# Compile the law: the data behind the book The open data of the book *Compile the law* (https://future-seems-so-good.com/blog/compile-the-law). In October 2026 a four-wave search by AI agents looked for software that Brazilian small businesses, professionals and consumers would buy, with a deadline, a regulation or a price gap as the reason to buy now. It wrote 93 research reports, 698 idea records, a ranking of 220 product families, and ten products were built from the top. This package holds all of it in a form other people and other agents can check, join and reuse. License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/), except the StatCounter tables (CC BY-SA 3.0). Cite as: Martin Roark, *Compile the law: the data behind the book*, future seems so good, 2026-10-12, https://future-seems-so-good.com/blog/assets/compile-the-law/data/. ## Start here - `tables/ideas.csv`: every idea at every stage, with lineage. `tables/families.csv`: the ranking. - `tables/findings.csv`: 1,008 findings, verbatim, classified by asymmetry type, polarity, persona and date. - `tables/obligations.csv`: the dated rules that force decisions. `tables/personas.csv`: who is forced. - `tables/economic-value-summary.csv`: what the products would be worth to buyers and to Brazil. - `datapackage.json`: every table with its schema and field descriptions. `compile-the-law-data.xlsx`: the same tables as one workbook. `compile-the-law-data.zip`: everything at once. - `MANIFEST.csv`: every file the search produced, published, merged into a table, or withheld with the reason. Join keys: `family_id` (families, ideas, r50-cards, products, economic-value), `report_id` (reports, findings, ideas, r50-cards), `finding_id` (findings, sources), `persona` codes (personas, findings, ideas, families, obligations), `url` (sources, findings, r50-cards). ## Tables ### ideas - `tables/ideas.csv` (698 rows): **Every idea, at every stage.** All 698 idea records of the search in one table: the 208 wave-1 cards, the 19 wave-2 deep dives, the 100 wave-3 re-scores of the earlier ideas, the 331 R50 variants of waves 3 and 4, and 40 ideas screened out in wave 4. Each row has its stage's own score and verdict, a percentile within the stage, an English gloss, buyer personas, the job the software does, whether a rule forces the purchase and its legal basis, and its lineage: the merged idea a card went into, the family a variant belongs to, the product it became. - `tables/wave1-opportunities.csv` (208 rows): **Wave 1 opportunity cards.** The 208 opportunity cards written by the 30 wave-1 agents on 11 Oct 2026, scored on the first rubric (demand, willingness to pay, competition gap, channel fit, build ease, low risk, recurrence, defensibility), with price, unit forecasts, target buyers, channels and the forcing function each card cites. - `tables/wave1-card-families.csv` (208 rows): **Wave 1 cards grouped into families.** The wave-1 cards grouped by the synthesis into idea families, with their weighted score, price and base units. - `tables/wave2-deep-dives.csv` (19 rows): **Wave 2 deep dives.** The 19 product families examined in depth in wave 2: verdict, revenue forecast (base, low, high, net), ten criterion scores, MVP weeks, fiscal and hardware needs, why each wins and fails, spot checks and the red team's verdict. - `tables/wave3-card-map.csv` (255 rows): **Lineage from wave-1 cards to wave-3 merged ideas.** How each wave-1 and wave-2 card was merged into a wave-3 idea (merged id), with the wave-1 family and score and the merged idea's wave-3 verdict and composite. - `tables/wave3-rescore.csv` (100 rows): **Wave 3 re-score of the earlier ideas.** The 100 merged ideas from waves 1 and 2 re-scored under the wave-3 rubric (paid demand, reach, price power, zero maintenance, zero support, competition gap, demo power, speed, ad safety, synergy), with the lite version, exclusions, evidence, kill reasons and why survivors survive. - `tables/wave3-candidates.csv` (132 rows): **Wave 3 candidates.** The 132 candidates written by wave-3 vertical agents, scored on the wave-3 rubric, with buyer, entry price, key evidence and the report each comes from. - `tables/idea-screen-mei.csv` (28 rows): **MEI 2027 idea screen.** 28 ideas screened for the MEI 2027 obligations, with outcome and reason. - `tables/idea-screen-ai.csv` (31 rows): **Vertical AI idea screen.** 31 vertical AI cloud ideas screened: buyer, AI role, hard exclusions, AI entrants found, prices, free floor and decision. ### ranking - `tables/families.csv` (222 rows): **Product families, ranked and tagged.