Your AI translation pipeline is losing context. Give every stage the same contract.

The glossary arrived. The meaning did not.

The translation uses the right language and the wrong product term. QA changes it back. A retry changes it again. Your team uploaded the glossary, style guide and reference files. Why is the pipeline still arguing with itself?

Usually, the missing piece is an explicit agreement about which resources apply, what each stage receives, and which revision is being processed. The model sees the request you assembled, not the project your PM has in mind.

Consider a fictional payments app. “Charge” means taking a payment on one screen and a fee on another. A glossary alone cannot settle both. Build the context contract before improving the prompt.

Resolve once. Record what actually won.

For each job, resolve the customer, language pair, content type and approved resource set. Decide whether a project glossary replaces the default or merges with it. For a merge, define how conflicting entries are handled. Stop for an owner’s decision when two mandatory instructions cannot both be satisfied.

A useful run manifest identifies the original file and its checksum; glossary and reference contents; prompt templates and code-owned instructions; output schema; model configuration; and segmentation rules. A configuration version without its resource contents is incomplete. A fingerprint of only the editable prompt misses instructions added by application code.

If the pipeline retrieves translation-memory examples, filter by permitted customer, locale, approval status and revision before selecting relevant entries. Record which examples were used. Never treat an uploaded TMX file as proof that retrieval is implemented.

Give each chunk a job and a boundary.

Segmentation defines the units of translation. Batching groups those units into a model request. Keep document, unit and segment identities intact across both; XLIFF defines these separately. See the XLIFF specification.

Send the current segments, relevant terminology and enough neighboring source text to resolve meaning. Label neighbors read-only and require output only for the requested IDs. For UI strings, use tagged screenshots and notes tied to the relevant string. Review source changes before reusing translations or approvals; an unexplained reference folder is not a context policy.

Budget the complete request: instructions, schema, terms, references, source and response allowance. Reduce optional neighbors or split the batch before dropping mandatory constraints. Define what every stage receives; style instructions used only in postediting will not shape the initial translation.

A second AI pass is another opinion.

Reject missing, duplicated or unexpected segment IDs before advancing. Check nonempty targets and protected placeholders in code. Terminology checks need a declared policy for inflections and context-sensitive alternatives. Passing these checks proves structural compliance, not linguistic correctness.

Reviewing only the first reviewer’s flagged segments can reduce work, but cannot find errors that reviewer missed. Use independent human checks on representative output and track missed errors alongside false alarms. Keep AI-complete and human-approved as separate states.

Retry the run you accepted.

Save a validated chunk and its completion checkpoint in one transaction. On resume, retain completed work and retry only unresolved units under the same manifest. A resource change starts a new revision; it must not silently alter the remaining half of a document.

Keep source, translated, reviewed and delivered revisions distinct. Before export, verify that the file contains the intended current targets and review state. Preserve the file skeleton and prove an unchanged round trip before trusting edited exports. The handoff checklist covers that boundary.