Every row is a server our prober reached itself, not an entry copied from a published list — an OpenVPN handshake in almost every case, and a plain connection check for the few whose control channel is encrypted end to end. Latency is the median of its successful probes rather than a single reading, so one lucky reply cannot flatter it. Uptime 7d is the share of the probes we sent over the last week that came back. Score folds those together: 50% reliability, 35% median latency and 15% how the server has done in the last hour, then 15% off for UDP on a random port above 1024 (the standard 1194 is exempt) — a risk our prober cannot see, because mobile networks drop idle UDP mappings and it does not sit behind one. The formula is public.
The list is ordered by score, highest first. It is computed when the page is requested rather than cached, so it always reflects the last probe round.
Of the 50 servers on this page, all of them speak TCP. UDP is faster and the right default on a fixed line; TCP is the safer pick on mobile data or behind a corporate firewall, because those drop idle UDP long before they interfere with TCP. The port matters as much as the protocol: none listens on 443 or 1194, the two least likely to be blocked, so a restrictive network may refuse the lot.
Churn is the thing to expect. 787 servers match this view at this moment, and the same request tomorrow will not return the same list: these are volunteer machines, not servers this site operates, and their operators restart, move and withdraw them without notice. Ones that stop answering keep their page and their probe history rather than being deleted. A handshake proves only that a daemon is listening — what the operator can and cannot see is a separate question.