# The seams of the law: the data behind the book

The open data of the book *The seams of the law* (https://future-seems-so-good.com/blog/the-seams-of-the-law). In October 2026 AI agents read Brazil's statute book all at once: 323,868 norms of the Union, 24 states and the Federal District, as collected by ctresb/leis (https://github.com/ctresb/leis). They classified every norm, measured the statute book as a space of vectors, built the graph of citations and revocations, and searched it for seams: boundaries across which the law treats economically equivalent situations differently. Two red teams checked the claims against the official texts. This package holds the tables, the module reports, 77 strategies with both red teams' verdicts, and the code.

License: the tables and texts under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/); the code in `code/` under the MIT licence. The corpus is ctresb/leis (MIT licence); the laws themselves are official acts, outside copyright under Lei 9.610/1998 art. 8 IV. Cite as: Martin Roark, *The seams of the law: the data behind the book*, future seems so good, 2026-10-16, https://future-seems-so-good.com/blog/assets/the-seams-of-the-law/data/.

Nothing here is legal or tax advice. Every row that a decision would rest on names the law it reads; read the official text before acting.

## Start here

- `tables/strategies/strategies.csv`: 77 lawful strategies ranked by urgency and size, with the legal basis, the window, the abuse line and both red teams' verdicts.
- `tables/clocks/calendar-2026-2040.csv`: the dated windows, each checked against the law that sets it.
- `tables/payers/who-pays.csv`: who loses from each seam, and the public lever against it.
- `tables/made/classes-corrected.csv`: what the statute book is made of, corrected with a held-out audit.
- `tables/wiring/status-summary.csv` and `tables/wiring/scheduled-revocations.csv`: what is in force, and what LC 214/2025 revokes on 1 January 2033.
- `datapackage.json`: every table with its schema and field descriptions. `the-seams-of-the-law-data.xlsx`: the same tables as one workbook. `the-seams-of-the-law-data.zip`: everything at once; `SHA256SUMS` holds its hash.
- `MANIFEST.csv`: every file the project produced, published or withheld, with the reason.

Join keys: in tables with one row per norm, `id` is the norm's key in the corpus, a path of sphere, folder, year and file name (for example `federal/leis-ordinarias/2021/lei-14133-2021-04-01.md`), and joins them to each other and to the corpus. `uf` is a state's two-letter code. The `id` of the strategies table is the `strategy_id` of the two red teams' audits.

## Tables, by chapter of the book

### The corpus

- `tables/corpus/federal-repairs.csv` (976 rows): **Federal norms whose corpus text was replaced or completed from the official compilation.** One row per federal norm whose corpus text was replaced or supplied from an official source: 976 norms, 19 of them new to the corpus. Where the corpus held only a fragment (a rectification notice, the promulgation of overridden vetoes, a partial republication), the full enacted text came from normas.leg.br, the federal legislation portal; for key laws, for fragments that portal could not repair and for amended laws whose text lacked later amendments, the compiled current text came from planalto.gov.br, the Presidency's legislation site. A text labelled 'original (reparado)' is the law as enacted, not as amended; a compiled text is current to the date its page was fetched (October 2026).
- `tables/corpus/federal-missing.csv` (214 rows): **Federal norms the official index lists and the corpus lacks.** One row per federal norm that the official index lists but the corpus downloader could not obtain, and that a second attempt tried to recover: 214 of the 261 the downloader logged. Each was requested again from normas.leg.br and then from planalto.gov.br, keeping a text only if its title carried the norm's number and year; 20 now have a text (18 added as new norms, 2 empty records filled) and 194, all from before 1988, were found nowhere. The 45 logged entries without a plain number (unnumbered imperial laws, Lei 312-A/1936 and similar) were not tried and have no row.
- `tables/corpus/federal-revocations-confirmed.csv` (125 rows): **Federal revocations confirmed against the official compilation.** One row per federal law that the corpus catalog lists as in force but whose compiled page on planalto.gov.br, the Presidency's legislation site, says was revoked in whole: 125 laws. Found while checking texts against planalto; the corpus flag comes from normas.leg.br, which marks only explicit 'not in force' records, so it misses these revocations. The list covers only the laws checked against planalto, not all federal law.
- `tables/corpus/census-by-part.csv` (119 rows): **Norms in the corpus by part (the Union, each state) and folder.** One row per part of the corpus and folder: how many norms of each kind the corpus holds for the 1988 Constitution, federal law and each of the 24 states and the Federal District. Counted in the corpus catalog, which matched the coverage table published with the source corpus to the unit. The counts are of the base corpus, 323,849 norms, before the federal repair added 19 norms (see corpus/federal-repairs), so federal law here is 19 short of the 323,868 total; Paraná and Rio de Janeiro are absent.
- `tables/corpus/census-reconciliation.csv` (29 rows): **Corpus counts reconciled against official indexes.** One row per part (every state and the Federal District, Paraná and Rio de Janeiro with zero, the Federal Constitution and federal law): the counts declared in the coverage table published with the source corpus, set against the counts in the corpus catalog. Columns starting with cob_ are the declared figures, the unprefixed columns are counted in the catalog, and d_ columns are catalog minus declared; every difference is zero. Counts are of the base corpus before the federal repair added 19 norms.
- `tables/corpus/completeness-by-series.csv` (53 rows): **Share of each numbered series present in the corpus.** One row per numbered series of laws: federal ordinary laws from 1946, federal complementary laws from 1967, the federal decree-laws of 1937-1946 and of 1965-1988, and each state's ordinary laws and, where numbered separately, complementary laws, from the start of the state's current numbering. Laws are numbered in sequence, so every number from 1 to the last should exist; the table counts the numbers present in the corpus and the gaps, with completeness measured from number 1 and completeness_in_window over the stretch the source actually covers. The table holds the uncorrected counts: counting numbers printed only in titles, Rio Grande do Norte has 32.7% of its ordinary-law numbers (16.2% here), and against Lei 4.778 of March 2026 Acre has 45% (99.0% here, measured against its last file, Lei 2.179 of 2009); corpus/informational-asymmetries applies both corrections. Paraná and Rio de Janeiro are absent.
- `tables/corpus/completeness-by-decade.csv` (332 rows): **Completeness of the corpus by decade.** One row per numbered series and decade: how many of the numbers issued in that decade the corpus holds. A missing number's decade is estimated by interpolating between the dates of the numbers present around it, so the split of numbers across decades is approximate. Same series and same uncorrected counts as corpus/completeness-by-series; Paraná and Rio de Janeiro are absent.
- `tables/corpus/text-fidelity.csv` (40 rows): **Corpus texts compared with the official source.** One row per norm in a 40-norm sample drawn in 15 strata (federal compiled and enacted texts; the scanned and text-layer PDFs of Paraíba, Rondônia, Piauí and Pará; HTML texts of SP, Minas Gerais, Goiás, Rio Grande do Sul and Bahia), whose official source was downloaded again on 15 October 2026 and compared with the corpus text. Similarities are character-sequence ratios from 0 to 1 on normalized text; scanned PDFs without a text layer could not be compared by machine and were read by eye. All 32 machine-comparable texts match their source, so the corpus's defects are the sources' own; 40 norms are too few to estimate error rates by state, and federal texts were compared with normas.leg.br only, the corpus's own source.
- `tables/corpus/quality-by-part.csv` (27 rows): **Empty, unreadable and readable texts by part.** One row per part (each state, the Federal District, federal law and the Federal Constitution): norms in each quality tier. Tiers were set by rules on the stored text: A analysis-grade; B usable with care (OCR, light noise, mis-decoded characters, article numbers out of order, duplicate records, missing dates); C unreliable as text (a revocation or veto note in place of the text, an empty body, header and summary only, noise above 10%, another law's text). Counts cover 323,850 norms, the base corpus plus LC 227/2026; a tier judges the text as stored, not whether the norm is in force. Paraná and Rio de Janeiro are absent.
- `tables/corpus/informational-asymmetries.csv` (28 rows): **Where the statute book cannot be read, ranked.** One row per state and the Federal District, ranked, plus a federal row at the end: the share of each part's ordinary laws that a machine reader cannot get as clean text. Coverage is the share of the ordinary-law numbers present in the corpus, clean the share of quality tier A (analysis-grade text) among the laws present, readable their product and asymmetry 1 minus readable. Unlike corpus/completeness-by-series, Rio Grande do Norte counts the numbers carried only in titles and Acre is measured against Lei 4.778 of March 2026, a lower bound for its last number; Paraná and Rio de Janeiro, with nothing collected, are set to 1 although their laws are on official sites for human readers.
- `tables/corpus/output-by-state-year.csv` (26 rows): **Statutes enacted by each state each year.** One row per part (24 states, the Federal District and federal law) and one column per year: the number of laws in the corpus dated that year, counting ordinary, complementary and promulgated laws (not decree-laws, amendments or constitutions). Counted in the corpus catalog by the norm's year; years in which no part has a law have no column, and the federal row reaches back to imperial laws of the 1820s. These are files present, so output is understated where files are missing (Acre after 2009, Tocantins before 2020, Rio Grande do Norte before 2017, Alagoas before 1999, Rio Grande do Sul's 5,569 undigitized laws); Paraná and Rio de Janeiro are absent.
- `tables/corpus/output-per-capita.csv` (27 rows): **Statutes per inhabitant by state.** One row per state and the Federal District (27 rows): laws in the corpus dated 2015 to 2024, per year and per million inhabitants, using the 2022 Census population from IBGE, the national statistics office. Laws are ordinary, complementary and promulgated laws counted as files in the corpus, so output is understated where files are missing (Acre has none after 2009, Tocantins none before 2020); Paraná and Rio de Janeiro show zero because nothing was collected.
- `tables/corpus/status-coverage.csv` (27 rows): **How far the corpus records whether each norm is in force.** One row per part: how far the corpus's own status fields record whether each norm is in force, and how they agree with revocation notes printed in the texts. The status text ('situacao') exists only for SP, Goiás, Minas Gerais, Espírito Santo, Sergipe and the Federal District, and the revoked flag ('revogada') for federal law and those six; the table holds Santa Catarina's flag as first downloaded, true for 22,252 of 22,254 norms because of a downloader bug, a flag that means nothing and has since been cleared. Where there is no flag, revocation notes in the text or the citation graph (wiring/status-summary) are the only machine-readable status.

