LLM translation validator
When an LLM translates copy, you usually cannot read the target language well enough to know whether it stayed faithful. Back-translation is the classic trick: translate to the target language, then translate the result back to the source. If the round trip drifts, the forward translation probably did too. This tool runs both legs with your own OpenAI key and shows them side by side.
How it works
On run, the tool sends two sequential requests to
https://api.openai.com/v1/chat/completions. The first asks the model to
translate your source text into the chosen target language and return only the
translation. The second feeds that output back in and asks for a translation
back into the source language. Both calls use temperature 0 for stable,
repeatable results, and everything happens directly between your browser and
OpenAI — there is no Gera backend in the loop.
Tips and notes
- Read for meaning, not wording. Synonyms and reordered clauses in the back-translation are fine. Watch for changed numbers, dropped clauses, or new claims that were not in your original.
- Test idioms and product names. These are where translation most often breaks; a back-translation that literalises an idiom is a clear warning.
- Short segments isolate problems. Validating one sentence at a time makes it obvious which line drifted, instead of one long blurry paragraph.
- Pair with a native reviewer for anything shipping. Back-translation narrows where to look but does not replace a human who reads the target language.