How every number in the atlas is produced

01Scope and refusals

The model answers one question: conditional on a given catastrophe occurring at a given severity, what is the probability that a prepared occupant of a given location survives the full arc of that event?

It does not estimate the probability that any catastrophe occurs. It is not a forecast, and the product says so on the page. Nothing here should be read as a claim about likelihood.

Three things the model deliberately refuses to do

02The stage-chain method

Every scenario resolves into three or four sequential, independently survivable stages. The composite index is their product, not a weighted sum.

This choice is the single most consequential design decision in the model, and it is why the atlas disagrees with almost every other survival map.

Independence assumption

Strict independence between stages does not hold. Fallout and famine share a cause; social collapse correlates with every preceding stage. The model treats them as independent, which biases composites slightly optimistic where stages are positively correlated in failure. Section 14 quantifies the direction. A copula-based correlation structure is the obvious next version and is not in v0.9.

03Shared physical layers

Four fields are computed once and reused by every scenario.

3.1 Population pressure

A Gaussian kernel over 120 metropolitan centers with populations from 0.24M (Reykjavik) to 37.2M (Tokyo). The 230 km bandwidth approximates the radius within which a metropolitan population can reach a location on foot within roughly a week, which is the horizon that matters for refugee pressure.

3.2 Strike node set

52 strategic facilities from unclassified public sources, plus 63 metropolitan areas above 4 million as countervalue targets. Tiering follows declared doctrine rather than speculation.

3.3 Continentality

Built at runtime rather than hardcoded. World coastline polygons are rasterised to a 288×144 grid, then a breadth-first distance transform propagates from every ocean cell. Continentality is distance to the nearest coast, normalized at 1,400 km.

This layer is what makes the model penalise Kazakhstan and central Canada and reward Tasmania and the Chatham Islands under any sunlight-reduction scenario, without any of those places being named in the code.

3.4 Protection factor

Fallout gamma attenuation through earth and rock. The tenth-value layer for mixed soil and concrete against fission-product gamma is approximately 0.32 m.

04Nuclear exchange Grade A

Four stages on four clocks: seconds, days, years, and as long as it takes.

4.1 Stage one — prompt blast and thermal

Overpressure radii use standard cube-root yield scaling from Glasstone and Dolan, calibrated against an optimal-height airburst of 1 Mt.

4.2 Stage two — fallout, first 96 hours

A downwind wedge model. For each tier-1 or tier-2 surface burst within range, dose accumulates according to distance along the plume axis and angular alignment with the wind.

450 rad is the approximate LD50/60 for acute whole-body gamma without medical support. The 180·√Y plume length reproduces the roughly 180 km extent of the 500-rad 96-hour contour for a 1 Mt surface burst under 24 km/h winds.

The decay rule the doctrine panel is built on

Fission-product activity follows an approximate t^−1.2 decay, which is the basis of the 7:10 rule: for every sevenfold increase in time since detonation, dose rate falls tenfold.

4.3 Stage three — crop failure over two years

The soot-injection famine layer, parameterised from Xia et al. 2022 in Nature Food, which modeled six stratospheric soot loadings and their effect on crop, fishery and livestock calories.

The atlas defaults to 90 Tg, between the regional and full-exchange cases, because a modern US–Russia exchange at current deployed inventories sits below the 150 Tg figure, which assumes larger arsenals than either state now fields.

4.4 Stage four — social collapse

The coefficients here are the softest in the model and are stated as judgment, not measurement. Section 14 says what would be needed to firm them up. Note the sign on D: population pressure is the dominant negative term, which is the model's way of encoding the finding that proximity to large unsupplied populations, not the absence of a weapon, is what makes a location untenable in year two.

05Single-city detonation Grade A

The most probable nuclear event and the one with the most divergent correct response. Three stages, no famine layer, no nuclear winter.

06Bioweapon Grade B and natural pandemic Grade A

Two separate models because the failure modes differ. The natural pandemic is calibrated on 1918 and 2020. The engineered case is an extrapolation and is graded accordingly.

