I haven’t checked what they do either, but I don’t see that as a reasonable assumption.

If they test for short burst performance specifically then the short runtime is the point. You then use multiple iterations of the test run to get a reasonable, reproducible result by averaging.