Storm case studies.

The 30% ProbHail reference is selected by Hail Sentinel for this comparison. It is not an official NOAA detection rule.

• This is an individual event illustration, not a population accuracy or false-alarm-rate measurement.
• Ground truth is a public NWS local storm report with estimated size, mirrored in IEM and SPC and counted once. Neither occurrence nor size is inferred from MESH.
• Native ProbSevere v3 boundaries, not nearest-centroid proxies, were tested against the report location. Its full object lineage was then retrieved without geographic restriction: 31 scans, 2026-08-16 23:18:41 through 2026-08-17 00:18:40 UTC; maximum ProbHail 21%.
• Anvil detection ENTER is 0.096083; the ProbHail comparison bar is 30%. These are different scales and operating points. Crossing a detector threshold does not establish that a phone notification was delivered.
• Anvil scores and physical textures were recorded live after model promotion. Images are those archived physical fields rendered by the current mobile contour shader; observations are not interpolated.
• MESH images use original NOAA grid cells, and displayed values are sampled at the report point. Object maxima and neighborhood maxima are different spatial quantities; a point underestimate is not proof the MESH system missed an entire storm.
• At 23:44:39, original MESH is 7.8 mm at the point, 13.9 mm in the 3×3 neighborhood and 32.4 mm in the 11×11 neighborhood. MESH did identify hail-sized values nearby and is credited accordingly.
• All three reports are estimated NWS local storm reports, counted once per event despite duplicate feeds. They match one tracked storm: each report lies inside its native footprint at the nearest scan (offsets +41, +39 and −21 seconds). At the selected 23:38:41 scan, all three locations also lie inside that storm and the same ProbSevere object. This selected scan is 2 minutes 41 seconds after the first report and 3 minutes 19 seconds before the second. Its Anvil median is 31.2 mm and same-object MESH maximum is 41.7 mm. These storm-level estimates and reported individual stone sizes are different quantities; this case does not establish population sizing accuracy. The point MESH readout remains anchored to the 23:45 report location, not to all three sites.
• Hail Sentinel selected this comparison threshold. The 30% ProbHail reference is used for this comparison; it is not an official ProbSevere alert threshold.
• Native contour colors use detection-gated MESH size classes, not the displayed trained-model size estimate. The separate size readout uses the storm-object hybrid product calculation.
• Archived Anvil sizing fields contain only the trained size model’s storm-object median (p50) and p90. NULL means that model did not emit a value. These raw model fields stay separate from the displayed hybrid product estimate.
• Anvil product sizing is reconstructed from archived inputs using the current production hybrid-size function: the same detection object’s maximum MESH, probability and any archived size quantiles. The gate is detection ENTER (0.096083), with a 40 mm MESH crossover. The default arm is a calibrated MESH band; the larger-hail arm uses the trained size model. This is a storm-object median, not a point prediction or largest stone. The upper estimate is p95 for calibrated bands and p90 for the trained model, not a guaranteed bound. It is not an archived delivered notification, and the MESH-derived arm is not independent evidence of a learned sizing advantage.

The 30% ProbHail reference is selected by Hail Sentinel for this comparison. It is not an official NOAA detection rule.

• This is an individual event illustration, not a population accuracy or false-alarm-rate measurement.
• Ground truth is a public NWS local storm report with estimated size, mirrored in IEM and SPC and counted once. Neither occurrence nor size is inferred from MESH.
• Native ProbSevere v3 boundaries, not nearest-centroid proxies, were tested against the report location. Its full object lineage was then retrieved without geographic restriction: 43 scans, 2026-09-12 20:10:42 through 2026-09-12 21:34:40 UTC; maximum ProbHail 21%.
• Anvil detection ENTER is 0.096083; the ProbHail comparison bar is 30%. These are different scales and operating points. Crossing a detector threshold does not establish that a phone notification was delivered.
• Anvil scores and physical textures were recorded live after model promotion. Images are those archived physical fields rendered by the current mobile contour shader; observations are not interpolated.
• MESH images use original NOAA grid cells, and displayed values are sampled at the report point. Object maxima and neighborhood maxima are different spatial quantities; a point underestimate is not proof the MESH system missed an entire storm.
• At the 21:10:39 report-time scan, original MESH is 18.3 mm at the point and throughout the sampled 11×11 neighborhood. Other overlapping storm objects have larger maxima at other times; this is not a blanket MESH miss.
• Hail Sentinel selected this comparison threshold. The 30% ProbHail reference is used for this comparison; it is not an official ProbSevere alert threshold.
• Native contour colors use detection-gated MESH size classes, not the displayed trained-model size estimate. The separate size readout uses the storm-object hybrid product calculation.
• Archived Anvil sizing fields contain only the trained size model’s storm-object median (p50) and p90. NULL means that model did not emit a value. These raw model fields stay separate from the displayed hybrid product estimate.
• Anvil product sizing is reconstructed from archived inputs using the current production hybrid-size function: the same detection object’s maximum MESH, probability and any archived size quantiles. The gate is detection ENTER (0.096083), with a 40 mm MESH crossover. The default arm is a calibrated MESH band; the larger-hail arm uses the trained size model. This is a storm-object median, not a point prediction or largest stone. The upper estimate is p95 for calibrated bands and p90 for the trained model, not a guaranteed bound. It is not an archived delivered notification, and the MESH-derived arm is not independent evidence of a learned sizing advantage.

The 30% ProbHail reference is selected by Hail Sentinel for this comparison. It is not an official NOAA detection rule.

• The ground size is a trained spotter’s estimate, mirrored in IEM and SPC and counted once.
• At 22:32:42 UTC, the native tracked footprint covers the report point. Its centroid is 1.09 km away; its stored live Anvil score is 0.126114 versus detection ENTER 0.096083. This does not establish that an alert was delivered.
• MESH values are sampled from the original NOAA grid at the report coordinate. At 22:32:42 the nearest pixel is 12.8 mm, 3×3 maximum 16.9 mm and 11×11 maximum 24.7 mm. MESH is an estimate; the spotter report supplies event truth.
• ProbSevere v3’s own polygon covers the report. It had crossed the 30% comparison threshold before the report: 40% at 22:20, 31% at 22:26, 24% at 22:32. This is not a ProbSevere miss.
• Anvil imagery uses archived physical textures rendered by the current mobile contour shader. Available observed images are 22:26:41 and 22:36:40 only. The missing 22:32:42 Anvil image must be shown as unavailable; never hold or interpolate another image while displaying that timestamp.
• All three products have different purposes and scales. This selected illustration is not a population accuracy or notification-performance measurement.
• Hail Sentinel selected this comparison threshold. The 30% ProbHail reference is used for this comparison; it is not an official ProbSevere alert threshold.
• Native contour colors use detection-gated MESH size classes, not the displayed trained-model size estimate. The separate size readout uses the storm-object hybrid product calculation.
• Archived Anvil sizing fields contain only the trained size model’s storm-object median (p50) and p90. NULL means that model did not emit a value. These raw model fields stay separate from the displayed hybrid product estimate.
• Anvil product sizing is reconstructed from archived inputs using the current production hybrid-size function: the same detection object’s maximum MESH, probability and any archived size quantiles. The gate is detection ENTER (0.096083), with a 40 mm MESH crossover. The default arm is a calibrated MESH band; the larger-hail arm uses the trained size model. This is a storm-object median, not a point prediction or largest stone. The upper estimate is p95 for calibrated bands and p90 for the trained model, not a guaranteed bound. It is not an archived delivered notification, and the MESH-derived arm is not independent evidence of a learned sizing advantage.