Design a translation review policy
Match review to the consequences of an error and make missing evidence, reviewer responsibility, and publication decisions explicit.
Review is useful when it answers the question that could change a release decision. A malformed catalog needs a structural check. An altered obligation needs examination of meaning. A disputed product claim needs its owner. One generic approval step cannot substitute for all three.
Start with the consequences of an incorrect translation, then identify the evidence and people needed to resolve them. This guide develops the team's policy; Studio review policy documents what the product's execution controls do.
Classify the actual message #
A settings page can contain an ordinary label and a destructive-action warning. Judge the claim or action, not just the directory containing it.
An illustrative policy could distinguish:
- Consent and legal obligations: retain the matching source, inspect obligations with a qualified locale reviewer, and involve the legal owner when meaning is unresolved. Require explicit acceptance before release.
- Payment or delivery commitments: check amounts, timing, certainty, conditions, and variables. The product owner resolves source ambiguity; a qualified reviewer examines the target claim.
- Routine instructions: combine structural, meaning, and terminology checks with the team's chosen sample or full-review requirement. Escalate new concepts and flagged exceptions.
- Low-risk labels: check required languages, catalog validity, and terminology under an assigned owner, with a route for ambiguous cases.
These are examples, not configured defaults. Choose a level of review the team can enforce and define what blocks publication when evidence is missing.
Test a quality gate level before relying on it #
If a quality gate determines which translations receive human review, compare its findings with known defects, acceptable paraphrases, and uncertain cases. Examine false alarms and missed consequential errors separately.
In Lingo.dev's asynchronous pipeline, Strict sends any finding to review, Medium sends major and critical findings, and Lax sends critical findings only. Everything turns the gate off so every new covered translation goes to a reviewer. A translation the gate could not check goes to review too. Human review explains the levels and severity definitions.
The gate's judgement is the same at every level; the level decides which findings need review. Coverage can bypass review independently, and the gate can miss an error. Those behaviors must fit the team's policy rather than define it accidentally.
Measurement design explains how to calibrate the check and preserve its denominator when reporting results.
Name coverage and responsibility #
Identify required locales and the qualified reviewer for each. A locale outside review coverage has skipped review, not approval. Matching also matters: a base language can cover variants where the contract allows it, while a specific region need not cover a sibling region.
When a reviewer is unavailable, choose a deferral, alternate reviewer, or release block based on the consequence. An empty Inbox is not enough evidence that all work is complete; filters, permissions, and readiness can limit what appears.
Decide what timeout and partial review mean #
A workflow can settle after timeout while retaining preceding AI text. That completes execution without satisfying a policy that requires a human decision. Keep accepted edits, indeterminate quality evidence, and undecided content distinct in the release record.
Later pipeline stages can also change reviewed wording. If the exact accepted text must reach readers, configure those stages deliberately and verify final output. Cancellation, fallback, timeout, and partial decisions need their own consequences rather than one “finished” status.
Control changes to the policy #
Assign authority to change quality gate levels, locale coverage, and review depth. Record the reason and affected content, then test representative cases. A temporary broad bypass for one release can affect later work if it is not restored and checked.
Keep run approval, translation acceptance, and publication or merge permission separate. CI automation covers repository delivery.
Test the policy on one real change and one exception. It is ready when an operator knows what evidence is sufficient, who can decide, and what happens when the evidence never arrives.