ISPs quote a number from a server sitting inside their own network, which measures the shortest possible path rather than the one your traffic actually takes. NetVerdict measures to a real internet edge and shows both figures, so "why is my speed test 57 but Netflix buffers" has an actual answer instead of a shrug.

Challenges (real ones, from this build):

• Chrome's HTTP/1.1 6-connections-per-origin cap was silently starving the bufferbloat probe — with all 6 slots held by transfer streams, a latency probe queued behind them for 45 seconds and returned only when the phase deadline killed it. Fixed by reserving a connection slot whenever bufferbloat measurement is active.
• The original byte-target sizing could end a transfer while it was still in TCP slow-start, reporting the ramp-up as the result. Rebuilt around a fixed-duration deadline instead of a guessed byte count.
• Kept the live particle-field visualization fully isolated in its own Web Worker with a transferred OffscreenCanvas, so heavy rendering can never contend with, or skew the timing of, the actual measurement.

Bufferbloat probe stalled 45s behind its own traffic. Chrome caps connections per origin at 6; the download phase used 7 streams, starving the latency probe. Capped streams to the real connection budget.

Download read 3x low — 9.7 vs 57 Mbps. Four compounding bugs: mistimed speed probe, byte-target sizing that ended transfers mid slow-start, a timeout below the target duration, warm-up discard anchored to the wrong clock. Rebuilt around a fixed-duration deadline. Fixed figure: 29.4 Mbps, verified independently.

A+ bufferbloat grade on a connection that was never loaded. A rate-limited download still scored a perfect grade because nothing checked if the transfer actually succeeded. Gated grading on the transfer's own outcome.

Canvas crashed on remount. transferControlToOffscreen() only works once; React reused the same DOM node on remount and threw. Fixed by tracking transferred canvases and forcing a fresh element.