HailScore v7 and a Complete NOAA Radar Archive: Why Some Colorado Scores Moved

On July 31, 2026 two changes shipped to HailScore™ on the same day: the scoring engine moved to v7, and a national backfill finished repairing a radar archive that had been quietly incomplete since it was first built. The 0–100 scale and the tier bands did not change, but many scores did — some down, some up. If you looked up a Colorado address before and after that date and saw a different number, this explains why.

The short version: v7 stopped over-crediting storms that were far from your property, and the backfill recovered 2,792,263 NOAA radar records we should have had all along. The full entry is on our public changelog.

What v7 actually changed

The most consequential change is how recency is weighted.

In v6, a storm three miles away counted almost as much toward your recency score as one directly overhead. That sounds minor, but it had a flattening effect: because the Front Range gets so much storm traffic, nearly every property ended up with similar recency points, and a channel that was supposed to separate addresses stopped separating them. v7 weights recency by distance, so that channel discriminates again.

The practical result is that scores driven mostly by far-off storms went down. That is not a downgrade of your property — it is the number reflecting the distinction it was always supposed to make.

Three other changes landed with it:

Radar readings are rolled up to one event per storm-day per roughly 0.7-mile cell, so a single physical storm can no longer be counted many times over.

Live MRMS radar is fused into every scoring surface. The free report, bulk lookup, the embed, and the v1 API now share one storm builder instead of three separate ones that could drift apart.

Roof age and material are bounded and disclosed. Self-reported details move the score within a published range, they are never a cheat code, and the report now states what was assumed when you leave a field blank.

Every score also carries explicit data-quality flags now. If a feed is degraded, or our archive is thin for your area, the report says so — rather than quietly showing you a reassuring low number. That last one matters more than it sounds, and the next section is why.

The archive was incomplete, and our old test said it wasn't

Here is the uncomfortable part, published for the same reason we publish everything else.

When the national radar archive was first built, the loader predated a guard we added later. Large-area fetches silently returned partial data instead of failing — the request came back looking successful, just short. Nothing errored, so nothing got flagged.

The scale: 40 of 48 states were affected. In the worst cases, more than 90% of a state's records were missing.

We found it with a new test that compares our row count against NOAA's for the same area and time window. Our old coverage check had been grading a state "complete" if its data spanned enough years — which is exactly the kind of check that feels rigorous and proves nothing. North Dakota passed that check while missing 88% of its records. Both tests now ship in our tooling and run against production.

The repair added 2,792,263 rows across all 48 states. It was append-only: zero rows were modified, zero were deleted. You can see current coverage on our data coverage page.

What this means for a Colorado address

Colorado was not among the worst-hit states — the Front Range is one of the most heavily observed hail corridors in the country, and our coverage here was in better shape than in the states that lost 90% of their records. But Colorado scores still moved, for two reasons pulling in opposite directions:

v7's distance-weighted recency pulled some scores down. This is the dominant effect along the Front Range precisely because the region gets so many storms. When a metro sees hail somewhere nearly every week in season, "there was a recent storm within a few miles" stops being informative. v7 asks a sharper question: how close, and how big?

Recovered radar pulled some scores up, wherever our archive had been thin for a particular area.

If you want to see the underlying activity rather than just the score, the Colorado hail statistics page tracks live statewide counts, and each address lookup shows the individual storm records behind its own number.

What did not change

The 0–100 scale is the same.

Tier bands are unchanged: 0–25 Low, 26–50 Moderate, 51–75 High, 76–100 Very High.

Scores stored before July 31, 2026 keep the engine version that produced them. An older saved report is not silently rewritten under a new engine — it stays what it was, labeled with what produced it.

The score still measures documented hail exposure near an address, not damage. It is not a damage diagnosis and not an insurance decision. Whether a roof is actually damaged is a question for a physical inspection by a qualified roofer or adjuster.

HailScore is built on 34.4M+ hail records across government sources, including 8.6M+ NOAA NEXRAD radar signatures, plus the NOAA Storm Events Database, SPC Local Storm Reports, NWS active alerts, CoCoRaHS volunteer observations, and live MRMS radar. Those totals are machine-derived from live production counts and move as the archive grows. (For historical context: the July 31 changelog entry recorded the public floor moving from 25.1M+ to 29.6M+ as of that ship.)

Why we wrote this down

A scoring change that moves numbers without explanation is how a data product loses the only thing it really sells. The methodology behind the score is public on our methodology page, the mechanics are in how HailScore works, and every shipped change lands on the changelog — including the ones, like this archive gap, that are not flattering.

If your Colorado address showed a different number after July 31, it was one of these two changes. Look it up again and the report will show you the storm records it used.