Independent back-translation review of machine-translated interface text: 98 strings across 15 languages, 12 August 2026

Flag rates when a second, independent language model back-translates interface strings produced by a first model. We localised a supplier-verification tool into fourteen languages nobody on the team reads; before shipping, a different model back-translated every string so we could see what our own words actually said. Frame: 98 short interface strings across 15 languages on a live supplier-verification tool, reviewed on 12 August 2026 in two rounds (42 interface strings, then 56 registry-status glosses). Results: 11 of 98 strings came back wrong (11.2%). Of the eleven, five were the wrong domain term, three were grammar, voice or collocation, and three were close but not exact. One would have sent the reader to ask for the wrong document: the Arabic gloss for the Chinese business licence meant work permit. Four automated string checks ran before the review — completeness, script and diacritics, forbidden wordings, and length — and caught none of the eleven, because every flagged string was correctly spelled, correctly accented, the right length and free of forbidden wording. Eight flags were corrected in the shipped strings; three were accepted with the drift documented. Limits: this is not a benchmark. 98 short domain-specific strings, one translating model, one reviewing model, one review date — it measures a workflow, not machine translation as a field. Neither model is a native speaker, and the reviewer stated it could not find authoritative sources for some phrases it judged, so its flags are findings to investigate rather than verdicts. 11.2% is a floor, not a ceiling: errors both models are confidently wrong about are invisible to this method by construction. The per-language split (six languages with flags, eight without, one to three each) is far too thin to say which languages a model handles worse. No personal data is included. Full method and write-up: https://currawongweb.com/research/#machine-translation-review