Here is what the data actually showed.

**Failure mode 1: terminology drift.** A character name spelled one way in chapter 3 and another way in chapter 28. A product term translated consistently for 200 pages and then re-invented. This is invisible if you read the translation alone; it is obvious in a bilingual file.

**Failure mode 2: silent omissions.** A sentence dropped during chunking. A table row lost. No error, no warning — the output simply reads fine with less information in it.

**Failure mode 3: structural damage.** Heading levels flattened, lists renumbered, links detached from their anchors. The text survives; the document does not.

**Failure mode 4: confident guessing.** When the model does not know a term, it invents one — fluently. Fluency is exactly what makes this hard to catch.

None of these are solved by a better model alone. They are solved by giving the reviewer something to check against. That is why every delivery ships with a whole-document glossary (locked before translation and held across chapters), a side-by-side bilingual review file (source and translation paragraph by paragraph), and an MQM-style quality report that names what still needs human attention.

The uncomfortable part: a quality report that only ever says "passed" is marketing. Ours reports the segments that scored low, the terms that drifted, and the checks that did not run for a given language pair. A report you cannot fail is not a report.

What this changes for someone publishing a book or a serious document: you stop asking "is this AI good enough?" and start asking "which paragraphs do I need to fix?". That is a question a human editor can answer in an hour instead of re-reading a whole manuscript.

The full data write-up, including the failure taxonomy and the cost math, is on our blog: https://goodtrans.app/blog/what-530000-units-of-translation-produced

If you want to test it on your own material, GoodTrans gives you a trial grant and shows the credit cost before you submit. Upload an EPUB, PDF, Markdown or TXT file and the delivery arrives by email.