** The 222 families of the R50 ranking (220 products, one module, one distribution channel) with rank, composite, verdict, the nine criterion scores, prices, the base, low and high 24-month gross revenue of the variant that sets the family's money, the dated window that forces the purchase, an English gloss, personas, job and legal basis, and the product built from it. - `tables/r50-cards.csv` (331 rows): **The 331 R50 cards as written.** Every R50 card as its agent wrote it, with the evidence behind its scores (key evidence, money method, bureaucracy items), the hand adjustments, and its bracketed sources resolved to URLs from the report's own sources section. - `tables/r50-families-raw.csv` (222 rows): **Product families as ranked (raw scoreboard).** The scoreboard the ranking was read from: 222 families (220 ranked products, one module, one distribution channel) with the R50 composite, verdict, best variant, delivery model, prices, base-case revenue over 24 months, the nine criterion scores, the variants merged into each family and the dated window that forces the purchase. - `tables/r50-variants-raw.csv` (331 rows): **Scored variants (raw R50 file).** All 331 variants scored on the R50 rubric: delivery, revenue model, buyer type, sales motion, prices, base, low and high 24-month gross revenue, the probability of the base case, nine criterion scores, reported and recomputed composites, hand adjustments and the bureaucracy items each variant carries. ### build - `tables/products.csv` (10 rows): **The ten products built.** The ten products built on 12 October 2026, read from the book's product table: what each engine checks or computes, price, where it runs, tests, the family behind it and the hours its build lane ran. ### insights - `tables/findings.csv` (1,008 rows): **Every summary finding of the 93 reports, classified.** 1,008 findings extracted verbatim from the summary or verdict section of each research report, with the URLs they cite and whether those answered, plus a classification pass: an English headline, the asymmetry type, polarity (opportunity, barrier, context), personas, the date a rule forces a decision, the kind of evidence and a strength from 1 to 3. - `tables/asymmetries.csv` (14 rows): **The fourteen asymmetry types.** For each type of asymmetry (forced decision, missing tool, price gap, demand signal...): its definition, how many findings carry it, their polarity and strength, the reports, personas and kinds of evidence behind it, and the three strongest examples. ### population - `tables/personas.csv` (28 rows): **Persona atlas.** The 28 buyer personas: population with its basis (CNAE subclasses of the open CNPJ data, MEI and Simples counts, research figures), MEI share, recent openings, findings by polarity and asymmetry, the next dates a rule forces a decision on them, obligations, families and top-50 families aimed at them, their money and the products built. - `tables/persona-cnae.csv` (1,337 rows): **CNAE subclasses assigned to personas.** All 1,337 CNAE subclasses with their active establishments and the persona each is counted under (first matching rule wins; blank when no persona fits), so that every persona population can be audited. - `tables/cnae-counts.csv` (1,337 rows): **Active establishments per CNAE subclass.** 1,337 CNAE subclasses from the Receita Federal open CNPJ data (August-September 2026): active establishments and head offices, MEI, ME, EPP and other sizes, Simples, branches, counts per region, e-mail domain type (none, free, own domain), establishments opened since 2024 and median capital of non-MEI firms. - `tables/cnae-openings-by-year.csv` (21,572 rows): **Openings per CNAE subclass and year.** Establishments opened per CNAE subclass and year, how many are still active, closed (baixadas) or unfit (inaptas), and the MEI share. - `tables/cnae-closures-by-year.csv` (43,793 rows): **Registration status per CNAE and year.** Counts per CNAE subclass, year and registration status (situação cadastral) from the open CNPJ data. - `tables/vertical-tam.csv` (24 rows): **Vertical sizes and one-time software TAM.** 24 verticals built from groups of CNAE subclasses: active companies and establishments, MEI share, ME/EPP, PC-equipped estimate, openings and closures in 2025, assumed need share, base one-time price and the low, base and high yearly market for one-time software. - `tables/cnae-trades.csv` (48 rows): **CNAE counts for auto and technical trades.** 48 CNAE subclasses in auto repair and the technical trades with active establishments (Aug 2026), MEI and Simples shares, monthly openings and closures and IBGE CEMPRE 2024 units, workers and employees. - `tables/ai-companies.csv` (90 rows): **Companies declaring AI activity.