← Check a claim

How a claim is checked

The method, the verdicts and the limits. Last updated November 17, 2026.

The short version

A claim is read for three things: what it measures, where, and when. That figure is then fetched from the agency that publishes it, and the claim’s number is compared with the agency’s. The answer shows the official figure, the rows it was read from and a link to the series, so you can check the check.

Step by step

  1. The text is cut into claims. A sentence that states two statistics (“unemployment was 3.6% and inflation was 4.1% in 2023”) is two claims. A comparison (“from 5.2% in 2021 to 3.6% in 2023”) stays one. If a year or a place is stated once for several claims, it is added to each, and the answer shows the claim as it was checked.
  2. Each claim is read by fixed rules, not guessed: the measure (the unemployment rate, CPI inflation, the national debt), the place (the nation, a state, a county, a city) and the period (a year, a quarter, a month, or “now”).
  3. The figure comes from the agency’s own data service (see the list of sources), never from a news story or a summary of one.
  4. The two numbers are compared by the rules below.
  5. The answer names its source: the agency, the dataset, the series or table, the period, and the published rows the figure was read from, with the cited row marked.

Text nobody picked sentence by sentence

A passage, or a page checked by its link, holds sentences that name a measure and contain a number without stating that measure’s value (“the Panic of 1837 brought unprecedented unemployment”). So in such text a sentence is checked only when its figure is written the way that measure is written: a percent sign for a rate, a dollar sign or “million” for money, never a bare year. And if the official figure found is twenty times the claim’s or more, no verdict is given, because the claim was probably about something else. The figure must also sit within a few words of the measure it is said to be a value of, and a figure that says how far a rate moved (“rose 0.3%”) is not taken for the rate. A page is read from the public web, as any visitor without an account would see it; sites that refuse automated readers are not read.

A month without its year. A news release says “in October” and leaves the year to its date. When a page states the day it was published, a month, a quarter, “last month” or “last year” named without a year is read against that day, and the result says so (“period from the page’s date”). This is not done for a page that gives no date, for one reworked long after it came out, or for the month the page itself came out in, which no agency had reported on yet. Where the text nearby names some other year, such a sentence is listed only if the official figure agrees with the reading. The same holds for a sentence that names no place when the text before it is about another country or a state: read alone it would be the United States’, and it may not be.

Which figure a claim is held to

Every answer shows the measure, place and period it was read as, so a wrong reading is visible.

What the verdicts mean

Supported

The official figure matches the claim to the precision the claim was written to. “9%” agrees with 9.06%; “3.7%” does not agree with 4.0%. For “over”, “under”, “at least” and “at most”, the claim’s number is a threshold: “the debt is over $33 trillion” is supported by $40 trillion. Rounding never covers a gap of more than 10%.

Accurate when published

The claim matches the figure the agency first released, and the agency has revised it since. Both figures are shown, with the date of the archive the first one was read from. This is checked for the unemployment rate, for the jobs added in a month, for real GDP growth by quarter and year, and for a year’s GDP in dollars. The first release is tried, then the two that followed it.

Mostly accurate

The claim is off by one step in its last written digit, and by no more than 5% of the official figure: 4.1% against 4.0%. The size of the gap is stated.

Contradicted

The official figure is further from the claim than that.

Misleading

The number is real, but the claim uses it in a way the record does not support: a figure that is right alongside a record (“the lowest since 1969”) that is not, for example.

Can’t verify

No verdict is given, and the answer says why: no official series covers the claim, the figure for that period is not published yet, the claim compares points in time that one figure cannot settle, or the claim is an opinion or a prediction. A source that failed to respond is reported as an outage, never as a judgment about the claim.

Confidence

High, medium or low, never a percentage. High means the agency’s own published figure for the period and place in the claim. Medium means the same with a note to read, such as a revision. For an answer found by search, high means the page of the agency that publishes the figure, giving the figure and its period, or a reliable page that a published fact-check agrees with; medium means a reliable page alone, or an official page that leaves out the figure or the period.

When no data service covers a claim

Most checks end at the step above and use no model at all. When the fixed rules cannot read a claim, or no wired series covers it, a single claim (not a passage) may be searched for in the agencies’ datasets and then in published sources, with the help of an AI provider named in the privacy policy. Those answers are marked “Found by web search” and link the pages they rest on. A verdict stands only if the page it rests on was really returned by the search and is official or reliable (a professional fact-checker, a research institute, a journal or a newspaper of record); otherwise none is given. A passage of several claims, and a shared link, are checked on agency data only.

What it does not do

Revisions and dates

Agencies revise what they publish. Every answer carries the date it was checked, and figures for an open period are fetched again within the hour. A check you saved earlier can be run again.

How it is tested

A set of plain claims whose answers were read from the agencies’ own data is run against the live service, on agency data only. The last run, on November 17, 2026, checked 28 claims: 28 were answered correctly, 5 of them by rightly declining, and no model was called. That set is small and its claims are simple, so it shows the checker reads and compares correctly on the kinds of claim it is built for; it is not a measure of how often any claim at all gets an answer. The claims, the expected answers and the full results are in the project’s public code, and the run stops without reporting a score if any source was down.

Corrections

If an answer is wrong, open an issue with the claim and what the agency’s record says. Fixes are made in the open, in the project’s public code, and each comes with a test that holds the corrected case.

Independence

Project Mirador is an independent project with no ads and no accounts. It is not a government service, and is not affiliated with or endorsed by any agency whose data it shows. The same rules are applied to every claim, whoever made it.