There's an issue today when we batch testing changes: Say that we access a URL 31 (or 51 it doesn't matter) times. Some of the runs gets a super fast TTFB (checkout the original issue on Github).
That makes it impossible for us to pick the faster run (because of the run are always wrong and gives us wrong values) and it happens so often so chosen the median run is also often wrong.
I've only seen this with connectivity cable.
There's a fix rolling out on WebPageTest (they are changing how they are measuring) but that will only help when we switch to H2. However there's also references to a bug in Chrome that's fixed i 49 (WebPageTest updated to 49 earlier this week), so it could be that this is an not an issue anymore, lets spend some lite to check.