### What the law is made of

- `tables/made/classes.csv` (3 rows): **Norms by functional class, as filed by the classifier.** One row for all norms and one for each sphere (state, federal): the share of norms in each functional class and class group, as filed by a rule-based classifier that reads each norm's ementa (official summary) or, failing that, its first article. Counted over the corpus's 323,867 norms (the Federal Constitution left out); bare amendments take the class of the law they amend. These are shares as filed: the classifier's share-weighted precision is 0.845 on a held-out sample of 280 hand-judged norms, and the economica class is right only 45% of the time, so general rules are 17.7% once corrected, not the 24.9% filed (see made/classes-corrected).
- `tables/made/classes-corrected.csv` (16 rows): **Class shares corrected with the held-out audit, with intervals.** One row per functional class and per class group: the number and share of norms the classifier filed in it, and the number and share estimated to truly belong there, corrected with the confusion matrix of a held-out sample of 280 norms judged by hand, with a 95% interval. Because the sample is stratified by predicted class, the true count of a class is the sum, over predicted classes, of the class size times the share of sampled norms truly in it. The filed counts are those of the build the sample was drawn from, which differs from the final build in made/classes by about 1,100 norms (0.35%); the correction applies to the whole corpus only, not to states or decades.
- `tables/made/classes-by-state.csv` (26 rows): **Class shares by state.** One row per part (24 states, the Federal District and federal law): the share of its norms in each functional class and class group, over all years and again for norms dated 2015 to 2024. Classes are as filed by the rule-based classifier (share-weighted precision 0.845) and are not corrected per state, so the general-rule and economy shares are upper bounds. Differences between states mix legislative custom and what each official portal publishes; Paraná and Rio de Janeiro are absent, and Acre has no norm after 2009.
- `tables/made/classes-by-decade.csv` (30 rows): **Class shares by decade.** One row per decade from the 1880s and sphere (state, federal): the share of norms in each functional class and class group, as filed by the rule-based classifier and not corrected for its errors. The number of states covered grows until the 1990s (see estados_com_50), so rows before then mix changes in composition with changes in coverage; before 1940 the state rows are essentially SP. Paraná and Rio de Janeiro are absent.
- `tables/made/classifier-precision.csv` (13 rows): **Precision of each class on the held-out audit sample.** One row per functional class, plus a share-weighted total: how often the class the classifier filed was right on a fresh held-out sample of 290 norms (280 judged; none used to write the rules), judged by hand from each norm's ementa and first article, with Wilson 95% intervals. Trust simbolica, pessoal, tributaria, concreta, patrimonial and organizacao (0.93 to 1.00); use orcamentaria (0.85) and revogacao (0.89, 9 norms) with care; read economica (0.45) and social (0.55) as candidate lists; outros (0 of 15) is a residue of misses. Class sizes are those of the build the sample was drawn from; the final build's total precision, after a last fix measured on the norms it moved, is the same 0.845.
- `tables/made/business-duties.csv` (5,242 rows): **Norms that impose a duty on a kind of business, with their sanction.** One row per general rule not known to be out of force that obliges or forbids something to a kind of business: 5,242 norms (4,545 state, 697 federal), 3,608 of them with a sanction clause. Selected by word patterns: an obligation or prohibition within 160 characters of a business word, the instrument 'obliges' or 'forbids', and a business actor named, with recognition seals, prizes and commemorative dates left out. A 15-row random check of an earlier version found 13 genuine duties on businesses; Paraná and Rio de Janeiro are absent, and state texts are often the original enactment, not the text in force.
- `tables/made/icms-incentive-norms.csv` (2,007 rows): **State norms that grant or change an ICMS incentive.** One row per state norm whose text mentions ICMS (the state tax on goods and some services) and that exempts, reduces the base or rate, defers, grants a presumed credit, amnesties or otherwise grants a tax benefit: 2,007 norms in 25 parts, 898 naming ICMS in the ementa, 1,885 not known to be out of force. Selected by text search and word patterns, with symbolic, budget, personnel, property, organization and revocation classes left out. A 12-row random check of an earlier version found 9 that grant or amend ICMS benefits and 3 that only mention ICMS. The corpus holds no decrees and no interstate agreements of CONFAZ (the council of state finance secretaries), through which many ICMS benefits are granted, and Paraná and Rio de Janeiro are absent.
- `tables/made/fund-creations.csv` (2,201 rows): **Norms that create a public fund.** One row per norm whose ementa (official summary) creates or institutes a fund: 2,201 norms, 2,105 state and 96 federal, 305 of them out of force. Selected by a word pattern on the ementa (a verb such as cria or institui followed within 80 characters by fundo); an 8-row random check found 7 genuine fund creations, so some rows only concern an existing fund (its council, for instance). Paraná and Rio de Janeiro are absent.

### No small skeleton

- `tables/skeleton/basis-federal-in-force.csv` (200 rows): **Greedy orthogonal basis of federal law in force, in order of pick, with residuals.** One row per pick of a greedy orthogonal basis of the federal law in force (13,051 normative federal norms): starting from the medoid, each step adds the norm whose direction is least explained by the norms already picked, so the residual column shows how much new direction each pick adds. Norms are compared as 200-dimensional vectors obtained by reducing (truncated SVD) a TF-IDF matrix of accent-folded words and word pairs of each norm's full text, after removing site boilerplate; norms with fewer than 100 content words could not be picked. The geometry is lexical: it finds laws written in distinct vocabularies, one per genre and era, not the most important laws; 'in force' means not known to be revoked or struck down.
- `tables/skeleton/basis-all.csv` (200 rows): **Greedy orthogonal basis of all normative norms.** One row per pick of a greedy orthogonal basis of all normative norms in the space (102,944: 14,443 federal and 88,501 state), built as skeleton/basis-federal-in-force is: each pick is the norm whose direction is least explained by the earlier picks. Norms are compared as 200-dimensional vectors obtained by reducing (truncated SVD) a TF-IDF matrix of accent-folded words and word pairs of each norm's full text, after removing site boilerplate; one-off administrative acts and short or noisy texts were left out. Two state-wide artifacts survived cleaning (a site notice in every Amapá text, a promulgation formula in most Maranhão texts) and may explain an Amapá law among the first picks; Paraná and Rio de Janeiro are absent.
- `tables/skeleton/basis-summary.csv` (54 rows): **Residual at each basis size, by corpus part.** One row per subset of the space (federal law in force, all norms, each state and the Federal District) and variant: the residual of the 10th, 30th, 50th and 100th pick of the greedy orthogonal basis, a measure of how slowly a few laws come to span the rest. A residual near 1 at k means the k-th pick is still almost orthogonal to the previous ones; 0.879 at k = 50 for federal law in force means the 50th law keeps 88% of its direction outside the span of the first 49. The 'direction' variant (unit vectors) is the one reported; the 'projection' variant favours short templated laws and is kept for comparison.
- `tables/skeleton/basis-themes.csv` (138 rows): **Themes of the first fifty basis laws against their share of the corpus.** One row per subset (federal law in force, all norms, and the five largest state economies in the corpus: SP, Minas Gerais, Rio Grande do Sul, Santa Catarina, Bahia) and theme: the theme's share of the subset against its share of the first 50 picks of the greedy orthogonal basis. An over-representation above 1 means the theme supplies more independent directions than its size suggests: laws imposing one duty on one kind of establishment do (4.5 times their share in federal law), tax and exemption laws, which share one vocabulary, do not (0.49). Themes are 20 clusters of the space, so the figures are lexical.
- `tables/skeleton/uniqueness-top.csv` (60 rows): **The most singular norms (largest distance to their nearest neighbour).** One row per norm among the 60 most singular normative norms with at least 200 content words: those whose nearest other norm in the corpus is furthest away. Uniqueness is 1 minus the largest exact TF-IDF cosine between the norm's text and any other norm's, the 20 nearest candidates having been found in the 200-dimensional reduced space. It is lexical: part of Paraíba's prominence (19 of the 60) comes from OCR noise, which makes scanned texts look unique.
- `tables/skeleton/axes.csv` (40 rows): **Principal axes of the space and the words that load on them.** One row per axis of the 200-dimensional reduced space (the first 40 axes of a truncated SVD of the corpus's TF-IDF matrix): the words that weigh most at each end and example norms that sit there. Axis 0 is the mean direction of the corpus; later axes oppose genres and eras (axis 1: authorizing and financial law against social-policy law). Terms are accent-folded, lightly stemmed words and word pairs recovered from hashed columns, and a few axes are artifacts (an Amapá site notice on axis 3, electronic-signature blocks on axes 8 and 9).

