Week reported gives an idea of the workload for the triaging team and developers. Due to the delay between a bug being introduced and reported, any potential effect of LLM review reducing the total number of bugs is also delayed.
Week introduced does not have this delay, but instead undercounts due to bugs yet to be found. Missed bugs tend to be reported later, so recent weeks overstate the caught share. An estimate to compensate for both the delay and caught share can be toggled.
Even compensating for delays, interpreting these numbers is tricky. There is impact from the release cycle, summer holidays, Winter of Quality, number of projects and features landed, large numbers of LLM-assisted fix commits, and changing numbers of users and developers.
By week reported
Moving average
Caught by reviewMissed by reviewOther: not reviewed, external, untraceablePre-existing (before 2026)Commits (right axis)
Week 5 · 26 Jan – 1 Feb
9-week average
Commits199.7
Caught1.4
Missed5.7
Other0.6
Pre-existing13.2
20.9 reports
By week introduced
Moving average
Caught by reviewMissed by reviewNot reviewedCommits (right axis)
Notes
Recent weeks undercount bugs: the median delay from commit to report is 31 days, the 90th percentile 142 days. "Adjust for reporting delay" divides each week's counts by the estimated fraction reported so far, derived from older weeks' delays. Caught and missed bugs each use their own delay distribution, since caught bugs are reported sooner. Bars are faded where under half is reported. It can't account for delays longer than this year's 43 weeks.
"Reported within" counts only bugs reported that many days after their commit, so weeks can be compared on equal terms. Unadjusted, weeks that haven't had that many days yet are hidden. Adjusted, their count is estimated.
Commits exclude merge commits.
Commit Review Stats
This is an analysis of the effectiveness of commit review catching high severity bugs. Do not expect the numbers here to be strictly accurate and every issue to be classified correctly, but from spot checking it appears generally correct.
This covers commit reviews from 2026-01-01 to 2026-10-29.
The model and review instructions were changed at week 21 and again at week 31, to catch more bugs.
High Severity Bugs
Period
Caught
Missed
Not reviewed
Skipped
Pre-existing
External
Untraceable
Caught rate
W31+
32
25
0
0
45
5
3
56%
W21–W30
26
70
0
0
80
6
0
27%
W01–W20
40
276
3
0
240
3
0
13%
Caught
Introducing commit was reviewed and the review flagged the same defect.
Missed
Introducing commit was reviewed but the review did not flag this defect, whether it flagged a different issue or nothing at all.
Not reviewed
The commit was not reviewed, typically release branch only bugs as only main was covered by commit review.
Skipped
Introducing commit was in a skipped-commits list.
Pre-existing
Introduced before commit review started on 2026-01-01, the review could not have caught it.
External
Root cause is a third-party library, compiler, or infrastructure issue outside the Blender repo.
Untraceable
Commit that introduced the issue could not be determined, and the bug is not clearly an in-window regression.
Caught rate
caught out of caught and missed.
Pre-existing, external and untraceable bugs have no introducing commit in a period, so they are grouped by the week the bug was reported instead.
Delays Between Findings and Reports
Different periods are not directly comparable. There is a delay between a bug being introduced and it being reported and fixed. And bugs caught tend to be ones that are reported earlier.
For the latest period the data is also not yet statistically significant (at the time of writing).
Here are some stats counting only bugs reported within a given number of days after the introducing commit.
Window
W31+
W21–W30
W01–W20
15 days
65% (15/23)
31% (14/45)
18% (20/111)
30 days
59% (16/27)
30% (17/57)
16% (22/135)
60 days
53% (10/19)
29% (22/75)
14% (26/185)
All
56% (32/57)
27% (26/96)
13% (40/316)
Notes
There are a number of reasons these stats don't represent the true effectiveness:
Some missed issues have not been reported or classified as high severity yet. Especially for commits that are not in an official release yet.
The commit review log still contains bugs that would be classified as high severity (regression or crash) if they were reported.
The commit review log was used to fix issues before they became bug reports, and some developers may use their own LLM review already for pull requests.