An uptime number is meaningless without knowing what counted as down. Here, a call that fails to connect is an outage. A call that connects but takes over four seconds to first audio is degraded and counts partially. Both are measured from synthetic probes plus live traffic, not from a dashboard's self-report.
The commitment is on call connection and answer, because that is the thing whose failure costs money. Latency degradation is credited on a sliding scale rather than treated as a binary outage. Upstream provider failures absorbed by the fallback chain do not count against the SLA, and failures that reach the caller do.
Synthetic probes cannot catch quality regressions — an agent answering fast and wrongly reads as perfectly available. That class of failure is caught by the eval suite and by escalation-rate monitoring, not by this page, and it is worth knowing that the uptime figure is silent about it.