Method

O Estado do País is written by artificial intelligence and directed by a person.

It is directed by Nuno dos Santos, who chooses the subjects, sets the rules of this method and answers, ultimately, for what is published here. The writing, the research and the verification are done by artificial intelligence models (Claude, by Anthropic), under that direction.

Every page declares this authorship in the footer. There are no exceptions: if it is published here, it was written by AI and reviewed according to this method.

We say this in the first paragraph because it is the right question to ask of this site. The answer that follows — the ledger, the public corrections, the declared limits — exists so that trust does not depend on who writes, but on what any reader can verify.

A study begins with a question — from the director, or proposed by the AI itself and approved by the director. From there:

1. Research in primary sources. Statistical institutes, official publications, original documents. A press summary is never the source.
2. Writing on top of the ledger. The text contains no hand-written numbers: each value is resolved from a ledger row at the moment the page is built. A number without a row is not published — it is flagged [a verificar] (to verify), or it is cut.
3. Verification by someone who did not write it. Central claims are verified by agents that took no part in the writing, instructed to refute them. Headline claims go through blind re-derivation: another agent redoes the calculation without seeing the text.
4. Review by the director before any publication.

What does not survive this process is not softened — it is removed, or published as declared uncertainty.

Every number on this site traces back to a row of a ledger of claims. Each row holds: the exact value as published by the source; the entity that produced it; the document and the edition; the URL; the access date; a textual excerpt from the source; and, when the value is calculated by us, the full derivation.

The provenance seal beside each number is the door to that row. The site build fails — deliberately — if a page tries to publish a number without a corresponding row.

The data behind each graphic is downloadable. If we cite, we show; if we calculated, we show how.

Correcting in silence is the cheapest way of lying. Here, every correction is public, dated and permanent: previous value, corrected value, reason, and a link to the ledger row that changed. This page’s corrections log lists them all, from the most recent to the first, and none is removed.

Anyone who finds an error can report it to [endereço a confirmar] (address to confirm). A confirmed error enters the log with credit to whoever found it, if they wish.

The most delicate part of this observatory is linking outcomes to those who govern. The rule is management accountability, not attribution of merit:

• Inherited, decided, left. Each term is read by what it inherited, by what it decided, and by the state in which it handed over — never by a single number of merit.
• Metrics announced before verdicts. The indicators used to read a term are published before the reading, not chosen afterwards.
• Confounders declared. Much of what happens in a municipality is decided outside it — in Lisbon, in Brussels, in demography, in the economic cycle. And there is institutional machinery that breaks naive readings: in the 2025 local elections, a substantial share of councils changed president through term limits, not through the vote [a verificar: número exato antes de publicar] (to verify: exact number before publishing). These limits are declared alongside each reading.
• No party rankings. Averages by party without controlling for the territory each party governs are misleading arithmetic. We do not publish them.
• Signed inference. Where this site concludes beyond the data, the conclusion is marked as inference and signed as such.

• This site does not know what the sources do not measure. Where the data does not exist, we say it does not exist — a null answer is an answer.
• The AI that writes this can err in ways of its own: inventing fluently is the central risk, and that is exactly why the ledger, adversarial verification and public corrections are not optional.
• Official data arrives late and is revised. The access date on each seal says when we read it; if the source changed afterwards, the discrepancy is treated as a correction, not as an argument.
• Independence: this site receives no funding from parties, governments or municipal councils. It is funded personally by the director, and has no advertising and no external funding.