### Same ground, opposite pull

- `tables/opposed/coupled-pairs-read.csv` (30 rows): **Thirty families of same-subject, opposite-action pairs, read by hand.** One row per law family read by hand: the top 30 entries of the three rankings of same-subject, opposite-action pairs (20 between states, 6 federal-state, 4 federal-federal), one entry per family when one model law produced several pairs, with the article and a verbatim quote from each side. Pairs were ranked by machine (see opposed/coupled-pairs-states); the verdict is a legal reading of the quoted texts. Of the 30, 9 are true oppositions (8 between states), 3 enforcement gaps, 4 the same rule in other grammar, 4 restatements of federal law, 3 complementary, 1 a rider and 6 false positives; the state texts were not re-checked on official sites, so a later amendment could change a reading.
- `tables/opposed/coupled-pairs-states.csv` (400 rows): **Cross-state pairs with high subject similarity and opposed action.** One row per pair of laws from two different states with a highly similar subject (cosine at least 0.6) and diverging action: the top 400 of 456 eligible pairs, ranked by opposition score. Built by machine: each norm's 15 nearest norms in a subject space (the text minus every action word) were scored by how far their action words (duty, prohibition, permission, exemption, taxation, benefit, sanction) oppose, keeping pairs of norms both in force, with enough action words and content, and not linked by an amendment, revocation or regulation. The lexicon counts words, not legal effect: of the 30 families read by hand (opposed/coupled-pairs-read) only 9 were true oppositions, so rows are candidates to read, not findings. Paraná and Rio de Janeiro are absent.
- `tables/opposed/coupled-pairs-federal-state.csv` (48 rows): **Federal-state pairs with high subject similarity and opposed action.** One row per pair of a federal law (side a) and a state law (side b) with a highly similar subject (cosine at least 0.6) and diverging action: all 48 eligible pairs, ranked by opposition score. Built by machine: each norm's 15 nearest norms in a subject space (the text minus every action word) were scored by how far their action words (duty, prohibition, permission, exemption, taxation, benefit, sanction) oppose, keeping pairs of norms both in force, with enough action words and content, and not linked by an amendment, revocation or regulation. The lexicon counts words, not legal effect: of the 30 families read by hand (opposed/coupled-pairs-read) only 9 were true oppositions, so rows are candidates to read, not findings. None of the six federal-state families read was an opposition: the state publicizes, executes or softens a federal rule.

### Copies

- `tables/copies/copy-families.csv` (1,474 rows): **Families of state laws copied across states (strict rule).** One row per family of near-identical laws enacted in two or more jurisdictions (states, the Federal District, the Union): 1,474 families under the strict rule, 1,455 policy families and 19 federal templates. Laws were compared by the exact Jaccard similarity of their sets of word 5-shingles after removing state names, dates, numbers and standard clauses; two laws of different jurisdictions are linked at J of 0.5 or more, and a family is a connected group of links. Federal templates (debt renegotiations with the Union and similar deals) are common shocks rather than copies and are kept apart from the network and the gap model; verbatim copying is a small slice of state law (3.4% of normative state laws have a twin), and Paraná and Rio de Janeiro are absent.
- `tables/copies/copy-family-members.csv` (3,546 rows): **Each law in each copy family, with its state and year.** One row per law in each strict copy family (3,546 laws: 3,479 state and 67 federal), with the earlier law it most resembles. The copy source of a law is the member of an earlier-adopting jurisdiction whose text is most similar to it, which is the likely model, not a proven one: the true source may be a bill, a national advocacy group or a third state. Status fields are the corpus header's and exist only for some states.
- `tables/copies/adoption-gaps.csv` (395 rows): **States missing a widely adopted law, ranked.** One row per pair of a copy family that is still spreading (79 families with members in 4 or more states and 2 or more state adoptions in 2024-2026, under the loose similarity rule, whose numbers differ from those of copies/copy-families) and one of the five states most likely to adopt it next. Probabilities come from a logistic hazard model fitted on adoptions up to 2025 (the share of states, neighbours and region that adopted, time since the first adoption, state and theme effects); it ranks well out of sample (AUC 0.80 to 0.81) but under-predicts families that are still spreading by a factor of about 1.97, which p3_scaled applies as a rough correction, not a refit. Many gaps are data gaps or reworded adoptions: weak_corpus and maybe_adopted_other_text flag them.
- `tables/copies/struck-law-copies.csv` (40 rows): **Copies of laws the STF struck, checked against the court's records.** One row per copy of a state law that a court annotation in the corpus marks as struck, suspended or narrowed by the STF (the Supreme Federal Court), found among the members of 17 copy families of the loose similarity rule (J of 0.3 or more), plus two rows that apply an STF thesis to a family. Each status was checked by hand against the court's and the assemblies' sites on 15 October 2026. After the check, 8 families have a copy whose model law the STF struck while the copy stands, 5 more are exposed only through a clause or an interpretation, and 2 flags were wrong (the soy-moratorium incentive laws were upheld; the action against SP's smoke-free law was dismissed as moot).
- `tables/copies/state-roles.csv` (25 rows): **Each state as first adopter, supplier and copier.** One row per state and the Federal District (25): how often it adopts a copied law first, how many copies others take from it and how many it takes from others, over the strict policy families. Counted in the corpus from the adoption sequences; the copy source of a law is the most similar earlier member, so 'supplied' and 'made' count likely models, not proven ones. Part of the lead of Minas Gerais and SP is drafting style (their assemblies rewrite what they borrow, so their own copies fall below the similarity rule); Paraná and Rio de Janeiro are absent.
- `tables/copies/copy-network.csv` (433 rows): **Who copies whom: state-to-state copy counts.** One row per ordered pair of states with at least one copy: how many laws of the follower state have the source state's law as their likely model, over the strict policy families (1,893 copies on 433 edges). The source is the most similar earlier member of the family, a likely model rather than a proven one. Geography does not decide who copies whom: the source is a bordering state in 21.3% of copies against 20.9% expected from timing alone; Paraná and Rio de Janeiro are absent.

### The wiring

- `tables/wiring/status-summary.csv` (40 rows): **Norms in force, revoked and struck by courts.** Headline figures of the citation graph and of the status of norms derived from it, one figure per row. Counted in the corpus: every reference a norm makes to another (citation, amendment, revocation, regulation, or a compiler's 'see' note) was extracted from the text with a grammar and matched to the catalog, and a norm counts as revoked when a total revocation in force by mid-October 2026 points at it. Court rulings are not in this table: the same final build (16 October 2026) counts 1,382 norms wholly struck down or suspended by a court and 279,795 norms neither revoked nor struck. A revocation the text does not record is not seen, so 'not revoked' means 'not known to be revoked'.
- `tables/wiring/load-bearing-federal.csv` (300 rows): **Federal laws ranked by the norms that cite them.** One row per federal norm (the 1988 Constitution included): the 300 most cited by other norms in the corpus, ranked by the number of distinct citing norms. Counted in the corpus on the citation graph: references that cite or regulate, found in a norm's own article text or preamble, excluding references inside new wording a law writes into another and preamble citations of a constitution (enacting formulas), collapsed to one per citing and cited pair; citing norms count whatever their status. References the grammar could not match to the catalog are not counted, and the match rate for statutes varies from 99.6% (Federal District) to 28.0% (Rio Grande do Norte), so counts are lower bounds. Decree-Law 1.202/1939 ranks high as an artefact: it is named in the enacting formula of the 1939-1945 state decree-laws.
- `tables/wiring/federal-cited-by-states.csv` (100 rows): **Federal laws cited by state norms, by state.** One row per federal norm: the 100 most cited by state and Federal District norms, ranked by the number of distinct state citing norms. It is the subset of wiring/load-bearing-federal (the 300 federal norms most cited overall) re-sorted by state citations, on the same citation graph and rules, so a norm much cited by states but outside that top 300 would be missing. State references the grammar could not match to the catalog are not counted, and the match rate is lowest in Rio Grande do Norte, Tocantins, Alagoas, Paraíba and Pernambuco, so counts are lower bounds.
- `tables/wiring/scheduled-revocations.csv` (85 rows): **Revocations scheduled for a future date, including 1 January 2033.** One row per revocation already enacted that takes effect after mid-October 2026: the norm revoked, the date, whether whole or in part, and the revoking act. Read from revocation clauses in the corpus text, the date taken from the clause, a list sub-heading or the act's own entry-into-force clause; 77 of the 85 rows come from LC 214/2025 (the Lei Complementar implementing the consumption-tax reform), which revokes PIS/Cofins laws (federal contributions on revenue) from 1 January 2027 and LC 87/1996 (the ICMS framework law), LC 116/2003, LC 160/2017, LC 192/2022 and Decree-Law 406/1968 from 1 January 2033; the rest are state acts. The norms listed are still in force. Citer counts include citing norms of every status, so they are higher than counts of citers not known to be revoked (352 state citers of LC 87/1996 here, 330 not known to be revoked).
- `tables/wiring/lei-8666-dependants.csv` (200 rows): **State norms in force that cite the revoked Lei 8.666/1993, by article.** One row per revoked norm that norms enacted after its revocation still cite: the 200 revoked norms, federal and state, with the most such later citers. Despite the table's name it is not limited to Lei 8.666/1993 (the procurement law revoked from 30 December 2023), which is one row; the counts of state norms citing Lei 8.666 and the articles they name are in wiring/status-summary. Counted in the corpus on the citation graph of wiring/load-bearing-federal, for norms revoked whole by an explicit text revocation, comparing each citing norm's date with the date the revocation took effect. The list carries noise: undated references to small numbers that resolved to 1937 federal decree-laws ('Decreto-Lei nº 2'), and portal-specific formulas (Sergipe's public-utility declarations citing its Lei 5.206/2003, 193 later norms).
- `tables/wiring/churn-top.csv` (436 rows): **The most amended laws.** One row per amended norm: the 15 most amended norms of each state and the Federal District, the 60 most amended federal norms, and the 1988 Constitution. Counted in the corpus as distinct amending acts: acts whose text amends the norm, plus acts outside the corpus (medidas provisórias, decrees, missing state laws) named in compiler notes such as '(Redação dada pela Lei nº ...)'; acts dated before the norm are dropped. Counts are lower bounds where portals publish only the original text, since an amendment is then seen only when the later act names the norm; much state churn is symbolic (consolidated laws of public-utility titles or commemorative dates take one amendment per new title or date).
- `tables/wiring/complexity-by-regime.csv` (25 rows): **Size and churn of the main tax regimes' laws.** One row per regime: a root law and how far its articles reach by following cross-references. Roots are LC 123/2006 (the Simples Nacional, the simplified regime for firms up to R$ 4.8 million of revenue), the CLT (the labour code), the CTN (the national tax code), LC 87/1996 (the ICMS framework law) and the main ICMS statute (ICMS: the state tax on goods and some services) of 21 states; Alagoas, Pernambuco and Tocantins have none in the corpus, Goiás's is a stub without text, and Paraná and Rio de Janeiro are absent. Counted in the corpus by breadth-first search over resolved citations in article text: a reference that names an article leads to that provision, one that does not leads to the whole norm and everything it cites. References to decrees, CONFAZ agreements (CONFAZ: the council of state finance secretaries that authorises ICMS incentives) and other acts outside the statute corpus are counted but not followed; LC 123's delegation to the resolutions of its managing committee is not captured, so its zero there is an undercount.

