A public-health department wants richer critique of triage drafts without exposing long reasoning chains to residents. Which approach fits?

A public-health department that wants richer critique of triage drafts without exposing long reasoning chains to residents should use chain-of-thought evaluation offline to critique drafts, not in resident-facing replies. Offline critique can use extended reasoning to catch missing symptoms, unsafe advice, or incomplete next steps, then produce a short, clean resident answer. Keeping CoT backstage preserves evaluation depth without dumping confusing or sensitive reasoning into the public health portal chat. Evaluating triage drafts offline with single-score LLM judging for rapid release decisions fails because a single score speeds release but gives no richer, case-specific critique of missing symptoms, unsafe advice, or incomplete next steps that the stem asks for. Aggregating offline resident preference comparisons on triage drafts to guide prompt revision cycles fails because preference rankings show which draft is liked, not why it is unsafe or incomplete, and they don't provide the richer, case-specific critique the stem asks for. Scoring triage drafts offline using embedding similarity search against curated reference responses fails because similarity scores only measure closeness to a reference, not missing symptoms, unsafe advice, or incomplete next steps, and they don't provide the richer, case-specific critique the stem asks for. Exam caveat: chain-of-thought belongs in offline evaluation/critique for civic triage quality—not as single-score judging for rapid release decisions. Operational check: run CoT critique on triage drafts in a non-resident pipeline, store critique artifacts for reviewers, release only the cleaned final reply to residents, and verify release decisions never rely on single-score judging alone without rich critique.