How we verify

1. Public, official sources only

We collect only documents published by public institutions: courts, prosecutors, the Federal Police, the Central Bank, the CVM and similar bodies. Each source has a declared list of allowed addresses; we honour every site's robots.txt and limit how often we visit.

Each document is archived when collected, with date, address and fingerprint (SHA-256). If the institution changes or removes the text, the earlier version stays on record and the change appears in Updates.

2. Extraction with a mandatory citation

An AI model reads each document and proposes people, organizations, events and relationships. Every proposal must come with the verbatim passage that supports it and the page it is on. A proposal without a passage is discarded.

Dates and amounts in reais are not interpreted by the AI: code reads them straight from the quoted passage.

3. Three independent checks

First, rules in code: the passage must exist word for word on the stated page; the names, dates and amounts must be in the passage; whoever makes the statement must be identified; words such as "guilty" or "criminal" pass only when the source is a ruling that uses them.

Next, automated evaluators (Jev) score whether the passage really supports the statement, who is making it, and whether an accusation is being presented as fact.

Finally, a second AI model acts as an adversarial reviewer: it sees only the passage and the proposal, never the earlier scores, and looks for reasons to reject. High-impact items (accusations, rulings, amounts, private individuals) always go through it.

4. What gets published

The final decision is a rule written in code, not the model's opinion. If any check finds a serious problem, the item is held and does not appear on the site. Inferences and unverified statements are never published automatically.

Next to every item there is a stamp ("Why is this here?"). Tap it to see the source, page, passage and which checks the item passed.

5. An accusation is not a conviction

An indictment, an investigation or a police operation records what an institution states or suspects. Here they always appear attributed: "according to the MPF", "the PF says". No one is called guilty without a court ruling that says so.

6. Continuous updates

Sources are checked automatically every day or every week, depending on each institution's pace. Published items are re-checked periodically against the current rules. Everything that changes goes into the Updates log.

7. Corrections

If you find a mistake, use "Report a problem" on any page. Withdrawn items are not deleted: they leave the site, and the withdrawal, with its reason, is public in the Corrections log.

8. The press, kept apart from the record

We follow outlets with named editors, a public corrections policy and their own reporting. What they publish appears in a layer of its own, always attributed ("According to the outlet, ..."), and never becomes an official fact. Subscriber-only articles keep only their headline, date and link; full text is never reproduced here.

Every press claim is compared with the published official documents: confirmed, partly confirmed, not yet confirmed, contradicted or not checkable. An automated evaluator (Jev) compares the passages and, before we say the official record contradicts an outlet, a second model must confirm it. "Not yet confirmed" does not mean false: much journalism comes before the documents, and we show how many days earlier.

9. The case map: the investigator's reading

An AI model reads the whole published record and the press layer and builds the case map: lines of inquiry, findings, hypotheses, open questions and possible outcomes. It is analysis, labelled as such, and never replaces the sources.

Every item cites the record items it rests on. Findings need official sources; hypotheses and scenarios say how likely they are and what would refute them. No one is called guilty: accusations appear only attributed to the official body that made them, and we give no probability for a person's conviction or acquittal.

Before publication, rules in code check citations, figures and wording; an automated evaluator (Jev) checks that findings really are in the sources; and a second, sceptical model looks for reasons not to publish and may lower the confidence. Each new reading rebuilds the map, and earlier readings are kept.

10. Ask the record: answers checked sentence by sentence

Answers come only from what is published: passages of official documents, the claims extracted from them, the press layer and the Case Map. Search combines words and meaning (a local language model running on the server itself), and the answering model can look up profiles, timelines and links in the graph.

Every sentence of an answer must cite its sources. Before it appears, rules in code check the citations, the voice (record, press or analysis, never mixed), the attribution of accusations and the figures; then Jev checks that the cited sources really say what the sentence says. Sentences that fail are removed, and the reader is told.

Questions about guilt, identifying victims and speculation about people outside the record are declined. Questions are not stored; we only count how many were asked. A control set of questions, with expected answers and manipulation attempts, is rerun every night something changes, and questions stay open only while it passes.