Brazil governments fact-check and X-ray: hash-chained public data to check a viral infographic and compare what it omits

Concept DOI 10.5281/zenodo.23604103 (unchanged); version DOI 10.5281/zenodo.23874109 (minted by the v1.1.0-rc1 release; the final tag carries it in the files). Changed (layout, after rc1) The four translated READMEs moved from the root to docs/<lang>/README.md, next to each language's METHODOLOGY.md; the root keeps a single README.md. The templates that render them moved from docs-src/ to scripts/docs-src/. The SIL Open Font License 1.1 texts of the two bundled fonts (Fahkwang, Noto Sans Hebrew) are now in LICENSES/ and named in NOTICE. Added Causal-evidence layer (amendments 14 and 15, 2026-10-29/30). For each of 27 policy episodes (plus the spending-discipline calculation as episode 28), what the literature measured causally: strength of identification in four levels, whether the study measures the episode's policy, and the direction it reports. Three layers never summed: 43 peer-reviewed Q1 studies, 24 working papers and 11 post-hoc exploratory ones. Two readers (Grok and Gemini) coded each study under a prompt sealed before any abstract was read; agreement is reported with Cohen's kappa and the tie-breaks were written beforehand. Result in the journal layer: 10 episodes without a study, 10 with a study that does not measure the policy, 7 with at least one that does (4 without caveat). New: code/govfc/causal.py, output/pesquisa-causal.json, tests/test_pesquisa_causal.py (every number the code computes is an assertion; effect sizes quoted from abstracts are not computed), collections literatura-fase3/{codificacao-forca,camada2,alavancas,tabela-final} and custeio-fase3. Lever map: who had the power to decide in each episode, by the norm, 124 lines of act or norm with the primary source (4 lines without a verified base). Spending discipline (cards plus travel over primary expenditure) per term, in BRL of September 2026 and as a nominal ratio, with the per-term table (Dilma 2 and Temer as one term; Lula 3 up to July 2026, not projected). The Portal da Transparência files are hash and URL only. The six figures of the article in five languages (scripts/make_figures_pesquisa_causal.py, output/figures/pesquisa-causal/) with a test that measures every text against its budget. The texts of amendments 14 and 15 (and annexes) and the reader prompt are redistributed in data/raw/preregistro/. Changed Tests of the causal layer now pin, one by one, the spending numbers per term and per year, the per-episode counts the article names and the direction totals of each layer (found missing by two independent reviewers before release). Figure labels of the spending article round half up from the published three-decimal value (2.795 prints 2.80, 2.025 prints 2.03); a binary float had printed 2.79 and 2.02. The IPCA is credited to IBGE, read through Central Bank series 433. scripts/prune_evidence.py no longer drops our own amendment texts, reader prompts and resultado/ reports (the previous rule would have deleted PREREGISTRATION-EMENDA-*.md).