An answer can be grammatically polished and still name the wrong contract, invert a funding direction or describe a feature that is not deployed. Treat an AI response as a lead to investigate, not a citation or a second analyst who has confirmed the claim.
Sort the claim before searching
First write one sentence that could be proven or disproven. Then identify what kind of evidence it needs. A live venue mechanic needs the current documentation for that venue, contract and account mode. A historical result needs the dataset, code, dates, assumptions and a reproducible calculation. A product claim needs the actual interface or deployed behavior. A legal claim needs a current primary regulator or statutory source and an expert review appropriate to the jurisdiction.
Do not let a broad article or search snippet stand in for the exact evidence. A page about perpetual futures does not prove a particular account’s funding debit. A generic description of stop orders does not tell you what happens in every venue’s matching engine. If the necessary source is unavailable, narrow the statement or mark it unresolved.
Recalculate a small example
Suppose a draft says: “Hyperliquid charges 0.01% funding every hour, and longs always pay.” That sentence combines a rate, a clock and a payer rule, so check each part against the current documentation. Hyperliquid describes funding as hourly. Its fixed interest input is 0.01% per eight hours, or 0.00125% per hour; a variable premium component also contributes to the total rate. Under its stated convention, positive funding means longs pay shorts, while negative funding reverses the direction. “Always” is wrong.
To audit the arithmetic under a deliberately narrow assumption, set the average premium to zero and hold fictional oracle-price notional at $10,000 for one full hour. Under that case, the documented interest input contributes 0.00125% for the hour:
$10,000 × 0.0000125 = $0.125 for that hour
At the same assumed notional for eight unchanged hourly intervals and zero premium, the interest input sums to $1.00. This is a formula and unit check, not an account payment forecast: the premium can change the total rate and payer, and the position’s oracle notional can change. The venue documentation defines the formula and conversion. Never annualize one observation as though it were guaranteed to repeat.
This check can be reproduced in a scratch sheet: preserve the original sentence, highlight “0.01%,” “every hour” and “always,” source each separately, convert percent to decimal, multiply by the stated notional and interval, and write the narrowest supported replacement. If one term cannot be verified, leave it unresolved instead of letting the other two make the sentence sound settled.
For code, test a small input by hand and then check boundaries: zero size, missing data, negative values where prohibited, timestamps at interval edges and a rejected or partial order. Save the source version, output and environment. A deterministic script can reproduce a wrong assumption perfectly.
Keep an evidence ledger
| Claim | Evidence to retain | Decision |
|---|---|---|
| “Funding is 0.01% each hour” | Venue, instrument, date, endpoint or contract page, timestamped observation and units | Accept only for the stated observation and venue scope |
| “This backtest returned 40%” | Dataset fingerprint, strategy version, execution model, dates, costs and rerun output | Label as simulated and disclose limits |
| “The app submits orders” | Deployed route, authorization, venue response, fill and account reconciliation | Reject unless the full live path is verified |
| “This is a legal requirement” | Current primary law or regulator source, jurisdiction and qualified review | Escalate if scope is unclear |
The ledger should also preserve the original AI prompt or draft, the correction, who checked it and when. Separate sourced fact, calculation, hypothesis and fictional example. If a later reviewer finds an error, record the correction instead of silently replacing the earlier text.
Use the free AI trading claim audit worksheet to keep the exact statement, source scope, independent reproduction, unresolved limits and versioned review together. Do not include credentials, account identifiers or private user data in the record.
Use AI risk guidance in the right way
NIST’s Generative AI Profile, published in 2024 and updated on its page in April 2026, is voluntary, cross-sector risk-management guidance. It can help an organization think about evaluating and governing generative AI, but it does not certify a trading answer, strategy or model. The CFTC separately warns consumers about AI-branded trading-bot pitches that promise extreme or guaranteed returns; its examples concern fraud claims and do not evaluate every AI tool or establish that a model can or cannot forecast prices.
For a current exchange claim, consult the exchange itself. For SuperQuant product boundaries, consult the research-methodology page and verify any time-sensitive statement against the deployed product. The local research workspace uses synthetic histories, simplified execution assumptions and manual paper practice; those are not live account results.
Decide whether to use, revise or reject
Keep a claim only when its source supports the exact wording, its calculation can be reproduced and its limits are visible. Revise it when evidence supports a narrower statement. Reject it when the source is missing, the numbers do not reconcile or the prose quietly turns a historical observation into a forecast. Ask a qualified human to review consequential financial mechanics before publication or operational use.
The backtest workflow, data-quality guide and trading journal show how to retain evidence across research stages. This is a claim-checking method, not financial advice or a guarantee against model error.