The project

Methodology

Every claim on the portal is a piece of evidence: a short sentence tied to a source and to the exact place within it (page, slide, paragraph, minute). The pages are assembled from these pieces of evidence.

The archive in numbers

307 sources · 454 pieces of evidence · 109 statements · 14 comparisons · 24 episodes · 65 amounts on the dashboard · 13 studies · 124 events · 67 people · 84 organizations · 120 links · 32 cases. Another 32 research records are kept off the site — internal agenda items without sufficient documentation or records that have not yet passed validation.

What each piece of evidence is

Each piece of evidence says what type it is. An accusation is always attributed to whoever made it; an estimate always states who estimated it and how.

Official document
Judicial or administrative decision
Official data
Study estimate
News report
Document described in a news report
Statement
Accusation
Leaked message
Response/defense

Degree of substantiation

Proven in a document
Several independent sources
Single source
Allegation
Disputed
Estimate

“Proven in a document” requires that the project has opened the primary document. When only a news report describes a document (a sealed decision, a PF report), it appears as “document described in a news report” and the portal does not say it checked it.

Quotes: verbatim checking

A statement enters the library only if the exact text was found in an open source. Paraphrase is not placed in quotation marks: when a news report only summarizes what someone said, the summary becomes evidence, without quotation marks. Each quote reports how it was checked:

Checked against the original video
Checked against the original audio
Checked against the text published by the author
Checked against an official transcript
Checked against the journalistic record
Checked against a case document

An automatic routine (conferir-citacoes) downloads the source of each quote and looks for the verbatim text. In the current version, 109 of 109 quotes have been checked beyond the researcher's own check. Leaked messages (such as those of Vaza Jato) are marked separately, with their origin, whether or not there was a forensic examination, and what the author said about their authenticity.

Comparisons

Literal contradiction: Two sentences by the same author that deny each other in their own terms.
Change of position: The author came to defend or do what they had previously rejected.
Change of alliance: The author broke with or allied with someone they had previously supported or accused.
Statement × document: What was said diverges from what a document records.

When the author maintained an accusation even after changing alliances, the comparison records this; we do not presume a retraction.

Decisions: which sphere

“Conviction” never appears alone. Each decision carries its sphere:

Criminal sphere
Civil sphere
Electoral Court
Disciplinary sphere
Court of accounts
Administrative sphere
Procedural decision
Constitutional review

Amounts: eight categories that are not added up

1. Improper payments and estimated losses. How much is estimated to have been diverted or overpaid. Estimates, with the method stated.
2. Amounts cited in investigations and decisions. Figures from indictments, expert reports, appellate rulings and sentences. They depend on the outcome of each case.
3. Amounts provided for in agreements. What companies and cooperating witnesses committed to pay. A promise, not money received.
4. Frozen assets. Assets and funds made unavailable by court order. They may be released.
5. Amounts actually paid. What actually entered the treasuries or court accounts, on the date reported.
6. Amounts returned to Petrobras and other recipients. What reached the final recipient: Petrobras, the Union, other victims.
7. Fines, indemnities and expenses paid by Petrobras. Agreements with foreign authorities and investors, fees and other expenses of the victim itself.
Petrobras accounting results. Loss and impairment losses: accounting measures, with several causes, that overlap.
8. Estimated economic impacts. Estimates from studies on investment, output, employment and tax revenue that ceased to exist. Counterfactual scenarios.

Rules: (1) each value has a reference date or period, a basis (nominal or adjusted), and who calculated it; (2) cumulative announcements are snapshots on a date, not installments to be added up; (3) an agreement signed with the MPF, the CGU and the AGU is a single agreement; (4) values that overlap carry a “do not add to” list; (5) we do not subtract costs from recoveries: they are incompatible quantities.

Economy: what is the effect of what

We distinguish three things: the effect of the corruption itself (overpricing, poorly planned projects), the effect of the way the investigations were conducted (companies barred from contracting, halted works, credit cut off), and other factors (oil prices, exchange rate, fiscal and political crisis). Studies rarely manage to separate the three; when they cannot, we say so.

Review statuses

Documented · review pending: the source was opened and read by the research team; a check by another person is still missing.
Verified by audit: a second process (another researcher or the automated routine) redid the reading.
Reviewed: human editorial review, carried out by the person responsible for the project (the first on October 24, 2026, covering the entire published version). No item is marked as reviewed by artificial intelligence.

Content without sufficient documentation stays as an internal agenda item and does not appear on the site.

Sources and rights

We prioritize original documents: Petrobras, CVM, SEC, B3, STF, STJ, TSE, CNJ, CNMP, MPF, PF, TCU, CGU, AGU, IBGE, the Senate and the Chamber of Deputies. News reports are used for facts, interviews and investigations — including when there was no conviction. For protected reports, we keep our own summary and short excerpts with attribution; we do not republish full texts. We keep a local copy of public documents where permitted, with the date of consultation.