Where the headline 99.94 percent uptime number comes from, what it includes, and what it excludes.
The 99.94 percent figure is computed over a 90-day rolling window, sampled per mirror, then aggregated into a platform-level availability percentage. A mirror is counted as “up” for a probe cycle if it returns a complete HTTP response within the 60-second probe budget. Degraded latency does not count against uptime; only failed responses do.
What's excluded from the figure: scheduled maintenance windows announced in advance via signed advisory; Tor-network-side guard rotations that affect probe circuits but do not affect on-platform sessions; and individual circuit failures that the probe retries successfully on a fresh circuit within the same cycle. The full computation is reproducible from the public probe log.
Successful probe responses divided by attempts over the window, per mirror. It counts reachability. It does not count whether pages rendered correctly, whether orders processed, or whether the service was usable rather than merely answering.
With several mirrors, the aggregate figure only needs one of them answering. That is the number people quote and it hides a mirror that has been degraded for weeks. The per mirror bars are the honest view and they are the ones worth reading.
Ninety days is long enough to smooth out noise and short enough that an outage two months ago has already started fading. A figure that looks stable can be sitting on top of a rough patch that is about to age out of the calculation entirely.
It tells you the operator can keep infrastructure running under pressure, which is genuinely useful for planning when to visit. It says nothing about escrow, arbitration or how the service ends. Every platform that ever disappeared had a good uptime record the day before.
Component states, the 90 day uptime history and the incident log live on the status dashboard. This page covers one question in depth rather than repeating the board.