Red build
40 failed tests
Mostly noise, with a few real regressions hidden inside
Step 1 · Deterministic rules
Rules decide what they can first
For example:failed, then passed in the same run→flaky
Rules can decide
Rules can't decide
Step 2 · Jev
Classifies only the ambiguous cases
{"category": "environment",
"p": 0.82}
≥ 0.75
Confidence ≥ 0.75?
Confident answers are labelled. The rest go to a person.
Labelled
Failure category
RegressionFlakyEnvironmentTest dataUnknown
< 0.75
Needs a person
Low-confidence calls never label a failure on their own