One bar for a red eval case
gstack 1.91.33.0: a red eval case gets one clear verdict
9/10meets8/10, reds explainedqualified7/10extend onceanything elsefixWhen /ship measures a red eval case, it now gets one verdict from one bar. Nine of ten trials meets it. Below that, the case is qualified, extended once, or sent to a fix round.
Eight of ten can qualify
Eight of ten passes as qualified when every red is explained
Eight of ten passes as qualified when every red is explained. A provider error, judge noise at the threshold, or a model miss with a cited transcript line. A hand picked label needs evidence.
One more batch, never a third
Seven of ten gets exactly one more batch
Seven of ten gets exactly one more batch. Same tree, one decision on all twenty trials. If anything changed, it refuses to pool, and there is never a third batch.
Fewer pointless fix rounds
A healthy case is sent to a pointless fix round half as often
a case that passes 95% of its trials
That means fewer pointless fix rounds. A healthy case used to land in one about one time in twelve; now about one in twenty-four. A flaky case still fails more often than not.
A weekly sweep on main
A weekly sweep measures main's flakiest gate cases first
Mondays 12:00 UTCor on dispatchtop 5 casesthe ones that cost the all-green odds most$150 a weeka hard capread-onlynever pushes or opens a pull requestEvery Monday, a new sweep measures the flakiest gate cases on main, before a ship hits them. It takes the five that cost the most, under a hundred and fifty dollars a week.
Smaller fixes
CHANGELOG.mdFull notes for 1.91.33.0Measure now accepts a case that shares its test file, runs native evidence cases outside CI, and the plan mode no-op case stops timing out on large pull requests. The change log has the rest.
