First honest example count¶
build run against the real October 2024 OpenStack snapshot, 476 of 1,678 changes
processed in 95 seconds before a time cap.
| examples produced | 208 |
| examples per change | 0.437 |
| projected for the full month | ~733 |
Drops, by reason:
| reason | count |
|---|---|
| comment on a metadata pseudo-file | 675 |
| comment anchored in no changed hunk | 191 |
| comment on the last patch set | 42 |
| comment with no line anchor | 5 |
This revises the yield estimate upward, substantially. The earlier projection chained two small-sample rates (23% of changes carry a code comment, 27% of those anchor) to imply roughly 6% of changes yielding an example. The real rate is 0.437 examples per change, because changes that do carry comments frequently yield several. At ~733 examples per month, a ten-month training window gives OpenStack alone something in the thousands, with Qt on top.
Commit-message comments are the largest single drop category, 675 against 208 kept. That makes the pre-registered exclusion of commit metadata a decision with real weight rather than a formality, and it belongs in the report's sampling section by number.
The strict anchoring rule looks viable. 208 anchored against 191 unanchored is roughly 52% of code-file comments landing inside a changed hunk, against the 27% the 15-comment sample suggested. The case for widening to a 10-line neighbourhood is correspondingly weaker; decide it on the pilot's minimum detectable effect, not on this, but the strict corpus is no longer obviously too small.