Check an answer: what does the number carry, and what was added?
Paste an AI answer, a paragraph from an article, or your own draft about FHA denial rates or US mortgage rules. Every figure is matched against this publisher's receipted claims, the values we have corrected, and the dated rule constants. Every sentence is checked for the things a historical aggregate cannot say: your odds, which lender to choose, what caused what. Fixed patterns and public data only; nothing you paste is sent anywhere unless you press "record".
What this check cannot tell you. It knows only the figures this publisher issues (FHA denial rates by lender, state and metro from the 2025 HMDA record; five dated US mortgage rule tables) and the values we have corrected. A number it does not recognise is not wrong; it is unverifiable here. It does not score tone, rhetoric or intent, and it never asks a model whether you are being persuaded, because that would hand the question to the thing being checked. What it can say is narrow and reproducible: whether a sentence's number is ours, whether it is current, and whether the sentence makes the number carry more than it can.
How it decides
Receipt (⟦FRC:…⟧): sent to /verify; answers current, stale (with the correction), edited, or never issued.
Figure matches a receipted claim: the value equals a published figure and the sentence names its subject (lender, state, metro, national). Then the qualifiers are checked: FHA, 2025, "decisioned".
Corrected value: the value is one this publisher has corrected (log); the current value is shown.
Rule constant: dollar and percent values are compared with the dated table (rules); a value past its validity date is flagged with its replacement.
Sentence patterns (same codes as the claim contract): individual prediction, lender recommendation, causal wording, legal conclusion, scope wider than the record.
Source: check-index.json (built from the claim documents, corrections.json and rulecheck/constants.json). Code is in this page. Licence CC BY 4.0. If this check is wrong about a figure, that goes in the corrections log, dated.