Method truth-1. Every verdict links to the records it used, so you can check it yourself.
Promises (a speaker will do something by a stated time), forecasts (a measurable quantity will reach a level by a stated time) and factual claims (a measurable quantity is or was at a level). We read SEC filings and press releases, members of Congress’s own press releases, the Congressional Record, the White House, governors’ newsrooms, Federal Register presidential documents, video transcripts and headlines.
A statement counts toward a speaker only when it is in their own document (their press release, their floor remarks, their company’s filing) or a quotation attributed to them by name with a speech verb (“said”, “pledged”) right next to it. Headlines count only in the “X says …” form.
Numbers are checked against the official series: FRED and BLS for the economy, and the company’s own XBRL filings with the SEC for revenue, earnings per share and headcount. A factual claim is compared with the latest value published by the day it was made; a forecast with the value for the period it named. Within the metric’s tolerance is true, within twice the tolerance (or on the wrong side by less than the tolerance) is partly true, and anything further off is false. Promises are checked against layoff notices, careers pages, bills sponsored and votes cast on Congress.gov, executive orders in the Federal Register and the news.
AI reads documents for statements rules miss and judges promises no official record settles. An AI verdict has to cite the evidence it used, its confidence is capped at 0.6, and anything under 0.5 is reported as unverifiable and left out of the score.
The Truth Score is the share of a speaker’s checked statements that held up, weighting each by the confidence of its check and counting “partly” as half. It is smoothed toward 50 with a Beta(2, 2) prior so a handful of checks can’t produce a 0 or 100, and shown with an 80% interval. A speaker needs at least 5 checked statements before a score is published. Unverifiable statements and ones not yet due are listed but never scored.
Speakers who say more checkable things get more checks, so scores compare best between similar speakers. Layoff checks cover the WARN states we track. Official series get revised; we compare with the current vintage and show the value we used. If you think a verdict is wrong, report it on the statement’s page; reports are reviewed and the correction is shown on the page.
| Measure | Source | Counts as true within |
|---|---|---|
| US unemployment rate | FRED UNRATE, or BLS LNS14000000 | 0.2 points |
| Inflation (CPI, year on year) | FRED CPIAUCSL, or BLS CUUR0000SA0 | 0.3 points |
| Core PCE inflation (year on year) | FRED PCEPILFE | 0.3 points |
| Real GDP growth (annualized, quarterly) | FRED A191RL1Q225SBEA | 0.5 points |
| Jobs added (nonfarm payrolls, thousands) | FRED PAYEMS, or BLS CES0000000001 | 15% |
| US regular gasoline price ($ per gallon) | FRED GASREGW | $0.15 |
| 30-year mortgage rate | FRED MORTGAGE30US | 0.15 points |
| Federal funds rate | FRED FEDFUNDS | 0.15 points |
| Federal budget deficit (fiscal year, $) | FRED FYFSD | 5% |
| Revenue (as filed with the SEC) | The company's XBRL filings with the SEC | 3% |
| Diluted earnings per share (as filed) | The company's XBRL filings with the SEC | $0.02 |
| Employees (as filed, where reported) | The company's XBRL filings with the SEC | 5% |
A speaker needs 5 checked statements before we publish a score. Until then the page lists the statements and their verdicts without a number.