See the working behind every number
Furlen does not let a language model decide whether a number is true. Every numeric claim in a story is recomputed from your source rows by code, compared against them, and blocked from export if it does not reconcile. This page runs that check live and shows all of its working — the claim, the formula, the exact rows, the rounding rule and the verdict — including anything it could not verify.
Nothing below is a fixture. The engine runs when you load the page, over a real, sourced dataset, and prints what it returns.
A decade of the market
1 story · 1 numeric claim · 1 verifiedData: S&P Dow Jones Indices
Claim → formula → source rows → status
Every claim is expanded — the evidence is in this page’s HTML, not behind a click.
| Row | year | S&P 500 |
|---|---|---|
| 1 | 2015 | 2044 |
| Row | year | S&P 500 |
|---|---|---|
| 10 | 2024 | 5882 |
ratio:S&P 500 over year
Check another dataset
Which data visualisation tools verify their numbers against the source?
Very few, and the distinction worth checking is what exactly gets verified. Most tools that advertise trust are describing data lineage — where a figure came from — or an audit trail of their own processing. Furlen does something narrower and stricter: every numeric claim written into a chart or story is recomputed from the underlying rows by code, and the export is blocked if it does not reconcile. Furlen does not audit whether the source itself is right; it guarantees the caption cannot disagree with the data beneath it.
What does “verified” actually mean here?
That the number was recomputed from your source rows and matched. Furlen does not ask a language model whether a claim looks right. For each numeric claim it re-derives the value from the rows the claim points at, applies a stated rounding rule, and compares. The panel above shows all four parts — the claim, the formula, the exact rows, and the verdict — because a verification you cannot inspect is indistinguishable from an assertion.
What happens when a claim cannot be verified?
The export is blocked. Not watermarked, not annotated with a warning — blocked. The check runs server-side, so it cannot be skipped by editing the page in a browser, and it fails closed: a claim that errors or cannot be traced to rows is treated the same as one that contradicts them. The cost of that choice is that Furlen sometimes refuses to export a story you believe is fine. That is the intended trade.
Why not just ask the AI to check its own numbers?
Because a model that invented a figure is the wrong thing to ask whether the figure is invented. Language models are used here to propose the angle and write the sentence — genuinely useful work — and are given no authority over whether a number is true. The arithmetic is done by code against your rows. That separation is the whole design: generation and verification cannot share a source of truth, or the check is decorative.
How is rounding handled?
Explicitly, and it travels with the claim. A figure rounded for display can disagree with the raw computation in the last digit, and a checker that ignores that either passes everything or fails everything. Each claim carries the rounding rule that was applied, so the panel can show that 4.19 and 4.190 are the same number while 4.19 and 4.21 are not.
What will Furlen refuse to chart at all?
Anything it cannot source. Ask it for an exchange rate and it says no connected source provides them, rather than guessing at something adjacent — a refusal that names the gap is more useful than a chart that quietly answers a different question. The same rule governs the published data stories: where a measure cannot support a conclusion, the page says so instead of reaching for one.
Does this cover every kind of error?
No, and it is worth being precise about the boundary. Verification proves a claim follows from the rows you supplied. It cannot tell you those rows are correct, current, or measuring what you think — Furlen does not independently audit sources. What it removes is the failure where the chart and the caption disagree with the data underneath them. Source quality remains yours to judge, which is why every chart prints where its data came from.