Review translations

Inspect the translation a workflow is waiting on, return an approval or edit, and follow the final text through delivery.

A translation review asks whether a particular result is acceptable for its intended use. Approve the displayed translation as-is or submit edited content when it needs a correction. The decision returns text to the waiting workflow; it does not change shared guidance or merge translated files.

Open Queue, use the Kind filter to show Human review items, select the item, and choose Open the review. You need access to the organization and engine:review_translations. Permission to read request history or edit the localization engine does not supply review authority.

An internal decision pool determines who is emailed about a waiting review. Pool membership does not grant review access or reserve the review for its members. See review emails for language coverage and notification settings.

Read the comparison the reviewer received #

Confirm the source, target locale, and localization engine, then compare the displayed source and generated target. The target baseline can already include a successful rephrase stage. Judge that version rather than assuming it is the first raw model output.

Check meaning, terminology, audience requirements, variables, and protected structure. For a promise or condition, verify what the reader is being told as well as its numbers and terminology.

When the quality gate found errors, the Queue shows its findings next to the translation: the quoted span, category, severity, and explanation. Use them to understand why the translation reached review, then check the text yourself. Quality gate levels explains which findings require review.

Opening the review does not reserve it for one person. Another reviewer can act on the same pending work. Decision submission is serialized; if the detail changes or the action is refused, refresh before submitting a decision based on an older view.

Return an approval or corrected text #

  1. Read the source and translation side by side. The review opens on units that still need a decision; use All units when available to inspect the rest.
  2. Edit any translation that needs a correction. Keep placeholder names, counts, and nesting unchanged.
  3. Choose Approve as-is for unchanged content, or Approve with edits after making corrections. The button counts the changes, such as Approve with 2 edits. Submission decides all units still needing review, using your edited text where changed.
  4. Check the decision and the job's final output before releasing the translation.

When handling multiple reviews, check each outcome. A completed decision on one item does not establish that the whole selection finished. Roles and permissions explains access when a colleague can inspect the request but cannot return a decision.

Check the consequences of external takeover #

For externally assigned work, inspect provider status before supplying your own approval or edit. That action can take over the vendor job. If a professional translator has already started, the external fee may remain due even though your team supplies the final text.

Vendor status and workflow status describe different work. A late vendor response is not necessarily the text the completed workflow returned. Inspect the retained output and fee information alongside the configured review policy.

Verify which text the workflow returns #

Review can be bypassed by locale coverage or the quality gate, and a wait can expire without a human decision. Timeout can continue with the earlier AI output. Closing the review view leaves the job waiting for a decision or timeout.

Establish whether human input arrived and which value was returned instead of relying on the status word alone.

Post-edit, content-length processing, and back-translation can modify text after review. Read the pipeline order and inspect final run output before release. If exact human-approved wording is required, configure later stages accordingly and verify the final value.

A correction can also generate a pending suggestion for a term, rule, or voice change. Examine it separately: a correct edit in this sentence does not automatically justify a requirement for other content or localization engines.