What our own gate did, on our own pull requests
This is dogfood. It is our own team gating our own pull requests with our own product, on one repository. It is not a customer's data, and it is not a benchmark.
align-stack, from 2026-08-04 to 2026-09-03 - measured from the commit that made the Decision Gate a required check.
The denominator
495 pull requests merged with the gate required
The gate flagged a conflict on 18 of them, about 1 in 28.
What a human then did about each flag
| Outcome | Count | What it means |
|---|---|---|
| Confirmed by a human | 1 | a person read the flag and said the conflict was real |
| Rejected with cause | 3 | a person read the flag and gave a reason it was wrong |
| Accepted as a supersession | 1 | the change was allowed to stand as replacing the earlier decision |
| Could not confirm | 13 | nobody rated it either way. Not a pass and not a failure |
The largest bucket is the one nobody rated. It is shown on its own line and is never counted as agreement. Four rated cases out of eighteen cannot support a false-positive rate in either direction, so this page does not give one.
What happened to the pull request
| Outcome | Count | What it means |
|---|---|---|
| Passed every time | 464 | the gate never failed on this pull request |
| Diff changed after a flag | 25 | the gate failed on an earlier commit and passed on the final one |
| Merged with the gate still red | 5 | a person merged anyway. The gate advises; it does not block a human |
| No gate result on the final commit | 1 | an anomaly, counted here rather than dropped |
How this was counted
Hand-verified from GitHub's own CI records for every flagged pull request, rather than read from our own database. A cross-check against the database found that only 6 of the 18 conflict events carry the platform tag that would let the product's own dashboard count them, so the dashboard cannot yet reproduce this table. That gap is in our instrumentation, and it is the reason these numbers were counted by hand.
A week with a zero, or a week that reads badly, is published the same way. A page that only shows good weeks is a marketing page.