Project / new run
Compare a candidate against a baseline
The run executes every case against both versions, scores each output, and ends with a release verdict — including hard gates that can block a high average.
A comparison needs at least two target versions. Create a second version on the target page first.