Model v3.7 · August 2026
How we score the Bolthole Index
The Bolthole Index scores every U.S. county 0–100 as backup land: somewhere reachable, private, self-sufficient, administratively quiet, and cheap to hold for years. It is not a livability index — jobs, broadband, and nightlife score nothing here. Every input is public data or our own county-by-county legal research. The versioned model specification, production pipeline, and release validator are identified below; the machine-readable specification used by this release is available for download.
The six factors
Water — rainfall adequacy, drought history, groundwater depth, mapped perennial-stream density, and USGS watershed supply stress — dimmed by legal water access where a county sits in a regulated, over-appropriated, or adjudicated groundwater basin. Workable land combines soil productivity with the share and acreage at 15% slope or less and NRCS conventional-septic limitations. Usable woodland, energy independence, and working-farm normalcy complete the factor. These components use a continuous harmonic blend (p = -1) so poor workable land cannot be averaged away, without threshold caps.
GHSL population reachable within 30, 60, and 120 minutes by car from representative 5 km private-land cells, plus a small wooded-cover modifier. PAD-US 4.1 manager-type rasters exclude mapped federal, tribal, state, local, district, NGO, joint, and territorial land. Straight-line 15, 50, and 100-mile catchments remain published as a comparison and per-band fallback. Percentile-ranked nationally.
Current FEMA expected annual loss per dollar of exposure, drought history, and operating-mine/federal-land pressure (65%), plus a distinct mid-century layer (35%). The forward layer combines projected days over 100°F, annual precipitation, longest dry spell, extreme 1-day/5-day rainfall, drought loss, and wildfire loss under moderate-emissions RCP4.5 / SSP2-4.5 scenarios.
The county’s own scouted legal regime — residential building permit, county-wide zoning, and owner-builder exemption — sourced to official county and state pages (65%), plus state build/land/water rules (35%). Unknown county rules stay in the denominator at neutral, and weaker evidence shrinks toward neutral. A v2 preview is adding inspections, occupancy certificates, actual septic requirements, off-grid sanitation, dwelling/lot minimums, RV/tiny/manufactured-home occupancy, and utility-hookup mandates.
The current headline uses state homestead property-tax relief and the Cato Institute’s Personal Freedom percentile. A source-cited v2 preview is replacing that coarse proxy with actual creditor homestead protection, civil-forfeiture safeguards, tax-foreclosure and surplus-equity protection, eminent-domain limits, right-to-farm protection, self-defense law, and a much smaller general-freedom component.
Land price per acre and home values on saturating bands — cheap enough is cheap enough; cheaper adds little — plus state property and income tax burden and observed Treasury/FIO homeowners-insurance premiums, growth, nonrenewal, and paid-loss ratios. A single published ZIP is shrunk toward its state median; counties with no published ZIP use the state median and disclose that fallback.
How the composite works
- Weak links matter without scoring cliffs. The headline score is a weighted negative-power mean with p = -0.5. Low factors pull harder than they would in an arithmetic or geometric mean, but the formula is continuous and monotonic: improving any factor can never lower the result, and crossing an arbitrary threshold cannot suddenly change it by ten points.
- Percentiles for relative quantities, real-world bands for absolute ones. Population catchments, soil productivity, and stream density are ranked against the national county distribution. Rainfall, temperature, and land/home prices use saturating physical or dollar bands: once land is cheap enough, making it cheaper does not manufacture quality; 55 inches of Gulf humidity is likewise not better bolthole water than 38 inches in the Ozarks.
- Missing physical data never scores; unknown legal claims stay neutral. A missing physical reading drops out and its weight redistributes. Legal missingness is different and demonstrably non-random: an unknown zoning rule cannot turn one known no-permit fact into a perfect regime. Every expected legal field therefore stays in the denominator at a neutral 50. County pages report the rules and source material that are actually known.
- The scale is calibrated for legibility. Averaging six factors that trade off against each other (remote counties are rarely also wet, safe, and cheap) compresses raw composites into a narrow middle band. We stretch the final score linearly around the release anchor so differences are readable. The transform is 58 + 1.43× (raw − 49.596), bounded to 2–98. Rankings are unaffected — it is the same order on a wider ruler, and no county receives a fake 100.
- Cheapness is deliberately light (5%). Price is a filter, not a virtue — an index that rewards cheap land ends up measuring poverty, not refuge. Carry cost also saturates: once land is affordable, cheaper stops adding points.
- No politics. Election returns feed nothing. Legal reality — permits, zoning, water rights, taxes — feeds a lot.
Production formula · model v3.7
raw = (Σ wᵢ · max(xᵢ, 2)-0.5 / Σ wᵢ)1/-0.5
The coefficients total 100%. They are design inputs, not claims that each factor explains that percentage of observed rank variation. Distribution and the power mean also determine empirical influence.
