Skip to content
Multigrid

Locale-correct output

Dates, numbers, plurals, addresses and names as a locale actually writes them, and why a model that gets the language right still gets the format wrong.

A model that writes fluent German will still write 3/4/2026 and 1,234.56 in the middle of it. That is not a language failure and no amount of prompting for “better German” fixes it, because the two things are learned from different evidence. Fluency comes from an enormous amount of German prose. Format comes from whatever numeric and date strings happened to appear in training data, which is dominated by English-language sources, code, logs and spreadsheets — none of which are German.

These pages are about that gap. Each one takes a formatting convention that a locale actually uses, shows what the convention is and where it diverges from the default a model reaches for, and then says which part of the job belongs to the model and which part belongs to a formatting library that has the table. The general answer is that the model should produce values and a renderer should produce strings; the useful part is knowing exactly where the seam is for each kind of value, because for names and honorifics it is not where you would guess.

Reordering Extracted Text From a Multi-Column Arabic Document

Why a two-column Arabic PDF extracts as interleaved half-sentences, and how to recover the real reading order from the glyph positions.

10 min read

Why Vector Search Ranks a Mistranslated Chunk Above the Correct One

An embedding scores topical and stylistic proximity, not truth, which is why a smooth mistranslation outranks an accurate awkward one.

9 min read

Formatting Numbers in the Indian Lakh-Crore System

The Indian grouping rule is three digits then twos, not threes all the way, and that single difference is where generated output goes wrong.

8 min read

Why AI Output Uses the Wrong Numerals for Arabic Text

Two different digit sets are both called Arabic, three if you count the Persian variant, and asking for the wrong one is why the digits come back Western.

9 min read

Fixing Date Format Confusion Between DD/MM and MM/DD in AI Output

Why an ambiguous numeric date is the most expensive locale bug in generated text, what ISO 8601 fixes, and the twelve days a year it does not.

9 min read

Where the Currency Symbol Goes in Different Locales

Symbol position, spacing and negative-value convention are per-locale data in CLDR, not a rule of thumb about Europe and America.

8 min read

Formatting an Address Correctly for a Given Country

Three worked postal formats — US, Japan, Germany — showing the field order a model defaults to and where each country actually diverges.

9 min read

Why AI Gets Name Order Wrong in Family-Name-First Cultures

The failure is not that models put the family name last, it is that they helpfully reorder a name that was already correct.

9 min read

Getting 24-Hour or 12-Hour Time Right in AI-Generated Text by Locale

Which locales prefer which clock is published data in CLDR, and models default to the US 12-hour form regardless of the language they are writing.

8 min read

Getting the Decimal Comma or Point Right by Locale in AI Output

The characteristic bug is half-right output — the decimal mark swapped to a comma while the thousands separator stays a comma too.

8 min read

Validating Postal Code Formats Across Countries

A sourced reference for postal code patterns in a labelled set of countries, including the ones that have no postal code system at all.

9 min read

Using the Correct Honorific and Title in AI-Generated Text

An honorific encodes a relationship, not politeness, and dropping it in translation deletes information the source sentence carried.

9 min read

Getting Metric or Imperial Units Right in AI Output by Locale

Why generated text arrives in the wrong measurement system for its reader, which locales are actually affected, and the exact conversion factors to use.

9 min read

Why AI Calendars Default to the Wrong Start of Week

The first day of the week is locale data, not a constant, and generated calendar code almost always hard-codes it.

8 min read

Formatting Phone Numbers Correctly by Country

There are two correct formats for every phone number — one to store and one to show — and AI output routinely produces a third that is neither.

9 min read

Using the Japanese Era Calendar in AI-Generated Dates

How gengo years map to Gregorian ones, why the first year of an era is not written 1, and what breaks when an era ends.

8 min read

Formatting Negative Numbers Correctly by Locale

Minus sign, parentheses or a trailing sign — which one is correct is decided by CLDR locale data and by whether the document is an accounting statement.

8 min read

Why AI-Generated Forms Break on Two-Character Surnames

The minimum-length rule that models add to every name field rejects millions of real surnames, and the character-class rule rejects millions more.

9 min read

Formatting Large Numbers for a Chinese-Reading Audience

Chinese groups digits in ten-thousands and hundred-millions, which means there is no word for million and no natural place for a Western thousands separator.

8 min read

Why AI-Generated Business Cards Get the Layout Backwards

A Japanese or Korean business card puts the organisation above the person, uses both sides for two languages, and is not the size a Western template assumes.

8 min read

Formatting a Salutation Correctly Across Languages

Whether a salutation takes a comma, a colon or no punctuation is a language rule, and in German it also decides whether the next word is capitalised.

8 min read

Why AI Gets the Week Number Wrong Across Locales

ISO 8601 and the US convention define week 1 differently, so around New Year the same date sits in two different weeks of two different years.

