Every hail accuracy number you have ever read is graded against paper. Somewhere under each published table — ours included — sits a stack of filed ground reports: a person stood under a storm, watched ice come down, and wrote it up. That record is the industry's yardstick for hail truth, and it is the right one, because it is the only one a human being actually witnessed.
It is also spectacularly sparse. This piece measures how sparse, on our own campaign data, and walks through what that does to every accuracy figure anyone publishes — which numbers it quietly floors, which it leaves honest, and why we grade against it anyway.
The yardstick, exactly
Our measurement campaigns score against cross-source-deduplicated ground reports drawn from public storm-report networks — the same class of record the rest of the industry verifies on. Over the campaign window (MAY 1, 2026 → JUL 26, 2026), that record contains 25,015 deduplicated reports of hail three-quarters of an inch and up, attaching to 6,762 storm tracks — a track being one storm object's life on radar, beginning to end.
The scoring contract is mechanical. A system's fire counts as confirmed when a filed report lands within 15 kilometers and 10 minutes of the track. Everything else is booked as a false alarm — including a storm that dropped hail on an empty field at two in the morning, forty miles from the nearest person inclined to file paperwork about it. That storm counts against every model in the table, equally, no matter which of them was right.
So a false-alarm column in a report-scored table measures two things at once: the model, and the odds that a witness existed. Nobody's published precision number escapes this — the interesting question is the size of the effect. That question has a measured answer.
The measured gap
Take one cohort — the campaign's 1,744,182 storm tracks, the same window and grain as every published detection figure — and count three sets. Tracks where radar's own hail algorithm saw a hail signature: 21,186. Tracks with at least one filed ground report: 6,762. Tracks with both: 866.
Run the division and the gap has a multiple: radar's hail algorithm flagged 3.1 times the storm tracks that a filed report covers. Not a folk statistic, not a remembered round number — one cohort, two counts, and the arithmetic between them.
And the overlap cuts both ways. Filed reports cover 4.09% of the radar-signature tracks — but only 12.8% of the reported tracks showed the radar signature. Each record misses most of the other. Two instruments, both partial, disagreeing about a thing neither fully sees: that is what hail truth actually looks like before anyone builds an accuracy table on top of it.
Here is what the sparseness does to the tables. At its shipped bar, anvil-detect-5 fired on 8,352 tracks across this cohort — the leanest alarm budget of the three systems in our benchmark table, and still more tracks than the entire filed record contains. Both operational baselines fired on more still. Every system on the table is therefore guaranteed unconfirmed fires by arithmetic alone, before a single modeling mistake enters the ledger. A precision number scored on this instrument has a floor built into it, and the floor belongs to the instrument.
The same gap, drawn again
A number this consequential should not rest on one draw. A separate measurement in the same campaign — assembled independently, for a different question, over its own track population — counts 22,574 radar-signature tracks, 7,016 reported tracks, and 969 carrying both: filed reports cover 4.29% of what radar flagged.
Two cohorts, two coverage figures — 4.09% and 4.29% — kept strictly apart, and telling the same story: about four percent. The gap is not an artifact of how one cohort was cut.
Written down, roughly
Sparseness is half the instrument problem. The reports that do get filed are imprecise in both axes — time and place — because they are written by people, after the fact, from memory and a kitchen window. Our campaign measured the time axis directly, on its own report cohort:
The place axis has its own literature. A field campaign that drove hail swaths and measured them directly found that in more than half of the areas it sampled, the largest stone in the filed record was smaller than what the survey itself picked up nearby (Ortega 2018). The report record is not just thin — where it exists, it runs small and it runs rounded.
None of this is unique to our cohort. The peer-reviewed record measures the same gap, on different instruments and different units:
The comparison behind that figure is a radar product's diagnosed severe-hail hours held against the filed-report record — a gap between two records, not an error rate of either one. The paper's own reading cuts the same way this piece does: the authors attribute a substantial share of that gap to under-reporting, and argue the radar view is most useful precisely where reports are sparse. Our cohort's multiple lands inside the range they published, measured with different instruments, units and windows — which is what you would expect if the gap is a fact about the record rather than a quirk of anyone's pipeline.
Why we grade against reports anyway
After all of that, the tempting move is to promote the radar signature to truth — it covers more, it never sleeps, it does not round to the nearest five minutes. We don't, and the reason is the same one that opened this piece: a radar hail signature is an instrument's inference about the air aloft. Grade a model against it and you are grading one algorithm with another — a closed loop that can be wrong everywhere, confidently, with perfect scores. A filed report is the only record in this business that a person actually stood under. It stays the yardstick.
What you do about a sparse yardstick is not swap it for a flattering one. You measure its sparseness and publish it — this piece. You match against it with a tolerance shaped by its measured imprecision, and test that your comparisons survive other tolerances. And where the record allows it, you grade on the strictest slice it offers: storms that at least two independent observers reported, which is the cohort our headline detection figure in the Gen 1 report card stands on. What "corroborated" buys, exactly, and what it costs, is its own piece — coming soon in this series.
A sparse yardstick honestly described beats a generous one you built yourself.
Notes and scope
- One cohort per figure. The coverage share, the multiple, the overlap and the alarm-budget comparison are all computed inside a single campaign cohort at track grain. The second cohort renders only beside its own counts. The two coverage figures are two measurements and are never pooled or averaged.
- A radar hail signature is not ground truth. It is an instrument inference, and some signature tracks with no report are radar false signals rather than unwitnessed hail. The gap measured here bounds the disagreement between two partial records — it is not a census of unreported hail.
- Nothing here renumbers anything. Filed reports remain the primary truth definition behind every accuracy figure we publish. This page explains the instrument's floor; it does not swap any published number onto a looser truth definition.
- The report record. Cross-source-deduplicated ground reports from public storm-report networks, at the severe floor of three-quarters of an inch. A report can appear in more than one network, so deduplication precedes every count on this page.
- Named baseline. The cited literature measures MESH, which belongs to its operator and is named here only as the radar product those papers studied. Our own cohort's radar-signature set is a generic radar hail signature, not a claim about any named product, and no endorsement by any operator is implied.
- Literature figures belong to their papers. The cited range is the peer-reviewed gap between a radar record and a report record — not an error rate of either, and not a verdict on either. Our multiple is compared to it for shape, not corrected by it.
- Generation scoping. The alarm-budget count was measured on anvil-detect-5, the generation the campaign scored, at its shipped bar. Scope: Contiguous United States only.