OpenRouter has added beta interfaces in Simplified Chinese, Spanish, Brazilian Portuguese, German, French, Japanese and Korean. Rather than translate the site once, it built a daily pipeline for 16,918 English strings that detects changes, creates replacements and submits them as code for review.
Each catalog entry stores a fingerprint of its source text. Changed entries are translated in batches with GPT-6.1 Sol, checked for broken placeholders, plural rules, markup and forbidden glossary forms, then reviewed by a second model call. Major meaning errors trigger bounded retries. A failed verifier leaves the interface in English, and every daily batch still requires a person to merge its pull request.
In OpenRouter’s 454-case comparison, the complete pipeline failed four authored glossary or format checks, versus 147 failures among 453 valid replies from a bare two-line prompt. Three outside model families counted 109 critical errors in pipeline output and 203 in the bare output. Back-translation scores were effectively mixed, showing that no single metric captures terminology and tone.
Readers can highlight a bad translation and report it. Confirmed issues become glossary rules, style notes, exact overrides or regression tests. The approach does not eliminate mistakes, but it gives frequently changing product text a traceable correction loop rather than treating machine translation as a one-time publishing step.