How the scores work

Score based on Energy Use Intensity (EUI) in kBTU/sqft/yr. The base EUI comes from NREL ResStock 2024 simulation medians (a real modeled distribution over ~550,000 homes), looked up by the property's building type, climate zone, and vintage (the 2021 IECC map, by county; e.g. 4A mixed-humid, 5B cool-dry). Keying on building type gives multi-family (2–4 and 5+ units) and mobile / manufactured homes their own curves instead of scoring every dwelling off the single-family-detached one, and the per-zone-per-vintage medians capture the large difference between, say, humid 3A and dry 3B that a leading-digit scalar missed. (Zones ResStock doesn't cover, e.g. interior Alaska, fall back to detached, then to the prior curve.)

Within-cell adjustments

Building type, climate zone, and vintage pick the ResStock cell; off that base the estimate is nudged for the specific home. The foundation and heating/cooling (HVAC) nudges are themselves ResStock-derived: each is the within-cell, climate-controlled median-EUI ratio for that feature (e.g. a heat-pump home uses ~22% less than its cell median; a heated basement lowers per-sqft EUI). Square footage (size bin) and construction/exterior-wall type remain engineering factors.

Construction modifiers applied to the energy-use estimate: ICF = 0.92× and SIP = 0.95× (tighter, higher-mass envelopes), and Passive House = 0.55×. A home in a multi-unit building is scored off the measured multi-family EUI for its unit-count band, which captures the shared-wall / smaller-unit reality directly, so there is no separate modeled shared-wall credit. (In ResStock, large 5+ unit stock runs lower per-sqft EUI than detached, but small 2–4 unit stock can run higher, an intensity picture a flat credit couldn't represent.) The building type and unit count come from the detected structure (or what you enter). Rooftop solar is credited to the Environmental footprint's operational-carbon leg, not to the energy-use score.

Cost Estimation

Monthly cost uses the property's state residential utility rates (EIA averages for electricity and natural gas), so the estimate reflects local prices, roughly $0.10/kWh in the lowest-cost states to $0.20–$0.30/kWh in the highest. The electricity/gas split follows the home's heating-system type.

Score Mapping

A component-lifespan / effective-age model: how much usable service life the building has left, tempered by the assessor's observed condition. (Distinct from Disaster Resilience, which measures hazard loss rather than wear-out.)

Component Service-Life Basket

Eight major systems each carry a typical service life and a replacement-cost weight; the long-lived shell dominates and short-cycle systems contribute less:

Structural shell 100 yr (0.30) · plumbing 55 yr (0.10) · electrical 35 yr (0.10) · roof covering 25 yr (0.15) · windows 25 yr (0.08) · interior finishes 20 yr (0.10) · HVAC 18 yr (0.12) · water heater 12 yr (0.05).

Each system's remaining-life fraction is clamp((service_life − effective_age) / service_life, 0..1), using the assessor's effective year built (which folds in major renovations) rather than the original year. The age score is the weighted mean × 100.

Condition & Adjustments

Condition comes from the assessor's condition rating (Excellent = 100, Average = 60, Unsound = 0) and is weighted slightly above pure age because an inspector's observation captures real-world maintenance. The base is then scaled by exterior-wall material and by construction grade (clamped to 0.90–1.12×).

Homes with neither a build year nor a condition rating (vacant / non-residential land) are left unscored and excluded from the composite rather than guessed.

Multi-Unit Buildings

For a multi-unit building (detected by the National Structure Inventory, or entered with its material), the structural shell is a shared, building-level element rather than one house's wood frame. A reinforced-concrete or steel mid-rise frame, or a load-bearing masonry shell, is a fundamentally longer-lived building element, so the representative unit's shell decays more slowly: its service life is lengthened to 110–120 yr (concrete/steel 120, masonry 110) from the 100 yr wood-frame baseline. Only the shared shell changes; the shorter-cycle unit-level systems (roof covering, interior finishes, in-unit HVAC and water heater) keep their per-unit schedules, and a wood-framed multi-family keeps the baseline. ISO 15686 / CIRIA design service lives InterNACHI / Fannie Mae structural schedules

A life-cycle CO₂e score blends three components (0.50 operational + 0.30 embodied + 0.20 water), each normalized 0–100 against published good-vs-poor benchmarks (log-linear; higher score = lower footprint).

