--- name: world-understood-checks description: 17 checks from the Noesa course "The World, Understood". For integrators and builders who ship one product to many countries — invoices, schedules, forms, reports — and direct AI to build the "global" parts. --- # The World, Understood — the checks 17 checks, taken from the course at https://noesa.leafsoft.online/c/world-understood Each one names a mistake a model makes by default and the check that catches it. Apply them to your own output before you present it as done. ## Draw the map you actually need **Watch for:** If you asked an AI to make an app "work everywhere", where would it quietly stop? It makes the code path work — a formatter here, a translation there — and tests it in its own default locale, so "everywhere" collapses to wherever the tests ran. The miss is durable because the AI has no idea where your customers are; nothing in the request told it. The check is to name the countries first, from your own list, and then ask for each one by name — a request for "India, Germany and the US" gets three answers where "everywhere" gets one. ## India: one clock, many everything else **Watch for:** If you asked an AI to draft an invoice for a customer in India, what would it likely get wrong? The parts that come from *where*, not from arithmetic: the amount grouped `2,499,000` instead of `24,99,000`; the date month-first; "this financial year" taken as the calendar year rather than April to March; the tax labelled VAT rather than GST with no GSTIN; a surname made mandatory. Each is right somewhere and wrong in India, and the total is still correct — which is exactly why nobody catches it. The check is to name the country before you ask, then read the invoice as its recipient would: grouping, date order, year, tax label, name fields. ## Time: the attribute that breaks first **Watch for:** If you asked an AI to schedule a launch email for "9am local time" in every customer's country, what would it likely get wrong? It maps each country to one zone — right for India, wrong for the United States with six, Indonesia with three and Russia with eleven. It computes the local hour from a fixed offset or an abbreviation, because that is the shortest code that passes a test on an ordinary day. And it tests on an ordinary day, so the skipped hour and the doubled hour never appear. The miss is durable because the request said "country" and supplied no zones, and because nothing errors: the send goes at eight or ten for half the year and the log says success. The check is a zone name per recipient rather than per country, the local time computed on the day of the send, and one test pinned to the two clock-change dates of every zone you sell in. ## South Asia around India **Watch for:** If you asked an AI to add a delivery-area map to a checkout page for customers in India and Pakistan, what would it likely get wrong? It drops in a map provider's default tiles, drawn for the provider's home audience, and the border those tiles show around Kashmir is a claim that one of your two markets reads as the wrong one — sometimes unlawfully so. The miss is durable because the AI treats a map as decoration, a neutral picture of the world, when a border is the most contested line on it, and because nothing in the request said the two markets disagree. The check is to render no map of a disputed area at all, to take any state or province list from the official list of the market you sell into rather than from the model's memory, and to store the customer's own choice of country as a code. ## Southeast Asia: one country, three clocks **Watch for:** If you asked an AI to localise an invoice template for a customer in Thailand, what would it likely get wrong? One of two things. It ignores the Buddhist Era and the customer reads a foreign year on a legal document; or it switches the template to `th-TH` and every date the rest of your pipeline reads back from that document is 543 years off — invisibly, when the year has two digits. The miss is durable because the AI holds one calendar in its head unless told otherwise, and because both renderings are correct somewhere, so no test written in the AI's own calendar catches the swap. The check is two labelled years on the document, the Buddhist Era for the reader and the ISO Gregorian date for the machine, and one test that round-trips 2569 to 2026. ## East Asia: one clock, five zones wide **Watch for:** If you asked an AI to generate the country dropdown for a product selling across Asia, what would it likely get wrong? It returns a list with labels lifted from whichever data file it learned from, and presents them as the standard names — because that file *is* the common pattern, and nothing in your request asked where the strings came from. The labels for Taiwan, Hong Kong and Macau differ between CLDR, ISO 3166 and commercial vendors, and each variant is a position, so shipping the default is taking one without deciding to. The same list will also fold jurisdictions your company may be barred from trading with into an ordinary option, since that is a sanctions question and not a display question. The check is to treat the ISO code as the only inherited thing, and to name an owner for the labels, the zone per country, and the answer to "may we sell here at all". ## Money: the same amount, eight ways **Watch for:** If you asked an AI to design the money columns for an invoice table that will bill customers in many countries, what would it likely get wrong? It reaches for the common pattern — an integer `amount_cents`, half-up rounding wherever a line is computed, a free-text `tax_id`, an FX rate stored as a bare number — because that pattern is correct in its home country and fills the code it learned from. Each choice is right for the dollar and the euro and wrong for the yen, the dinar, a GSTIN and a prescribed tax rate. Nothing in the request named a currency, so it assumed one. The check is to list the currencies you bill in and ask for the schema per currency: exponent beside amount, rounding rule with its stage, a tax id shape per country, a date and source on every rate. ## The Gulf and the moving month **Watch for:** If you asked an AI to schedule a launch email for customers across the Gulf, what would it likely get wrong? It treats "the Gulf" as one place: one weekend — Friday and Saturday everywhere, or Saturday and Sunday everywhere, depending on which page it learned from — one "Gulf time", and Ramadan forgotten or copied from last year's dates. The miss is durable because the request named a region, and a region is not a calendar: nothing in it said which of six weekends, which of two offsets, or which year's Hijri table. The check is a per-country row — weekend, zone name, this year's Ramadan from the calendar — and a rendered Arabic line read in a right-to-left client before the send. ## Europe: four sets, not one **Watch for:** If you asked an AI to write the "ship to Europe" rules for an invoice — currency, VAT line and data handling — what would it likely get wrong? It treats Europe as one list, usually the EU: Switzerland, Norway and the UK get sorted inside or outside by habit rather than by