How the Clariva Safety Score Works

The Clariva Safety Score is a single number, 0 to 100, describing conditions around a specific address. It combines seven factors drawn from public government data: violent crime, non-violent crime, sex offender proximity, flood risk, property conditions, Superfund sites, and air quality. Each is scored independently, then weighted into one composite.

The weighting is an editorial judgment, not a statistical model. We weight toward personal safety. Violent crime, non-violent crime, and sex offender proximity together carry 70% of the score, because those are the conditions people are actually asking about when they ask whether a place is safe.

Everything below describes exactly how each number is produced, what it measures, and where it falls short.

Each factor is scored 0 to 100 on its own curve. A curve maps a raw measurement, say 4.5 violent incidents per 1,000 residents per year, onto a score. The curves are calibrated to the range of conditions that actually occur in the cities we cover, not to a national distribution or a normal curve.

The score is not a percentile. A 70 does not mean "better than 70% of addresses." Our curves are built so that ordinary urban conditions score well above 50, because ordinary urban conditions are livable, and a scale that called them failing would be useless. An address scoring 72 on violent crime is in a normal Philadelphia neighborhood, not a bad one.

Each factor also carries a band, a short label describing the condition itself: Very Low, Low, Moderate, Elevated, High, Severe. The band describes the measurement; the score describes where that measurement falls on the curve. They are two views of the same thing, which is why a factor can read "Elevated" and still score in the 60s. Elevated property crime is a real finding, and it is also survivable.

Some factors use their own vocabulary where the generic ladder would say something false. Property Conditions runs from Well Kept to Severe, because "Low property conditions" would mean the opposite of what we measure. Flood risk runs Minimal to Severe, plus Undetermined for zones FEMA has not studied. Air quality uses EPA's own words (Good, Moderate, Unhealthy for Sensitive Groups, Unhealthy, Very Unhealthy) because it is EPA's index, and inventing our own labels for it would only cause confusion.

The composite uses a different ladder: Excellent, Good, Average, Below Average, High Concern. It describes a score rather than a condition.

We count reported incidents, not convictions or arrests. Every crime figure on a Clariva report comes from the police department's own incident data: what was reported to police and recorded. Unreported crime is invisible to us, as it is to any data source built on police records.

Violent crime means the FBI's violent crime index: homicide, rape, robbery, and aggravated assault. This is a narrower definition than most people use in conversation. It is the standard used for national comparison, which is why we use it.

Simple assault is counted as non-violent. This surprises people, and it is worth being direct about. The FBI's index classifies simple assault, an assault without a weapon and without serious injury, outside the violent crime category. We follow that classification rather than inventing our own, because a score that used a private definition of "violent" could not be compared to anything. Simple assault is counted, scored, and weighted. It sits in the non-violent factor.

We score the two separately, and never blend them. Non-violent crime is roughly ten times more common than violent crime in most urban neighborhoods. A single combined "crime score" would be dominated by property offenses, and an address with high theft and almost no violence would look the same as one with the reverse. They are different questions, so they get different numbers.

Crime data lags. Police incident data is published on a delay, and incidents are sometimes reclassified after the fact. A report generated today reflects the most recent complete data available, which is not the same as the most recent week.

The trend chart measures something slightly different from the score. The score uses the past 365 days as a rate per 1,000 residents. The chart shows monthly incident counts over 36 months: raw counts, not rates, because a rate computed against today's population and a three-year-old incident count is a number with no honest meaning. The chart answers "which direction," the score answers "how much."

Most factors are expressed per 1,000 residents. A raw count is not comparable between a dense block and a quiet one. Twenty incidents among 300 people is a different place than twenty among 8,000. Dividing by population makes two addresses comparable.

The population figure is an estimate. We use U.S. Census American Community Survey 5-year estimates, apportioned to the circle around the address. ACS estimates are themselves survey-based, published with margins of error, and updated annually. They are the best available figure at this geography, and they are not a census count.

This is the score's most significant distortion, and it affects commercial areas. The denominator counts people who live in the circle. It does not count people who work there, shop there, commute through, or go out at night. A busy commercial corridor can have a small residential population and a large number of incidents involving people who do not live there, which produces a high per-resident rate that overstates the risk to someone living on that block. The opposite is also true: a purely residential neighborhood's rate is a fair reflection of conditions, because the people counted and the people present are the same people.