Operational

The energy a home consumes is valued at the grid average, the property's EPA eGRID2023 Rev 2 subregion emission rate (looked up nationally by county, US-average fallback), so a home on a coal-heavy grid carries more operational carbon than one on a clean grid. But the energy it avoids (rooftop solar + a high-performance envelope / passive-house efficiency, measured against the same home with a standard envelope and no solar) is credited at the long-run marginal rate: a kWh you no longer draw turns off the generation on the margin, not the average mix. That marginal rate is the county's NREL Cambium 2023 LRMER (Long-Run Marginal Emission Rate, mid-case, levelized over 20 years, Combustion CO₂e so it shares eGRID's stack-emissions basis), looked up via a bundled county→GEA-region crosswalk. Cambium's GEA regions are CONUS-only; outside them (Alaska, Hawai‘i, Puerto Rico) the marginal rate falls back to the average, so the credit term vanishes and the formula reduces to consumed_kWh × average, today's number. Energy use is the same model the Energy dimension produces; the gas factor is from the EPA GHG Emission Factors Hub.

Embodied & Water

Embodied: a material/size estimator keyed on exterior wall and construction grade, calibrated to the verified US single-family band of ~39–121 kg CO₂e/m², then amortized over the shell's expected service life (EN 15978 / RICS 60-year reference study period), so more durable construction spreads its upfront carbon over more years.

Water: EPA WaterSense usage benchmarks, with a national-average estimate for the energy embedded in water supply and wastewater treatment. Indoor use scales with occupancy and fixtures; outdoor use is modeled from the irrigable lot area (lot minus building footprint).

Multi-unit buildings: a unit in a stacked or attached multi-unit building (detected from the National Structure Inventory, or from a unit count you enter) carries no private-yard irrigation (any shared landscaping is common area, not the unit's own load), so its water footprint is indoor-only, which is why an apartment or condo unit scores greener on water than a detached home of the same size.

Homes with no living area (vacant / non-residential) are left unscored.

The resilience score uses an Expected Annual Loss (EAL) framework, the standard approach used by FEMA's National Risk Index, Hazus, and the insurance catastrophe-modeling industry.

Four Hazard Pillars

Disaster Resilience covers four perils: flood, tornado, earthquake, and fire. (It does not measure general wear-out or longevity; that is the separate Durability dimension.)

Flood: FEMA NFHL zone codes map to annual exceedance probabilities. Zone AE = 1% annual chance × 28% damage ratio = 0.28% EAL rate. Zone X (minimal) = 0.04% × 5% = 0.002% EAL rate.

Tornado: the location's FEMA National Risk Index tornado EAL rate (TRND_AFREQ × TRND_HLRB, expected annual building loss as a fraction of value), resolved at the census tract with county/national fallback from a bundled crosswalk, the same location-aware treatment as wildfire. This replaced the earlier SPC path-strike model so all four perils now ride one consistent expected-annual-loss basis.

Seismic: the USGS National Seismic Hazard Model supplies the local peak ground acceleration (PGA) for any U.S. address, run through simplified fragility curves. High-seismic regions (the West Coast, and the New Madrid zone in the Mid-South) get proportionally higher seismic loss than the low-hazard interior.

