Notifications

Storm notifications are opt-in, per device, per place. You pick the places that matter and the band each one is allowed to wake you for, and nothing else gets through.

This page is about the part that is harder than it sounds: not deciding whether a storm matters to you, but deciding how many times to say so.

Why episodes exist

Two deduplication mechanisms already run before anything reaches a person:

Neither touches the third axis, and it is the one a person actually feels.

Observed live in one 90-minute window, seven storms crossed the alert threshold over one lake shore. Two of them were 51 km apart on near-parallel tracks about an hour apart along the line - one weather event to anyone underneath. Two others shared an identical bearing with one roughly 370 km and seven hours behind the other - two separate events that happen to follow the same road.

Under per-storm notification, somebody under that line is woken three or four times for the first group and again that night for the second.

This cannot be resolved at the storm. Whether two cells are “the same event” depends entirely on whether both reach you. It is only answerable per person, per place.

The episode

An episode is one weather event as one person experiences it. It is hysteresis again, moved from the storm to the person:

episode = (device, place)

Per device rather than per account: two handsets on one account are two people’s pockets as often as not, and the one that was switched off should still be told when it comes back. place is either the device’s own position or a saved place, so somebody watching their house and their parents’ house gets an episode for each.

The place is snapshotted, not looked up. A phone that moves 200 km should not retroactively change which storms were judged to have covered it.

What opens one

A storm qualifies for a place when all of:

1. Its cone at the notification lead covers the place - the cell’s cone for a thunderstorm tier, the rotation’s own cone for a tornado one, capped at 25 minutes whatever the lead says.
2. The tier it would raise is at or above the floor you selected.
3. It is genuinely firing, not merely touching a threshold for one scan.

The first qualifying storm opens the episode and sends one notification.

The notification lead is 30 minutes, and it is a product decision rather than the map’s lead selector. The cone’s area grows 5.3× between 10 and 60 minutes, and everyone inside it gets woken.

The bands

Five bands on the tornado axis, and the one you choose is the floor: nothing below it can open an episode. The labels track the measured rate rather than how a number feels, because the label is a claim.

The tornado column is year-round with quiet days included; the severe column is measured on an outbreak day, so read down a column rather than across a row. Both bases, and why they differ.

Significant is the default and the one the rest of this documentation is written around. It is where the multi-hazard claim is strongest: better than four in five of those storms produce damaging wind, hail or a tornado.

What escalates, and what stays quiet

A second storm entering the episode, or an existing one intensifying, sends a new notification only when it raises the episode above what has already been delivered. Everything else updates the existing notification silently.

The ladder

Everything WXA/SA can tell you is one integer on one ordered ladder, so “have I already told this person something at least this serious?” stays a single comparison. Two axes are deliberately kept apart: the hazard (wind, thunderstorm, tornado) and the evidence tier (how we know, and how strongly).

Detected outranks Risk within a hazard; the hazard class outranks the tier across them. A located rotation whose cone covers your point is more actionable than a probability on the parent cell.

So the ladder is not primarily a description of likelihood - it is the escalation key. A cell crossing from Elevated to Significant is worth waking somebody for. Drifting around inside Significant is not. A thunderstorm alert becoming a tornado alert is just a tier change, which is what makes the upgrade fall out of the same mechanism rather than needing one of its own.

This also handles the growth problem with no state on the phone at all. Cells grow abruptly - every couple of minutes somewhere, a cone’s footprint and the population under it multiply:

• Someone newly swept in has no open episode - first notification.
• Someone already told, at the same tier - silent update.
• Someone already told, at a lower tier - escalation breaks through.

Nothing depends on a phone remembering what it received. A handset that was off, reinstalled, or restored from a backup would otherwise re-alert or stay silent for the wrong reasons.

Time-to-arrival deliberately does not escalate. It changes constantly and would re-wake you every two minutes; it rides in the payload so the notification’s content stays current without alerting again.

What closes one

No qualifying storm for three hours.

Short quiet windows do not work here. Closing after six minutes of quiet would let the next cell in the same line open a fresh episode, which is the double notification this exists to prevent. The observed spacing is what sets the number: cells an hour apart are one event, cells seven hours apart are two, three hours sits between them.

How many times one storm wakes you

Replayed against planted devices across three real events of very different shapes:

One episode per device in all three, with zero repeats at the same tier. Devices under the supercell walked Thunderstorm Risk, Thunderstorm Detected, Tornado Risk, Tornado Detected as one storm matured across them, the western ones catching it young and the eastern ones getting the full ladder. The squall line arrived 15 to 33 minutes ahead of the first severe report near each device and correctly never reached a tornado tier on a wind event.

A new storm can break a long silence

Escalation is only an escape hatch while there is somewhere left to escalate to. Under a hurricane the top tier is reached almost immediately, and after that nothing is worse - it is equally bad, repeatedly, for hours. At one point under a landfall, seven distinct qualifying storms crossed over four hours, three of them deep in the Extreme band, and escalation alone would have gone quiet for four of those hours.

So there is a third reason to notify, beside open and escalate:

A different storm reaching the episode’s current tier, when the last notification was more than 30 minutes ago.

Never the same storm twice. The cooldown is not a way around the property this design exists for: a storm that has already woken you stays silent for as long as it lasts, however long that is. Under that landfall the longest silence falls from 6.3 hours to 2.0, while a single supercell sitting overhead still holds its four-hour silence, because it is one storm and it has already said what it has to say.

How often will it alert me?

Nobody can answer that, and it is the wrong question. Severe weather volume changes by the minute and by the season, which is why the National Weather Service makes no such promise either. A rate would be a forecast of the climate at your house rather than a property of this system.

The answerable question is the one above: when a storm does come, one episode, and two to seven notifications across an event lasting five to ten hours.