TW + Localization alignment on AI-only translation merge approach for launch

Background

Each file in Phrase goes through a multi-step workflow, and we receive different versions of the same file in separate MRs depending on which step they've completed. This is a high level overview of what happens, more details can be provided if needed:

1. "Translation" step – AI translation script generates initial output and marks the step as "completed" in Phrase
   • This triggers an MR to our fork with the AI-only version
   • AI-only MRs can be filtered by the translationMR-status:: AI Complete label. Example here.
2. Linguist review step – The same file continues in Phrase where a JA linguist reviews and refines the translation.
   • Once the linguist completes their work a separate MR with the human-reviewed version of that same file gets created. (Note: the MR with the humanly reviewed file only gets created when the AI Translation MR is merged.)
   • These Translation MRs with humanly-reviewed files can be filtered by the translationMR-status:: Complete label. Like here.

Then, a specific MR gets created by an engineer to merge translations upstream. Example here.

Merge criteria of AI-only translations for launch

For Phase 2+3 (lower-traffic pages), we're proposing to publish the AI-only versions (from the first MR) to production before the human-reviewed versions are ready.

Pipeline validation will serve as our quality gate for AI-only versions: if the MR passes all pipeline checks on the fork, we'll merge it. We expect to follow the same process for upstream merges to main/master branches.

The human-reviewed versions will arrive later in subsequent MRs as files complete the linguist review step.

Question

1. How involved would the TW team like to be in reviewing upstream MRs containing AI-only files? We recognize that AI-only quality will not meet the standards of human-reviewed content.