Fire: two contributions summed as one EAL rate. A structural/electrical base rate (≈0.02%/yr, from NFPA home-fire loss data) plus the location's wildfire EAL rate from the FEMA National Risk Index (WFIR_AFREQ × WFIR_HLRB), resolved at the census tract (county/national fallback) from a bundled national crosswalk. Both are scaled by a fire-specific modifier: wiring era (a continuous curve: 1950 knob-and-tube = 1.5×, 2002 = 1.0×, 2010+ post-AFCI/NEC 2002 = 0.85×), construction combustibility (wood frame = 1.10×, masonry/concrete = 0.80×), and condition (same factor as the other perils). Residential sprinklers cut the peril ~60% (0.40×) in the simulator. The wildfire term makes fire genuinely location-aware: near-zero across much of the country, but the dominant peril in the fire-prone West (e.g. Los Angeles County, “Very High”).

Building Resilience Modifier (BRM)

Raw EAL is adjusted by a construction-quality multiplier based on four factors:

Code era (year built): a continuous curve interpolated between code-milestone anchors: 1940 = 1.6× (balloon framing, no engineered connections), 1970 = 1.3×, 1990 = 1.1×, 2003 = 1.0×, 2010+ = 0.85× (modern IBC / ASCE 7). Interpolating rather than stepping means a 1969 and a 1970 home no longer differ by a sudden 0.2× jump; the curve clamps flat before 1940 and after 2010.
Construction type: Wood frame = 1.2×, ICF = 0.25× PCA racking data FEMA MAT Joplin/Moore
Foundation (flood only): Full basement = 1.4×, Slab = 0.7×
Condition: Unsound = 1.5×, Excellent = 0.8×
Fire uses its own wiring-era curve (also continuous): 1950 knob-and-tube = 1.5×, 2002 = 1.0×, 2010+ (post-AFCI/NEC 2002) = 0.85×.

There is no upper cap on the BRM, so old, poorly-built, and poor-condition stock compounds well above the code-current baseline (a pre-1940 unsound frame house exceeds 2.5×). Construction-type-specific floors still prevent over-crediting: wood frame floor = 0.50, ICF floor = 0.15 (85% max EAL reduction). The offline batch scorer and the live single-address model apply this identical BRM.

Multi-Unit Buildings

When a building is a multi-unit building, two things change. The building can be detected by the National Structure Inventory, or declared by entering its unit count together with its material and height (useful when detection misses a garden-apartment complex NSI models as single-family structures). First, the building's actual structural material (reinforced concrete or steel for a mid-rise, load-bearing masonry otherwise) drives the construction factor instead of the often-defaulted single-family type, because a concrete or steel frame is far more wind-, seismic-, and fire-resistant than wood. Second, flood exposure becomes floor-aware: flood damage concentrates on the lowest floors (FEMA P-259 depth-damage), so a representative unit averaged over the building's height carries roughly 1/stories of the ground-floor exposure, floored at 0.15 (lobbies, parking, and mechanicals never reach zero). A wood-framed multi-family keeps the single-family factors. FEMA Hazus building types FEMA P-259

Above-Code Feature Modifiers

20+ toggleable features with literature-backed multipliers:

Wind: Hurricane straps (0.70), hip roof (0.55), impact garage door (0.75), sealed roof deck (0.80), metal roof (0.75) IBHS research
FORTIFIED: Roof tier (0.35), Silver (0.25), Gold (0.20) 73–76% claim reduction, Hurricane Sally actuarial data
Seismic: Cripple wall bracing (0.45), seismic retrofit (0.75) PEER-CEA research
Flood: Elevation +1ft above BFE (0.15), +2ft (0.08), +3ft (0.04) FEMA depth-damage curves

Score Mapping

Total EAL rate maps to 0–100 via log-linear interpolation: score 100 at EAL ≤ 0.001%, score 80 at 0.02%, score 60 at 0.1%, score 40 at 0.3%, score 20 at 1%, score 0 at ≥ 2%.

Models the annual municipal cost to serve each parcel using a density-based cost allocation approach, inspired by the Halifax cost-of-sprawl study, Strong Towns fiscal analysis, and municipal density-budget research.

