Browser congestion-control rate
Writes datagrams as fast as the browser will accept them and measures what actually arrives at the server. A rate on its own is ambiguous, so the run also records what was offered here and, at the server, the loss split: application packets that never arrived while the QUIC packet numbers around them all did were dropped inside the browser — proof that its congestion control, not the path, was the limit. The verdict below says which of the four possible bottlenecks (browser CC, path, this page, the server) the run measured. Other pages: WebTransport probe, WebRTC probe, packet trains, RTT monitor.
This saturates your uplink for the whole run. On a phone it uses real data — a 10 s run at 20 Mbps is about 25 MB — and it will slow down anything else on the link. Runs are capped at 120 s.
Offered vs delivered
The page's offered rate against what the server actually received, second by second. In saturate mode the gap between the two lines is what the browser refused to send. In ramp mode the knee — where delivered stops following offered — is the rate its congestion control settled at.
RTT while saturating
The server's own QUIC RTT to this browser. If it climbs as the rate rises, the bottleneck queue is filling — the classic bufferbloat signature, and the reason a loss-based controller finds this rate at all.