Distinct installs that contacted the server in the hour (or the 24 h) ending at each point — deduplicated by install across all four registries, so a phone holding a device token, a push-to-start token and a widget token counts once. Any contact counts, foreground or background: an install woken by a nudge is still one that exists, has push enabled and is running our code. Note this is a floor, not a ceiling — the app fetches its weather straight from Open-Meteo and SHMÚ, so someone can read a full forecast without us hearing about it. Recorded from 2026-08-24; earlier points are absent rather than zero.

Windy meters per request against our key, and every call here comes from one place: a device opening the Webcams screen somewhere the SHMÚ cameras don't cover. Calls per hour is the last 24 h of that log; Calls over time follows the Stats range picker like every other trend. Windy publishes no request quota — their free-tier limits are paging depth, image URL lifetime and image size — so there is deliberately no ceiling drawn here and none enforced in the server either. What holds the number down is the cache: one upstream call per 0.25° area per 10 minutes, however many devices ask.

Each forecast is scored an hour later against the radar at that point. Covers both running Live Activities (LA) and the nowcast computed for every registered install (app) — the second is what the app's rain card would have shown that user, sampled hourly. Two different questions, kept in separate charts on purpose. Outcomes grade what we SAID — of the forecasts we issued, how many the radar confirmed, so false alarms are the number to watch there. Coverage grades what we DIDN'T say — of the rain worth announcing that fell, how much we called in advance. Silence is only counted against us above the 0.8 mm/h announce floor, because the app deliberately never predicts drizzle; grading it at the bare detection threshold once turned every intentional silence into a "miss" and made a drizzly spell look like a total failure. Radar is the reference, so a radar that over-reads flatters these figures.

Nudges = silent refresh pushes · Widget push = WidgetKit rain transitions · Registrations show guard rejections in red.

Last reported position, rounded to ~1 km. Click a cluster to split it; click a legend entry to hide that category. Colour = highest-priority registration at that spot.

One count per install, not per request. "Unknown" is pre-1.3.2, which didn't report a version.

Devices get silent nudge pushes; push-to-start tokens let the server start a Live Activity without the app running.

Cached ~1 h, per storefront. Downloads and revenue need App Store Connect and aren't available here.

"Tripped" = the limit fired. Trips and warnings are in memory only, so they reset on restart.