Per-Service Cost Model

Each service is split into a shared component that amortizes with density (dwelling units per acre) and a per-capita residual that does not. The shared, linear-network services (roads, water/sewer) fall along a continuous log-log cost curve: the published band costs placed at each band's geometric-mean density and extended through high-rise densities (to ~200 DU/acre), so per-household cost keeps amortizing as the same frontage and mains are shared across more homes rather than flooring at a small multiplex. Fire and sanitation also amortize: a large building is one address on one hydrant within a fire station's existing coverage, and a dense building uses shared collection (one compactor stop), so both fall with density to a per-capita floor (call volume, disposal tonnage). Parks stays flat (per-capita). The net effect: a 157-unit tower is no longer billed like a quadplex, and its per-unit cost-to-serve keeps falling until the per-capita floor dominates (~48 DU/acre and up).

Roads: ~$2,400/yr at <1 DU/ac → ~$400 at 12 → ~$150 at 48 → ~$60 at 200 DU/ac (floor)
Water/sewer: ~$1,500/yr at <1 DU/ac → ~$350 at 12 → ~$135 at 48 → ~$90 at 200 DU/ac (treatment is per-capita)
Fire/EMS: ~$800 base × distance modifier × density modifier (down to 0.60× for a high-rise in-district)
Police: ~$1,200 base × density modifier (most efficient at 16+ DU/ac)
Sanitation: ~$500/unit × density modifier (shared collection, down to 0.60×)
Parks: $300/unit

Local Calibration (per county)

These per-household figures set the cost-to-serve shape (how cost falls with density). The cost level is then scaled to each county's actual local-government spending using the U.S. Census of Governments: per-capita direct expenditure by function (roads, water/sewer, fire, police, sanitation, parks). So a county that spends 2× the reference per-capita rate on roads gets 2× the road cost. Los Angeles County, for example, runs ~2.0× on roads and ~2.6× on water/sewer (2022 census). Counties not in the crosswalk fall back to the national average.

Fiscal Ratio

A ratio > 1.0 = net contributor to the city. Both sides are localized per county and cover the same services, on a like-for-like non-school basis.

Cost side: Census of Governments municipal spending, schools excluded.

Revenue side: starts from each county's effective property-tax rate (median real-estate taxes ÷ median home value, U.S. Census ACS), then removes the school-district portion so it matches the school-excluded cost side. This is done two ways. In Texas the school tax an owner-occupier actually pays is computed from the county's school operating and debt rates and the $100,000 homestead exemption, and subtracted. Everywhere else it is estimated, by netting out the county-wide school share of property tax (~41% nationally; the national-average share is used where a county funds schools through its general government rather than a separate district).

The distinction matters because the ACS rate is measured over owner-occupied homes while the county-wide school share is measured over all property. Where a state gives owner-occupied homes school-specific relief, the ACS rate has already lost most of its school component and netting the share removes it a second time — understating municipal revenue, and so the score, for every home in that state. Texas is the largest such state at 9.2% of the population; Michigan, Arizona, South Carolina, South Dakota and Vermont carry the same distortion for a further 7.4% and still use the estimate. It runs one way, so those states are scored conservatively rather than unpredictably.

Why user fees are counted: the cost side includes water, sewer, and trash, but residents pay for those through utility bills and a monthly fee, not property tax. Measuring that cost against property tax alone compared unlike things and made every home look like a fiscal drain. So each service's modeled cost is multiplied by that county's actual fee-recovery rate — current-charges revenue ÷ direct expenditure, from the same Census of Governments file — and the result joins property tax in the numerator. Nationally, water/sewer recovers ~100% of its cost from charges and solid waste ~75%, while fire and police recover 0%: the Census classification has no current-charge code for either, so property tax is genuinely the only thing paying for them. Recovery is capped at 100%, so a utility running a surplus (Memphis's MLGW does) is credited at break-even and never above it.

