METHODOLOGY 1.1

The math behind the questionable judgment.

Reproducible arithmetic. Editorial assumptions. No divine peer review.

The Apocalypse Meter currently reads 64%, from an unrounded 64.424508. Every number on this page is generated from edition edition-2026-11-01-r2 and the same engine the site uses. The code is the source of truth, so where it differs from the original brief, this page follows the code and says so.

Notation and categories

Five kinds of number

• Raw factual measurements come from named sources, in their original units.
• Editorial transformations are our anchors, rubrics and reference periods. They turn facts into game scores.
• Fictional catalog points are invented and describe nothing real.
• The aggregate percentage combines the above, and only exists inside this game.
• User-created scenarios are game inputs, kept in your browser and your link.

The model

For each component, a transformation produces a score. A headline is then a weighted mean of the scores that are eligible.

Reproducibility and validity are different things. Anyone can reproduce this number exactly, and that says nothing about whether it measures anything. The aggregate is a satirical editorial construction, not an extinction probability. A satirical score, not a scientific prediction. Real sources. Questionable species.

Display rounding

The engine rounds to nine decimals to remove floating-point noise, then rounds half up.

Applying this, 64.424508 displays as 64. The status text comes from the displayed integer.

The eight components

Two sections are open by default, and the rest expand.

Aggregation and missing data

Configured weights are stored as 15, 15, 10, 10, 10, 15, 10, 15. The engine divides by their sum, so they act as fractions that sum to one. They are an editorial decision, not an estimate. With every component eligible:

When a component is unavailable, its weight is redistributed in proportion to the others.

• Coverage is not confidence. $C_t$ is the share of weight available. It says nothing about statistical uncertainty.
• Renormalizing changes the composition. The mean then answers a slightly different question, with more weight on what remains, including fiction.
• Coverage alone can move the score. The worked example below changes with no observation changing.
• Missing is not zero. A missing component leaves the mean. It does not enter with a score of 0.

Publication gate

Having $C_t>0$ is not enough. The code requires at least 3 of the 5 documented components to be scored. Otherwise it uses the newest earlier edition that qualifies, with its own date, or shows “No valid score yet.” Synthetic editions can never become the headline.

Worked example, hypothetical

Suppose nuclear were unavailable and nothing else changed. Coverage would be 85%, and the mean would change from 64.425 to 61.676, displayed as 62%, because a component scoring 80 left the mean. If nuclear, climate and food were all unavailable, 2 of 5 documented components would remain, and the gate would block the headline, whatever $P_t$ computed.

Calculation ledger

Generated from the same edition and configuration the page displays.

Unrounded total

64.42450824527648

Displayed percentage

64%

Weighted coverage

100.00%

Publication status

Published: 5 of 5 documented

Edition

edition-2026-11-01-r2

Method and catalog

1.1, 2026.11

Table cells are rounded independently. Adding the contributions rounded to two decimals gives 64.43, while the total rounded once gives 64.42. Any gap comes from rounding cells, and the totals above come from unrounded values.

Both files carry full-precision values, the edition hash and the method and catalog versions. The JSON also includes every raw input and evidence record.

Playing God, formally

Baseline: the edition behind the headline. Scenarios never touch that edition. For each eligible component:

Why +20 is not +20 percentage points. Impacts act inside one component, and that component only holds its effective weight of the total. For common sense, +20 times 15.0% gives 3.000 points, unless clipping reduces it.

Properties and edge cases

• Boundedness. Each score is in [0, 100] and effective weights are non-negative and sum to 1, so $0\le P\le 100$.
• Monotonicity, fixed composition. With weights and eligibility fixed, P never falls when a score rises. Anchored and percentile scores never fall as x rises.
• Intensity. Each catalog event’s impacts all have the same sign, so P′ moves in one direction as a rises, with other inputs fixed.
• Piecewise linearity. P′ is continuous and piecewise linear in each a. The linear score is piecewise linear in x.
• Rubric discontinuities. Rubric scores jump by multiples of 20 as evidence enters or expires.
• Percentile steps. A percentile is a step function of x. It jumps by 2.70 points per reference value for conflicts, and half that when it lands on a tie.
• Saturation. Beyond a clip, extra impact has no effect. Linear scores and simulated scores saturate, and catalog sums are capped.
• Rounding boundaries. R jumps at each half. Values within 5e-10 below a half round up, because of the nine-decimal step.
• Determinism. Scoring reads no clock and no random source. The only randomness is the seeded sensitivity generator.

Monotonicity holds only for a fixed composition. If a component with a low score becomes unavailable, P can rise with no score rising, as can happen when weights change.

Validation, as implemented

Sensitivity

These formulas describe how the existing engine responds. They assume fixed eligible components and a positive total weight. One more point in a component adds its effective weight to P:

For an event on components that are not clipped, the slope is the weighted sum of its impacts:

Before clipping, “Give an AI the nuclear passwords.” has a slope of 14.500 points per unit of intensity. Clipping changes the slope: a saturated component contributes nothing, and at the exact boundary only a one-sided derivative makes sense.

Arbitrary weights are normalized, which gives:

So raising a weight pulls P toward that component’s score. With the configured weights, one extra unit on nuclear changes P by 0.1558, and one on food by -0.3317.

