Method · Language
Localising is adapting with invariants of meaning
A localisation decides what has to stay identical and lets everything else move; a substitution changes words and hopes the meaning survives.
The quickest way to ruin a localisation is to treat it as a word-for-word replacement with a better model. The text comes out fluent and the problem appears weeks later, in a legal nuance or a commercial promise that no longer says the same thing.
We have started working the other way round. Before translating anything, we write the list of invariants with the client: figures, proper names, units, regulatory references and controlled terminology. That list is short, it is checked automatically, and it defines the perimeter of what cannot move.
Everything outside it can be reformulated, and in fact has to be able to be. A sentence that keeps the syntactic structure of the original usually reads worse and communicates less than one rewritten for the target language.
What we still watch uneasily is register. It is not lexical, it does not appear in the glossary, and it changes the effect of the text as much as a wrong figure does. The same notice can come across as courteous in one language and distant in another without any check flagging it.
For now we treat register as an editorial decision agreed at the start of the project and reviewed by hand. We have not found a way of automating it that convinces us.