Pre-register a strategy

Declare a strategy before it runs and we publish its record whatever happens. That earns Grade A−, the highest grade available without per-decision attestation, and it costs nothing.

Registering a strategy that is already running

You can, and most people will, nobody pre-registers a strategy they have not yet decided to run. It is worth doing and it is worth being precise about what it buys.

A declaration attests forward, never backward. If the vault has been trading for a year, the model, the prompt and the parameters could have changed any number of times in that year and no signature covers it. From the moment you sign, they cannot change without a new declaration, and the registry shows exactly where that line falls, with the unattested period hatched and counted in days.

We take the start date from the venue’s own creation timestamp where there is one, and from the earliest data we hold otherwise. The second case is a floor: it started no later than that, possibly earlier, and it is labelled as such rather than presented as the answer.

Why we ask for this before rather than after

Every leaderboard in this industry shows the strategies that worked, because those are the ones somebody chose to publish. The failures are not hidden so much as never submitted, which makes the denominator unknowable and every average drawn from it wrong.

Medicine had the same problem and fixed it with mandatory trial pre-registration, not by asking researchers to confess. Nobody volunteers a confession; anybody will fill in a form about something they are optimistic about. So: tell us now, and the record counts either way.

This is a commitment, and it is meant to be. A declaration cannot be edited, any change breaks the signature, and it is never deleted. A strategy that quietly disappeared when it went wrong would recreate the exact bias this exists to remove. 3 strategies have been declared so far.

Registering Topiary Capital on hyperliquid/hyperliquid. The address, network and venue below are filled in from that vault, check them anyway, since the signature covers them.

Vault or leader addressNetworkVenue

The same address exists on every EVM network and means a different vault on each, so the network is part of what you sign.

Who decides the trades

Autonomous agentHumanHybrid

A model decides and executes with nobody in the loop.

We infer this independently from execution timing and publish both. A claim that disagrees with the behaviour is shown as a disagreement rather than resolved silently, that is more useful than either alone.

Strategy name

What it is meant to doIn your words. It is published as written, and it is what your record will be read against.

Instruments, comma separatedMax leverageMax drawdown %

Leave a ceiling blank rather than inventing one. “Not committed” is honest; a limit you exceed in month two is worse than none, because the record will show it.

Strategy hash, optionalEmail, optional, only so we can reach you

What is deciding

Without this we would be rating “an AI agent” without knowing whether that means a frontier model, a linear regression, or a person with a spreadsheet. Every field is optional and every omission is published as an omission; “not disclosed” is an answer, and readers can weigh it.

Model

The buttons fill the field; the field is what gets signed. Type the exact name if the button is only close, a family is a weaker claim than a version, and the specific one is the one worth making.

Pinned version or snapshotSystem prompt hashConfig hash, temperature, tools, limitsCode referenceWhat it is allowed to read

Changing any of these is a new strategy, not an edit. The declaration you sign says so. File a new one pointing back at this, the old record stays exactly as signed, which is the only reason it is worth anything.

Or lodge the real thing with us and let us compute the hashes

A hash you compute proves the thing did not change. It does not prove it is anything at all, it could be an empty file and neither we nor a reader could tell. Paste the real text and we compute the hash ourselves, so we can say we hold something that matches the published value.

Never published, never returned through the API. The cost is that you have to trust us with it, which is why all three are optional and empty by default. Publishing them openly is stronger than either.

System promptConfigurationStrategy

Declared on 2026-10-05, today, and not selectable. A declaration whose date you could choose would be one you could backdate, and a backdated pre-registration is worth nothing at all: the whole claim is that it was made before the results were known. The date is inside the signed text for the same reason.

An address, a name and a thesis first.

What we do and do not see

• The strategy hash is optional and we never see what it hashes. It only lets you show later that the model, prompt or parameters are the ones you declared.
• Signing costs nothing and moves nothing. It is a message, not a transaction, there is no gas and no approval of any kind.
• Everything except your email is published, including the declaration text and the signature, so that anyone can check it themselves rather than take our word for it.

Already declared? The registry lists every strategy and where it got to.

Registering a strategy that is already running

You can, and most people will, nobody pre-registers a strategy they have not yet decided to run. It is worth doing and it is worth being precise about what it buys.

Why we ask for this before rather than after

What we do and do not see

• The strategy hash is optional and we never see what it hashes. It only lets you show later that the model, prompt or parameters are the ones you declared.

• Signing costs nothing and moves nothing. It is a message, not a transaction, there is no gas and no approval of any kind.

• Everything except your email is published, including the declaration text and the signature, so that anyone can check it themselves rather than take our word for it.

Already declared? The registry lists every strategy and where it got to.