The home value defaults to the county median when you don't enter one, so the revenue estimate reflects the local market. These remain county-level estimates (a county median, not a parcel's actual millage or value).

Score Mapping

These breakpoints are anchored to the national distribution of fiscal ratios (a population-weighted reference over U.S. counties × residential-density archetypes), so the score tracks national percentile rank (A = top ~20%, B = 60–80th, C = 40–60th, D = 20–40th, F = bottom ~20%). The national median fiscal ratio is ~0.69 — the typical U.S. home covers about two-thirds of its non-school municipal cost — so a typical home scores around 50 (C), and roughly 22% of homes clear 1.0. Treat the score as a rank, not a verdict: a home can score an A and still not fully pay its way, because most homes don't.

Density on this parcel (the density dividend)

Because the cost model is density-driven, the same lot scores very differently as a single-family home, a duplex, a triplex, or a quadplex. The density comparison holds the location and lot fixed and varies the number of dwelling units, keeping the per-unit value constant (so total value scales with units). As units increase, the same land and services are shared across more homes, so per-unit cost-to-serve falls and both the fiscal ratio and the Infrastructure Burden grade improve. That gap is the density dividend.

The per-unit fiscal ratio understates the case for infill, though, because the headline gain is on the revenue side. So the comparison also reports total revenue per acre (property tax + user fees, matching the cost-per-acre figure beside it). On a fixed Memphis lot at constant per-unit value, going from a single-family home to a quadplex raises it about 4.6×. The two legs move for different reasons:

Property tax per acre: ~6.4×. Four times the units, times another 1.6× because a parcel with 2+ rental units moves from the 25% residential assessment to the 40% commercial one (see Tax Classification of Rental Housing below). Outside Tennessee this leg is ~4×.
User fees per acre: ~2.0×. Fee revenue rides on modeled cost rather than on value, so it amortizes with density instead of scaling with units — which is correct: shared mains and one collection stop genuinely cost less to serve per home.

The net figure is the blunt one. On that same lot, net fiscal productivity goes from about −$6,900/acre as a single-family home to +$15,900/acre as a quadplex — from a net drain to a net contributor, on identical land and identical shared infrastructure. That is the lens that makes small-scale infill's productivity visible even when each individual unit still doesn't fully cover its cost. Try it with the “What if this parcel were denser?” button under the address search on the home and examples pages.

Multi-Unit Buildings

When a building is a multi-unit building (detected by the National Structure Inventory, or from a unit count you enter), the number of units drives the density calculation, so the building's shared land and services are amortized across its actual units rather than scored as a single detached home on the lot. The value side is also building-aware: instead of the single-family owner-occupied median (wrong for a rental building or condo), the per-unit value is an income-based “value-per-door” estimate, the way apartments are actually valued, from local rent: value_per_door = annual rent × occupancy × (1 − operating expense ratio) / cap rate. Rent is the county median gross rent (ACS B25064); occupancy (0.93), operating-expense ratio (0.40), and cap rate (~5.5%) are national constants. This is a neighborhood-average estimate, not an appraisal, so the fiscal and dollar figures remain approximate for a specific apartment or condo. ACS B25064 rent CBRE / Statista cap rate Census HVS occupancy

Tax Classification of Rental Housing

Some states put rental housing above a unit threshold into the commercial tax class, where it is assessed at a higher fraction of its value. Tennessee — the pilot state — does this in its constitution: residential property is assessed at 25% of value, “provided that residential property containing two (2) or more rental units is hereby defined as industrial and commercial property,” which is assessed at 40% (Tenn. Const. art. II, § 28; Tenn. Code Ann. § 67-5-501(11), § 67-5-801). The count that matters is rental units, not dwelling units: a single-family home rented long-term stays residential, and so does an owner-occupied duplex, because each contains only one rental unit (Tenn. Att'y Gen. Op. No. 25-016, Aug. 25, 2025).

