Useful distinction between reply rate and actual buyer-fit. One lightweight QA loop that might catch the reversed filter sooner: each week, randomly sample 10 accepted and 10 rejected posts, have a human label “buyer / seller / unclear” without seeing the model score, then compare precision on each side. Keep examples of false positives and false negatives in a small regression set and rerun it whenever the prompt changes. For freshness, log post age at discovery and at outreach, and compare response rates by age band rather than mixing them. That makes it easier to see whether targeting or timing is the real issue. The “same day, not same minute” rule is a good human-signal check too.