After an elections chatbot gives a wrong answer, a clerk's office must show which open-data CSV fed the embedding index. What practice enables that investigation?

Showing which open-data CSV fed the embedding index after a wrong elections chatbot answer requires data lineage and provenance tracking for civic AI inputs and indexes. Lineage records how source files move into preprocessing, embedding, and retrieval so clerks can reconstruct what knowledge the bot used. Retrieved passage citation graphs for civic AI inputs and indexes fail by concept because citation graphs map relationships among retrieved passages at answer time, not a data-lineage record; they don't enable clerks to reconstruct which specific open-data CSV was ingested into the embedding index, so they can't support the error-review trail. Embedding model and chunk config history for civic AI inputs and indexes fails because recording which embedding version and chunk settings were used tracks processing method, not provenance; it provides no mechanism to identify which open-data CSV sourced a particular vector or answer. Voter prompt and moderation audit trails for civic AI inputs and indexes fail because prompt and moderation trails track user interactions and safety signals, not dataset provenance; they still leave clerks unable to show which specific open-data CSV fed the embedding index after a wrong answer. Exam caveat: lineage metadata must itself be protected and accurate; incomplete logging can create false confidence about which version was live during the bad answer. Operational check: tag each embedding batch with dataset ID, publisher, hash, and load time; expose those tags in admin traces for missed answers; and pause the bot’s elections answers until the offending source is corrected or removed.