Every claim.
Checked. Sourced.
Paste a speech, briefing or article. Each claim that official figures can test comes back with a verdict, the numbers behind it and a fairer wording. Numbers come from code, not from a model.
This sheet
Draft- Data as of
- Method version
- Checked
Every figure keeps its claim ID and the release it came from, so each can be traced and reproduced as of the same date. Parts of a claim we cannot test are named on the card, never dropped.
Numbers from code. Words from agents. Every figure keeps its source.
A claim is only called accurate or inaccurate by a fixed rulebook running on official statistics. A language model never decides a verdict and never supplies a number.
- 01
Find the claims
One pass reads your document and quotes each claim exactly as written. Opinions, forecasts, causes and anything vague are listed as such, not tested.
- 02
Say what would make it true
Each testable claim becomes a precise test, such as "fell over twelve months" or "above the 2% target". Who said it is never included, so the speaker cannot change the result.
- 03
Check it in code
The test runs against a versioned store of official figures from the ONS, the Bank of England, NHS England, HM Land Registry and others. Same claim, same data, same answer.
- 04
Show the working
A verdict, the figures behind it, the release each came from and a fairer way to word the claim. Every number is checked against the computed figures before it is shown.
- Neutral by designThe checker is never told which party or person made a claim, and the same rule applies to every party.
- ReproducibleEach result records the method version and data date, so it can be rerun and traced.
- Honest about gapsIf we hold no suitable official figure for a claim, it is marked Can't check, never guessed. Each pass takes up to 8,000 characters and 12 claims.