How a reportis made

A Refugia report is about one point, not a town or a postcode. You drop a pin on the map, search for an address or paste coordinates, and everything that follows is measured at those exact coordinates and in rings around them: the ground under the pin, the next two kilometres, the next fifty.

The first step takes under a second. The point is reverse geocoded with OpenStreetMap’s Nominatim service to find its country, region and nearest locality. The country matters more than it looks: it decides which regional sources apply. A pin in Florida brings in the FEMA flood map; a pin in the Alps does not, and leans on the global flood model instead.

Refugia does not hold one big database. It holds a catalogue of models, and each model is a small manifest: which signals it produces, where on Earth it is valid, how authoritative it is, how fine its resolution is, how old its data is and what its licence allows. A raster model reads one pixel from a cloud-hosted map. A live service model calls an API. A calculator derives new numbers from other models, such as terrain prominence from the elevation model.

All models that cover the point run in parallel and report signals: elevation in metres, river flood depth in a 100-year flood, distance to the nearest hospital, solar energy per square metre per day. Today the catalogue holds 31 models, and the methodology page lists every one with its licence and its weight.

Because models are only manifests, adding a better source is a configuration change, not a rewrite. When a new national flood map is added, every report in that country improves on its next run; when a source is removed, reports keep working and their confidence drops where it was needed.

A signal on its own means nothing to a reader: 4.6 kilowatt-hours per square metre a day is good for solar panels, but you should not have to know that. Each criterion in a report reads one or more signals through a curve that turns the measurement into a score from 0 to 100. The curves are written down in the criteria files and are the same for every place on Earth, so two reports always compare fairly.

Between the marked points the score runs in a straight line. Most criteria read several signals. Food production, for example, reads the frost-free season, yearly rain, whether there is soil, the share of cropland within two kilometres and the longest dry spell, each with its own curve and its own weight.

Scores are then given a grade, with the same bands everywhere:

When two models measure the same thing, Refugia does not pick one. It blends them, and each one’s share of the result is its weight. The weight of a model is worked out from its manifest, with no hand tuning:

• Authority. Official agencies count 1.0, academic datasets 0.9, commercial 0.85, values Refugia derives 0.8, community data such as OpenStreetMap 0.6 and web research by AI agents 0.45.
• Resolution. A 30-metre grid counts in full, a 1-kilometre grid about 0.6 and a 10-kilometre grid about a third.
• Age. Each year since the data was made takes 4% off, down to half.
• Specificity. A source built for one country gets a quarter on top of a global one of the same quality.

These four numbers are multiplied, then multiplied again by the criterion’s own weight for that signal and by how sure the model was of its reading at this point. A criterion may override a model’s weight where experience says it should, and that override is written in the criteria file where anyone can read it.

Every criterion carries a confidence next to its score. It answers a plain question: how much of the evidence this criterion asks for is actually here, and do the sources agree?

• Coverage is the share of the criterion's signals that were measured, counted by weight, so missing the main signal costs far more than missing a minor one.
• Evidence is how much total weight stands behind the score.
• Disagreement lowers confidence: when two models measuring the same thing land far apart, the score is less certain, and the report says so.

Two rules keep weak evidence from steering the result. A criterion below 30% confidence is shown, with its evidence, but does not move its category score. And a category in which fewer than 40% of criteria could be scored stays ungraded instead of being graded on a handful of numbers.

Criteria are grouped into categories, and categories into layers. Each layer answers a different question about the same point.

Today

What is it like to live here now? Acquisition, access to services, community, self-sufficiency and environment, including floods, earthquakes, volcanoes and wildfire.

Grid down

What holds up if the power goes out for good? Energy you can make yourself, water and food without supply chains, your position relative to cities and strategic sites, and how easily a newcomer fits in. These are factors set by geography that preparation alone cannot change.

Scenarios

Optional layers model specific catastrophes under stated parameters: a large nuclear exchange, and two contested theories of a rapid crust shift, ECDO and a solar micronova. Each scores the point during the event and after it. A scenario is a model of a theory, not a forecast, and the report keeps it apart from the overall score.

Some things no dataset holds: what property costs nearby, how deep the water table is, how reliable the local grid has been, how much crime there is. For these, five research agents search the web in parallel, each on its own subject: property and acquisition, community and safety, land and water, community resilience, and healthcare, grid and connectivity.

Every figure an agent returns comes with the source it was found in, and the report links to it. Agent findings carry the lowest authority in the blend, 0.45, so they fill gaps without overruling a measured dataset.

Last, an editor reads the scored evidence and writes the summary, the strengths and the concerns. The editor writes about the numbers; it does not change them.

People choosing a place to fall back to rarely want it published. A report on your exact coordinates is private to your account and is never shown to anyone else.

What joins the public map is a generalised summary of the surrounding area: category scores for a hexagonal cell of about five square kilometres, with no exact point. That shared library is what lets anyone compare regions before buying a report of their own.

If you watch a place, Refugia checks live hazard feeds from USGS, GDACS, NASA EONET and NOAA around it every ten minutes, and on the Steward plan also tells you when its scores change.