BACKDROP · ILLUSTRATION
Licensed data CURRENT 2026.11.03

Data changelog.

Every release of the licensed hail and wind record, and what changed in it.

Licensed hail and wind data DATA VERSION 2026.11.03

How data versions work

  • A data version names one release of the archive (YYYY.MM.DD) and its three components: the archive build (date of the last rewrite of historical hail data), the model generation scoring new radar scans, and the segmentation version for new scans.
  • Every API response carries X-Data-Version; bulk exports and GET /v1/verify/coverage carry the full stamp.
  • A historical rewrite ships as a new release, listed here before or when it is served. Work still running is listed as in progress and is not part of the release that lists it.
  • Live answers read the current archive. To hold a fixed copy, use a bulk export: its file never changes, and an export read from a release snapshot can be regenerated byte for byte. See the methodology.

Releases

2026.11.03 Storm dates 2026-04-27 to 2026-07-31: detections the current model makes added

CURRENT
RELEASED
2026-11-03
ARCHIVE BUILD
2026-11-03
MODEL
anvil-detect-6
SEGMENTATION
seg-2
AUDITED THROUGH
2026-10-23
SNAPSHOT
none
  • REPAIRED hail On storm dates 2026-04-27 to 2026-07-31 the record now holds the detections the current model, anvil-detect-6, makes from the stored storm cells where the models in service at the time wrote none. 6.36 million detections at or above the detection threshold were added, in 24,768 storm cells on 89 storm dates: 27% more detections and 32% more storm cells than that window held before. Of storm-day map cells of about 5 km with a detection in the window, 27% are new (85,047 of 315,181). Detections that already existed are unchanged, and the confidence alerted at the time is still kept as the alert record and is not returned. Measured on storm dates 2026-06-06 to 06-12, the share of hail reports under hail-size radar echoes that have a detection within 5 miles rose from 47% to 90% (93% is typical for June). Answers on these dates can change, mostly from no hail found to hail found. 2026-08-01 to 08-04 needed no additions.
  • CHANGED hail Correction to release 2026.10.25: which detections exist on storm dates 2026-04-27 to 2026-08-04 is no longer limited to what the earlier models' screen recorded. Every storm date in that window now measures complete at the current model's detection threshold.

2026.11.01 Disclosure: which method sets a hail size class

RELEASED
2026-11-01
ARCHIVE BUILD
2026-10-26
MODEL
anvil-detect-6
SEGMENTATION
seg-2
AUDITED THROUGH
2026-10-23
SNAPSHOT
none
  • ADDED api Documentation and one additive field; no stored data or existing value changes. Hail size classes and calibrated sizes (size_category, ml_predicted_size_mm, max_size_mm) are set by one of two methods: the trained size model's median, or empirical radar-to-report size bands. Below a radar size estimate of about 40 mm the bands are always used. At and above it the model is used where it produced a size for the cell, and the bands everywhere else. The model's size is missing for most of the archive before July 2026. Measured on the licensed record for cells of about 5 km per storm day with a radar estimate of at least 40 mm: the model's size is present on roughly half of cells from March to September (40-65% by month) and under a quarter in most October to February months for storm days from 2022-07-21 to 2026-03-31, on under 6% for 2026-04-01 to 2026-06-30, and on about 99% from storm day 2026-07-01. Archive sizes are not rescored. The methodology page explains how to read a size class by era.
  • ADDED api size_method (model or radar_band, null when there is no size) is returned beside the size it describes: GET /v1/hail/current (cells[].hail.size_method, summary.max_size_method), POST /v1/hail/current/batch (results[].max_size_method), GET /v1/hail/history (events[].hail.size_method) and GET /v1/portfolio/impacts and its webhook (impacts[].max_size_method). Bulk exports and GET /v1/storms are unchanged and carry no size_method.

