They compute as total minutes down as a fraction of total time. What this means is that being down, say, 55min during peak-use counts the same as being down 55min when nobody is trying to use it. And congruently it counts being up when nobody is trying to use it as the same as being up when everyone is trying to use it.

Over 90 days though. They had a lot fewer users in February. (And even then, these outage durations seem to add up to more than the error budget 99.26% implies...)