A 311 chatbot audit captures chat UI screenshots but omits model version, prompt template, and retrieval metadata. What completeness issue should be raised?

Raising incompleteness of technical evidence attributes is correct because 311 chatbot screenshots of the UI rarely capture model version, prompt template, and retrieval metadata. Without those attributes, auditors cannot tie a visible answer to the configuration that produced it, reperform under comparable conditions, or compare behavior across releases. Civic GenAI evidence packages need lineage fields alongside resident-facing output.

Citing chat transcript retention in the 311 ticketing system as support for audit review of response quality trends misses the completeness issue: transcript retention tracks what residents said, not the model version, prompt template, or retrieval metadata needed to attribute an answer to its configuration. Relying on ticket classification history to evaluate prompt template effectiveness does not resolve the audit gap: classification history shows how tickets were routed, not which model version, prompt template, or retrieval settings produced a specific answer. Pointing to session timestamps combined with agent handoff logs as documentation of retrieval sources still leaves the evidence incomplete: those logs do not capture the model version, prompt template, and retrieval metadata required to tie a visible answer to the configuration that produced it.

Exam caveat: even complete metadata does not prove answer quality; it enables attribution and reperformance that must still be paired with substantive evaluation of referrals or resolutions. Operational check: extend evidence collection to export model/prompt IDs, retrieval traces, and timestamps with each sampled 311 chat; reject screenshot-only packages for conclusions that depend on version-specific behavior.