### Lines

- `tables/lines/threshold-catalog.csv` (137 rows): **Numeric lines in the law (revenue, headcount, area, value), verified.** One row per statutory line the essay relies on (a number in the law where treatment changes: revenue, headcount, area, value, age), with the norm, the article, the verbatim words and what changes past the line. Each quote was read by hand and checked verbatim against the corpus text; all 137 were found. versao_texto says which version of the text the corpus holds: federal texts are mostly compiled current texts, state texts often original enactments, so a state line may have changed since and should be confirmed with the official source; Paraná and Rio de Janeiro are absent.
- `tables/lines/bunching-revenue.csv` (203 rows): **Firms by revenue bin around the regime lines.** One row per year and revenue bin: the national count of firms in the Simples Nacional (the simplified tax regime for small firms) by annual gross revenue, with their revenue, jobs and payroll, 2009-2019. Taken from the Receita Federal's public workbooks 'Grandes Números do Simples Nacional' and summed over states: equal bins of R$ 120,000 up to the R$ 2.4 million ceiling in 2009-2011 and of R$ 180,000 up to R$ 3.6 million in 2012-2017, each year with an open top bin, and only the seven statutory brackets (up to R$ 4.8 million and above) in 2018-2019. Against a log-linear trend the last bin below the ceiling holds 49% to 99% more firms than predicted every year, against 3% to 9% in placebo bins. These are aggregate counts, not microdata: real size cannot be told apart from under-reported revenue, and firms excluded during the year drop out of the lower bins.
- `tables/lines/bunching-headcount.csv` (40 rows): **Establishments by headcount around the headcount lines.** One row per year and headcount tested: whether private establishments pile up at the last headcount free of a legal duty (16 'line' rows, such as time records above 20 workers or the disability quota from 100 employees) or at a headcount with no line (24 'placebo' rows). Counted in the public, anonymised RAIS establishment files (RAIS: the labour ministry's yearly register of formal employment) for 2018 and 2024: private establishments by active employment links on 31 December. Two measures: the ratio of the count at L to the count at L+1 against the same ratio at neighbouring headcounts, and the count at L against a smooth polynomial fit. The file has no firm identifier, so firm-level lines (99 and 200) are tested on establishments, and above about 50 workers there are too few establishments per headcount to detect effects under 10%.
- `tables/lines/mei-exit.csv` (6 rows): **Tax before and after the MEI line.** One row per annual revenue level: the year's tax for a service MEI (the one-person firm regime, up to R$ 81,000 of revenue a year) as its revenue crosses the ceiling. Computed from LC 123/2006 art. 18-A with 2026 values: a fixed DAS (the single monthly Simples payment) of R$ 86.05 a month, 5% of the R$ 1,621 minimum wage plus R$ 5 of ISS (the municipal services tax); between R$ 81,000 and R$ 97,200 the Simples rate on the excess only; one real above R$ 97,200 the exit is retroactive to 1 January and the whole year is taxed as a microempresa in Anexo III (6%). Costs that also start at the exit (bookkeeping, the partner's social-security contribution) are not included.
- `tables/lines/simples-marginal.csv` (30 rows): **Marginal rates inside the Simples bands.** One row per Simples Nacional annex and revenue band: the band's statutory rate, which is the marginal rate on revenue, and the effective rate at the band's edges. Computed from the tables of LC 123/2006 (Anexos I to V, in force since 2018): the effective rate is (RBT12 x nominal rate - deduction) / RBT12, so each extra real of revenue pays the nominal rate, while the deduction keeps the effective rate continuous across band edges. In band 6 (R$ 3.6 to 4.8 million) ICMS and ISS are paid outside the Simples and are not in these rates.
- `tables/lines/fator-r.csv` (8 rows): **The payroll-to-revenue ratio rule and its effect on the rate.** One row per revenue level: what the fator R rule is worth to a service firm in the Simples Nacional. Under LC 123/2006 art. 18 §5-J and §5-M a firm whose payroll (partners' pro-labore included) is at least 28% of revenue pays the rates of Anexo III instead of the higher ones of Anexo V; computed from the statutory tables. R$ figures are yearly; the cost of raising payroll (the partners' income tax and social-security contributions on a higher pro-labore) is not in the table.
- `tables/lines/presumed-versus-actual-profit.csv` (12 rows): **Lucro presumido against lucro real at several margins.** One row per kind of firm and profit margin: the federal tax of a firm with R$ 78 million of yearly revenue, the ceiling of the lucro presumido (the regime that taxes income on a fixed share of revenue), against what it would pay under lucro real (tax on actual accounting profit), compulsory above that line. Computed from the statutes for IRPJ and CSLL (the corporate income taxes) and PIS/Cofins (federal contributions on revenue) only: presumido uses presumption rates of 32% for services and 8% (IRPJ) and 12% (CSLL) for trade, raised by LC 224/2025 on revenue above R$ 5 million, and cumulative PIS/Cofins of 3.65%; lucro real taxes profit at 15%, plus 10% above R$ 240,000, plus 9%, and revenue at 9.25% net of credits on inputs. Margins and creditable input shares are assumptions, not data; ICMS, ISS, payroll charges and compliance costs are left out.
- `tables/lines/irpf-2026-marginal.csv` (37 rows): **Marginal income-tax rates under Lei 15.270/2025.** One row per gross monthly salary, from R$ 3,000 to R$ 12,000 in steps of R$ 250: an employee's monthly income-tax withholding (IRPF, the personal income tax) and social-security contribution (INSS) in 2026, with average and marginal rates. Computed from the 2026 monthly IRPF table with the reduction of Lei 15.270/2025 (no tax up to R$ 5,000 a month, the reduction phased out linearly up to R$ 7,350) and the 2026 INSS table, for a worker with no dependants who deducts the larger of the INSS contribution and the simplified deduction of R$ 607.20. Marginal rates are the tax on the next R$ 10 of salary; the employer's own charges are not included.
- `tables/lines/small-claims-lines.csv` (25 rows): **State small-claim (RPV) payment lines.** One row per state (the Federal District included): the line below which a court award against the state is paid within months as an RPV (requisição de pequeno valor) rather than through precatórios (court-ordered debts paid in a queue over years), in 2026 reais. Read in the corpus: the newest state law that sets the line for the state's own debts, converted with 2026 values of the minimum wage (R$ 1,621) and of state fiscal units; where no law was found (Alagoas, Pernambuco, Piauí, Sergipe, Tocantins) the constitutional default of 40 minimum wages applies (ADCT art. 87 I; ADCT: the Constitution's transitional provisions). Only SP, Minas Gerais and the Federal District were checked against official sites; the other texts are often original enactments, so the current line should be confirmed with the state court; Paraná and Rio de Janeiro are absent.

### Words

- `tables/words/definitions-top20.csv` (183 rows): **Twenty terms with the most legal definitions, and the definitions.** One row per selected legal definition. For each of the 20 economically loaded terms whose definitions diverge most (ranks 1-20, by div_overall in words/concepts-ranked) and six terms added on request (ranks 21-26: startup, empresa de base tecnológica, inovação, data center, renda familiar per capita, entidade sem fins lucrativos), it keeps up to five federal definitions, the four state definitions least similar to federal law (one per state), and the definitions holding the lowest and highest value of each numeric parameter. Definitions were extracted from article text in the corpus with patterns such as 'considera-se ...' and 'para os efeitos desta Lei, entende-se por ...' and compared as vectors of words and word pairs; terms that set the coverage of a statute rather than eligibility (servidor público, família, criança ...) are left out of the ranking. Definition texts are cut to 450 characters; Paraná and Rio de Janeiro are absent.
- `tables/words/concepts-ranked.csv` (61 rows): **Defined concepts ranked by how far their definitions diverge.** One row per curated concept: 61 economically or legally loaded terms, each gathering several wordings (baixa renda, for example, also covers 'família de baixa renda' and 'pessoa hipossuficiente'), with how many norms define it and how far the definitions diverge overall, within federal law, across states and between federal and state law. Definitions were extracted from article text in the corpus; each distinct defining text became a vector of words and word pairs weighted by rarity, texts were compared by cosine similarity, and 'meanings' are clusters of texts cut at similarity 0.25. Scores are heuristic and grow with the number of texts; concepts with very many texts are compared on a sample of them, numeric parameters are read by regular expressions, and Paraná and Rio de Janeiro are absent.
- `tables/words/disability-census.csv` (134 rows): **State laws that extend the legal meaning of pessoa com deficiência.** One row per state law that extends the legal status of pessoa com deficiência (person with disability, which federal Lei 13.146/2015, the LBI, ties to a biopsychosocial assessment) to a named condition such as autism, fibromyalgia or unilateral deafness: 134 laws in 23 states. Found in the corpus by definition patterns in article text and by a scan of the laws' official summaries (24 laws were found only that way and read in full); what the law requires before the status applies was read by hand for 34 laws and otherwise taken from the extending article or a keyword screen of the full text. The 63 laws marked 'screen: none found' were screened, not read; Paraná and Rio de Janeiro are absent.

