Human-review policy

Decide which translations need review, who can provide it, and what happens when a decision does not arrive.

A review policy connects a translation workflow to the people who can examine its output. It determines eligible languages, the reviewer provider, quality-gate strictness, and how long the work waits. The right policy depends on what the team needs to establish before using the text.

Set that expectation before enabling review. The current workflow can continue with AI output after timeout, and later stages can rewrite a reviewed result. A strict publication requirement for human-approved text therefore also needs controls in the consuming workflow and a check of the final output.

For setup instructions, see Human review and Decision pools. They cover the Human review step's settings, pool management, and routing by language and directory.

Choose the reviewer and languages #

Use internal review for decisions made by your team in Inbox reviews. Use supported external professional review when that provider should handle the work. Provider availability depends on organization entitlement, and the person returning a decision needs review authority.

Choose target-locale coverage. An empty list covers all locales under this policy; a request outside a specified list bypasses review and its dependent post-edit. Coverage follows locale resolution, like the rest of the engine configuration: de covers de-CH, and de-DE covers a de request, but de-DE does not cover de-CH. These are execution settings, separate from the display filters used to find pending reviews.

External review sends content to another provider and can incur per-word charges. Confirm the material, supported payload, and commercial terms before enabling it. An internal decision and an external result have different provenance even when both produce usable text.

Choose how strict the quality gate should be #

When the job's pipeline includes human review, Review level applies to new translations that match a decision-pool route:

  • Everything: Every new translation covered by a route goes to a reviewer.
  • Strict: Review anything a careful editor would change.
  • Medium: Review translations that get the meaning or your guidelines wrong.
  • Lax: Review only broken or misleading translations.

Everything turns the gate off. Strict routes any finding, Medium routes major and critical findings, and Lax routes critical findings only. Human review defines the severities and what reviewers see in the Queue.

A missing assessment or gate error keeps review eligible. This protects the bypass decision, but does not guarantee that review eventually happens: coverage exclusions and timeout remain independent. A finding is an automated signal, not a person's decision.

Set the wait and takeover behavior #

Choose a positive whole-hour timeout that fits the reviewers' response window. The default is 48 hours. If review does not complete, the workflow can retain its preceding AI translation and continue. A longer timeout or notification does not change that completion behavior.

For external work, check vendor status before taking over. If a professional translator has started, the review fee can remain due even when your team supplies the final decision. Taking over the work and cancelling charge liability are different outcomes.

Verify the policy against final use #

The relevant order is rephrase, human review, post-edit, content length, then back-translation. Post-edit requires human output; shortening and reverse-translation correction can also change later wording. Inspect the final returned text before release, especially when exact human-approved wording is required.

Review authority differs from permission to read history or edit a localization engine. Check Roles and permissions and Feature access when someone can see the request but cannot act on it.

After changing the policy, inspect a covered sample and an excluded sample through run evidence. Verify the actual stages and retained values, then keep publication and repository-merge decisions in the workflow that consumes the result.