The sales agent reads replies to cold outreach and stops contacting anyone who asks. Missing an opt-out is the costly error, so we wrote 36 new replies, 10 of them opt-outs, and scored the rules once before changing anything.
The numbers. 26 of 36 correct. Four opt-outs fell to "unclear", which hands them to a person instead of suppressing them: "take this address off whatever you use", "cease all correspondence", "I'd appreciate it if you struck my name from your prospect database", and a CCPA request to delete and not sell personal information. We added those phrasings. The set now scores 30/36 with all 10 opt-outs caught.
What it does not show. The replies were written by an agent that had read the earlier sets, and the set is now tuned on, so 30/36 is a regression figure, not a fair one. The misses were all in the safe direction (a person reads them), but a real inbox will phrase things we have not thought of. Real human replies are still the open item.