So a Memphis apartment building generates 1.6× the property tax per dollar of value that a flat residential assessment implies — a large correction that had been missing for exactly the buildings the density model treats most favorably. A condominium building is unaffected: each unit is its own parcel containing at most one rental unit. Where tenure isn't known, a multi-unit building is assumed to be rental, which ACS 2024 (table B25032) supports for 86% of units in 2+ unit structures and 88% in 5+ unit structures.

How the correction is applied depends on which revenue basis is in play, and the two are never combined. Inside the pilot county the model applies a statutory assessment ratio against a statutory millage, so classification swaps the ratio outright (25% → 40%). Everywhere else the revenue side is an ACS effective rate measured over owner-occupied homes, which already carries the residential class in its denominator — so classification instead multiplies that rate by the ratio between the classes (×1.6). Both routes move the tax leg by the same factor; they simply start from different baselines.

What is not classification. Assessment caps and homestead exemptions — Florida's Save Our Homes, Texas's homestead cap, California's Proposition 13 — also open a large gap between owner-occupied and rental property, but they are keyed to ownership tenure rather than to a property class, and the ACS rate already embeds them for owner-occupied homes. The rental difference there is value-dependent rather than a fixed ratio, so encoding it as a class multiplier would over-correct. Those states are recorded as researched with no correction, which is deliberately distinct from not yet researched.

Coverage: 49 of 51 scorable jurisdictions (50 states + DC), 99.4% of the US population — all nine Census divisions have been worked through, and the two jurisdictions left out, the District of Columbia and Hawaii, are deferred rather than unexamined. Eight carry a correction: Alabama and West Virginia at 2.0×, New York City at 1.81×, Tennessee at 1.6×, Mississippi and South Carolina at 1.5×, Minnesota at 1.25× and North Dakota at 1.11×. Minnesota and North Dakota both reclassify at four units but count differently — Minnesota counts units held for rent, North Dakota counts units the structure accommodates — so an owner-occupied fourplex is commercial in one and residential in the other. Most classify by tenure rather than unit count, so in those states even a rented single-family home is taxed as commercial. West Virginia splits by tax rate rather than assessment ratio — the same economic effect by a different mechanism. New York is different again: the rule is set by the city rather than the state, applies only above 11 dwelling units, and its size comes from the city's own published effective-tax-rate study rather than from the statute — reading the statute alone gives 4.70×, which would over-correct by a factor of 2.6, because assessors value large rentals well below sales-based market value. The other forty-one are recorded as researched-with-no-correction, which is deliberately distinct from not yet researched — and two of those, Rhode Island and Connecticut, do classify rental housing, but per municipality in states whose counties are not governmental units, so the rule cannot be applied at the county granularity this model uses. Louisiana and Ohio are the instructive cases: both have a real two-class split, but it keys on use rather than tenure, so an apartment building sits in the same class as a detached house and no correction applies. Michigan is a third kind again — its owner-occupied exemption is large (18 mills) but falls entirely on the school levy this dimension already excludes from both sides. Four Mountain states are a fourth: Utah's 45% residential exemption, Montana's homestead rate and Colorado's residential rate all read as owner-occupied preferences in a secondary source, but each keys on how the home is occupied rather than on who owns it, and long-term rentals share the preferential treatment. The two deferrals are deliberate and differ in kind. The District of Columbia restructured its classes for tax year 2025 and sources conflict on where a multifamily rental lands, so the multiplier is genuinely ambiguous. Hawaii is the opposite problem: its four counties are the taxing units and their classes really do split an owner's principal residence from rented housing, but the implied corrections run from 1.97× to 3.56× — two of the four above the ceiling this table applies as a research-error tripwire, and Honolulu's is a value-tiered bracket above $1,000,000 rather than a fixed class ratio. Rental housing in those two jurisdictions is therefore still scored as though taxed like an owner-occupied home, which for Hawaii is known to understate it. Puerto Rico and the four smaller territories are out of scope entirely: they carry no Census of Governments data, so this dimension cannot score them at all.

