Qt's ban lifts, and the corpus is complete¶
scripts/resume_when_allowed.sh probed Qt's Gerrit with five requests every twenty minutes
from 20:29. Fourteen probes answered nothing. The fifteenth, at 01:45, answered all five, and
the script built 2025-10 at one request a second, finishing 04:30: 826 examples, and Qt's
thirteenth month. The snapshot had been on disk since the fetch that preceded the ban, so
nothing was refetched; only the build, which needs a comments request per change and a diff
request per comment, had been blocked. Total ban duration from the first refused handshake:
about thirty-three hours.
| Qt | examples | after dedup | pilot | train | dev |
|---|---|---|---|---|---|
| 12 months (before) | 10,622 | 9,918 | 1,300 / 403 ch | 8,254 / 3,160 ch | 364 / 175 ch |
| 13 months | 11,448 | 10,695 | 1,301 / 404 ch | 8,443 / 3,208 ch | 951 / 462 ch |
The dev window is where the missing month lived, so it nearly triples in changes and Qt's dev window is now larger than OpenStack's 235, reversing the relationship every result so far was measured under.
The censoring figures move a third time, and settle. Qt's dev was being reported as 70.4% missing, which was one absent month plus real censoring of the other. With the month built it is 28.8%, all of it censoring.
| window | OpenStack missing | Qt missing | differential | with Qt's month absent |
|---|---|---|---|---|
| train | 5.0% | 3.0% | 2.08 pt | 0.83 pt |
| dev | 39.1% | 28.8% | 10.33 pt | 31.26 pt |
| test, fetched 2027-02 | 1.1% | 0.7% | 0.36 pt | 0.54 pt |
Qt's estimator check passes at 0.019 against a two-standard-error band of 0.050. OpenStack's still fails at 0.086 against 0.071, which is the cohort-trend limitation and is unaffected by any of this.
The decision has now survived three revisions of the numbers that justified it. Creation windows stand: the confirmatory contrast carries 0.36 points of differential censoring against the dev window's 10.33, a factor of 29. The ratio has been 26, then 58, now 29, while the underlying figures moved by tens of points. That the conclusion never depended on the figures being right is worth more than any one of them.
What it costs the windowed RQ1 result. Job 144345 measured Qt's contrast on 175 changes, +0.011 [-0.0208, +0.0433]. Job 145084 reruns it on 462, which should narrow that interval by roughly a third without touching OpenStack's. Nothing about the earlier run was wrong; it was measured on a corpus that was missing a month, and said so.