### Borders

- `tables/borders/icms-incentive-programs.csv` (32 rows): **ICMS incentive programmes by state.** One row per main ICMS incentive programme of a state (ICMS: the state tax on goods and some services): its base law, mechanism, ceiling, what the firm gives in return, maximum term, a verbatim quote and whether the benefit can be compensated by the federal fund for the 2029-2032 cut. Read in the corpus from each programme's base law and its amending laws, with a few rows from official web pages; Paraná and Rio de Janeiro appear only as gap rows. Several base laws are missing or garbled by OCR in the corpus (see evidence and gaps), and most percentages sit in decrees outside it, so ceilings are incomplete.
- `tables/borders/icms-incentive-value-by-year.csv` (5 rows): **Value of an ICMS incentive through the 2029-2032 phase-down.** One row per kind of ICMS incentive: its value in each year from 2027 to 2033 as a share of its 2028 level, under the phase-out of the consumption-tax reform. Computed from the Constitution and the statutes: ADCT art. 128 (ADCT: the Constitution's transitional provisions) cuts ICMS rates and benefits to 9/10, 8/10, 7/10 and 6/10 in 2029-2032, ADCT art. 129 abolishes ICMS in 2033, onerous benefits granted by 31 May 2023 are compensated for the cut by a federal fund (EC 132/2023 art. 12), and commercial, port and farm benefits had been on a 20%-a-year cut under LC 160/2017 art. 3 §2-A before the reform replaced it. Values are before discounting and assume the fund pays in full; the §2-A path assumes the linear reading of '20% ao ano'.
- `tables/borders/icms-incentive-present-value.csv` (13 rows): **Benefit-years and present value of an incentive from 2027.** One row per kind of ICMS incentive and fund payout ratio: the present value on 1 January 2027 of the incentive's 2027-2032 flows, per unit of annual benefit at its 2028 level, at three real discount rates. Estimated from the yearly shares in borders/icms-incentive-value-by-year, in constant real terms with mid-year discounting, assuming the compensation fund pays 100%, 75% or 50% of each claim. The discount rates are assumptions, and values are in years of annual benefit, not reais.
- `tables/borders/compensation-fund-sufficiency.csv` (24 rows): **Whether the ICMS compensation fund covers the claims.** One row per scenario: whether the federal fund that compensates holders of onerous ICMS incentives for the 2029-2032 cut can pay their claims. Estimated: the fund's deposits (R$ 160 billion in 2023 reais, paid in 2025-2032 under EC 132/2023 art. 12) are set against claims of 10%, 20%, 30% and 40% of onerous benefits in 2029-2032, with onerous benefits taken as a share (theta) of the R$ 333.7 billion of state tax expenditures declared for 2026, with or without 2% real growth, and with or without 2023 inflation in the indexation. The onerous share is published nowhere and is the weakest number; amounts are R$ billion in 2026 reais.
- `tables/borders/icms-standard-rates.csv` (36 rows): **Standard ICMS rates by state, 2023 and 2026.** One row per state and law that changed its modal (standard) ICMS rate between 2022 and 2026; a state with no change, or whose law is not in the corpus, has a single row without law details. Read in the corpus from the states' ICMS laws of 2022-2026, with the rate clause matched to each state's main ICMS statute and quoted verbatim; Acre, Paraná, Rio de Janeiro and Sergipe rest on web sources (Paraná and Rio de Janeiro have no laws in the corpus), and the eight unchanged states rest on their main statute and on finding no later amending law. 19 of 27 states raised the rate; the mean rose from 17.6% to 19.3%.
- `tables/borders/itcmd-rates.csv` (27 rows): **Inheritance and gift tax rates by state.** One row per state (all 27, Paraná and Rio de Janeiro included): how its ITCMD (the state tax on inheritances and gifts) is charged, its top rates and its fit with LC 227/2026, which now requires marginal progressive brackets up to the Senate's 8% maximum. Taken from a third-party state-by-state table (TaxUp, CC BY 4.0) checked on 8 October 2026, with the federal frame read in the corpus; state laws were not re-verified one by one. 13 states are out of line with the marginal method: 6 charge a flat rate and 7 apply one bracket's rate to the whole base.
- `tables/borders/ipva-rates.csv` (27 rows): **Vehicle tax rates by state.** One row per state (all 27): IPVA (the state tax on motor vehicles) rates for cars, motorcycles and trucks, the state's age exemption and the treatment of electric and hybrid cars. Taken from two third-party compilations of official sources (simularipva.com.br, checked by the site on 5-6 October 2026, and recarregados.com, 3 October 2026), with the constitutional frame read in the corpus; rates were not re-verified law by law, and secondary sources disagree for a few states (see note). Rates are text with a comma as the decimal mark ('2,5'; ranges as '2,5 a 3').
- `tables/borders/state-waivers.csv` (27 rows): **ICMS waivers declared in the states' 2026 budget guidelines.** One row per state (all 27): the tax expenditure (revenue waived through state tax benefits, mostly ICMS, also IPVA and ITCMD reliefs) that it declared in its budget-guidelines bills for 2025 and 2026 (PLDO: the yearly bill that sets the next budget's targets), with ranks. Taken from the FGV-Ibre compilation of the PLDO annexes (October 2025), not from the corpus; the 2026 total is R$ 333.7 billion. Each state measures its waiver differently, so the figures compare declarations, not like-for-like costs; Espírito Santo's jump from R$ 3.6 billion to R$ 14.9 billion is more likely a change of method than of policy.
- `tables/borders/state-fees-and-funds.csv` (17 rows): **State fees and funds charged on top of ICMS.** One row per state fee or fund contribution charged on top of ICMS: mining and water-resource inspection fees charged per tonne or cubic metre, an oil-and-gas fee, and the infrastructure-fund contributions that states demand as a condition for ICMS benefits, with the charging base and the outcome at the STF (the Supreme Federal Court). Bases were read in the corpus; court outcomes come from STF news and decisions on the web. The last row describes ADCT art. 136, which lets Goiás, Mato Grosso and Mato Grosso do Sul turn ICMS-linked fund contributions into stand-alone ones until 2043.

### The federal menu

- `tables/menu/federal-regimes.csv` (44 rows): **Federal tax regimes with their legal basis, beneficiaries and end dates.** One row per federal tax regime or incentive in the catalog: 44 entries, from the Simples Nacional (the simplified regime for firms up to R$ 4.8 million of revenue) and the MEI (the one-person firm regime) to the Zona Franca de Manaus (ZFM, the Manaus free-trade zone), the Sudene/Sudam regional income-tax cut, investor exemptions and the rules of the 2025 tax reform (LC 214/2025, the law that creates the IBS and CBS value-added taxes). Each row was read by hand: a verbatim quote from the current text (from the corpus, or from planalto.gov.br where the corpus held only a fragment), eligibility, benefit, end date, the instruments the corpus lacks, and the status under LC 224/2025 (the 10% linear cut of federal tax benefits) and under the reform; sizes come from the DGT 2026 (the Receita Federal's statement of tax expenditures attached to the 2026 budget bill) and from declared or reported use. Sizes are R$ millions a year; the mapping of DGT lines to regimes is approximate (ZFM: R$ 35.2 billion here, R$ 36.0 billion in menu/tax-expenditures-2026-largest), and end dates are those of the current law, which differ from the terms the DGT 2026 prints for Sudene/Sudam, Pronon/Pronas, Padis, petrochemicals and sport. Free-text columns are in Portuguese; rows of group 'reform' and 'boundary' describe rules of the reform and the 2027 ban on new benefits rather than benefits.
- `tables/menu/federal-regimes-ranked.csv` (44 rows): **Federal regimes ranked by size and underuse.** The 44 regimes of menu/federal-regimes sorted by their size in the DGT 2026 (the Receita Federal's statement of tax expenditures attached to the 2026 budget bill), with the other measures of size beside it. Derived from menu/federal-regimes without new data. Ranks run 1-23 over regimes with a DGT value above zero; Perse and FIP-IE (DGT value 0) and the 19 regimes the DGT does not count have no rank. Values are R$ millions a year; size_best_rs_mi takes the larger of the DGT estimate and the other measure, so for the ZFM, the food-and-farm PIS/Cofins zero rate, Lei do Bem and Perse it holds the declared or reported figure, and for the Imposto Seletivo it holds planned tax revenue, not a benefit.
- `tables/menu/tax-expenditures-2026-by-tax.csv` (216 rows): **Federal tax expenditures in the DGT 2026 by tax.** One row per tax and expenditure item in table VII of the DGT 2026, the Receita Federal's statement of federal tax expenditures attached to the 2026 budget bill: 216 items over 15 federal taxes and contributions, summing to R$ 612.8 billion. Taken from the official spreadsheet without change. A benefit that touches several taxes appears once per tax (the Simples Nacional appears under IRPJ, CSLL, PIS, Cofins, IPI and social security), so a benefit's total needs a sum across taxes. Values are the Receita's estimates for 2026, in reais (not millions).
- `tables/menu/tax-expenditures-2026-largest.csv` (65 rows): **The largest federal tax expenditures in the DGT 2026.** One row per expenditure group in table X of the DGT 2026, which ranks the main groups of federal tax expenditures across all taxes, plus a closing TOTAL row (R$ 612.8 billion, share 1). Taken from the official spreadsheet without change; do not add the TOTAL row to the others. The groups are the DGT's own (e.g. 'Agricultura e Agroindústria', 'Desenvolvimento Regional'), not the regimes of menu/federal-regimes, so values differ from that table where regimes were mapped line by line (ZFM: R$ 36.0 billion here, R$ 35.2 billion there). Values are the Receita Federal's estimates for 2026, in reais.
- `tables/menu/simples-hybrid-2027.csv` (60 rows): **Who gains from the Simples hybrid choice for 2027.** One row per annex of the Simples Nacional (the simplified regime for firms up to R$ 4.8 million of revenue), revenue band and assumed 2027 tax rate: 5 annexes, 6 bands, 2 rates. Computed from the 2027-2028 rate tables of LC 123 (the Simples law) as rewritten by LC 214/2025 art. 517 and LC 227/2026, it estimates whether a Simples firm gains by paying the new IBS and CBS value-added taxes in the regular regime (the 'hybrid' option, chosen by 30/10/2026 for January-June 2027) instead of inside its single monthly payment, the DAS. All rates and gains are shares of revenue or price, from 0 to 1, computed at the top of each band. These are estimates: the 2027 CBS rate was not yet fixed (the Senate sets it by 15/12/2026), and the model assumes business buyers keep their net cost, consumers pay the same price and the firm's inputs bear the full rate.
- `tables/menu/sudene-active-2026.csv` (159 rows): **Active Sudene incentive holders in 2026.** Counts of the holders of Sudene's 75% income-tax cut whose benefit runs in 2026. Sudene is the federal development agency for the Northeast, northern Minas Gerais and Espírito Santo; its laudo (approval certificate) grants a 75% cut of the corporate income tax on operating profit for 10 years from the year it is issued, so laudos issued in 2017-2026 run in 2026. Counted from Sudene's open file of approvals since 2011 (6,459 rows, downloaded 15/10/2026, last laudo 08/10/2026), checked against Sudene's announced 2025 totals; laudos later revoked or replaced are not netted out. Two counts are stacked: 'firms by UF' counts distinct firms (by the first 8 digits of the CNPJ, the company tax number, transfers of laudos included) per state of the project, and a firm with projects in several states counts in each, so its 11 rows add up to 2,864 against 2,527 distinct firms; 'laudos by sector' counts 3,596 of the 3,597 running laudos of the 75% cut by Sudene's sector label (one has no sector, inferred).
- `tables/menu/pis-cofins-to-lc214.csv` (37 rows): **Where each PIS/Cofins benefit lands under LC 214/2025.** One row per PIS/Cofins tax expenditure in the DGT 2026 (the Receita Federal's statement of tax expenditures attached to the 2026 budget bill), with the treatment that the 2025 tax reform (LC 214/2025, compiled with LC 227/2026) gives the same goods, services or entities from 2027. PIS and Cofins, the federal contributions on revenue, end on 31/12/2026 and are replaced by the CBS. The 37 lines with value add to R$ 167.0 billion; each was mapped by hand after reading the cited articles. The mapping covers the federal side only, and 'kept_partly' does not say how the base splits between zero and reduced rates, which is why the essay gives retained relief as a range (62-93% outside the Simples).
- `tables/menu/lc214-dates.csv` (408 rows): **Every date in LC 214/2025.** One row per date written in the own articles of LC 214/2025 (the law that creates the IBS, the CBS and the Imposto Seletivo), compiled with LC 227/2026: 408 matches of 'dd de mês de yyyy', 'a partir de yyyy' (from yyyy), 'até yyyy' (until yyyy) and 'em yyyy' (in yyyy) across 140 articles. Found by pattern in the repaired compiled text; text that the law writes into other laws stays inside the article that writes it. The matches include the dates of laws cited by name (e.g. '14 de dezembro de 2006', the date of LC 123), not only deadlines and starts, so read the excerpt before counting deadlines.
- `tables/menu/lc214-thresholds.csv` (24 rows): **Every numeric line in LC 214/2025.** One row per amount in reais written in the own articles of LC 214/2025, compiled with LC 227/2026: 24 matches across 13 articles, found by pattern in the repaired compiled text. They include the revenue lines that decide who pays IBS/CBS (R$ 3.6 million for rural producers, R$ 240,000 for individual landlords), the real-estate social reducers, the car price caps for disabled buyers, and the LC 123 amounts that art. 517 rewrites (R$ 81,000 for the MEI, the one-person firm regime). Amounts are nominal reais as written, and some are not thresholds: art. 341-C fixes the R$ 200 tax unit, art. 484 the federal funding of the IBS managing committee in 2025-2028, and art. 537 per-cubic-metre rates on ethanol.

