Both corpora windowed, and the collection etiquette that has to change¶

datasets/results/window-report-{openstack,qt}.json.

Both train windows are complete; Qt's dev holds 2025-09 only. Both dev windows clear the bootstrap's ten-change floor. Equalized training is 4,327 examples per organization.

Cross-window near-duplicate rate, Qt: pilot into train 0.0000 / 0.0005 / 0.0011 / 0.0028 and train into dev 0.0000 / 0.0055 / 0.0082 / 0.0302 at Jaccard 0.8 / 0.7 / 0.6 / 0.5. OpenStack's train into dev is 0.0000 / 0.0106 / 0.0142 / 0.0177. A threshold registered at 0.7 must sit above roughly 1.1%, the larger of the two, to be a bound both organizations can meet.

Collection etiquette, which the transport got wrong twice. One comments request per change plus one diff request per surviving comment is roughly 5,000 to 6,000 requests per month of Qt data, about 70,000 for thirteen months, issued at the 5 requests a second the 0.2-second interval allows. Both servers tolerated that for about three hours and then stopped completing handshakes. The first incident was read as transient and answered with pacing tuned to make the symptom go away; the correct reading was that the rate itself was wrong for a volunteer-run Gerrit. Future collection runs at about 1 request a second and stops rather than retrying when handshakes begin to fail, since the frozen corpus is fetched once and reused.