How a measurement becomes registry data
You finish a lab run and tick "Contribute anonymously" under the result. Only the summary metrics shown on screen are sent.
The server validates every number against a per-kind whitelist and range, hashes your connection with a daily salt for rate limiting, and stores the row with the model identity.
Once a model has 30 or more measurements it gets a public page with medians, percentiles and connection splits, regenerated with every release.
Why the threshold matters
A model page with three samples would mislead. Thirty independent measurements give a stable median and meaningful quartiles, so the registry waits. Until then the model only appears in the counters above and in the percentile sentence shown to contributors of that model once ten samples exist.
Model identity
- Controllers: the USB vendor and product ids the Gamepad API reports, mapped to names such as Sony DualSense or Xbox Series X|S controller. Unknown pairs keep the raw name.
- Mice and keyboards: WebHID ids when the visitor identifies the device, otherwise the name they typed, normalised to a key.
- Phones: the model from browser client hints on Android; Safari hides the model, so iPhones aggregate under a generic Apple entry.
Reading the numbers
Drift, latency, jitter and gap shares are better when lower; sampling rates, refresh rates, stability and coverage are better when higher. Every model page states the direction next to the metric. Percentiles are computed from the stored samples for that model only, so a DualSense is compared with other DualSense units and never with an Xbox controller.
Measurement boundary: every value in the registry is browser-observed. Bluetooth and wired connections are counted separately because they behave differently, and a large sample of browser measurements still describes what browsers see, not a laboratory bench.
Abuse controls
Twenty-four submissions per connection per day, identical results suppressed for ten minutes, strict numeric ranges, no free text beyond an 80-character device label that is stripped of markup. Medians and quartiles are used instead of means so a handful of absurd values cannot move a model's headline.