The existing weight experiment

The engine has one seeded experiment. Each of 1,000 trials multiplies every weight by an independent uniform factor from 0.5 to 1.5, using a fixed seed of 20261101, and records P. Scores and eligibility stay fixed. This edition gives a range of 57.54 to 70.02. That range only shows how the result reacts to these weight changes. It is not a confidence interval, and the 50% spread is itself an arbitrary choice.

Dependence and compensation

The components are not independent. Conflict can damage food security and health systems, and climate pressure can feed both. Adding related dimensions can count linked consequences more than once, and the model applies no correction for that.

Addition also lets low values offset high ones. These two hypothetical profiles use the configured weights:

The totals match and the worlds do not. In the second, four documented pressures are at maximum and it would still read the same as the first. Always check the components, not just the headline.

Time, provenance and revisions

Edition date 2026-11-01, method 1.1, catalog 2026.11. An edition is immutable, and the sources update on different cycles: ERA5 daily, UCDP and FAO yearly, and rubric evidence by event.

Staleness. An expired observation makes its component unavailable, and expired evidence stops counting. These checks use the edition date, so a published edition keeps its score until a new one is built. This edition’s scores hold through 2026-11-24. After that, the hero shows a review-overdue notice, and the score stays as published.

Comparability. Editions compare only when method version, catalog version, weights, the synthetic flag, catalog selections, frozen references, anchors, scoring types and eligible components all match. The application cannot split a score change into observation changes, source revisions, method changes and coverage changes. It offers no such attribution, and any you see elsewhere would be invented.

Limitations

• Construct validity. “Apocalypse progress” is not an observable quantity, so the number can’t be checked against any ground truth. No result here can be called accurate.
• Normative weights. Fiction holds 40% of the configured weight because of a product decision. A different editor would get a different number from the same facts.
• Editorial anchors and rubrics. Moving the climate ceiling from 2 to 3 °C would lower that score by a third with no change in the climate.
• Coverage and selection. We scored what has usable public data. Things without a comparable series, such as AI risk, are not in the score at all.
• Different scopes. A ten-year mean, a one-year rate and a single dated statement are combined, so one number mixes slow trends with last month’s news.
• Correlated components. One underlying crisis can raise several scores, which adds emphasis without adding information.
• Fictional inputs. Changing a joke can move the headline more than a real improvement would, because fiction holds real weight.
• False precision. We print decimals so you can audit the arithmetic. They imply no accuracy, which is why the headline is a whole number.
• No probabilistic calibration. A reading of 80 doesn’t mean 80% of anything, and no outcome could ever score this model.
• No causal interpretation. Event impacts are game values. Nothing here estimates what a policy would do.
• Correct arithmetic, debatable inputs. Every step can be verified and every assumption can still be wrong. A clean calculation only moves the argument to the assumptions.

The model is still useful as transparent satire. It puts real facts next to its own arbitrary choices, shows exactly how those choices drive the result, and lets you change them. Most indices with one number don’t offer that.

Reproduce it

Using the real engine, from the project folder:

npm run reproduce

Or from the downloaded JSON, with no project code:

import json
d = json.load(open("apocalypse-loading-edition-2026-11-01-r2.json"))
rows = [r for r in d["ledger"] if r["eligible"]]
C = sum(r["configuredWeight"] for r in rows)
P = sum(r["configuredWeight"] / C * r["score"] for r in rows)
print(P, d["summary"]["unrounded"])

Verification statement

Automated tests confirm that the examples on this page match engine output, ledger contributions reconcile with the unrounded score, partial examples follow the current eligibility rules, simulation examples match catalog impacts, display rounding agrees with the status table, and downloads match the page. This covers arithmetic and consistency only. No external statistician or scientist has reviewed this methodology.

Known inconsistencies

Found while writing this page from the code.

• Published editions do not recalculate over time. Validity and evidence windows use the edition date, so results stay reproducible and never depend on the visitor’s clock. A score therefore stays as published after its evidence lapses. The hero shows a review-overdue notice from the day after the review date, and npm run check:data warns.

Resolved in method 1.1

None of these changed a score.

• An edition with no eligible component now falls back to an earlier edition, or to “No valid score yet,” instead of stopping the build.
• Editions now sort by date and numeric revision, so r10 follows r2.
• The engine now rejects custom impacts and dials that are nonfinite or outside -100 to 100, and rejects unknown component ids.
• Comparability now also requires matching catalog selections, frozen references, anchors, scoring types and eligible components.
• The schema now rejects reference dates that do not match the number of reference values.
• Climate anchors, validity rules and the UCDP release date moved to src/config/method.ts. The climate source date is now the newest date in the ERA5 file, and the food source date comes from the World Bank response.
• Nonfinite observations and invalid anchors now make a component unavailable instead of throwing.
• The hero and share-card bars now use the same number of cells.

Where the shipped code differs from the briefs

• The brief’s simulation formula has events only. The code also adds custom factors, which have no intensity, and the “Turn the dials” offsets.
• The brief allowed an explicitly synthetic full sandbox. None shipped, so the fallback is the last valid edition or a “no valid score” message.
• The brief asked for food to use a historical position only with a qualifying series. The Global Report on Food Crises did not qualify, so the score uses FAO undernourishment and the crisis figure stays as a fact.
• The brief said structural conditions use the documented current state. The code represents that with a manual review date on each record.
• The weight sliders accept whole numbers from 0 to 100 and are normalized, so they do not need to sum to 100.

Events reference