Review localization-engine suggestions

Decide when feedback supports reusable guidance, apply the proposed change, and inspect its effect on a fresh translation.

A correction can reveal knowledge that future translations need. It can also be specific to one sentence. Suggestions propose reusable glossary, rule, or voice changes from feedback, Jira tickets, low-scoring work, and human edits, so the team can decide which lessons should affect later requests.

Applying a suggestion saves guidance for future translation requests. Use the steps below to test the change on fresh output.

Read the proposed change and its evidence #

Open the suggestion in Queue. The six action types add or update a glossary term, rule, or brand-voice variation. Inspect the proposed text, explanation, locale scope, target resource, and originating evidence. For an update, compare the current record with the replacement.

Read the cited correction and check that it supports the proposed change and locale scope. A narrow terminology problem may justify a term change without justifying a broad voice change.

Check the localization engines attached to the target resource. A suggestion shown under one localization engine can edit shared guidance used by several. Accepting the corrected output and authorizing that wider change are separate decisions.

Apply or dismiss the proposal #

Choose Apply for a useful pending suggestion, then open the saved term, rule, or voice variation to verify the change.

Choose Dismiss for a suggestion that is irrelevant, redundant, too broad, or unsupported by its evidence. Applying or dismissing requires org:suggestion:edit together with edit access to the engine, or an explicit grant to that engine for an organization member. Suggestion permissions explains both access paths.

If the target was deleted or a write failed, inspect the error and current suggestion status. A failed mutation can leave the proposal pending for repair. An already applied proposal cannot be applied again as new work.

For bulk actions, keep the successful IDs and inspect each refusal or unsent item. Correct the cause and retry only the unresolved set. Stopping a batch does not undo changes already saved.

Check whether the reusable change helped #

  1. If the suggestion came from a pending human review, resolve that translation through Reviews. Low scores or explicit feedback can produce suggestions without a review to complete.
  2. Inspect the generated candidate, evidence, locale scope, and shared target; narrow or dismiss an unsupported generalization.
  3. Apply the useful proposal and locate the saved resource.
  4. Deliberately request a fresh translation of the same representative source and locale. Unchanged-content reuse could otherwise retain the old result.
  5. Compare the new translation and context through request evidence, using a defined measurement where appropriate.

One improved example supports that case. To establish a recurring improvement, use a broader selection and examine mistakes as well as successes. This workflow changes guidance; it does not train the model on the correction.

Generation and application policy #

Open the localization engine's Reviewers tab. Under Suggestions, enable Propose improvements and choose Languages, or leave that selection empty for every language. Use the same switch to turn off automatic generation.

Automatic generation produces proposals for review in Queue. Low-score evidence is batched and debounced, and configured locale coverage limits eligibility.

For manual generation, use the Engine Suggestions API to request proposals from recent work or written feedback. Supplied feedback takes precedence over recent history. A request can produce no suggestions.

Generated proposals start pending. Choose Apply to save the proposed change or Dismiss to decline it. Generation alone leaves shared guidance unchanged. When a Jira ticket closes, its pending suggestions are dismissed unless another open ticket still cites them.

Disabling automatic generation stops new proposals from low scores and human edits. A generation already in progress may still finish, so check pending proposals afterward. Earlier guidance changes remain saved. Localization engines covers the configuration, and Decide which corrections become rules develops the judgment behind applying a proposal.

The API reference also covers listing proposals with pagination and applying or dismissing them from a server workflow.