Three layers, each mapped to a 0–100 sub-score against the national distribution of US census tracts and blended (PM2.5 0.45, ozone 0.25, radon 0.30; radon’s weight is redistributed for the ~0.2% of counties with no EPA zone):

• Fine particulate (PM2.5): the annual mean concentration (µg/m³) at the census tract from the CDC Environmental Public Health Tracking downscaler model (a monitor + CMAQ fusion with full coverage, including areas with no monitor). PM2.5 is the ambient pollutant most tightly tied to mortality; the US annual standard is 9 µg/m³, the WHO guideline 5.
• Ozone: the annual mean of the daily maximum 8-hour ozone (ppb) at the census tract, same CDC downscaler basis. Ground-level ozone drives respiratory harm and is elevated across much of the sunny, warm West.
• Radon: the EPA Map of Radon Zones class for the tract’s county (Zone 1 = highest predicted indoor level, ≥4 pCi/L; Zone 2 = 2–4; Zone 3 = <2). Radon is the leading cause of lung cancer among non-smokers. This is a county-level dataset (there is no finer public source), so it is the same across a county’s tracts.

PM2.5 and ozone are read at the tract (~84,000 US tracts), falling back to the county where a tract isn’t modeled. Because the breakpoints are anchored to national tract quantiles, the score reads directly as a national percentile (cleaner than N% of US tracts), comparable across locations. Data is bundled and versioned (CDC Tracking PM2.5/ozone 2021, tract-level; EPA radon zones), so the live scoring path needs no API key.

Uses the US DOT Bureau of Transportation Statistics National Transportation Noise Map (via the census-tract National Transportation Noise Exposure Map, Seto & Huang 2023), which models combined aviation + road + rail noise nationwide. For each census tract it gives the population exposed in dB bands (45–50, 50–60, 60–70, 70–80, 80–90, 90+ LAeq). The metric is the share of residents exposed to ≥ 60 dB, roughly the loudness of a busy restaurant, the level at which sustained transportation noise becomes a recognized nuisance and sleep/health concern. That share is mapped to a 0–100 score against the national distribution of US tracts (more exposure → lower score), so it reads directly as a national percentile of quiet, with a tract → county-mean fallback.

Uses the EPA National Walkability Index (NWI), a public-domain national index covering every US census block group, built from three built-environment measures: street-intersection density, proximity to transit, and diversity of land use. Its 1–20 index is scaled to 0–100 (higher = more walkable) and aggregated from block groups to the census tract (household-weighted).

Because the NWI is a national index (0 = car-dependent, 100 = highly walkable), the national grade is directly comparable across cities, and no per-address API call is made, because the value is a bundled, storable lookup. This replaces the Walk Score API, whose Terms of Use prohibit storing scores and whose free tier caps at ~5,000 calls/day.

Sub-county downscaled climate-hazard projections from the USGS CMIP6-LOCA2 threshold/extreme-event metrics: the Weighted Multi-Model Mean (an ensemble mean over CMIP6 LOCA2-downscaled models, ~6 km), sampled at each census tract's internal point (county = the mean of its tracts). Four hazard legs are blended equally into a 0–100 score (higher = less projected hazard): extreme heat (annual days > 95°F and > 100°F), heavy precipitation & flood (days > 1″ and the annual max 5-day total), drought (max consecutive dry days), and fire weather (the 95th-percentile Fire Weather Index from Argonne ClimRR). Each metric is scored against national breakpoints anchored to the SSP2-4.5 mid-century distribution — taken over census tracts weighted by the households in them, which is both the geography the dimension resolves at and the population the percentile claims to describe. Ranking tracts against a distribution of counties used to compress the extremes: county-averaging put the drought p95 at 56.8 consecutive dry days, pinning 15% of US households at a flat score of 0 with no way to tell them apart, where the household-weighted tract p95 is 148 days. The dimension is reported as a low/high band from SSP2-4.5 → SSP5-8.5 (a lower- vs higher-emissions future) at mid-century (2040–2069), with the SSP2-4.5 value as the headline.

