Validation

A synthetic population is only worth asking if it gives back what the real one would. We check Humanity in two ways. The statistics have to add up, every part to its whole, every time. And the attitudes have to match a survey the model has never seen.

Test one: every part adds up

Official statistics are published as totals: how many people live in a region, how old they are, how many households there are, what households spend. Humanity has to give every one of those totals back, exactly, from the bottom up. Not as an average that comes out roughly right, but as a sum of individual records that lands on the published number.

That is a hard constraint, and a useful one. If a single household were missing, or a budget line were invented, some total somewhere would stop adding up. So every release is checked from the smallest unit to the largest, and a build that misses does not ship.

People sum to households, households to square kilometres, and those to municipalities, regions and the country. At every level the total is the published figure.

Every household has 47 spending lines, and they sum to its budget exactly. Where a line is split further, like the grocery basket, the parts sum to the line.

A market sized in households and the same market sized in people give the same answer. Nobody is lost between the two, and nobody is counted twice.

Every person, children included, belongs to a household, and every household to one square kilometre.

Household income follows the published figures for its region, in local currency.

Cut by age, then income, then place, add the pieces back up, and you get the official figure again.

Test two: a survey it never saw

Values and attitudes are the hardest part of a population to get right, and the easiest to fake. Every person in Humanity carries around sixty of them, from trust and life satisfaction to how much the environment matters, and they come from the European Social Survey.

Which is exactly why we do not test them against the European Social Survey. A model that gives back its own source has only proved it can copy. So we test against Eurobarometer, the European Commission's own public opinion survey, which is not used anywhere in the build. Different people, different interviewers, different years, different wording, overlapping themes.

If the synthetic population answers like Eurobarometer's respondents, country by country, the attitudes are real structure that carries over, not an echo of the source.

in parliament, the legal system, politicians and other people

with life, and with how democracy works

how much climate and nature matter

on the economy, on Europe, on the future

The share of people in each answer is set side by side, per country. Eurobarometer is published by the European Commission and is used for testing only.

Other things it was never told

A check only counts if the model could not have learned the answer.

Second homes in France. Humanity derives a 10.3% share of secondary dwellings from the housing stock and the resident population. INSEE publishes 10%.
Education in Dutch neighbourhoods. Neighbourhoods ranked by education level correlate at 0.59 with CBS neighbourhood data, which was not an input.
Gradients that averages erase. French heat pump use ranges from 2.6% to 35% across regions, matching the official gradient. Households run larger in Paris than in Burgundy, as the census says.

Where a number needs care

Regions and countries are precise. One square kilometre with five filters stacked is an estimate, and every answer says how confident it is. Anything under 100 people is not shown at all. Money stays in local currency, so compare across countries by rank or income quartile. The reference year is 2024. More on Accuracy, with the publishers behind each layer on Sources.

How the population is constructed is shared under agreement for procurement, due diligence and audits. Write to hello@hiasynth.co and say what the review needs to cover.