The hail year has a second act. The season most people picture — towering supercells over the Plains in May and June — tapers through late summer. It does not stop. What changes in autumn is not whether the atmosphere can make hail but how it makes it, and a detection system that treats September as the off-season has misread the physics.
This piece is about both halves of that: the meteorology of fall hail, and the discipline a detection system owes its users once the cold air arrives — because the cold months are where an undisciplined system does its most expensive damage, one shrugged-off alert at a time.
Fall runs on different fuel
Summer hail is thermodynamic. A hot, humid boundary layer is the fuel: the surface heats, the air rises, and the updraft gets strong enough to hold a growing stone aloft. By October most of that surface fuel is spent — and the atmosphere compensates from above. The winds aloft sharpen as the hemisphere cools, cold fronts steepen, and the temperature contrast across a front starts doing the organizing that raw heat did in June.
Autumn hail is dynamic. Cold air arriving aloft steepens the temperature profile — the drop-off of temperature with height — so a storm needs less surface warmth to keep its updraft alive. Wind shear tilts the updraft so it does not choke on its own rain. And the freezing level comes down, so a stone that forms has less warm air to melt through on the way to the ground: hail that would have arrived as cold rain in July reaches the grass intact in October. The season does not end so much as migrate — south and east through the fall, riding the fronts.
None of this makes autumn a rerun of spring. Fall events are fewer, briefer, and more tied to fast-moving frontal lines than to slow supercells. But a roof does not grade a storm on seasonality — a stone the size of a quarter does the same damage in November that it does in May. The long tail is thin. It is not empty.
Most cold-air ice is not hail
Here is the trap in the second act: the same cold air that makes autumn hail possible fills the sky with things that resemble it and are not. Graupel — soft, rimed pellets that crush between two fingers — falls by the shower-load behind cold fronts. Sleet rides the shallow warm layers of early winter. Small soft ice below the damage line rattles on skylights and means nothing. On radar, all of it can wear hail's costume.
A detection system tuned for June that meets November has one seductive failure mode: fire on the look-alikes. Every one of those alerts is technically defensible — there was ice, it did fall — and every one of them teaches the person holding the phone the same lesson: this thing cries. By the time a real November storm arrives, the alert that matters gets the swipe that the graupel alerts earned. Cold-season discipline is not a refinement. On quiet cold-air days, silence is the product.
The cold-season record, correctly scoped
We have measured this, and the scope matters as much as the result. The record below is pooled across Nov 2025 – Feb 2026 at track grain — one radar object's life — and it was measured on an earlier Anvil detect generation, not the model now serving. It is a different window and a different probability scale from our main summer campaign, so it is labelled with its own window everywhere it appears and never averaged with the summer figures.
A cold-season catch rate means nothing without its price, so the chart prints both: the share of hail-producing truth tracks each system caught, and — as the half-weight track under each bar — the false fires each system spent per confirmed catch to get there.
Through the pooled cold months, the earlier generation caught 54.3% of the truth tracks — the largest share of the three systems — while holding the strongest precision and spending the fewest false fires per confirmed catch: 2.53 against 5.79 for ProbSevere v3 and 14.32 for MESH. That combination is the cold-season job description in one row: catch more of what is real, and pay less noise for it.
Holding silence on quiet ground
Two more rows speak to the discipline half directly. Inside our main summer campaign — measured on anvil-detect-5 — sits a cohort of 820 storm tracks whose radar presentation matches the winter-graupel profile: cold-air ice, no hail truth. On those tracks, anvil-detect-5 fired 0 times. So did both baselines. That row is a floor any serious system should clear, not a bragging right — which is exactly why we publish it as a floor.
The battery figures belong to anvil-nowcast-2, the address-level nowcast: before any generation of it ships, it replays 1,041 real storm-and-point cases from historical data through the production alerting path. Winter cases and clear days are in that battery precisely so that a model cannot buy its catch rate by chattering through the cold months.
What we have not measured
On a page like this, what we cannot tell you is as load-bearing as what we can. So, plainly:
The generation now serving has no cold-season record. Our main campaign window runs May 1, 2026 to July 30, 2026 — high summer. Inside it, cold-regime radar scans are 484 of 6.88M object-scans, roughly 0.01% of the window, carrying zero hail truth. The campaign artifact declares that slice void rather than printing a flattering zero over it — cool-season discipline simply cannot be measured from a summer window, at any threshold. The cold-season chart above belongs to an earlier Anvil detect generation; anvil-detect-6's own cold-season record will exist when a cold-season campaign has been run against it, and not before.
No cold-season outlook claim is available either. The multi-day outlook's verification panel runs March to July, so no cold-season claim is computable from it — and none is made here or anywhere else. A summer campaign cannot back an autumn promise, and this page does not try.
Before the season turns
The service half of this piece has nothing to do with software. Autumn is the last calm window of the hail year, and the highest-value work a property owner can do in it is record-keeping — because every later question about hail damage becomes a question about what the roof looked like before.
- Put the roof's age on paper. Installation year, material, and the last inspection date. Age drives how damage is assessed and what a repair is worth; a roof whose age is documented argues for itself.
- Photograph before the season turns. Date-stamped photos of each roof plane, the gutters, downspouts, flashing, and any skylights. This is the baseline every future claim is judged against, and it costs ten minutes.
- Read the soft metals after any storm. Gutters, downspouts, and window flashing dent at stone sizes shingles can hide. Dents in soft metal are a durable record that hail fell here, even when the roof looks fine from the ground.
- Log every storm, even the small ones. A dated note of what fell and roughly how big separates one storm's damage from a season's — the distinction that decides most disputed claims.
Notes and scope
- The cold-season arm. The Nov 2025 – Feb 2026 figures were measured on an earlier Anvil detect generation — a different window and a different probability scale from the main summer campaign, and not the generation now serving. They are labelled with their own window wherever they render and are never averaged with the summer figures.
- Pooled only. The cold-season window is quoted as one pooled figure because single cold-season months carry too few truth tracks to stand alone. No per-month cold-season number is publishable, and none is published.
- The winter-graupel row is a floor. All three systems hold silence on it, which is the point: it certifies a baseline of discipline, not an advantage. It was measured on anvil-detect-5 inside the main campaign window, over regime-classified tracks with no hail truth.
- The battery is offline. The nowcast battery replays real historical cases through the production alerting path, deterministically. It is not a measurement of live production behaviour.
- The void is deliberate. The main campaign's calendar-winter slice carries zero hail truth, so the artifact declares it void rather than publishing a rate over nothing. A zero that cannot be tested is not a result, and we do not print it as one.
- No forward promise. Every figure here is a scored record of a past window. Nothing on this page converts one into a promise about what the coming autumn will do, or about what any model will do in it.
- Named baselines. MESH and ProbSevere v3 belong to their respective operators and are named here as the measurement baselines the cold-season and regime figures are compared against — same window, same truth definition, same grain in each comparison. They are never inputs to Anvil, and no endorsement by their operators is implied.