Download the exact machine-readable model →Evidence strength
Every county publishes a separate 0–100 evidence-strength value. It uses fixed expected-source denominators, so a fallback, a partially rated soil survey, failed travel samples, or missing forward-climate components reduce the value instead of disappearing when the remaining inputs are reweighted. Evidence strength is not a score, probability, error bar, or promise that a source is correct; it tells you how directly and completely the model observed the inputs behind that score. County pages also show the value for each headline factor.
Code and release provenance
The scoring contract is src/lib/scoreModel.json. The production computation is implemented in scripts/merge_data.py, and scripts/validate_release.mjs independently recomputes every public county before a production build. The public download above is generated from that same tracked specification. We do not label the repository open source or link to a code host unless an unauthenticated visitor can actually reach it.
Data sources
| Source | Feeds | Vintage |
|---|---|---|
| County legal scout (first-party) | Building permits, zoning, owner-builder, septic authority for all 3,143 counties, cited to official county/state pages | 2026, ongoing |
| FEMA National Risk Index | Expected annual loss rate (hazard durability) | 2023 release |
| Climate Mapping for Resilience and Adaptation | County mid-century heat, annual precipitation, dry-spell, and extreme-rain projections under LOCA RCP4.5 | 2036–2065 projection period |
| FEMA Future Risk Index archive | SSP2-4.5 mid-century drought, extreme-heat, and wildfire multipliers applied to current NRI exposure-normalized loss | 30-year average centered on 2050 |
| U.S. Treasury Federal Insurance Office / NAIC | Published ZIP-level homeowners premiums, premium growth, nonrenewal, and paid-loss ratios, aggregated to county medians | 2018–2022 |
| NOAA nClimDiv county normals | Annual precipitation and mean temperature | 1991–2020 normals |
| U.S. Drought Monitor | Share of weeks in severe drought | 2010–2024 |
| USGS groundwater observations | Observed depth to water where monitored | latest year |
| USGS National Hydrography Dataset | Mapped perennial-stream density | Legacy national map; accessed Aug 2026 |
| USGS National Water Availability Assessment | HUC12 surface-water supply and use stress | 2010–2020 model, 2025 assessment |
| USGS national water-table model (Zell & Sanford 2020) | Modeled depth to water for every CONUS county; Nantucket uses a labeled nearest-county edge fallback | 1985–2015 average |
| Groundwater governance scout (first-party) | All 50 states plus DC reviewed for managed, over-appropriated, adjudicated, and critically-overdrafted basins and domestic-well rules; county records appear only where a special regime applies | 2026 |
| USDA NRCS SSURGO | Acreage-weighted NCCPI, representative slope, and septic absorption-field limitations | current Soil Data Access |
| USFS NLCD Tree Canopy Cover + LCMS | Mapped share of county area with at least 10% tree cover; LCMS supplies the full-Alaska layer | 2023 TCC; 2025 LCMS |
| USFS Forest Inventory & Analysis | Optional timberland and standing-biomass detail where county inventory estimates exist | 2015–2019 estimates |
| USDA Census of Agriculture | Farm counts, sizes, and land values; suppressed county values use a clearly labeled nearest-county proxy | 2022 |
| PAD-US 4.1 | Manager-type raster used to exclude mapped non-private analysis cells; county federal share retained for other factors | March 2025 release |
| European Commission JRC GHSL GHS-POP | Gridded population inside road-network and straight-line seclusion catchments | 2025 population epoch; R2023A release |
| OpenStreetMap / Geofabrik + Valhalla | Drivable road graph and 30/60/120-minute auto isochrones | August 2026 OSM extract; Valhalla 3.8.3 |
| Mine Safety and Health Administration | Current active/intermittent mines by county, type, and commodity | Weekly |
| ACS B25040 heating fuel + NREL solar summaries | Wood/propane heating share; solar resource (GHI) | ACS 2019–2023; NSRDB 1998–2009 |
| Zillow ZHVI + Census ACS B25077 | Typical home value; ACS median owner-occupied value fills counties Zillow omits, with four clearly labeled nearest-county fallbacks | Zillow monthly; ACS 2023 |
| Cato Institute — Freedom in the 50 States | Personal Freedom percentile | 2023 edition |
| Tax Foundation / Lincoln Institute | State property, income tax, homestead exemption tiers | 2024–2026 |
| Census population + gazetteer | County population, density, and nearest-metro context shown outside the score | 2023 estimates |
Known limitations
- Western water rights are now modeled, at county grain. A first-party groundwater-governance layer dims a county’s water self-sufficiency where it sits in a regulated, over-appropriated, adjudicated, or critically-overdrafted basin — cited to each state water agency, and deliberately household-first (a basin closed to new irrigation while a domestic well stays permit-exempt is only lightly dinged; the harsh factors are reserved for where a newcomer genuinely cannot get household water without buying an existing right). Two honest caveats remain: only counties inside a specifically mapped regime receive a sub-state adjustment, and the answer is county-level — basins follow the aquifer, not the county line, so the final call is always parcel-level. The national depth-to-water model is also a 1985–2015 average, so pumping depletion (Ogallala, Central Valley) still reads shallower than today. The water-rights hub maps it by state and county.
