Find every number an LLM asserted, in one table
Language models state statistics with total confidence whether or not they are true. “Revenue grew 47% to $3.2M, beating the 12% industry average” reads authoritatively — and may be entirely invented. This tool scans LLM output, extracts every numeric claim with the sentence around it, and lays them out so you can fact-check fast.
Numeric hallucinations are particularly dangerous because they are easy to miss on a first read — a plausible percentage in a business summary, a market size figure in a research brief, a conversion rate in a marketing analysis. The number looks credible next to an authoritative-sounding sentence. Pulling every figure into a table makes the act of verification systematic rather than incidental.
How it works
The extractor matches several families of numbers: percentages, currency amounts, multipliers (“3x”), years, and plain counts, including those written with thousands separators or decimals. For each match it captures the surrounding sentence as the claim, labels the number by type, and puts it in a table. Everything runs locally in your browser, so you can paste confidential drafts safely.
Number type labels and what they mean
| Type | Examples caught |
|---|---|
| Percentage | 47%, 3.2%, 100% |
| Currency | $3.2M, £500, €1,200 |
| Multiplier | 3x, 10x, 2.5x |
| Count/Quantity | 1,000 users, 42 days, 500 employees |
| Year | 2019, 2024, 1998 |
| Measurement | 150 km, 32°C, 4.5 kg |
Each type is labelled in the table so you can sort or filter by the categories most relevant to your fact-check. Percentages and currency figures are usually the highest-risk claims in business and research text; start there.
Tips for fact-checking output
- Check the high-impact figures first — percentages and currency amounts drive decisions, so verify those before counts.
- Read the claim column, not just the number. A correct figure attached to the wrong subject is still a factual error.
- Watch for suspiciously round or precise numbers. “Exactly 1,000,000 users” and “a 73.6% improvement” are both classic hallucination tells. Real-world data rarely lands on such clean values; very round or very precise figures deserve extra scrutiny.
- Years deserve a check too. An LLM citing “a 2019 study” may be conflating sources and the study may be from a different year — or may not exist.
- Pair with the date extractor to verify the temporal claims in the same text.
- Copy the table for a review audit trail. Pasting the extracted numbers into a document with a “verified / unverified” column makes the fact-check reviewable by a second person without re-reading the full source text.
- Numbers in parentheses — for example
(47%)in a disclaimer — are often important context, not primary claims; read the surrounding sentence carefully before marking them as verified.