Methodology

Each cataloged source is downloaded from its official publisher and normalized to a stable schema with consistent dates, amounts, currencies, and entity keys. The normalized layer preserves provenance to the published dataset: which file, which snapshot, and how many rows.

A normalized layer is not enough. Before publishing, each source must pass blocking gates: structural integrity of the manifest and its rows, valid periods and dates, useful labels and names, a global duplicate rate below the threshold, and a reviewed editorial presentation (title, default question, value formatting, and partial-period policy). A withheld source keeps its layer, but its records and charts do not publish until approved. If a later refresh fails the checks, the last approved version stays published and the quality page says so.

Each signal is calculated with a reproducible query over a dated normalized snapshot. The verified editorial snapshot is preserved even if the source publishes an update, and the page says so. Every signal publishes its chart, the result table, the query scope, how it was calculated, and what it means and does not mean.

Editorial selection ranks by impact, novelty, investigative potential, recurrence, and confidence, then seeks diversity of countries and topics. Editing does not change results or rewrite evidence.

Regional series come from the CEPALSTAT, UNESCO Institute for Statistics and World Bank APIs and are normalized to one table per source with an indicator key, a country and a period. Comparisons use each country’s latest available figure within a declared age window; a year older than the rest is marked in colour. Macro findings (the region’s highest, lowest or largest change) are published only when the automated editor reproduces the value, the rank, the runner-up and the country count against the live source, and are re-checked daily.

A profile consolidates source keys within one country and role only through exact identifiers or an unambiguous normalized legal name; similar names are not assumed to be the same organization. Currencies and unit scales remain separate. Names published in all caps are displayed with typographic capitalization; record tables keep the source-exact string.

Yearly files can be partial. Colombia and the Dominican Republic are queried through bounded API windows and are not a complete historical index. We do not convert between currencies. Concentration or a process modality does not, by itself, prove wrongdoing.