DN evidence method · Updated 18 November 2026
What matters
Never act on a financial number you cannot trace to a primary source. AI can write compelling citations, calculations and conclusions while getting the underlying facts wrong. NIST describes this risk as confabulation, including false citations and confidently incorrect claims.
The eight checks, with worked examples
| Verification habit | Illustrative failure (hypothetical) | Action |
|---|---|---|
| 1. Pin the claim to a date | An assistant says a token has a $2 billion market cap. Ask: as of what UTC date, from which supply definition and price feed? | Record time, currency, unit and reporting period. |
| 2. Inspect the original document | A model claims a company holds 50,000 BTC. Find the original filing or treasury disclosure, not a repost. | Save the filing URL, section and disclosure date. |
| 3. Rebuild the arithmetic | AI says a 12% annual yield earns $1,200 monthly on $10,000. Recalculate: 0.12 × $10,000 ÷ 12 = $100 before fees, if using a simple monthly equivalent. | Compute manually or with a second calculator; distinguish APR and APY. |
| 4. Check the denominator | A firm claims 95% trading accuracy based on 19 correct examples from 20 cherry-picked signals. | Find sample size, out-of-sample performance and fees. |
| 5. Verify entity and instrument | An AI quotes a company’s revenue but accidentally uses the parent group’s consolidated revenue. | Check the exact issuer, ticker, subsidiary and reporting currency. |
| 6. Separate observed data from an estimate | A dashboard prints a $200 million on-chain market size from extrapolated addresses. | Label the number estimated; disclose model assumptions and exclusions. |
| 7. Challenge stale or inaccessible sources | An agent cites an old terms page and says a product is available in every country. | Open official current terms and regional eligibility before use. |
| 8. Independently confirm material decisions | The model recommends routing an expensive cross-chain payment through a bridge. | Have a human verify recipient, chain, current fees, destination acceptance and failure/recovery procedure. |
DN Financial Claim Verification Scorecard
This is a free editorial checklist, not a numerical prediction of truth. Four or fewer checks means the claim is not ready for a consequential decision.
Check the evidence you actually possess:
Worked example: the $1,200-per-month yield claim
Suppose an assistant states: ‘A 12% return on $10,000 produces $1,200 per month.’ The base arithmetic is wrong: simple annual return is $1,200 a year, or approximately $100 per month if spread evenly, before tax, costs, slippage, variable rates and compounding assumptions. A rate quotation is not a promised return. A displayed APY also must not be treated as fixed or risk-free.
Save four audit fields: original statement; primary source; recalculation; final verdict. A second AI answer is a lead, not a substitute for evidence.
When to escalate to human review
- Before any transaction, leverage increase, bridge transfer, portfolio allocation or credit decision.
- When the source is a secondary article rather than a company filing, contract, public ledger or regulator notice.
- When the claim depends on fast-changing prices, jurisdiction or instrument-specific restrictions.
- When an unverified answer might be redistributed as a headline, social post or investment memo.
Methodology, sources and limitations
- NIST AI Risk Management Framework, including its generative-AI profile and discussion of fabricated content.
- SEC EDGAR, an example primary-source destination for public-company filings.
- South Africa FSCA, regulatory source for relevant local financial-service questions.
The eight examples in this guide are illustrative hypotheticals, not allegations about named AI providers or products. This is an editorial method, not a controlled head-to-head model benchmark. No assistant accuracy statistics were invented.
Disclosure: No provider ranking or affiliate commissions determined this checklist. This article is for educational purposes and is not investment, tax or legal advice.
Frequently asked questions
Does a citation make an AI financial answer correct?
No. A citation can be irrelevant, outdated, invented, or misread. Verify the source passage and calculations.
Can a second AI model validate the first?
It may flag inconsistencies, but shared training data and common sources can reproduce the same mistake. Independent original evidence matters more.
What should I do when I cannot verify a financial number?
Mark it unverified, do not act on it, and obtain the original filing, transaction record or first-party measurement.