2026.10.31 Every hail output at the detection threshold

RELEASED
2026-10-31
ARCHIVE BUILD
2026-10-26
MODEL
anvil-detect-6
SEGMENTATION
seg-2
AUDITED THROUGH
2026-10-23
SNAPSHOT
none
  • CHANGED api Every hail output now serves only detections at the model's event-start confidence: the detection threshold of the model generation that scored each detection (anvil-detect-6: 0.096), the test GET /v1/hail/swaths and GET /v1/events adopted in release 2026.10.27. It now also applies to bulk exports (detections and swaths, including exports pinned to a release snapshot), verification at every grain (POST /v1/verify, /v1/verify/area, /v1/verify/portfolio and the reports built from them), GET /v1/hail/history (events at any min_confidence, and daily_peaks), GET /v1/hail/current and POST /v1/hail/current/batch, portfolio impacts and the portfolio impact webhook, and the assistant. 15.1% of stored detections (41.1 million of 272.7 million) are below the threshold and are no longer returned by any output. Of storm-day map cells of about 5 km, 8.5% (252,605 of 2,964,024) lose every detection and 7.5% keep a lower peak; of cells whose peak estimate was 19 mm or more, 8.7% (72,289 of 830,318) no longer reach 19 mm. Whole-US monthly detections exports shrink by 7.2% (May 2024), 9.9% (June 2023) and 21.1% (July 2026).
  • REPAIRED hail A recurring radar artifact over east-central Nevada, around Ely, which the model scored below the detection threshold on every detection (on most days under 1%), is no longer served by any output; release 2026.10.27 had removed it from swaths and events only. In that area 66 storm days had held detections; the 7 summer storm days the model did call hail (up to 26%) remain. Address history there no longer reports daily peaks of 136.2 mm (2024-11-21) or 35.7 mm (2025-01-29), and an export of that area for 2025-01-29 returns no rows instead of 16,058.
  • CHANGED api GET /v1/hail/history daily_peaks[].detection_count counts only detections at the threshold. below_bar_detection_count, below_bar_estimated_size_mm and the possible_signal outcome are unchanged: they already counted only those. Regenerating a detections or swaths export from release 2026.10.26's snapshot now returns only the at-threshold rows; a file generated earlier keeps its stored copy until it expires.

2026.10.30 Ground storm reports: relayed copies of corrected reports removed

RELEASED
2026-10-30
ARCHIVE BUILD
2026-10-26
MODEL
anvil-detect-6
SEGMENTATION
seg-2
AUDITED THROUGH
2026-10-23
SNAPSHOT
none
  • REPAIRED hail SPC storm-report copies of a Local Storm Report that the issuing office later corrected are removed when the corrected report stands within 0.1 degrees and 4 hours of the copy. Once the superseded original was removed in release 2026.10.28, such a copy no longer matched any Local Storm Report and was counted as a second report of the same event beside the correction. 160 hail copies were removed (129 storm days, 2020-02 to 2026-09). Copies with no corrected report nearby are kept, and future SPC updates no longer re-add these copies.
  • REPAIRED wind 547 SPC wind copies were removed the same way (255 storm days, 2022-07 to 2026-10), 514 of them damage reports without a measured speed.

2026.10.29 Wind swaths: United States only, and cold-season summit gusts removed

RELEASED
2026-10-29
ARCHIVE BUILD
2026-10-25
MODEL
anvil-detect-6
SEGMENTATION
seg-2
AUDITED THROUGH
2026-10-23
SNAPSHOT
none
  • REPAIRED wind Radar-indicated wind areas no longer extend over Canada, Mexico or open water. Before this release 377 of 6,000 radar-indicated polygons (206 storm days) lay wholly outside the United States, 129 of them (78 days, 895,621 km2) wholly north of 49 degrees N, so an address in Canada could be given a radar-indicated severe band. The layer is now cut to the contiguous-US boundary the other licensed outputs use; no polygon lies wholly outside it. 519 storm days were rebuilt (2022-12 to 2026-10).
  • REPAIRED wind Winter mountain-summit and ridge-top gusts filed as thunderstorm wind no longer paint gust-report areas. From November to March, a report whose remarks place the station at 6,000 ft or higher, or on a ridge top or summit, is excluded: 12 reports on 10 storm days, among them 146 mph at Snowbasin, UT (8,999 ft) on 2022-12-31 and 104 mph at Ogden Peak, UT (9,570 ft) on 2022-12-30, which had each set the 80 mph band. Summer mountain thunderstorm gusts are unchanged. Winter gusts at ridge-named stations that publish no elevation are kept.
  • CHANGED api GET /v1/wind/swaths is clipped to the United States on every layer, including the default URMA analysis, and refuses a point or box wholly outside it with 422 outside_service_area (not charged). Responses carry spatial_clip. Its analysis note now states the full-archive capture rates: of 20,648 measured severe thunderstorm gusts, 36.1% fall inside any band and 5.1% inside the 58 mph band (earlier text quoted a smaller first sample). at_point.evidence.nearest_gust_report is deterministic: reports at the same distance resolve to the higher gust, and the summit reports above are excluded from it.

