Methodology

A documented achievement is worth more than an exaggerated claim. These are the rules the archive follows, and that the system verifies before publishing.

Every claim has evidence

The archive is made of claims. Each one points to one or more pieces of evidence: the source (institution, title, address), the exact place where the information appears (article of the law, page, table, minute of the video), the date it was consulted and the literal excerpt.

How checking works

Checking has separate stages, counted separately. Finding the words and numbers in a document does not prove that the claim means the same thing as the source.

1. Document retrieved. The system downloads the cited document from the official address and records the date, the HTTP status code and the cryptographic digest (sha256) of what was consulted.
2. Excerpt located. The recorded literal excerpt must appear in the text of the document. If it does not, the evidence does not count.
3. Element matching. Every number, date, proper name and legal act number in the claim must appear in the located excerpts or in the document header. This ensures traceability, not meaning.
4. Contextual support. An automated review, assisted by artificial intelligence and given the context of the document around each excerpt, checks who performed the action, what the action was, on what and for whom, when and where; whether there is a negation, condition or reservation; whether the document reports an announcement, forecast, target, execution or completion; whether a number is a target, a budget, a delivery or a result; whether a date refers to the event, the publication or an update; and whether causality is demonstrated or only an association in time. This review is subject to error, is recorded together with its analysis, and is valid only for the text and evidence examined: if either changes, the review lapses and must be redone.
5. Editorially eligible. In addition to the previous stages, the source must be suitable for the type of claim (see below), and causal or superlative claims, or claims that add up numbers, require additional documentation. When in doubt, the text is corrected to say only what the document says, or the claim stays pending.
6. Publication authorized. Only what is in a publication batch approved by the editor in charge appears on the site, with the list of dossiers and the version of each.

A second artificial intelligence response agreeing with the first is not proof and is not used as such. When an identified person reviews a claim, the review is recorded with name and date; no human review is recorded unless it took place. Titles, summaries, cards, classifications (created, expanded, resumed) and fact sheets display only what is eligible.

A source suited to the claim

• Award or title: the granting institution. Editorial list: the outlet itself.
• Government announcement or statement: a government document, and the text says that it is the government reporting.
• Money actually spent: a budget execution database (SIOP, Treasury, Transparency Portal) or an audit.
• Completed works or a service in operation: a delivery record, statistic or audit; official communication supports "the government reported that it delivered," not the delivery as an independent fact.
• Impact: a suitable evaluation study. The evolution of an indicator during a term does not prove the effect of a policy, and a national result is not the president's personal merit.

States of a claim: Lead, Extracted; excerpt not located, Excerpt located; elements incomplete, Literal match only; no contextual review, Supported by contextual review (automated), Reviewed by an identified person, Sources conflict, Pending: insufficient evidence or not supported in context, Withdrawn. "Not located" does not mean "does not exist," and "not available" never becomes zero. Conflicting claims, or claims without sufficient evidence, are blocked.

Attribution: what belongs to the Lula government

Each initiative (program, policy, government strategy, plan, fund, institutional reform, public work or diplomatic initiative) can have several relationships with each term in office, with a period and evidence:

• Created under Lula: only with an origin milestone in the term and a documented investigation of precursors concluding that the structure is new.
• Merged earlier initiatives: when it unified or absorbed earlier initiatives, which are recorded.
• Expanded under Lula
• Resumed under Lula: requires a previously recorded interruption or replacement.
• Reshaped under Lula
• Continued under Lula
• Shared implementation

Precursors are classified by what the source says: unified, replaced, absorbed, legal basis, technical basis, or a similar initiative with no declared succession. "Something similar existed before" does not cancel a creation, and a launch during the term is not enough to attribute creation. Initiatives with no law of their own bearing the name (such as government strategies) have the origin milestone recorded as an official launch or an institutional document, with the scope of the search for the law.

We separate achievements of the Lula government, Lula's personal achievements and results for Brazil that occurred during the term, which may have other causes. The responsibilities of ministries, Congress, states, municipalities and operators are recorded in each dossier.

Budget, beneficiaries and deliveries

• Announced, authorized, committed, liquidated and paid are different figures and appear with the type indicated; amounts are nominal unless stated otherwise.
• Beneficiaries are an annual flow, a stock in a given month or a cumulative total, and are not added together.
• Announcement, contracting, completion and operation of public works are different phases. A delivery is not a social result.
• Where data is missing, the gap is stated explicitly, together with the research task that would resolve it.

Indicators

• Each series has a methodological sheet checked against the official metadata: table, variable, unit, territory, population, frequency, nominal or real values, seasonal adjustment and revisions.
• The series appears in full within the documented window, including the unfavorable years.
• Surveys with different methodologies stay in separate series, with the break explained; comparisons use the start and end of the same series and state the rule for choosing the years.
• Partial, preliminary or projected data does not become an annual result. Rates change in percentage points; levels, in percent.
• An observed result, an interpretation and evidence of causal impact are different things.

Speeches

Only statements proven in an official record are included. Prepared text, text in an official publication, transcript, translation and video are distinct records; the English in UN records is text or interpretation supplied by the delegation, not a transcript. Quotations preserve the original words and indicate where they are.

International recognition

Award, decoration, honorary title, tribute, nomination, editorial list, praise and recognition of a Brazilian policy are different categories. A grant requires a checked official source (the granting institution, an official record, an international organization or a government); press alone is not enough, except for an editorial list, whose source is the outlet itself. Recognition given to Brazil or to a program is not a personal honor.

Sources and blocks

Government communication proves what was announced; results are checked against statistical databases, budget execution, audits and evaluations. The archive does not get around CAPTCHAs, anti-bot challenges, logins or firewalls: when a source blocks automated access, this is recorded and an alternative public route is sought.

Translation

This is the English edition of Exército do Lula (exercitolula.com.br), the original archive, which is maintained in Portuguese. Every text shown here has a translation tied to the original text: if the original changes, the old translation no longer matches it, and the English edition is not republished until that text is translated again. Translations are produced with the help of artificial intelligence and are identified as such. Excerpts from sources stay in their original language. If the two versions differ, the Portuguese original prevails.