** Brazilian companies with AI-related activity in the CNPJ data (June 2026), by dimension (state, size, CNAE). - `tables/os-share-brazil.csv` (57 rows): **Operating-system share in Brazil.** Monthly StatCounter share of operating systems across all platforms in Brazil, 2022-2026. - `tables/desktop-os-share-brazil.csv` (57 rows): **Desktop operating-system share in Brazil.** Monthly StatCounter desktop OS share in Brazil, 2022-2026. - `tables/windows-version-share-brazil.csv` (57 rows): **Windows version share in Brazil.** Monthly StatCounter Windows version share in Brazil, 2022-2026. - `tables/mobile-os-share-brazil.csv` (57 rows): **Mobile operating-system share in Brazil.** Monthly StatCounter mobile OS share in Brazil, 2022-2026. - `tables/platform-share-brazil.csv` (57 rows): **Desktop, mobile and tablet share in Brazil.** Monthly StatCounter platform share in Brazil, 2022-2026. - `tables/desktop-resolution-brazil.csv` (21 rows): **Desktop screen resolutions in Brazil.** StatCounter desktop screen-resolution share in Brazil, 2024-2026. ### regulation - `tables/obligations.csv` (78 rows): **Dated obligations that force decisions.** The 78 dated events of the two calendars (tax reform and other forced demand) in one table: date, event, legal basis, who is forced and how many (as written and parsed), penalty, the free government tool and its quality, paid tools, the software job, personas and the gap index. - `tables/reform-calendar.csv` (42 rows): **Tax-reform calendar and gap index.** 42 dated events of Brazil's consumption-tax reform and related rules (2026-2033): legal basis, who is forced, population, what changes, penalty, the free government tool and its quality, paid tools and prices, the software jobs it creates, and a 0-100 gap index built from population, pain, urgency, free-tool weakness and low bureaucracy. - `tables/forced-demand-calendar.csv` (36 rows): **Forced-demand calendar.** 36 dated obligations outside the tax reform that force decisions on small firms, professionals or consumers: legal basis, who is forced, population, penalty, enforcement reality, the free government floor, paid tools today, the software job, bureaucracy for a seller, fit and candidate products. - `tables/regulation-matrix.csv` (442 rows): **Regulation matrix (archetype x obligation).** 442 rows: for each of 26 product archetypes, every legal obligation a seller of that software takes on (LGPD, consumer law, tax, invoices, sector rules, health, AI, accessibility...), its severity 0-3, what to do, legal basis and status in October 2026. - `tables/regulation-archetypes.csv` (26 rows): **Regulation summary per archetype.** The regulation matrix summed per product archetype: severity per obligation code, counts of severity 2 and 3, a bureaucracy index, the R50 bureaucracy score it suggests, the key constraint and the design rule that keeps the product light. - `tables/reform-credit-gap.csv` (50 rows): **Simples Nacional credit gap under CBS/IBS.** For Simples Nacional bands and annexes: effective rate, CBS/IBS share, credit passed to a business buyer, the full-rate 'por fora' alternative and the price cut a Simples seller needs to match a regular-regime seller. - `tables/tax-table-versions.csv` (46 rows): **Versions of the reform's code tables.** Publication history of the CST, cClassTrib and related code tables of the reform: table, version, date, source. - `tables/nfe-technical-notes.csv` (35 rows): **Cadence of NF-e technical notes.** 35 NF-e technical notes with publication date, title, summary and relevance to a small retail emitter: the rate at which invoice software must change. ### evidence - `tables/sources.csv` (8,709 rows): **Every cited URL and whether it answered.** All 8,709 distinct URLs cited in the published research, with host, kind of source, citations, the documents and findings that cite them, and the outcome of a link check on 12-13 October 2026 (ok, blocked by a bot wall, gone, server error, unreachable). - `tables/reports.csv` (93 rows): **The 93 research reports.** One row per research report: wave, title, topic, words, when the agent ran, findings and their polarity, main asymmetry, URLs cited and how many answered, idea records, families and top-50 families it fed, and products built from it. ### process - `tables/agent-runs.csv` (103 rows): **Research agent runs.** The 103 agent runs of the search (93 research reports and 10 R50 re-scores): start, end, minutes and words. - `tables/build-lanes.csv` (12 rows): **Build lanes.