Debug a bad translation
Trace a specific wording problem to the source, delivery, selected guidance, or generation, then test the narrow repair.
An unexpected translation needs an explanation before it needs a broader rule. The application may have loaded an old catalog, the source may be ambiguous, or the model may have changed a clear meaning. A low score can identify a case worth examining without distinguishing those causes.
Keep the source and target revisions, locale, content ID, expected meaning, and actual wording. Start with one defect that can be inspected rather than a general complaint about a language.
Confirm the text and meaning being discussed #
Inspect the file or record the application actually loaded. Match it to the source revision. Stale content or fallback to another language can look like a generation error when the problem is delivery.
Read surrounding content and product requirements. If “available after approval” is unclear about whose approval is needed, resolve that with the source owner. Translation guidance cannot responsibly decide a product policy the source team has not established.
Inspect evidence from the original request #
Use request evidence, results, or the account MCP tools to inspect the supported input/output, localization engine, actual model and fallback, execution status, and measurement reasoning. Keep the submitted context and engine configuration alongside those records; the model-request view does not contain a complete historical prompt. Local catalog inspection and account records provide different evidence.
Give an assistant the exact record and a narrow date or locale scope. Continue pagination and say when only a sample was inspected. Missing traces, cost fields, and source context remain gaps; they are not values to infer. Agent delegation explains how to keep observation separate from authorized changes.
Compare possible causes #
| Possible cause | Deciding evidence | Repair |
|---|---|---|
| Wrong source revision or locale loaded | Source/target identities and loading path. | Correct synchronization or application routing. |
| Ambiguous product term | Definition, locale wording, context, and selected term. | Define its meaning and boundaries. |
| Conflicting guidance | Applicable rules, voice, glossary, and source requirement. | Resolve scope. |
| Missing context | Actual hints and surrounding content available to the request. | Supply the information that changes interpretation. |
| Generation changed a clear requirement | Exact input, selected guidance, output, and fallback evidence. | Correct the output and test a bounded model or instruction change if recurrence warrants it. |
Retrieval proves guidance was supplied, not that it was followed. A measurement's reasoning is also something to inspect. Model identity, duration, tokens, and average scores do not explain one sentence by themselves.
Correct at the scope the evidence supports #
Use feedback triage to assign the repair. A one-off edit can remain local; correction-to-rule decisions determines whether the reason should govern later work.
For shared guidance, verify the saved change and its attached localization engines. Submit a separate bounded rerun through the actual execution workflow: applying guidance does not regenerate every earlier translation.
Test the repair against an unchanged case #
Rerun the defective example and a counterexample the change should leave intact. If the problem was a possibility turned into a promise, retain a definite promise where the source really makes one. A general instruction to weaken every claim would repair one case by damaging another.
Inspect the returned target and application-loaded text. Keep the original evidence, proposed cause, correction, and comparison together. If output improves but causation remains uncertain, say so. One successful rerun supports that case, not a measured general reduction in defects.