk of loss, and automated actions may execute incorrectly, partially, or with delay, particularly in tail events, venue outages or oracle failures. Targets are not guarantees and past performance is not indicative of future results.