2026.10.28 Ground storm reports: NWS corrections applied, and duplicates and mislabels cleaned

RELEASED
2026-10-28
ARCHIVE BUILD
2026-10-25
MODEL
anvil-detect-6
SEGMENTATION
seg-2
AUDITED THROUGH
2026-10-23
SNAPSHOT
none
  • REPAIRED hail When the National Weather Service corrects a Local Storm Report, the corrected report now replaces the original instead of standing beside it. 609 superseded hail reports and 130 relayed copies that still carried a superseded size were removed (2020-01 to 2026-10), among them a 7.0 in report that the issuing office had corrected to 0.7 in. SPC storm-report hail sizes are now labelled method unknown: that feed does not say whether a size was measured or estimated, and every one of its 55,457 reports had been labelled estimated. SPC hail reports for storm dates 2026-05-20 (14) and 2026-06-27 (39), missing from the archive, were restored.
  • REPAIRED wind 1,966 superseded wind reports and 66 relayed copies carrying a superseded speed were removed the same way. 10 reports typed as gusts under 10 mph (snowfall totals and remarks filed under the wrong type) and 6 reports whose coordinates contradict their own location were removed. 81 marine reports whose state field held a marine-area code (Lake Superior read as Louisiana, Atlantic and Pacific coastal waters) now carry the state they lie in, or none offshore.
  • CHANGED api Ground-report de-duplication now treats an SPC report and the Local Storm Report it relays as one report when they agree on place and size or speed but differ in time by up to 3 hours, and a cluster holding an NWS correction reports the corrected value. A wind report's type now prefers the thunderstorm report when one report was filed under two types. GET /v1/ground-reports is clipped to the United States and its long-record aggregate states the years actually covered (window_years, window_start, archive_start; the ground-report archive starts 2020-01-01). Verification report counts no longer include app-user submissions or no-hail entries.

2026.10.27 Swaths and events at the detection threshold, and one model in portfolio impacts

RELEASED
2026-10-27
ARCHIVE BUILD
2026-10-25
MODEL
anvil-detect-6
SEGMENTATION
seg-2
AUDITED THROUGH
2026-10-23
SNAPSHOT
none
  • CHANGED api GET /v1/hail/swaths and GET /v1/events use only detections the model called hail: a detection scored below the detection threshold of the model generation that scored it (anvil-detect-6: 0.096) is no longer drawn or counted. This is the same per-detection test GET /v1/hail/history applies before a below-confidence detection counts toward possible_signal. Across the archive it removes 91,920 of 872,500 storm-day map cells (about 5 km) with an estimated size of 19 mm or more (10.5%), including every cell whose highest model confidence was under 1%: 34,130 cells on 650 storm days, most of them a recurring radar artifact over east-central Nevada that no ground report supports. 3,915 of the removed cells had a ground hail report within 10 km that day.
  • CHANGED api GET /v1/events cell_count counts distinct tracked storms. It counted per-scan storm labels, which repeat from scan to scan and undercounted busy days: 2024-05-28 counted at most 375 and now counts 820.
  • CHANGED api GET /v1/hail/current and POST /v1/hail/current/batch are clipped to the contiguous United States and use only detections that carry a model probability, like every other hail output. From 2026-08-05 to 2026-10-24, 4.2% of the detections they read lay outside the United States.
  • CHANGED api GET /v1/storms and GET /v1/storms/{track_id} return each storm-cell observation once. 411 stored duplicates (2026-06 to 2026-09) had been returned twice.
  • CHANGED hail Portfolio impacts (and the portfolio impact webhook) now report anvil-detect-6 confidence on storm dates 2026-04-27 to 2026-08-04, like every other output, instead of the confidence stored for each row.
  • REPAIRED hail Correction to releases 2026.10.24 and 2026.10.25: on 14 storm dates between 2026-05-13 and 2026-08-04 the stored detections had been regenerated with anvil-detect-6 (2026-07-27 to 2026-08-04 entirely, part of 2026-05-13, 05-21, 05-22, 07-12 and 07-26). On those days the stored rows did not keep the model in service when they were first written, and portfolio impacts already reported anvil-detect-6 there. The detections as originally served are preserved separately and are not returned by any output.