** The twelve build lanes of Monday (two platform versions and ten products), from first to last commit. ### value - `tables/value-estimates.csv` (50 rows): **Time saved per buyer, top 50 (estimates).** For each top-50 family: how the buyer does the job today, uses per year, minutes saved per use and hours saved per year (low, base, high), the main value the model leaves out (fines, money, sales), an incumbent price when the card names one, and a confidence. Written by a language-model pass from each card; estimates, not observations. - `tables/economic-value.csv` (220 rows): **Value to buyers per family.** For each of the 220 product families, in a low, base and high scenario: 24-month revenue, revenue per buyer, buyers, hours saved, the hourly value of the buyer's time, the value of that time and what buyers keep after paying. Everything the model leaves out (fines, extra sales, money saved) makes these floors. - `tables/economic-value-summary.csv` (9 rows): **Value to Brazil, summed.** The value model summed over the ten built products, the top 50 and all 220 product families, per scenario, with the split of revenue between the state (tax), payment platforms, ad platforms and the seller at the base shares. - `tables/economic-value-split.csv` (3 rows): **Where the revenue goes.** The low, base and high shares of revenue that go to the state as the seller's tax, to payment platforms and to ad platforms, with the table each comes from. - `tables/forced-decisions-stake.csv` (24 rows): **What an hour of each forced decision is worth.** The 24 obligations dated in the next 12 months whose population was counted: how many must decide (corrected where the source counts something else), whose time it is and what one hour of each decision is worth across them at PNAD hourly earnings. - `tables/value-assumptions.csv` (17 rows): **Assumptions of the value model.** Every parameter of the value model with low, base and high values, unit, what it applies to, whether it is observed, derived or a judgment, and its source. ### demand - `tables/search-keywords.csv` (13,511 rows): **Search autosuggest keywords.** 13,511 autosuggest completions from Google, Bing and YouTube for 178 niche seeds and buying patterns ('sistema para', 'planilha', 'app'...), each labelled with intent (buy software, free, template, how-to, piracy...), engine and whether it stays on topic. A place an engine appended to a completion reads : engines localize suggestions to whoever asks. - `tables/search-niche-metrics.csv` (178 rows): **Search demand per niche.** Per niche seed: breadth of on-topic completions per engine and the share of software, free-seeking, template, how-to, purchase, desktop, mobile, online and AI intent. - `tables/search-demand-index.csv` (27 rows): **Uniform demand index (wave 2).** A uniform demand index across 27 product families and niches built from Google, Bing and YouTube autosuggest breadth, commercial modifiers and marketplace review counts. - `tables/search-seed-metrics.csv` (144 rows): **Demand metrics per seed (wave 2).** The 144 seeds behind the uniform demand index with their per-engine breadth and modifier shares. - `tables/search-niches-wave1.csv` (214 rows): **Autosuggest niches ranked (wave 1).** 214 niches ranked by a demand score from Google autosuggest breadth, probes, organic hits and YouTube, with modifier counts (grátis, offline, sem mensalidade, Excel, online, baixar, nota fiscal, Mercado Livre...). Places in the example completions read . - `tables/search-autosuggest-docs.csv` (121 rows): **Autosuggest summary: documents, real estate, legal, office.** Autosuggest results per seed for the wave-3 documents, real-estate, legal and office niches, with free, template, tool and professional shares. Places in the completions read . - `tables/search-autosuggest-simples.csv` (102 rows): **Autosuggest summary: Simples Nacional and the reform.** Autosuggest completions per seed for Simples Nacional, hybrid regime and reform pricing queries. Places in the completions read . - `tables/marketplace-listings.csv` (3,322 rows): **Mercado Livre software listings (deduplicated).** 3,322 deduplicated Mercado Livre listings of software, spreadsheets and kits observed live and in archived search pages (2019-2026): title, prices, reviews, rating, dates, niche, job, kind (self-developed, resale, gray market), estimated lifetime units. Seller names are replaced by stable pseudonyms. - `tables/marketplace-observations.csv` (4,207 rows): **Mercado Livre listing observations.** 4,207 individual observations of listings (page, date, position, price, reviews, sponsored, best-seller label) behind the deduplicated listings. Seller names are pseudonyms. - `tables/marketplace-niche-demand.csv` (61 rows): **Mercado Livre demand per niche.