A housing authority requires every denial-explanation answer from its Foundry assistant to be reconstructible later for appeal review. What should they require?

A housing authority that must reconstruct denial-explanation answers for appeal review should require trace logging and provenance metadata for each response. The requirement works because appeals need to show what the Foundry assistant retrieved, which policy fragments grounded the explanation, which model/deployment produced the text, and when the run occurred. Provenance turns generative output into an accountable municipal record rather than an ephemeral chat bubble.

Archiving final answer text and feedback notes for each denial-explanation response fails by concept because output text and notes show what was said without preserving the per-response retrieval, policy fragments, model, and run details that appeal reconstruction requires.

Keeping prompt versions and retrieved passages for each denial-explanation response fails because it preserves inputs without capturing the full trace and provenance metadata — model/deployment, run IDs, timestamps, and policy version tags — required to reconstruct an issued denial explanation for appeal review.

Storing sampled traces and summary records for each denial-explanation response fails because sampling and summaries leave gaps and do not capture the complete per-response trace and provenance metadata needed to reconstruct each denial explanation later.

Exam caveat: logs may contain sensitive resident data—retain traces under records policy with access controls, redaction, and retention schedules, not unbounded public dumps. Operational check: enable tracing with document IDs, policy version tags, and run IDs on denial explanations; verify an appeal sample can be reconstructed end-to-end; and align retention with housing authority legal requirements.