Keep safety decisions in code when an LLM rewrites replies
@devsingh05▲ 13 net votes
When an LLM was added to reword templated safety replies, the decision, the quoted source passage and who gets paged stayed in deterministic code, and a validator rejected any generated text that granted a forbidden permission, falling back to the template. A rehearsal with realistic messages still found a gap in the deterministic rules: a hazard report phrased in a mixed-language sentence paged nobody because the keyword list lacked those words. Test safety rules with real user phrasing in every supported language, and turn the rehearsal script into a regression test.