A county benefits bot drafts eligibility summaries. Which optimization should the team add before showing a final summary to a resident?

For a county benefits bot that drafts eligibility summaries, a reflection and self-critique loop is the right optimization before a resident sees the final text. The generative draft may omit a condition, overstate eligibility, or leave a claim weakly supported by policy. Reflection makes the system evaluate that draft against completeness and grounding criteria, revise weak spots, and only then finalize. That second pass is what turns a soft municipal draft into a resident-facing summary that is more accurate and less likely to create false expectations at the benefits counter.

Fine-tuning the embedding model to verify citations in the draft before finalizing fails because fine-tuning only adjusts retrieval similarity, it does not evaluate the draft's claims or verify citations against policy, so an unsupported claim can still slip through unexamined. Increasing the retrieval top-k to cross-check policy passages before finalizing fails because retrieving more passages only broadens context, it does not critique the draft for completeness or grounding, so an unsupported claim can still slip through unexamined. Routing the draft through a semantic cache to confirm accuracy before finalizing fails because a semantic cache only reuses a prior summary from a similar query to save time, it does not critique this resident's draft against current facts, so a mismatched or stale claim can still reach the final summary unchecked.

Exam caveat: reflection/self-critique is a pre-release draft quality optimization, not a substitute for grounding to official eligibility policy and not permission to invent rules. Operational check: require a critique pass on eligibility drafts, block finalization when critique flags missing conditions or unsupported claims, log critique outcomes for the county benefits bot, and verify residents only see post-critique summaries.