Can Jev Translate? No, and What To Do Instead

No. Jev, TypeSafe AI's decision model, picks from answers you define; it never writes text, and translation is text-writing with extra steps. Ask it to translate and there's simply no output shape for that job to fit into.

The more useful news: most people asking "can Jev translate?" don't actually need translation. They need a judgment about non-English text, and that is a different question with a friendlier answer.

Why translation is the wrong shape for a decision model

Translation is open-ended generation: the output space is every possible sentence in the target language. A decision model's output space is a closed list you wrote, like "billing, bug, refund, other." That's the whole design, and it's why the limitations page puts all generative work (translation, summarization, drafting) in the same "not this tool" bucket.

When a pipeline genuinely needs translated text, name the split: a chat or translation model produces the prose, and Jev handles the closed decisions around it (which language is this, does it need a human translator, is the translated reply on-policy). Each model does the job it was built for.

What to do with non-English text instead

If your real goal is "classify, route, or score text that isn't English," skip the translation step. The rule the ecosystem converged on, covered in depth on the multilingual page: write your questions and choice labels in English, and leave the evidence in its source language. Translating first adds cost, latency, and a second model's mistakes before the verdict even runs.

Builders are already doing this. One Google ADK guard write-up scored confidence on the original Traditional Chinese text and explicitly kept it "as-sent rather than translating" (build). A Chinese-pronunciation harness chained Jev decisions over Chinese characters and, as reported, beat DeepSeek at one-seventh the cost and one-thirtieth the latency. Neither translated anything.

Calibrate per language, though. Agreement in English tells you little about agreement in Thai, and low-confidence non-English verdicts should route up the cascade to a frontier model rather than being forced.

Frequently asked questions

Can Jev translate text between languages?

No. Jev only returns answers from a closed set you define, so it cannot produce translated sentences; use a chat or translation model for that step.

Should I translate text into English before sending it to Jev?

Usually not. The common pattern is English questions over source-language evidence, calibrated per language, which avoids the cost and errors of a translation pass; details on the multilingual page.

Can Jev detect which language a text is in?

Language identification is a closed-set question ("which of these languages?"), so it fits the model's shape, but verify it against labeled samples before trusting it; no official accuracy figures exist.

What's the best way to combine Jev with a translation model?

Let the translation model write, and let Jev make the decisions around it, such as whether a translation needs human review. Keep irreversible actions behind a human or a second check, never a lone verdict.