Design a translation pipeline
Connect source changes to usable translations with explicit ownership, review, delivery, and recovery.
A translation pipeline is complete when a source change becomes usable content in the system that needs it. Generation is one part. The workflow also has to identify the right revision, retrieve its targets, preserve accepted work, and decide what can be published.
Design around one real content class first: repository catalogs, CMS articles, or application records. They have different owners and delivery behavior. A workflow that treats all three alike can hide the decisions that determine whether a result is safe to use.
Identify who owns each transition #
A repository-catalog design might make these choices:
| Transition | Decision |
|---|---|
| Source becomes eligible | Translate reviewed source at a recorded commit, from a trusted trigger that excludes generated targets. |
| Work begins | Use the explicit localization engine, locales, format, and accepted baselines; obtain run approval where required. |
| Targets become acceptable | Assigned reviewers inspect the required changes and unresolved cases remain visible. |
| Output returns | Downloaded, validated resources enter a translation pull request. |
| Content reaches readers | The application release accepts the change and passes its resource checks. |
| A result needs reversal | Restore known-good resources and the relevant configuration or baselines. |
This is an illustrative workflow design, not a product default. Name the actual mechanism enforcing each decision. A team policy without an automatic control needs an identified person to check it.
Choose tools by what they own #
A native integration can own provider triggers and writeback under its installation authority. A CLI runner owns local retrieval and commits. An API caller owns its stored revisions and activation of output. Choose the supported contracts that fit those responsibilities.
Studio maintains translation configuration and decisions; DevTools supplies programmatic and file execution; Radar provides available measurement and investigation evidence. They are capabilities to combine, not mandatory consecutive adoption steps.
Read the chosen provider's manual/automatic modes, draft handling, entitlements, and disconnect consequences. A connection that publishes to a CMS and one that opens a repository pull request do not share the same acceptance boundary.
Keep the revision relationship intact #
Retain content identity, source revision, target locale, output revision or artifact, configuration, and acceptance together. Decide what happens if a newer source arrives while translation or review is running. The older result must not become current merely because it finishes last.
Choose where reuse preserves accepted targets and where fresh generation is intended. Support for the same format does not establish identical options across the API, CLI, and App. Keep the caller's exact baseline and force behavior in the design.
Plan the unresolved cases #
Define what happens after input rejection, a failed target, interrupted retrieval, a review wait, a local conflict, or an uncertain response. The next action should recover known work under its contract, not start everything again because a receipt was lost.
Some completed targets may be staged while another remains unresolved. Decide whether the release requires all of them or can deliberately use a fallback. Review policy assigns consequences to missing decisions and timeout. Keep execution controls with the tool that owns the work rather than inventing a universal retry action.
Make quality evidence interpretable #
Define the questions that matter for the release, including for translations produced outside Lingo.dev. A measurement can identify work needing review without becoming the system that owns accepted content or decides publication.
Measurement design explains criteria, evidence, answer states, and reporting populations. Assign the people who interpret that evidence and resolve uncertain source meaning as part of the workflow.
Prove a complete cycle before expanding #
Take one source change through submission, retrieval, validation, review, delivery, and consumption. In a safe trial, change the source during processing and simulate one failed target. Verify that the current revision remains identifiable and recovery preserves accepted output.
Use a workflow trial to record what actually ran and what remains untested. The design is ready to expand when another operator can explain both how a translation becomes usable and how to restore the previous useful state.