** Per niche and dimension: listings, listings with reviews, total reviews, self-developed share, top listing, review-weighted median price, concentration and estimated lifetime units. Seller names are pseudonyms. - `tables/marketplace-price-bands.csv` (32 rows): **Mercado Livre price bands.** Listings and reviews per price band and group, with median price and top listing. - `tables/marketplace-seller-leaderboard.csv` (71 rows): **Mercado Livre seller leaderboard.** 71 sellers of software on Mercado Livre ranked by reviews: listings, reviews, estimated lifetime units, top listing, price range, niches and dates. Seller names are pseudonyms. - `tables/marketplace-search-pages.csv` (70 rows): **Archived Mercado Livre search pages.** 70 archived Mercado Livre search pages used for the listing dataset: date, slug, parser, items, reviews and total results. - `tables/marketplace-latam-listings.csv` (102 rows): **Software listings in other Latin American marketplaces.** 102 listings from Mercado Libre sites in other countries, priced in local currency and in reais at October 2026 exchange rates. Seller names are pseudonyms. - `tables/marketplace-listings-wave2.csv` (806 rows): **Mercado Livre listings used by the demand index.** 806 listings with units, reviews, price and observation dates used by the wave-2 demand index. - `tables/app-store-apps.csv` (1,380 rows): **Google Play apps for small-business jobs.** 1,380 Google Play apps found by job queries: developer, genre, installs, score, ratings, price, in-app purchase range, last update and the queries that found them. - `tables/app-store-complaints.csv` (48 rows): **Complaint themes per app (Google Play).** 48 small-business apps with installs, score, the share of negative reviews among the newest and the top complaint themes. - `tables/app-store-complaint-themes.csv` (23 rows): **Complaint themes across app groups.** 23 complaint themes counted across negative Google Play reviews in four app groups (POS and stock, fiscal and MEI, food delivery, beauty services). ### supply - `tables/global-analogs.csv` (43 rows): **Proven foreign products and the Brazilian gap.** 43 foreign micro-SaaS and one-time tools with revenue evidence (annualized USD), what they do and cost, Brazilian equivalents and prices, the gap status, the Brazilian demand signal, localization needs and the decision. ### channels - `tables/sales-pages.csv` (40 rows): **Sales pages of Brazilian software sellers.** 40 sales pages of software and kits sold in Brazil: channel, product, checkout, price and anchor, installments, headline, demo, proof, guarantee, bonuses, order bump, upsell, delivery and support promises, tracking, software claims and compliance flags. Seller names and contacts are removed. - `tables/checkout-platforms.csv` (21 rows): **Checkout and infoproduct platforms.** 21 checkout platforms compared for selling software: fees by method, effective fee on a R$97 sale, payout, reserves, refunds, chargebacks, whether software is allowed, affiliates, order bumps, upsells, webhooks, native NFS-e and its cost. - `tables/checkout-fee-per-sale.csv` (21 rows): **Checkout fees per R$97 sale.** Effective fee per R$97 sale per checkout, including dispute fees. - `tables/nfse-issuers.csv` (21 rows): **Cost of issuing NFS-e per sale volume.** Monthly and per-note cost of issuing service invoices automatically at different sales volumes. - `tables/selling-stack-costs.csv` (63 rows): **Cost of the selling stack per volume.** Total cost of checkout, invoicing and licensing stacks at different monthly sales volumes. - `tables/payment-providers.csv` (12 rows): **Payment providers for recurring billing.** 12 payment providers compared for subscriptions: fees for card, Pix, Pix Automático and boleto, payout, dunning, chargeback terms, built-in NFS-e, split, and the cost of charging R$49 and R$99. - `tables/fiscal-api-vendors.csv` (24 rows): **Fiscal invoicing API vendors.** 24 fiscal API vendors (NF-e, NFC-e, NFS-e, CT-e, MDF-e): reform readiness, pricing, cost per document, certificate custody, contingency, status pages, incidents, SLA and contract terms. - `tables/fiscal-api-cost-model.csv` (5 rows): **Cost model for invoicing through an API.** All-in monthly cost of issuing documents through a fiscal API for five buyer profiles. - `tables/code-signing-options.csv` (22 rows): **Code-signing options for Windows software.