### Eighteen ways to buy an hour

- `tables/work/cost-wedge.csv` (108 rows): **Employer cost and worker net under each way of buying work.** One row per way of buying work and monthly gross pay: 18 legal arrangements at six levels from R$ 1,621 (the 2026 minimum wage) to R$ 100,000 a month. The arrangements are CLT employment (the labor code's formal employment) under several employer regimes, through a temporary agency or an outsourcing provider, domestic, apprentice and intern contracts, a work cooperative, an autonomous contractor paid by RPA (the receipt for paying a self-employed individual), a MEI (the one-person firm regime) and PJ companies (a worker billing through his own firm) in the Simples Nacional or Lucro Presumido, and athlete image rights. Computed under the 2026 rules read from the corpus and the 2026 INSS (social security) table: contribution and income-tax tables, the 2026 income-tax reduction of Lei 15.270/2025, and the Simples annexes of LC 123. Estimates: accountant fees, the cooperative fee and the agency and outsourcing markups are assumptions, CLT benefits (meal, health, transport) are left out, and FGTS (the employee's severance fund) counts at face value as the worker's money; a CLT year has 13.33 monthly pays (12, the 13th salary and the vacation third), the other forms 12.
- `tables/work/cost-wedge-equal-budget.csv` (54 rows): **The same budget spent under each arrangement.** One row per arrangement and budget: the 18 arrangements of work/cost-wedge, each given the annual budget that an employer in the regular regime spends on a CLT employee earning R$ 3,000, 10,000 or 30,000 a month (R$ 55,600, 185,333 and 556,000), with the gross pay that budget buys and what the worker nets. Computed by bisection under the same 2026 rules and assumptions as work/cost-wedge (accountant, fee and markup assumptions; no CLT benefits; FGTS at face value).
- `tables/work/cost-wedge-public-take.csv` (54 rows): **What the Treasury and social security collect under each arrangement.** One row per arrangement and budget, as in work/cost-wedge-equal-budget, with what the Treasury and social security collect: public take is the budget minus the worker's net (FGTS included) minus private costs (accountant, agency or cooperative fees, markups). Computed with the same 2026 rules and assumptions. The public take of the Simples employer leaves out the contribution paid inside the firm's DAS on revenue, and that of the payroll-relief employer leaves out the CPRB on revenue, so both are understated.
- `tables/work/six-by-one-sectors.csv` (65 rows): **Workers and cost of the end of the 6x1 week by sector.** One row per sector (a division of CNAE, the national classification of economic activities) in Ipea's 2026 technical note on working hours, with the cost of ending the '6x1' week (six working days for one of rest). The constitutional amendment PEC 221/2019, approved by both houses and awaiting promulgation on 15 October 2026, caps the normal week at 42 hours 60 days after publication and at 40 hours a year later, with no pay cut. Job counts, the 40-hour hourly rise and labor's share of costs are Ipea's figures (RAIS 2023, the employers' annual employment return, and IBGE sector surveys), parsed from the note's tables; the 42-hour step is the 40-hour figure scaled by 4.76/10, exact for 44-hour contracts and an upper bound where 41-43 hours are common. Total-cost effects exist only for the 31 sectors with a labor share; health has none in the note, and RAIS 2023 predates three years of wage growth.
- `tables/work/six-by-one-effects.csv` (23 rows): **Hourly cost and staffing effects of the 42- and 40-hour steps.** One row per computed effect of the 42- and 40-hour steps of PEC 221/2019, the constitutional amendment ending the 6x1 week: the hourly cost at constant monthly pay for contracts of 41 to 44 hours, the hourly minimum wage under the new monthly divisors, the cost of keeping 44 hours of work through overtime, staffing of round-the-clock 12x36 posts (12 hours on, 36 off), and the hours to replace. Arithmetic on the text both houses approved (the Câmara substitute, passed unchanged by the Senate on 14 October 2026); the formula of each row is in note. The overtime reflex on paid rest with two rest days (2/5 instead of about 1/6) and the 12x36 reading are legal inferences, the hourly minimum wage assumes the divisor follows the cap, and the hiring shares are estimates.
- `tables/work/labor-corpus-map.csv` (88 rows): **Where each work arrangement lives in the statute book.** One row per provision that defines, limits or prices a legal way of buying work: 88 provisions from 18 federal norms and the Constitution, among them the CLT (the labor code), LC 123 on the Simples and the MEI, the laws on interns, temporary work and outsourcing, cooperatives, domestic and rural work and athletes, and the laws on social security contributions, the FGTS and income tax. Each phrase was searched in the corpus text (the repaired text where one exists), extended to the end of its clause and tied to the article heading before it; all 88 were found. Quotes come from compiled current texts except where versao_texto says 'original'; the Constitution's art. 7 XIII and XV are quoted before PEC 221/2019, which will change them.