The 2.2 multiplier on CFR in the natural case accounts for excess mortality beyond direct infection: care displacement, supply interruption, and untreated chronic conditions. It is calibrated so that a 2.5% CFR pathogen with R₀ 2.8 produces roughly the 1918 excess-death profile in a dense population, and it is the coefficient in this section I would most want independent review of.

07Grid collapse and EMP Grade B

Graded B because the historical analogues (Ukraine 2015, Texas 2021, Puerto Rico 2017) are all far smaller in scope than the modeled event, and scaling them is an assumption rather than a measurement.

The historical record the durations are anchored on

The 36-month ceiling on the duration slider is set by the upper bound of that last row, not by imagination.

08Carrington-class event Grade A

Graded A rather than B because the physical mechanism, geomagnetically induced currents in long conductors, is well characterised and the latitude dependence is measured rather than assumed.

The 38° onset and 68° saturation reflect the corrected geomagnetic latitude bands in the Lloyd's North American risk assessment, which places the highest exposure along the Washington to New York corridor, the upper Midwest, and comparable bands elsewhere.

09Supervolcano and impact winter Grade A

Shares the sunlight-reduction machinery with the nuclear famine layer, because the physical mechanism is the same: stratospheric aerosol loading that reduces insolation for years.

The 0.62 factor on stratospheric loading reflects that volcanic sulfate aerosol has a different optical depth and residence time per teragram than black carbon soot. Sources modeled: Yellowstone, Toba, Campi Flegrei, Taupo, Long Valley, Aira, Valles.

10Climate breakdown Grade A

The only scenario already underway, and therefore the only one a mortgage can be planned around.

The [|φ| > 40] bonus on agriculture encodes poleward migration of viable cropland. It is a coarse proxy for a process that is regionally specific and is the term most in need of replacement with actual gridded crop-suitability projections.

11The two Grade C models Speculative

AI takeover and alien contact are in the product because users ask for them. They are scored on generic obscurity and disconnection because nothing else exists to score them on, and both panels say so in their first line.

12Monte Carlo

1,500 runs per query. Two independent sources of uncertainty are sampled.

13Validation against published rankings

The only directly comparable published work is Boyd and Wilson 2023 in Risk Analysis, which applied crop models under nuclear-winter conditions to 38 island nations and ranked eight of them across eleven resilience dimensions.

The Iceland divergence, stated plainly

Boyd and Wilson rank Iceland third. The atlas ranks it mid-table under nuclear winter. This is a genuine methodological disagreement, not an error in either.

Their paper itself notes the limitation: "not all data available for all islands, average rank taken only across available data," with no predetermined weights. An unweighted average lets eight strong dimensions carry two fatal ones. That is the correct framing for asking which nation rebuilds fastest. It is the wrong framing for asking whether a specific household eats in year two.

14Known limitations

Ranked by how much they would change an answer.

14.1 The limitation that is a refusal, not a gap

The model does not simulate post-collapse armed conflict, faction formation, or territorial control, and will not. Three reasons, in order of weight.

The premise fails empirically. A 2025 systematic review in BMC Public Health screened 2,467 records and found that post-disaster looting is "very rare, and done by isolated individuals or very small groups," while 70 to 90% of affected populations report having heard looting stories. Sixty years of Disaster Research Center field studies find the same pattern: spontaneous prosocial behavior and emergent volunteer groups are the norm, and the documented deaths come disproportionately from authorities acting on the assumption of disorder rather than from disorder itself.

The modeling would be fabrication. There is no dataset. Any coefficient would be invented, and inventing coefficients and presenting them alongside Glasstone blast scaling would corrupt the grades that carry real content.

It would produce worse decisions. The same review found that high social capital correlates with reduced looting, and that weak state capacity plus low social capital is what predicts post-disaster crime. Aldrich's work across multiple disaster recoveries finds social capital to be a stronger predictor of household recovery than wealth or damage severity. A product that directs a user's preparation budget toward isolation and armament and away from the network of neighbors is directing them away from the best-evidenced survival factor available.

15Sources

Physical and epidemiological

Social and behavioral

Facility and geographic data