Multilingual pipelines

A pipeline that works in one language is not a pipeline that works in eleven with a translation step bolted on. The parts that break are not the parts anyone expects: the instruction that carried an English idiom into Japanese, the schema whose enum values were localised and now no longer match the branch in your code, the retrieved documents that quietly overruled an explicit output-language directive, the OCR stage that reported ninety-something percent accuracy on a corpus whose hard cases were four percent of the characters.

These pages work through those mechanisms one at a time. The prompting half covers where language belongs in a template, why a model drifts back into the user’s language, and how to test that one prompt behaves the same way in ten. The document half covers why a script’s own structure — cursive joining, stroke density, absent word spaces, fused conjuncts, a syllabary of several hundred symbols — sets a floor on recognition error that no amount of training data removes. Almost none of these failures raise an exception, so each page ends with what to measure.