9 min read

Formatting Fiscal Year Dates Correctly by Country

Fiscal years start in different months in different countries, and the label FY2026 is named for the ending year in some of them and the starting year in others.

9 min read

Why an AI-Generated Table Sorts Accented Names Incorrectly

A default string sort orders by code point, which puts every accented name after Z, and even a correct locale sort gives different answers in Swedish and German.

9 min read

Why Language Detection Fails on Code-Switched Text

Detectors score a whole string at once, so a sentence containing two languages returns whichever one supplied more characters — and no signal that anything was mixed.

9 min read

Setting a Confidence Threshold for Automatic Language Detection

How to pick the score below which you stop trusting a language detector, using your own labelled strings rather than a number from a blog post.

10 min read

Telling Closely Related Languages Apart Automatically

Why character-level detectors collapse on language pairs that share most of their orthography, and which features actually separate them.

9 min read

Detecting the Dominant Language in a Mixed-Script Document

A block-by-block detection pass for documents that switch script mid-page, using the Unicode Script property to partition before any classifier runs.

10 min read

What to Do When Automatic Language Detection Is Wrong

A ranked ladder of signals to use when detection confidence is low but a response still has to go out, and the rule for when to ask the user.

9 min read

NFC and NFKC Normalization: What Actually Differs

Canonical normalization preserves the text and compatibility normalization deliberately destroys distinctions — which is why one belongs in storage and the other in search keys.

9 min read

Why Vietnamese Diacritics Break Search Matching

Two identical-looking Vietnamese strings can hold different code points because the language stacks two marks on one vowel — here is the byte-level reason and the fix.

9 min read

The Turkish Dotted and Dotless I Problem in AI Text

Turkish has four I letters and default case conversion knows about two, which corrupts Turkish text and breaks protocol comparisons on Turkish machines.

9 min read

Normalizing German Umlauts and Eszett for Search

German search has to match Müller, Mueller and Muller, and the ß-to-ss expansion that makes it work changes the string's length.

10 min read

Stripping French Accents for Search Without Losing Meaning

Accent-insensitive French search merges real word pairs like où and ou, so folding belongs in the recall layer and never in the ranking layer.

10 min read

Handling the Spanish Ñ in Text Processing Pipelines

Ñ is a letter of the Spanish alphabet, not an N with a decoration, and pipelines that treat it as a variant produce wrong sort order and embarrassing search results.

9 min read

Handling Nordic Characters Å Ä Ö Correctly in AI Output

In Swedish, å ä ö are the last three letters of the alphabet and sort after z — the same code points that German sorts next to a, which is why locale-less sorting is always wrong for somebody.

10 min read

Building Diacritic-Insensitive Search Without Breaking Precision

A decomposition-and-strip fold that makes cafe match café, plus the languages where stripping marks changes which word you asked for.

9 min read

Combining Characters and Precomposed Characters: What Actually Differs

Why two strings that render identically compare unequal, worked through one character at the byte level.

8 min read

Normalizing Accented Names for Database Matching

How to fold accented names into a matching key without merging two families whose surnames differ only by a diacritic.

9 min read

Why Polish Diacritics Break Naive String Sorting

Why ł, ż and ś end up after z in a code-point sort, and the three places you have to change to get Polish order back.

8 min read

Why AI-Generated Filenames Break on Accented Characters

A file written with an accented name on one operating system cannot be opened by the same name on another, and the reason is which normalisation form reached the disk.

9 min read

Normalizing Right-to-Left Text Before Search Indexing

A pipeline for Arabic and Hebrew that strips diacritics, unifies letter variants and removes bidi controls before a single token is indexed.

10 min read

Why Emoji Skin Tone Modifiers Break Text Comparison

A toned emoji contains the untoned one as a prefix, so equality fails and substring matching succeeds — and normalisation does not fix either.

8 min read

Normalizing Curly and Straight Quotation Marks Across Languages

Why a single find-and-replace on quotation marks is wrong in German, French and any language where U+2019 is a letter.

9 min read

Why Korean Text Comparison Fails Without NFC Normalization

One Hangul syllable is either one code point or three jamo, and the arithmetic that connects them explains every Korean equality bug.

9 min read

Arabic Dual and Plural Noun Forms in AI-Generated Text

Arabic counts in singular, dual and three kinds of plural, and the rules invert above ten in a way generated text routinely gets wrong.

9 min read

Russian Noun Case Agreement in AI-Generated Text

The specific error where an adjective stays in the nominative while its noun is inflected, traced through all six cases of one phrase.

9 min read

How ICU MessageFormat Plural Rules Work

The six CLDR plural categories, what each one means in a language that uses it, and the four things about them that are not obvious.

10 min read

Other topics