- County legal data is best-effort. Every stated rule cites an official source and samples are independently re-checked, but 3,143 counties change their rules constantly. Treat it as a scouting report, not legal advice, and verify locally before you buy.
- State-level inputs are uniform within a state. Homestead exemptions, tax tiers, and state-law factors can’t distinguish two counties in the same state; county-level inputs do that work.
- Workable land is a screen, not a parcel approval. The county estimate combines land at or below a 15% representative slope with NRCS conventional-septic ratings. Where SSURGO cannot rate a county, those physical inputs remain missing and evidence strength falls; missing is not converted into bad soil or a zero-workability claim. A survey, driveway plan, local setbacks, and a perc or site test still decide whether a particular parcel can be built on.
- Climate scenarios are planning cases, not forecasts. The forward layer uses county means for a moderate-emissions pathway and a mid-century holding period. It separates absolute future burden from projected change, but does not resolve parcel-scale topography, urban heat islands, drainage, vegetation, defensible space, or a particular well. The Southeast can score poorly on heat and extreme rain even where annual rainfall rises; the Mountain West can score poorly on dry spells, drought loss, and wildfire even where extreme rain also intensifies. Chugach and Copper River, Alaska postdate the CMRA county vintage and therefore use current-FIPS FEMA projections only; that narrower basis is labeled on their county pages. Where current FEMA wildfire exposure is explicitly zero, the forward wildfire score uses the published hazard-change multiplier instead of inventing a loss rate. An incomplete forward profile is shrunk toward neutral in proportion to component coverage, so one available hazard never masquerades as the entire climate layer.
- Insurance is observed market stress, not a quote. Treasury publishes a ZIP-year only when at least 10 insurers and 50 policies support it. The county metric is an unweighted median of published ZIPs. Two or more ZIPs produce a direct county median; a one-ZIP estimate is shrunk one-third local and two-thirds toward the median of well-covered counties in its state; a county with no published ZIP uses that state median. The basis is shown on county pages. The source excludes flood, residual-market, surplus-lines, and some smaller-insurer business; flood insurance and parcel eligibility still require separate diligence.
- Travel-time catchments are screening estimates, not parcel directions. GHSL distributes modeled 2025 population on a 1 km grid, aggregated here to 5 km cells. Valhalla routes over an August 2026 OpenStreetMap graph from representative PAD-US-eligible cells; county values are medians across successful samples. The model sees mapped drivable roads and ferries, but not guaranteed legal access, current closures, congestion during an emergency, private gates, or seasonal passability. Counties without a routable sample retain the published 15/50/100-mile radius fallback instead of receiving a favorable missing-data score. PAD-US is not a parcel-title system: unclassified cells are treated as potentially private and still require title, easement, water, and access diligence before purchase.
- Surface water is landscape-level. Perennial-stream mapping and watershed stress say whether dependable surface water is common nearby; they do not prove that a specific parcel has access or a legal right to withdraw it. Stream density is approximate because short reaches crossing a county edge count whole. We display it nationwide, but it affects the water score only when at least one other physical-water source is available.
- Puerto Rico is excluded — too many missing sources to score honestly.
Versions
v3.7 · Aug 2026 — replaced threshold caps with a continuous negative-power composite; put Python and TypeScript on one versioned model specification; made public scores exactly reproducible from precise axes; separated evidence strength from score; and stopped treating unrated SSURGO land as physically bad land.
v3.6 · Aug 2026 — added nationwide terrain and conventional-septic feasibility to workable land, using both suitable share and screened acreage so rugged Appalachian counties can no longer win on water and woodland alone.
v3.5 · Aug 2026 — added a distinct mid-century heat, water, extreme-rain, drought, and wildfire layer to Durability, and replaced its insurance proxy in Carry Cost with observed Treasury/FIO insurance-market outcomes where county coverage is sufficient.
v3.4 · Aug 2026 — replaced county population and threshold-city distance in Seclusion with cross-boundary population catchments, stopped unknown county legal fields from inflating scores, and added direct research on local administrative rules and state property autonomy. Legal additions remain out of the headline score until the national research pass is complete.
v3.3 · Aug 2026 — replaced stream-gauge presence with mapped perennial-stream density and watershed supply stress; replaced mineral occurrences and regulated oil/gas facilities with current MSHA operating mines.
v3.2 · Jul 2026 — sub-state groundwater governance: a legal water-access factor dims physical water where a county is in a managed / over-appropriated / adjudicated basin (cited to each state water agency, household-first), and a state domestic-well dimension joins the Admin factor.
v3 · Jul 2026 — six-factor weighted model; woodland/biomass added; precipitation saturation; state-law double-counting removed; fatal-flaw caps; median-anchored display calibration.
v2 · Jul 2026 — exposure-normalized hazard loss; percentile normalization; county legal layer wired into scoring; geometric-mean composite.
v1 · Jun 2026 — first public release, seven equal-weight sub-scores.