2026.10.26 Storm-cell fragment recovery complete

RELEASED
2026-10-26
ARCHIVE BUILD
2026-10-25
MODEL
anvil-detect-6
SEGMENTATION
seg-2
AUDITED THROUGH
2026-10-23
SNAPSHOT
storm days through 2026-10-23
  • REPAIRED hail Storm cells that an earlier segmentation defect stored as fragments were re-derived on 18,957 radar scans across 368 storm dates (2022-07 to 2026-04), and their detections regenerated. The recovery covered every scan with a hidden fragment of at least 260 grid cells on a storm date with a ground hail report of 1 inch or larger. Recovered fragments now appear in every output; the fragments still excluded fell from 229,583 to 187,452. Smaller fragments, fragments on other storm dates, and storm dates from 2026-04-27 on stay excluded from every output, as before.
  • AUDITED hail Re-audit on 2026-10-25 over 1,556 storm days (2022-07-21 to 2026-10-23): no gap days, no duplicate rows, no rows without a model score; 99.9% of radar scans that held a hail-capable storm hold hail detections (249,424 of 249,621). The other 197 scans hold no storm cell with an extractable outline.

2026.10.25 One detection model across the archive, and ground reports in address history

RELEASED
2026-10-25
ARCHIVE BUILD
2026-10-24
MODEL
anvil-detect-6
SEGMENTATION
seg-2
AUDITED THROUGH
2026-10-19
SNAPSHOT
none
  • CHANGED hail Address history, swaths, verification and exports now report the current detection model's confidence, anvil-detect-6, on storm dates 2026-04-27 to 2026-08-04 too, so those outputs read the whole archive at one model. Portfolio impacts (and the portfolio impact webhook) still report the confidence recorded when each row was written. Those rows were re-scored from the same storm cells; the confidence alerted at the time is kept as the alert record and is no longer returned. Which detections exist on those dates is unchanged: it is still what the earlier models' screen recorded, and on May–June 2026 the current model flags about 25% more storm scans than that screen kept. Answers on those dates can change where a detection now passes or misses the confidence bar.
  • ADDED api GET /v1/hail/history lists vetted ground hail reports (NWS storm reports, Local Storm Reports and CoCoRaHS observers) inside the radius on each storm day as ground_reports, with size, distance, time and source; one report relayed by two feeds is counted once. A window with such a report and no radar event answers outcome: report_signal, and radar_outcome gives the radar-only answer. A report is a person's observation, never a radar measurement. In the October acceptance cases, all 15 report locations radar missed now answer report_signal, and none of 50 cases without a report does.

