Anyone who has driven the same road in morning traffic and again after midnight knows that is a compression. The question is whether the compression can be undone with measurement, and what falls out when it is.
What a risk index is, before anything else
A risk index is the number of driving events in a region, normalized by exposure. Exposure means actual observed driving, and it is the denominator underneath any rate. The available denominators are miles driven, hours driven, or cameras deployed in that region, computed at state, county or city level.
The events counted are inertial hard braking, cornering and acceleration, plus verified collisions that have been annotated as such. The view shows the last month by default, with month-by-month history behind it.
It exists because insurers and autonomous-vehicle teams asked for regional risk that changes, rather than the static factors that stay the same for years. Behind it are more than 100,000 cameras active every month.
Only we know
Risk can be calculated separately for daytime and for nighttime, using only the incidents and only the miles driven inside that window.
That is a small sentence with a large consequence. It is not a daytime figure adjusted by a nighttime coefficient, and it is not a model estimating what night probably looks like. Both the numerator and the denominator are re-measured inside the window, which means the two numbers are independent observations of the same place. A region can sit mid-table by day and move sharply at night, and seeing that requires having watched both.
Nobody can produce that split from claims history, because a claim records when a loss was paid and not how much driving happened in the hours around it. It comes from measuring the driving.
Below the city line
Aggregating to a city is still a compression, just a smaller one. The heat map renders risk on H3 hexagons, a standard grid that divides the map into equal hexagonal cells, rather than on Google's S2 cells. That puts a resolution below the state, county and city aggregates, showing which zones inside a city carry the most risk.
One honest note on the grid choice: H3 is the standard in use across these products, and the trade-off against alternatives is not something anyone here has argued from first principles. It is a convention, and it works.
The distinction that keeps the number honest
The index is a relative ranking of US regions against each other. It is not an absolute risk figure.
That governs every sentence anyone writes about it. A region ranked above another is riskier in this ordering. It is not "twice as dangerous", and it does not carry a probability that a given trip ends in a collision. The scoring approach is also described by the people who built it as a way to start, which is the correct level of confidence to hold about a first version.
Rate claims about small places deserve particular suspicion, including our own. A region with very few observed hours behind it can produce a dramatic number that one unusual month would erase. Before believing any ranking, ask what the denominator was and whether it is large enough that a single month could not have produced the result. Any county-level ranking published here clears a minimum observed-exposure floor for exactly that reason.
Two buyers, one dataset, different products
Autonomous-vehicle developers largely do not want the index. They want the exposure data underneath it: hours driven, miles driven, events by type. The reason is specific. They are trying to measure how autonomous driving compares with human driving, and to understand the conditions in a state before deploying a fleet into it. For that job the denominator is the deliverable and the ranking is noise.
Different profiles of driver are separable in the data by their signature rather than by any declaration. Rideshare driving looks like many short trips inside a confined area. Truck driving runs between interstates. Personal driving covers the fewest trips and hours.
Limits
The index ranks US regions. Geography resolves to state, county and city, plus the hexagonal heat map; there is no postcode, census tract or road segment level. Event types are currently limited to the inertial braking, cornering and acceleration set plus verified collisions. The refresh cadence of the underlying data is not something this piece states, because it is not settled. Day and night is the single environmental dimension implemented so far.
Measured, not imagined. Ask which hours the number was measured in.