Rules and the AI translator
What a rule is, how the translator turns your words into one, and why it can only add protection.
A rule
A rule is a trigger and one or more actions, optionally limited to a fixed window:
| Triggers | Actions |
|---|---|
| Buffer below a line | Alert |
| Account down by a share since the window opened or since you signed | Move idle USDC into the pool (the amount you set) |
| A market moving up or down by a share | Cut a position by a share, or close it |
| A position's leverage above a level | Trim until the buffer is back at a line, or a market back to a leverage |
| Cancel orders that would add to a position |
Windows are a fixed table, not numbers: weekend (Friday US close to Monday US open), overnight (US close to the next US open) and US session (9:30–16:00 New York time on weekdays).
Once per fall, or every time
Every rule carries one more setting, which you choose and the app never pre-selects:
- Once per fall: act once when the trigger is met, then leave the rest of the fall to the backstop. It acts again only after the market recovers to where it was when it acted.
- Every time: act each time the trigger is met again, including after the guard's own trim lifted your buffer.
Optionally: at most N actions in H hours, with both numbers typed by you. How each behaves.
In a policy this is the rule's repeat field: { "mode": "oncePerBreach" | "everyCrossing", "limit"?: { "times": N, "perHours": H } }. A policy version can't be signed until every rule has it.
Every number is yours
The app never pre-fills a number and never suggests one. Every line, share and amount in your rules is a number you typed.
The AI translator
You can write a rule in a sentence, for example "close gold if it falls [your number]% overnight". An OpenAI model (GPT-6.1 Sol, through OpenAI's API) turns it into one exact rule. The app labels it as the AI translator and names the provider next to it.
- What it sees: your sentence, the market list and your current rules, sent to OpenAI for this one request and not stored there for later retrieval (the request is made with storage off). No keys, no account data beyond those rules, no tools, and no access to the exchange. See Privacy.
- Numbers: every number in a draft must be one you typed; otherwise the draft is refused before you see it.
- Only adds protection: a draft can add a rule. It cannot change, loosen or remove an existing rule or your slippage limit. Loosening is a manual edit with its own signature.
- Once or every time, only from your words: the translator sets this only if your sentence says it ("only once", "every time"). If it doesn't, you get a question back, whatever the model would have picked. A plain "once the buffer drops…" or "whenever…" doesn't count, because those usually mean "when". A limit's numbers must be ones you typed.
- Asks or refuses: if a number, market or action is missing, it asks one question. If the request isn't a protective rule (raising leverage, opening positions, sending funds), it refuses.
- What you see: the rule shown back to you is generated from the rule's structure, not the model's wording, so every number on screen comes from the rule.
- You sign it: nothing applies until you sign the new policy version in your wallet.
Status
The translator is switched off. It switches on in the testnet preview only after it passes our full test set (including every adversarial sentence) with the live model. Until then, use the stage editor.
Signing a policy
A policy is your full set of rules plus your slippage limit, with a version number. You sign its hash (EIP-712) in your wallet. The guard acts only under the policy whose hash matches your stored signature (check I4).