2026.10.24 Storm-cell segmentation update and the last October repairs

RELEASED
2026-10-24
ARCHIVE BUILD
2026-10-23
MODEL
anvil-detect-6
SEGMENTATION
seg-2
AUDITED THROUGH
2026-10-19
SNAPSHOT
none
  • CHANGED hail From about 01:46 UTC on 2026-10-24, new radar scans are divided into storm cells by segmentation seg-2: a cell is split wherever its echoes are separated by more than a few empty grid cells, so distant specks are no longer merged into one large cell. Scans before that time keep seg-1 and were not re-run. Whether hail is detected is rarely affected: in a replay of two active storm windows, the same hail reports were detected at the detection threshold before and after the change (435 of 443). Storm-cell size, shape, centroid and track can differ across the change, so do not compare cell-level statistics from before and after it as one population. A full re-segmentation of the archive is planned with the next detection-model generation.
  • REPAIRED hail Storm-cell outlines were re-derived for a further 15,424 radar scans (2022-07 to 2026-04) whose cells had no usable outline, and their detections regenerated.
  • REPAIRED hail Hours missing from the storm-cell record were filled: 3,035 radar scans across 101 hours on storm dates between May 2023 and January 2025, with their detections regenerated.
  • REPAIRED hail Storm dates 2026-05-01 to 2026-10-19 were regenerated to fill radar-metadata columns. Each day keeps the detection model that was in service when it was first written.
  • REPAIRED wind The thunderstorm wind layer was rebuilt on the repaired storm-cell record. Every storm day from 2022-07-21 to 2026-10-19 now has complete radar input; before the rebuild, 189 days answered insufficient_coverage.
  • CHANGED api GET /v1/hail/history date windows are storm days (12:00 to 12:00 UTC), as coverage and daily_peaks always were. Every detection count counts a radar pixel once per scan. The 51 days repaired in October (2023-06-23, 2024-03-27, 2024-04-13, 2024-05-18..19, 2024-05-21..31 and 2026-04-30..06-03) are reported as covered.
  • ADDED api outcome: possible_signal on GET /v1/hail/history: no event, but the radius holds detections the model flagged below the confidence bar.
  • IN PROGRESS hail Recovery of storm cells that an earlier segmentation defect stored as fragments, on storm dates up to 2026-04-26, is running. It ships as a new data version; until then those fragments stay excluded from every output.

2026.10.21 October 2026 archive repair, re-audited

RELEASED
2026-10-21
ARCHIVE BUILD
2026-10-21
MODEL
anvil-detect-6
SEGMENTATION
seg-1
AUDITED THROUGH
2026-10-19
SNAPSHOT
none
  • REPAIRED hail The gap in the radar hail record from 2024-06-05 to 2026-04-26 was filled. The record is continuous from 2022-07-21.
  • REPAIRED hail Storm dates 2022-07-21 to 2026-04-26 were scored with the current detection model, anvil-detect-6. That model was trained on reports through May 2026, so accuracy measured on these dates is not independent validation.
  • REPAIRED hail 4,745,496 exact duplicate detection rows were removed, and 1,999,231 rows on 2026-04-27..30 that carried no model score were replaced with scored rows.
  • REPAIRED hail Storm-cell outlines were re-derived for 53,513 radar scans in which segmentation had produced one oversized cell, and 83 storm dates with no storm-cell records were re-run; their detections were regenerated.
  • AUDITED hail Re-audit on 2026-10-21 over 1,552 storm days (2022-07-21 to 2026-10-19): no gap days, no duplicate rows, no rows without a model score; 99.9% of radar scans that held a hail-capable storm hold hail detections.

Segmentation versions

VersionRadar scansWhat it does
seg-2 from 2026-10-24 01:46 UTC seg-1 plus a cluster split: each cell is split by connected group of echoes, so echoes separated by more than a few empty grid cells become separate cells.
seg-1 from the start of the record until 2026-10-24 01:46 UTC Watershed segmentation of the radar hail field into storm cells, with re-splitting of oversized cells. Every scan before 2026-10-24 01:46 UTC, including the October 2026 re-derivations.

Earlier data changes

Before data versions were stamped on responses.

  • CHANGED 2026-10-20 Every hail output is clipped to the contiguous United States and uses only rows that carry a hail-detection model probability. Responses name the public detection model behind every row they read (ml_provenance). The thunderstorm wind layer (GET /v1/wind/swaths?layer=convective|all) is available.
  • REPAIRED 2026-10-17 A storm-cell labelling defect that stored some cells as small fragments was fixed for new scans from 01:54 UTC. The 229,586 fragment cell records written before the fix (2022-07 to 2026-10-17) are excluded from every output.
  • CHANGED 2026-08-05 Detection model anvil-detect-6 in service for new radar scans.
  • CHANGED 2026-04-27 Storm dates 2026-04-27 to 2026-08-04 were written with the detection model, and its threshold, that was in service at the time; that model flagged about 25% fewer storm scans than the current one on May–June 2026. Since data version 2026.10.25 address history, swaths, verification and exports report the current model's confidence on those dates (portfolio impacts keep the recorded confidence); which detections exist is still what the earlier screen recorded. ml_provenance names the model on every row.