Buying evidence, not opinions.

Why a research agent

A strategy can only be as good as the facts behind it. When Crosswake evaluates a coin, some of what matters is already in our own data and some of it is not.

Rather than assume a missing metric must be bought, Crosswake first works out what it genuinely cannot measure itself. That list is short, it is written down, and it shrinks as our own coverage improves.

How a purchase would be decided

Every step below is ordinary code, not a language model’s judgement. The model may ask for a purchase. It can never make one.

• What is missing: the metric needed is compared against evidence already held, and anything already covered and fresh is ignored.
• Who sells it: only providers an operator has registered and verified are considered. Prices come from the seller’s own quote, never from a model.
• Is it worth the price: the offer is filtered against fixed caps and allowlists, then scored on relevance, freshness, coverage and reliability. Cheaper is not automatically better.
• Who authorises it: a person approves each purchase. Hard caps apply per payment, per research run and per day, and an emergency stop halts spending immediately.
• Was the data any good: a paid payload is checked for schema, asset, freshness, duplicates and hidden instructions before it is used. A settled payment does not prove the data is correct.

Where Hyperliquid fits

Hyperliquid is an on-chain perpetuals exchange whose public data needs no key and no payment. Crosswake already reads its prices, funding and open interest, and now reads its order book as well.

That matters for cost. Order-book depth and market impact were two of the metrics Crosswake could not measure itself, so they sat on the list of things worth buying. With Hyperliquid’s book they no longer do: the research run measures its own liquidity, and the buy list shrinks to what genuinely cannot be obtained for free.

One caveat, stated plainly. Hyperliquid’s order book is for perpetuals, so it is a cross-venue proxy for spot liquidity rather than a measurement of the spot market. Our evidence records it that way.

What is live, and what is not

Live: Hyperliquid market data feeds research, including prices, funding, open interest and order-book depth.

Not live: the purchase path. It runs against local fixtures in simulation. No real money has moved, no on-chain payment has been made, and no purchase has been authorised by a person.

Buying data never changes a signal on its own. Paid evidence is recorded separately from our own measurements, and it cannot raise confidence without a rule that can be tested. No such rule exists yet.

Hard limits

These are enforced in code, independently of anything the model is told.

• The model cannot sign a payment, move funds, change a spending cap, change a recipient, or approve its own request.
• Only registered providers, approved recipients, one blockchain network and one payment asset are reachable.
• Every purchase attempt is written to an append-only ledger, including the attempts that were refused and the reason.