Methodology

1) Every 30 minutes we crawl hundreds of outlets, respecting robots.txt. 2) We group articles telling the same story. 3) We compare who reports it, which figures each gives and when, what most report and some omit, and the loaded words each outlet uses. We only compare figures about the same fact: the headline figure counts, and two figures more than three times apart are different facts unless both are attributed ("according to…"). 4) We extract only explicit numeric claims about official indicators and compare them with the Eurostat, INE or World Bank series for the period the sentence cites. To find more, an AI reads sentences with figures and proposes indicator, territory and period; our rules check that the sentence and figure are in the text and that territory and period fit, or it is not judged. The verdict is always given by code against the official data, never by the AI. 5) Verdicts: True (up to 0.2 points or 3%), Imprecise (up to 0.6 points or 10%), False (more), Misleading (claimed trend contradicts the data) and Not enough data. We never judge with data older than 400 days. Every verdict links its source. Same rules for every outlet. We respect robots.txt. Only exception: when checking a specific figure needs a specific document (e.g. on a transparency portal) because outlets disagree or proof is missing, and it cannot be obtained otherwise, we reach that document in at most 5 pages, with our real identifier, stopping at any block; every visited page and the reason are kept in the evidence.

True Imprecise Falso Misleading Not enough data

Advertising plays no part in the method: it does not change verdicts, checked claims or the order of the news, and it is chosen only from the page content, with no cookies or profiles. Advertising on nosmienten.com.

What doesn't add up

Weight = verdict × deviation from official data × outlets repeating the figure × recency (halved every 30 days). The same person, party or outlet gives different figures for the same indicator, territory and period.

People and entities

Automatic extraction: names from our curated registry and proper names found in the text (a discovered entity is published once at least 2 outlets name it). A sentence is attributed to a speaker only with a reporting verb or "according to". Archive stories come from the GDELT archive and are only cited: headline, date and link.