set, so the Swiss customer is charged EU VAT, the Norwegian one is left out of the GDPR, and Sweden is billed in euro. The miss is durable because the four sets are written as one word — "Europe", "the EU" — in most of the text an AI learned from, and the request did not name a set either. The check is a table with four boolean columns per country and a last-checked date, and every rule naming the column it keys on: currency from euro, VAT from EU, data from EEA, Schengen from nothing on an invoice. ## Europe east, and Russia's eleven clocks **Watch for:** If you asked an AI to generate the country dropdown and the region field for a signup form sold across eastern Europe, what would it likely get wrong? It copies a list, and the list carries its source's decisions without saying so: "Czech Republic" from older data, Kosovo present or absent depending on which list it saw most, Crimea filed under whichever country its training text used, "Kiev" for the capital, and one time zone for Russia. The miss is durable because a generated list looks complete — there is no blank where a decision should be — and the request never said which source to follow. The check is to name the source yourself, diff it against ISO 3166 and your payment provider's list, and put every entry the sources disagree on into a reviewed table with a date. ## Africa: fifty-four countries, not one dropdown **Watch for:** If you asked an AI to add "Africa" to your product's region list, with prices and a checkout, what would it likely get wrong? It treats the word as a market: one entry, one currency conversion, one tax rate, cards first. The miss is durable because the label did the thinking — nothing in the request named Kenya, Nigeria or Ethiopia, so the AI reached for the pattern it uses for a country and applied it to fifty-four. The check is to expand the label before you ask: list the countries you sell to, and for each one ask for the currency code, the tax label and rate with its year, the leading payment method, whether a postcode is in daily use, and the calendar the customer reads. "Africa" gets one answer; Kenya, Nigeria and Ethiopia get three that differ on every line. ## North America: one sign, three dollars **Watch for:** If you asked an AI to write the daylight-saving rules for a meeting scheduler covering the United States, what would it likely get wrong? It recalls a rule rather than computing one — often the European last-Sunday dates, or a fixed UTC time for a wall-clock meeting — and it treats "Mountain Time" as one behaviour, so Phoenix moves with Denver and Honolulu is forgotten. The miss is durable because nothing in the request said which year, and the AI has no calendar to check against; it pattern-matches to the commonest rule it has seen. The check is arithmetic you own: compute the second Sunday of March and the first Sunday of November for the year, print the meeting for one January date and one July date in every zone you serve, and store `09:00 America/New_York`, never `14:00 UTC`. ## Latin America: Portuguese, two surnames, one comma **Watch for:** If you asked an AI to translate your launch email "into Spanish for Latin America", what would it likely get wrong? It treats the region label as a language. One Spanish comes back — often a neutral or European register — and the send list built from "Latin American countries" includes Brazil, whose 215 million people read Portuguese, plus Guyana, Suriname and Haiti, which read English, Dutch and French or Creole. The miss is durable because the request named a region, not a list of countries, and the AI has no way to know which of its readers say *vos* and which say *tú*. The check is to name the countries first, map each to a locale — `pt-BR`, `es-MX`, `es-AR`, `es-CO` — and let the language follow the country rather than the region. ## Oceania: the day that exists twice **Watch for:** If you asked an AI to schedule a launch email for "Monday 9 a.m. local time" across Australia and New Zealand, what would it likely get wrong? It collapses each country to its biggest city. Sydney's clock stands in for all of Australia, so Brisbane receives the email at 8 in summer, Perth at 6 in summer and 7 in winter, Adelaide at 8:30; and New Zealand is taken as "Sydney plus two", which fails for the week each spring when New Zealand has already moved clocks (last Sunday of September) and Australia's five states have not (first Sunday of October, as of 2026). The miss is durable because a country name looks like a zone, and the AI has no list of which states move. The check is to schedule by IANA zone per recipient — `Australia/Brisbane`, never "Australia" — and to print the send time for one January date and one July date before you trust the plan. ## Names, addresses and languages **Watch for:** If you asked an AI to generate a signup form for customers in twenty countries, what would it likely get wrong? The validators. It writes rules that look like hygiene — letters A to Z in the name fields, digits only in the postcode, a `(XXX) XXX-XXXX` phone mask, a mandatory last name — and each is the habit of one place, tested with names and addresses from that place. The failure is durable because a rejected customer never files a bug: the form told them their own name was invalid, and they left. The check is to push one real-shaped name, postcode and phone number per country on your list through the form before it ships — and to brief the AI with those countries by name, so it validates for them instead of for home. ## Disputed places, sanctions and where data may live **Watch for:** If you asked an AI whether you can sell to, and store data for, customers in a given country, what would it likely get wrong? It answers — fluently, completely, and with no source. A sanctions status for a whole country, when the lists name counterparties and change without notice; a data-law summary from memory, when the regime and its transfer rules carry a date; a storage region chosen by default. The failure is durable because the question sounds like a fact and the model answers in the grammar of facts, and nothing in the request forced a citation. The check is to demand three things in the answer — the list, the regime, the date — and treat any sentence without them as unverified; and when the decision is about a place someone will object to, stop and hand it to a human owner. ## Audit the world the AI built **Watch for:** If you asked an AI to review its own "global" invoice template for country problems, what would it likely miss? Everything that is absent. A review reads what is on the page — arithmetic, spelling, consistency — and the misses for a country are mostly things that are not there: no GSTIN, no CPF, no ABN, no era date, no comma decimal, no Sunday-to-Thursday week. Nothing on the page points at them, and nothing in the request named the countries, so the review confirms the draft. The failure is durable because an absence has no text to check. The check is a per-country list of what must be present — the scorecard — read against the draft by someone who holds the map; and the shortest way to seven different answers is to ask for seven countries by name.