Emergency management wants alert severity decided deterministically, then rewritten in plain language for the public. Which pattern should they implement?

Emergency management that wants alert severity decided deterministically, then rewritten in plain language for the public, should use rules for severity and an LLM only for the plain-language rewrite. Severity is a checklist decision with life-safety consequences; letting an LLM decide it is unsafe. Rules (or equivalent deterministic classifiers) assign severity from sensor/policy thresholds; the LLM then explains that fixed severity in clear public wording. That split keeps authority where it belongs and still improves readability. Using rules for alert severity and an LLM for the plain-language rewrite of readings fails by target: it rewrites raw sensor readings instead of the decided severity alert, so the public wording never communicates the deterministic severity decision. Using rules for alert severity and an LLM for the plain-language template selection fails by function: selecting a template is not rewriting the decided severity in plain language, and letting the model choose risks pairing the fixed severity with the wrong public wording. Using rules for alert severity and an LLM for the plain-language intake summaries fails by stage: it summarizes incoming intake instead of rewriting the decided outgoing alert, so the deterministically decided severity is never translated for the public. Exam caveat: for public alerts, deterministic rules own severity; the LLM owns wording only—do not let the model decide or invent severity. Operational check: encode severity rules from emergency policy, feed the decided severity into an LLM rewrite step, forbid model overrides of severity, and sample public posts to confirm severity labels match the rules engine output.