### Clocks

- `tables/clocks/calendar-2026-2040.csv` (120 rows): **Dated windows from 2026 to 2040, checked against the laws.** One row per dated legal boundary (a window, start, end, deadline or transition step) from 2026 to 2040, with its legal basis and a verbatim quote: 76 boundaries read by hand in the law (origin 'curated') and 44 federal tax-expenditure items of the DGT 2026 (the Receita Federal's statement of tax expenditures attached to the 2026 budget bill) whose end date was re-read in the current law (origin 'dgt'). Every quote was checked automatically, word for word, against the corpus text, the repaired text or a saved official page (all 120 verified), and the article holding it was read back. 118 rows fall inside 2026-2040 (the 2045 and 2073 rows are kept); 13 dates are estimates or computed from relative terms (column est); for DGT rows the date is the term under the current law, which replaces the DGT's printed date for the 13 items marked 'changed', and the amount is the DGT's 2026 estimate in R$ a year. State-level windows and regulations outside the corpus (CGSN, CONFAZ, ANEEL) are not covered, and the dates of PEC 221/2019 (the end of the 6x1 work week) move with its publication.
- `tables/clocks/closing-dates.csv` (11 rows): **Dates on which a regime, benefit or option closes.** One row per day on which three or more of the boundaries in clocks/calendar-2026-2040 end between 2026 and 2040 (the close date, or the date of a single event). Counted from that table: 31/12/2028 closes 19, 31/12/2029 16 and 31/12/2026 15. The amount is the sum of the DGT 2026 tax-expenditure items ending that day, R$ billions a year; boundaries read in the law carry no amount here.
- `tables/clocks/tax-expenditure-sunsets.csv` (63 rows): **Federal tax expenditures by end year.** One row per tax and item of the DGT 2026 (the Receita Federal's statement of tax expenditures attached to the 2026 budget bill) whose printed end date (prazo de vigência, term) falls between 2026 and 2075: 63 items, R$ 92.0 billion a year, of which R$ 57.0 billion end by 2040. Taken from the DGT's legal-description tables without change, so the dates are the DGT's; clocks/tax-expenditure-term-check re-reads those of 2026-2040 in the current law, where 9 of 27 have moved or no longer end. The 2050 items are the free-trade areas and the 2073 items the Zona Franca de Manaus; values are the Receita's 2026 estimates in reais.
- `tables/clocks/tax-expenditure-term-check.csv` (27 rows): **End dates in the DGT checked against the laws.** One row per DGT 2026 expenditure line dated between 2026 and 2040, with its term re-read in the current law: 27 lines (the taxes of the same item merged), R$ 57.0 billion a year. Each term was read by hand and its quote checked automatically against the corpus, the repaired text or a saved official page. 17 lines are confirmed (R$ 29.7 billion), 8 were changed by a later law (R$ 26.3 billion), the sport incentive no longer ends and Perse ended when its R$ 15 billion cost cap was reached; the 2038 end of the four Sudene/Sudam lines is an estimate (approval by 2028 plus 10 years of enjoyment).
- `tables/clocks/timing-ranking.csv` (26 rows): **Windows ranked by what acting inside them is worth.** One row per dated window ranked by what acting inside it is worth, as of 15 October 2026: the 26 boundaries of clocks/calendar-2026-2040 that carry arbitrage metadata. Group A (ranks 1-17) holds windows where acting before the date captures or locks value (registering, locking in a treatment, using a benefit window, front-loading before a sunset); group B (ranks 18-26) holds dated changes that create cost or demand (phase-outs, obligations, market openings). Within a group, score = log10(stake per year in R$) + urgency − 0.5 × (risk − 1), a transparent heuristic rather than a valuation, and the 12 windows without a national figure follow the scored ones by days left. Stakes come from the DGT 2026 or from amounts written in the law, with totals spread over their years; the stake of the 6x1 row is a cost to employers.

### Markets the law forces

- `tables/forced/forcing-laws.csv` (8,171 rows): **Laws that force a purchase or a service, with object, organ and sanction.** One row per norm that a text scan of the corpus flagged as possibly forcing someone to buy, install, keep, hire, train, certify, insure or report something: 8,171 federal and state laws (ordinary, complementary, promulgated and delegated; LC: Lei Complementar) and decree-laws, 7,685 of them with at least one obligation paragraph (constitutions, constitutional amendments and organic laws were skipped). Made by searching the corpus for forcing wording in the ementa (the official summary) and obligation phrasing in the articles, then tagging object, addressee, action, sanction, enforcing organ, need for regulation, deadline and mentions of unconstitutionality with regular expressions over accent-free lowercase text. This is the raw scan before any filter: the headline count of 2,291 forcing base laws comes from a stricter census (private or unspecified addressees, not known to be revoked or declared unconstitutional, amendments counted apart), and in a hand sample of 60 of those laws 77% truly oblige private actors. The status columns are the source headers as they stood at scan time: every Santa Catarina row reads revoked because that source's flag was broken, most states carry no status at all, and Paraná and Rio de Janeiro are absent.
- `tables/forced/forced-objects.csv` (5,720 rows): **The goods and services the law forces, counted.** One row per pair of a norm in forced/forcing-laws and an object class tagged in it: 5,720 rows covering 4,106 norms and 41 object classes. Made by the same text scan, which matches each object class with regular expressions in the ementa (the official summary) or in the paragraphs that carry obligation phrasing. Like forced/forcing-laws this is the raw scan, before the filters on addressee, revocation, unconstitutionality and amendments that give the census counts used in forced/forced-markets-sizing, so per-object counts here differ from those (mostly larger: fire prevention, for instance, is tagged in 918 norms here and counted in 38 laws there). Santa Catarina's status_flag is unreliable, and Paraná and Rio de Janeiro are absent.
- `tables/forced/forced-markets-ranked.csv` (24 rows): **Forced markets ranked.** One row per business or investment opportunity opened by the forcing laws or by a federal framework of 2019-2026 (sanitation, low-voltage electricity, crypto-asset providers, mutual vehicle protection, biomethane, betting and others): 24 rows. Each is scored 1 to 5 on size, timing, access for a small team, access for an investor, and risk; the scores are judgements argued from press, industry and government sources and from forced/forced-markets-sizing, not measurements. small_score = size + timing + 2 x access_small - risk and investor_score = 2 x size + timing + access_investor - risk, so a user can reweight them. The compliance_kit row's legal_basis cites 2,329 forcing laws, the census count before its final recount; the reconciled count is 2,291.
- `tables/forced/forced-markets-sizing.csv` (22 rows): **Size of the twenty largest forced markets.** One row per purchasable object among the 22 most legislated (the title says twenty), with the spending that full compliance with its forcing laws would require. Law counts and enforcement shares come from the final census of 2,291 forcing base laws (in force or of unknown status, private or unspecified addressees, amendments apart); in_scope = national count of the target x applicability x share of population (or of shopping centres) in states with at least one law, and the money is in_scope x units x unit cost, kept apart as one-off capex and yearly opex. These are estimates of full-compliance spend, not revenue anyone can capture: counts and prices come from cited public pages or are marked 'estimate', 15 of 22 rows are low confidence, and the state shares are inflated by classifier errors and deflated by counting Paraná and Rio de Janeiro as having no law. The table holds the reconciled figures after the census recount: R$ 1,599 million one-off and R$ 4,562 million a year in total.

### The Union and the states

- `tables/union/stf-adverse-state-norms.csv` (707 rows): **State norms the STF struck, suspended or narrowed.** One row per state norm whose official text carries a note that the STF (the Supreme Federal Court) struck it, suspended it, or narrowed it by interpretação conforme (a ruling that keeps the text but fixes its reading), in whole or in part: 707 norms in 21 of the 25 states present. Made by scanning the state corpus for control markers ('declarada inconstitucional', 'ADI', 'eficácia suspensa' and the like) and classifying each note line by court, outcome and scope; theme, vertical and private_duty come from keywords in the ementa (the official summary). 94% of the rows (663) sit in the nine states whose sources annotate control outcomes (DF, SP, ES, MT, MG, SC, GO, MS, PE), so elsewhere a missing note means unknown; the scan reads notes in the body, not the ementa, and a note can be outdated (suspensions later lifted still count). The table holds the raw flags: a later web review of 17 copy families found 2 of them wrong (SP Lei 13.541/2009, whose ADI 4249 was dismissed as moot, and MT Lei 12.709/2024, upheld in August 2026) and 2 strikes the scan missed (RO Lei 2.026/2009 and PB Lei 11.782/2020).
- `tables/union/struck-by-state.csv` (25 rows): **Struck state norms by state.** One row per state present in the corpus (24 states and the Federal District), counting its norms of every kind by the notes of constitutional control their texts carry. Made from the same marker scan as union/stf-adverse-state-norms: each note is classified by court (the STF, the state's Court of Justice, or unknown), outcome (struck, suspended, narrowed, upheld, no decision, cited) and scope (whole or part), and a norm counts once in each column that applies; the counts use the raw flags, including the two that a later review found wrong (see union/stf-adverse-state-norms). The counts measure annotation as much as litigation: the nine annotating states hold 663 of the 707 STF-adverse norms, elsewhere a missing note means unknown rather than unchallenged, and Paraná and Rio de Janeiro are absent.
- `tables/union/struck-by-theme.csv` (24 rows): **Struck state norms by theme, per thousand laws.** One row per theme of state law (24 codes, from keywords in the ementa), counting the laws of the nine annotating states (DF, SP, ES, MT, MG, SC, GO, MS, PE) by the notes of constitutional control in their texts, overall and for the laws that impose duties on private firms. Laws here are ordinary, complementary and delegated laws and decree-laws; the marker scan is the one behind union/stf-adverse-state-norms, and the theme is the first keyword match in a fixed order, so it is noisy ('transito', traffic, also catches traffic-department careers and IPVA laws). Rates per 1,000 laws compare themes; struck_share_of_decided is biased upward because annotators record strikes more often than upholds.
- `tables/union/risk-by-theme.csv` (24 rows): **Untested state laws on federal matters and their modelled risk, by theme.** One row per theme (23 themes and a TOTAL row) of state laws that sit on a federal matter or impose duties on private firms, counting how many carry no control note and are not known to be revoked, and how a constitutional-risk score rates those untested laws. The score is a logistic regression on ementa words, theme, decade and the private-duty and vertical flags, trained on the laws of the nine annotating states with an STF-adverse note as the label and no state feature (leave-one-state-out AUC 0.73-0.96, median 0.89); 'revoked' uses only the catalog flag and status text, with Santa Catarina's flag ignored. For non-vertical themes only private-duty laws enter and the TOTAL row sums all 23 themes, so the headline figures are the sums of the sixteen vertical rows (16,267 unchallenged laws not known to be revoked, 1,370 of them with private duties, 4,082 in the top risk decile); the later revocation data from the citation graph, not applied here, lowers the first two to about 15,800 and 1,345. The expected-strikes range applies the 75-82% rate at which published studies find the STF strikes the state laws it judges; the headline estimate instead uses a deliberately conservative 50-75%, or 2,000 to 3,100 of the 4,082.
- `tables/union/tax-theses-2026.csv` (14 rows): **Settled tax theses that still pay in 2026.** One row per tax thesis (a legal question settled or pending before the STF, the Supreme Federal Court, or the STJ, the Superior Court of Justice) that settles a federal-state or tax-base conflict, with its status in 2026 and who gains or pays: 14 rows. Status, value and winners were read from court, CONFAZ (the council of state finance secretaries) and specialist pages on 15 October 2026, with URLs in sources; the corpus columns count state norms whose text carries a clause tied to the thesis, found by keyword search and recounted after the revocation data were revised on the evening of 15 October. Corpus counts are mentions, not distinct taxes (a code, its amendments and budget laws that name a fund all count), money figures are third-party estimates as cited, and readings marked 'my reading' are the analyst's own and were not checked against case law.
- `tables/union/rental-car-ipva-siblings.csv` (14 rows): **State IPVA rules on rental cars like the one struck in ADI 4376.** One row per state law found to carry the rental-car clauses of IPVA (the state tax on motor vehicles) that the STF struck in SP's Lei 13.296/2008 in ADI 4376 (judged in August 2026): 14 rows, namely SP's own law, mirrors in ten other states (two laws each for ES and SC), and one false positive, GO Lei 11 of 1892 on hiring of services, which should be dropped. Found by a keyword search of the state corpus for the three clause types the STF examined (tax due when a car is rented or made available for rent in the state; tax due at the renter's domicile; joint liability of partners, managers, renters or public agents), quoting the first match of each. The texts are the corpus versions, often the original enactment, and their current wording was not checked on the assemblies' sites; none of the mirror laws carries a note of the 2026 judgment.

