Inside the pipeline: how a Trados file survives AI translation

Ask any LSP what scares them about AI translation and it is not word choice. It is opening the returned SDLXLIFF in Trados and watching the import fail — because somewhere in segment 900, a tag went missing.

Tags are the deliverable

Inline tags carry the formatting, links, and placeholders of the original document. Different tools speak different dialects — Trados, memoQ, Phrase and Wordfast each encode them their own way. A language model, left alone, will translate the sentence beautifully and eat the tags, duplicate them, or invent new ones. The translation is fine; the file is dead.

How the pipeline protects them

• Parse with a dialect-aware codec — each CAT format gets its own, so tags are understood, not regex-guessed.
• Shield before translation — tags are replaced with stable placeholders the model is instructed around, and locked or 100% TM segments never reach the model at all.
• Repair after translation — tag counts are compared per segment; missing tags are restored in reverse order of loss, hallucinated ones stripped.
• Verify before delivery — the file must round-trip: if the target tool would reject it, the pipeline fails the job instead of shipping it.

Every layer validates independently and rolls back rather than making things worse. That is the unglamorous machinery behind a one-line promise: the file you upload is the file you get back — translated, tags intact.