This smells to me like the server is struggling under load. I have a couple of servers in the pool in underserved regions that are doing thousands of requests per second, and I had forgotten to disable ip_conntrack on one of them. The table would fill up and drop connections, so monitoring would take it out of the pool, and the table would drain and it’d come back in.
I hadn’t read the scoring code before, but I agree with your reading: 0.731 / 2.308 = 0.3167… seconds seems to be the point where step = 0. That’s higher than I would have thought, but I guess it’s the answer for what’s considered “Good Enough.”
More generally, there was a pretty similar discussion a while back:
For my own servers, I usually configure the list of servers by hand and expect offsets to be on the order of a couple milliseconds or less, but they’re generally running in well-connected data centers, or my home LAN where I’ve ended up with a growing collection of GNSS-locked stratum 1s. I assume some pool clients are things like a low-power device connected through an intermittent cellular connection which will limit how close they could get.