Transliteration, romanization and script handling

Transliteration looks like a translation problem and is not one. Nothing is being interpreted: the letters of one script are being mapped onto the letters of another, and the only question is which mapping. That sounds mechanical until you notice that almost every language with a non-Latin script has more than one published mapping, that they disagree on ordinary letters, and that a language model will pick one per sentence without ever saying which. A list of two hundred names can come back in three conventions and look entirely consistent to anyone who does not read the source script.

These pages work from the mapping outwards. Each one takes a script, says what information the Latin alphabet cannot carry for it — a tone, a vowel that was never written, a consonant distinction English does not make — and then follows that loss into the place it actually hurts: a search index that will not match, a form a government rejects, a sort order that puts the same person in two places. Where a standard exists it is named with the body that issued it, because “the correct romanisation” is not a thing that exists and asking for one is how you end up with four.