The core number is the specific yield (the annual energy a standard 1 kWp rooftop array produces, in kWh per kW installed per year), modeled by the EU Joint Research Centre’s PVGIS v5.2 performance model on the PVGIS-NSRDB satellite database (the same NREL NSRDB solar resource PVWatts uses, covering the Americas). PVGIS is queried at the parcel itself for a building-mounted array at the optimal tilt facing south with 14% system losses; a bundled per-county table (each county queried at its Census-gazetteer internal point) is the fallback when the address is scored offline, sits outside PVGIS-NSRDB coverage, or PVGIS is unreachable. Querying the parcel matters because a single county can hold both ends of the range — in Los Angeles County the coast scores 90.3 and the high desert 99.3, against 80.7 at the county’s centroid. The yield is mapped to a 0–100 score against the household-weighted national distribution, so it reads directly as a national percentile (sunnier than N% of US homes). Weighting by households is what makes that sentence true: counties span five orders of magnitude in population, so their unweighted distribution is not the one US households live in. A sunny Southwest county (~1,700+ kWh/kWp) roughly doubles a cloudy Pacific-Northwest one (~950).

The drill-down scales that yield to a representative 6 kW residential array: annual production = yield × 6 kW; bill savings = production × the location’s residential electricity rate (EIA); and CO₂ avoided = production × the grid’s marginal emission rate (NREL Cambium 2023 LRMER, since rooftop solar displaces the generation that actually ramps down; eGRID average where the marginal rate is unavailable). This reuses the same rate and grid factors as the Energy and Environmental dimensions.

Uses the EPA Safe Drinking Water Information System (SDWIS) federal-reporting data. For each county it looks at every active community water system (the systems that serve year-round residents) and the share of the county’s served population on a system that had a health-based violation whose non-compliance period began within the trailing 5-year window. A health-based violation is a contaminant exceedance (lead, nitrate, arsenic, disinfection by-products, coliform…) or treatment-technique failure that can affect health, as opposed to a paperwork/monitoring lapse. That exposure share is mapped to a 0–100 score against the population-weighted national distribution (less exposure → higher score), so it reads directly as a national percentile of drinking-water safety.

Uses CDC PLACES census-tract-level health estimates. Seven measures: obesity, diabetes, high blood pressure, asthma, coronary heart disease, mental distress, and physical inactivity.

Each address is assigned to its census tract, and the health index is the average of the seven measures' inverted percentile ranks against the full national distribution of US census tracts (population-weighted). Lower prevalence = higher score. Because it is a national percentile rather than a within-county rank, a score means the same thing in every metro, so neighborhoods are comparable across cities. The reference distribution is bundled and versioned (CDC PLACES 2023).

Not available in Kentucky or Pennsylvania. CDC PLACES publishes none of these seven measures for either state — both appear only in the 2022 vintage and only for five unrelated prevention measures, and are absent from the 2023 release entirely. Puerto Rico has never been covered. In those places Health Impact is left unscored and the composite is the mean of the remaining dimensions, rather than filled with the national average — which would present an unmeasured neighborhood as an average one. Scoring those states on the five measures they do have would make the index mean something different there than everywhere else, defeating the national comparability above.

Census ACS 5-year estimates at the census tract level, blending five measures: poverty rate (B17001), median household income (B19013), housing-cost burden (B25106), educational attainment (B15003, the share of adults 25+ with a bachelor's degree or higher), and unemployment (B23025). Each measure is turned into a household-weighted national percentile and oriented so higher is better (income and education direct; poverty, cost burden, and unemployment inverted), then averaged, so a score is comparable across locations rather than ranked within the address's own county. Sourced from the keyless ACS 5-year Summary File (2020–2024), so the live scoring path needs no Census API key.