** 22 ways to sign Windows installers (OV, EV, Azure Trusted Signing, Store...): validation, key storage, price, SmartScreen and Smart App Control behaviour, availability to Brazilian firms and CI fit. - `tables/marketplace-rules.csv` (27 rows): **Marketplace rules for selling software.** 27 marketplaces and channels: whether downloadable software or kits are allowed, delivery rules, automation, limits, fees, net on R$97 and R$147, discovery value, manual work per order and policy risk. ### economics - `tables/seller-tax-model.csv` (18 rows): **Tax on software revenue for a Simples seller.** Effective tax (DAS, pro-labore, INSS, IRPF) at different monthly revenues for a Simples Nacional software seller. - `tables/evidence-costs.csv` (42 rows): **Costs of building and running local software.** 42 observed costs (certificates, signing, hosting, APIs, hardware) with date and source. - `tables/salesability.csv` (63 rows): **Salesability per product archetype.** 63 product archetypes: typical price, first-year revenue per buyer, the maximum acquisition cost for each salesability score, low, base and high CAC with method, CAC-to-revenue ratios and the resulting score. - `tables/salesability-benchmarks.csv` (51 rows): **Acquisition and conversion benchmarks.** 51 benchmarks for paid acquisition, conversion, refunds and churn, with sample, geography, date, grade and source. - `tables/salesability-cases.csv` (21 rows): **Cases of small software sold to SMBs.** 21 companies that sold small software to small buyers: segment, price, entry offer, sales motion, channels, traction, outcome and the lesson. - `tables/paid-acquisition-simulation.csv` (36 rows): **Paid acquisition simulation.** Monthly Monte Carlo simulation of ad spend, orders, gross sales and cash flow (mean, P10, P50, P90) for selling software through paid ads. - `tables/money-models-simples.csv` (72 rows): **Monthly revenue model for reform products.** Monthly new units, renewals and gross revenue for the Simples-hybrid and reform products of wave 4. ## Documents, models and code - `research/`: the four briefs, the 93 reports, the orchestrator's notes and the ranking write-up (103 documents, 1,071,934 words), plus the agents' working notes in `research/notes/`. - `models/`: 126 JSON files the agents wrote: forecasts, unit economics, prices, rules, calendars, cost models. - `code/`: 370 Python scripts: the agents' collection and analysis scripts and the release pipeline (`essay_ctl_findings.py` extracts the findings, `essay_ctl_links.py` checks the URLs, `essay_ctl_derive.py` builds the analysis tables, `essay_ctl_anim_data.py` the figures' data). - `methods/`: the specs given to the language-model passes that classified the findings, tagged the ideas and estimated time saved. ## How it was made Each research agent got a brief, a topic, a shared rate-limited fetcher and one file to write; it worked only from public evidence (marketplace listings and archived review counts, search autosuggest, app stores, Receita Federal's open CNPJ data, regulators' pages and the official gazette). The cards, scores and money estimates are the agents'. Three passes were added for this release: a classification of every summary finding, tags for every idea, family and obligation, and estimates of the time each top-50 product saves; they were run by language models under the specs in `methods/`, from the verbatim text only, and their fields are marked as such. Everything else here is computed by the scripts in `code/`. ## Limits - Classifications, tags and time-saved estimates are model judgments; the verbatim text and the sources sit next to them so that each can be checked. - The link check ran once from one machine. `blocked` means a bot wall answered (Reclame Aqui blocks every automated reader), not that the page is gone. Archived copies (web.archive.org) often time out. - Money figures are 24-month forecasts from the cards, not sales. None of the ten products was on sale on 2026-10-12. - The value model counts only time saved. Fines avoided, extra sales and money saved are left out, so its buyer value is a floor; its low and high scenarios combine every low or every high input. ## Privacy Every text file passed one scrub: local paths, private individuals' names, registration numbers, e-mail addresses and phone numbers removed; marketplace sellers replaced by stable pseudonyms (`seller-xxxxxxxx`) in every table and text. Search engines localize suggestions to whoever asks, so a place an engine appended to a completion reads ``. Times are on the book's clock and carry no zone, except the link check's, which are UTC. Page copies, caches, reviews with reviewers' names and run logs are withheld and listed in `MANIFEST.csv` (6,285 files).