Every figure on this page was measured by replaying real days through the same pipeline that runs live, and scoring the output against Storm Prediction Center storm reports and archived National Weather Service warning polygons. A hit means a report within 40 km of the storm and within the next 30 minutes.
How early the alerts are
Scored against the NWS tornado warning on the same storm:
| Supercell day | Second supercell day | Cold-season squall line | |
|---|---|---|---|
| Our lead on the tornado | 93 min | 67 min | 24 min |
| NWS tornado-warning lead | 35 min | 15 min | 10 min |
| We alerted first | 86% | 100% (23 of 23) | 86% (19 of 22) |
| Median gain | +45 min | +53 min | +10 min |
Measured a second way, against the tornado reports themselves, the Significant band precedes the report by a median 62 minutes against 21 minutes for the tornado warning on the same events.
What the alert actually tells you
Of the severe storms already warned tonight, these are the ones becoming tornado-capable, a median 45 to 54 minutes before a tornado warning is issued on them.
It is an upgrade signal, and three things follow from that.
The NWS is already on these storms. On the largest tornado day of 2026, 91 of our 94 cell firings already had a Severe Thunderstorm Warning in force, a median 99 minutes before we fired. The forecasters had the correct product out long before us. What they did later was upgrade to a tornado warning, and the upgrade is what we are early to.
Lead on the storm is not lead on your place. One tracked storm traveled 168 km between our alert and its tornado. Nobody standing where that tornado landed was under the storm when we fired, which is why notifications come from the cone rather than from the storm’s current position.
It does not catch everything. On one squall-line day the rotation tier reached 29 of 68 tornado reports and the cell tier 24, against the NWS’s 46. Days like that are not rare enough to treat this as complete coverage.
What it looks like on one storm
Nine hours of scans on a single supercell, condensed:
19:32 ── NWS SEVERE THUNDERSTORM WARNING
21:00 Watching severe 50% shear 0.009 hail 0.6" an ordinary strong storm
21:16 below band severe 41% shear 0.003 hail 0.7" nearly falls apart
21:28 Watching severe 65% shear 0.007 hail 0.9" reorganizing
21:42 Elevated severe 92% shear 0.007 hail 2.8" big hail, modest spin
21:46 Significant severe 92% shear 0.021 hail 2.8" ◀── WE ALERT
▲ shear tripled in one two-minute scan
23:16 Elevated severe 93% shear 0.012 rotation, 48 kt
23:42 Significant severe 95% shear 0.012 rotation, 57 kt
00:04 Significant severe 94% shear 0.011 ◀── NWS TORNADO WARNING
00:13 ◀── TORNADO
168 km from where we alertedIt starts as a hail storm, with hail size climbing past 2.8 inches while the rotation never leaves Watching. The rotation then appears in a single scan - shear triples and the storm crosses from Elevated to Significant in one two-minute step. That abruptness is the thing worth catching, and it is followed here by two hours and eighteen minutes of sustained rotation before anything touches the ground.
When the feed falls behind
Every response the system serves carries the age of the scan it was built from, and that age is used rather than displayed for decoration:
- Past 15 minutes, the map says the layer is stale instead of presenting old positions as current.
- Nothing may notify on a stale scan at all. A 30-minute cone drawn on a 15-minute-old position is a 15-minute cone drawn in the wrong place, because the storm has traveled 15 to 20 km since.
The good news is that it is late, not lost. Every scan is published, sometimes twenty minutes after the fact, and the ingest refetches the whole backlog the moment it appears, so tracks stay continuous through a stall. What you lose is warning time, not data.
Where it cannot see
- Far from a radar. The beam spreads about 1° with range, so the same rotation reads weaker at 200 km than it does at 50, and a fixed threshold under-detects at distance.
- Below the beam. About half of tracked rotations have no low-level dual-pol coverage at all, which is why debris confirmation is a badge on an alert rather than a level of its own.
- Anywhere the radar quality is too poor to trust. Quality gates whether a rotation is drawn at all. A badly covered place is told that we cannot see it, rather than shown a map identical to “nothing is happening”.