We flag this directly in the report when a non-violent crime reading is elevated, because commercial corridors are where it bites hardest. But it applies to every per-resident rate on the page.

Very small populations are excluded rather than scored. If the residential population inside a factor's radius falls below 500, the rate becomes unstable, and a single incident can swing it by a wide margin. In that case we do not score the factor. It is excluded, the remaining weights are redistributed, and the report says so.

Offender counts are approximate by source. Our sex offender data comes from Offenders.io, a commercial provider that aggregates public registry information and reports counts as approximate figures. We use them as published, rounded. We do not receive and do not display names, addresses, or any identifying detail.

Two different things can prevent a factor from being scored, and we treat them differently.

Some factors have no answer by design. FEMA Zone D means the area has not been studied: its flood risk is undetermined, not low. A population below 500 in the violation radius makes a per-resident rate meaningless. In these cases there is nothing to score, and scoring anything would be an invention. The factor is excluded, the remaining weights are redistributed proportionally, and the report states which factor was dropped and why.

Some factors fail to retrieve. A government data source can be slow, down, or return an unexpected response. This is a temporary condition, not a fact about the address. When it happens, we do not substitute a guess and we do not quietly drop the factor. We withhold the composite score and mark the report pending. We retry, and when the data arrives the report completes and you are notified. If it still has not resolved after 24 hours, we finalize the report with the factor excluded and disclose that on the face of it.

The distinction matters. A redistributed weight changes what the composite means, and you are entitled to know whether it was redistributed because the world has no answer or because our request failed.

We never write a zero for missing data. A zero is a measurement: it means we looked and found nothing. Absence of data is not a measurement, and it never renders as one.

Different sources refresh on different schedules, and we cache to keep reports fast. What that means in practice:

Crime and property conditions are queried live for every address report. The figures are as current as the city's published data.

Flood zone and Superfund records are cached up to 90 days. Both change rarely. A FEMA remap or a new listing is a matter of years, not weeks.

Sex offender counts are cached up to 30 days.

Air quality is a 30-day rolling average, updated nightly.

The trend chart is updated monthly, with the right edge set to the last complete month. Because crime data is published on a lag, the most recent month shown is generally two months behind the current date.

A report is a snapshot. It reflects conditions at the moment it was generated and is not updated afterward. Sources refresh on their own schedules and conditions change; for current figures, generate a new report.

Upstream agencies publish on their own timelines, which we do not control. When a source changes its format or its availability, we adapt as promptly as we can.

It is not a prediction. Every figure describes what has already been recorded. Nothing here forecasts what will happen on a block, and no score should be read as a probability of anything happening to you.

It is not a substitute for looking. Data cannot tell you whether a street feels right, whether the neighbors are good, or whether the corner you would walk past at night is one you would want to. Visit at different hours. This report is a starting point for that, not a replacement.

It is not address-specific in every dimension. Flood zone is genuinely parcel-level. Crime, violations, and offender counts describe a circle around the address, and conditions can differ materially from one end of that circle to the other. Block-by-block variation inside it is invisible to us. The trend chart is measured at the neighborhood's center, not at the address.

It does not include everything that matters. Schools, transit, noise, traffic, housing costs, and the quality of city services all affect whether a place is right for you, and none of them are in this score. We chose seven factors and weighted them toward physical safety. That is a judgment, and reasonable people would weight differently.

The offender data comes from a commercial aggregator. We query Offenders.io rather than state registries directly, and we cannot independently verify which registries it draws from or how current its records are. We report its counts as approximate because that is how they are published.

Superfund counts include sites at every stage. A facility assessed and cleared in 1998 counts the same as an active cleanup. Proximity matters as much as count, and the EPA ECHO database lists each site's contaminant profile and current status, worth checking directly if a count here concerns you.

Scores are algorithmic estimates, not professional assessments. They are not an inspection, an appraisal, an environmental assessment, or legal advice, and they should not be the only basis for a decision about where to live or what to buy.

Our full terms and disclaimer are on the legal page.