### Seventy-seven strategies

- `tables/strategies/strategies.csv` (77 rows): **Seventy-seven strategy records, sixty-nine lawful strategies and eight anti-stacks, ranked by urgency and size, with both red teams' verdicts.** One row per lawful strategy for using a seam of the law, 77 in all, written after the analysis by four strategy lists (products and services a small team can build; capital: structure, location, time windows, vehicles, recovery and litigation, new markets; public-interest actions; combinations of seams that add up or exclude each other) and merged with the verdicts of two red teams (adversarial reviews that tried to refute each premise), one checking tax claims and one checking labour, regulatory and constitutional claims against official texts. Rows are sorted by urgency (days from 16 October 2026 to the first date in window_end) and then by size_mid_brl. Sizes are estimates in yearly reais unless size_basis says otherwise, with the arithmetic in size_basis; the red-team verdicts and issues were given before the fixes, which every list has since applied (fixes_applied), so they describe problems the other fields already correct. abuse_line states where each strategy would cross from planning into abuse.
- `tables/strategies/strategy-calendar.csv` (44 rows): **The capital strategies' dated steps.** One row per dated action or deadline that the capital strategies (structure, location, time windows, vehicles, recovery and litigation, new markets) depend on, from 16 October 2026 to 1 January 2040: 44 rows. Dates were read from the statutes, regulations and official notices cited in legal_basis, as they stood on 16 October 2026. The dates that follow from the end of the 6x1 working week (six days of work for one of rest) run from the publication of constitutional amendment EC 141/2026 in the DOU, the federal official gazette, which had not happened when the table was written, so they may shift.

### Stacks

- `tables/stacks/stack-compatibility.csv` (72 rows): **Which strategies combine, which exclude each other, and why.** One row per pair of seams (tax regimes, incentives, legal forms or rules) checked for whether a firm or person can use both: 72 pairs, 26 compatible, 18 exclusive and 28 conditional. Each pair was read in the statutes, with what each side acts on, the legal reason, the effect on the combined gain, the article and a verbatim quote, and every quote was checked automatically against the corpus or a saved official copy. The table holds the version after the fixes required by the reviews in line/tax-strategy-audit and line/labor-strategy-audit (before them the split was 27 compatible, 20 exclusive and 25 conditional); the gains in stacking_effect are estimates with their arithmetic.

### The line

- `tables/line/tax-claims-checked.csv` (31 rows): **Tax claims checked against the official text, with the corrected fact.** One row per factual claim from the analysis that the tax red team (an adversarial review) checked against official texts on 16 October 2026: 31 rows, 23 confirmed and 8 corrected. Sources were official first (planalto.gov.br, normas.leg.br, the corpus with its repaired federal texts, STF and STJ pages, the Receita Federal, CGSN (the Simples Nacional management committee), CONFAZ, state finance and assembly sites, Congress pages), and a point resting on secondary sources was accepted only when two independent ones agreed. Where a claim is corrected, correct_statement holds the verified fact, which supersedes the claim.
- `tables/line/tax-strategy-audit.csv` (68 rows): **The tax red team's audit of each strategy.** One row per strategy or compatibility pair whose tax premises the tax red team (an adversarial review) audited on 16 October 2026: 68 rows (14 small-team products, 25 capital strategies, 18 combinations of seams and 11 pairs of stacks/stack-compatibility), 26 found sound and 42 needing a fix, none unsound. Each row was tested for verbatim quotes against the compiled law, window dates, arithmetic, an abuse line consistent with the rules on simulation and evasion, and fair risk and confidence ratings. The verdicts were given before the fixes; the strategy lists applied the corrections afterwards, so strategies/strategies and stacks/stack-compatibility already hold the corrected fields.
- `tables/line/labor-claims-checked.csv` (74 rows): **Labour and regulatory claims checked against the official text.** One row per factual claim on labour law, regulation or federal-state conflict that the labour and regulatory red team (an adversarial review) checked against official texts on 16 October 2026: 74 rows, 37 confirmed, 32 corrected, 3 refuted and 2 unresolved. Sources were taken in order: the law on planalto.gov.br or normas.leg.br or in the corpus; STF, TST (the Superior Labour Court) and STJ pages; Senate and Chamber pages; state sites; ministry, agency and Central Bank pages. Press and law-firm pages served only to locate a primary source; where none existed two independent sources had to agree or the verdict is unresolved, and 'no ADI found' means a negative web search, not a query of the STF case registry.
- `tables/line/labor-strategy-audit.csv` (96 rows): **The labour red team's audit of each strategy.** One row per item whose labour, regulatory and constitutional-conflict premises the labour and regulatory red team (an adversarial review) audited on 16 October 2026: 96 rows covering the strategies of all four lists, the rows of payers/who-pays, 8 pairs of stacks/stack-compatibility and the written reports of three lists; 68 found sound and 28 needing a fix, none unsound. Items with nothing in the team's scope (20 rows) are marked sound with a note saying so, their tax side being left to the tax red team (line/tax-strategy-audit). The verdicts were given before the fixes; the lists applied the corrections afterwards.

### Who pays

- `tables/payers/who-pays.csv` (31 rows): **Who loses from each seam, and the public lever against it.** One row per seam of the law with who loses from it, how much, who could fix it and with what public instrument: 31 rows. Compiled from the analysis and the checks in line/tax-claims-checked and line/labor-claims-checked: magnitudes come from the analysis tables (tax expenditures from the DGT, the Receita Federal's yearly statement of tax expenditures, and the states' 2026 budget-guideline bills) or are estimates marked as such, and who loses is often an inference, since distribution data do not exist. The money columns are yearly reais and are filled only for the 9 rows with a national yearly figure, and tax expenditures there are gross revenue forgone at current behaviour, not what a repeal would raise; the other rows give their magnitude per worker, per firm or as a count in magnitude and magnitude_basis.

## Module reports, code and method

- `modules/`: the twenty folders of the research (fourteen modules, two red teams and four strategy tracks), each with its report, its README and its other outputs (371 files). A file a report names that is also a core table is under `tables/`; `MANIFEST.csv` maps every original file name to its published path.
- `code/`: the 214 scripts that built the corpus tables, ran every module, merged the strategies and generated the book's figures, with `requirements.txt` (Python 3.12) and the MIT licence.
- `methods/`: the research plan, the rules every agent followed, and the synthesis brief with its reconciled facts.

## How to rebuild

1. Clone https://github.com/ctresb/leis and check out commit `51dc03feb0d1d3b1ce2ae340cba84e23864d2273`, whose texts were downloaded from the official sources on 13 October 2026.
2. Create a Python 3.12 environment and install `code/requirements.txt`.
3. Run `code/build_corpus.py` to load the corpus into tables, then the repair (`code/00-repair/`), and the modules in order, `code/01-verify/` to `code/13-conflicts/`. Each module's README lists its scripts in the order they run.
4. The strategy tracks (`code/S1-builder/` to `code/S4-stacks/`) and the red teams (`code/R1-redteam-tax/`, `code/R2-redteam-labor-reg/`) read the module outputs; `code/report/merge_strategies.py` merges them into the strategies table.

Web checks (court dockets, the official gazette, government pages) were made on the dates their rows give, between 13 and 16 October 2026; a rebuild on a newer snapshot will differ where the law has changed.

## Limits

- Coverage: statutes only. Paraná and Rio de Janeiro are missing; decrees, provisional measures, CONFAZ agreements, normative instructions and municipal law are absent, except where a module fetched one and says so. State texts are mostly the original enactment, so a state law may have been amended or revoked without its file showing it.
- Status: 279,795 norms are in force by the best reading of the corpus, the official compilations and the courts, but tens of thousands of state norms carry no status at all, and revocations go unrecorded.
- Classes come from a classifier with a held-out precision of 0.845; the corrected shares carry 95% intervals. Rows marked as estimates are arithmetic on stated assumptions, not observations.
- The strategies are lawful readings checked by two red teams, not advice; each names the step that would cross into abuse.

## Privacy

Every text file passed one scrub: local paths and e-mail addresses removed, a few names replaced by a description, and clock times given without a zone. Ementas and text excerpts lose the clauses that name a bill's author and the names in governors' signatures. Copies of third-party and official pages, raw API responses, embedding matrices, intermediate Parquet tables and run logs are withheld and listed in `MANIFEST.csv` (1,857 entries).