A synthetic Britain, and how you know it is wrong.
Fifty thousand synthetic adults, each drawn from a real anonymised Census record rather than invented. The point of building it this way is that it can be checked against figures it was never fitted to — and the check is reported here whether it flatters the model or not.
24 August 2026Feb 2024 – Mar 2026Public sources throughout
Every synthetic adult is a real, anonymised, disclosure-controlled Census record, reweighted so that region, sex and age reproduce published Census totals exactly. Nothing is generated by a language model. No attribute is inferred from a stereotype about the person holding it.
Two things worth knowing before trusting any figure from it. It is Great Britain, not the UK — Northern Ireland is absent. And "adult" means 16 and over, because the published age bands cannot separate 16 and 17 year olds, while the surveys it will be calibrated against sample 18 and over.
02
Validationi
Synthetic population against published Census figuresnot fittedi
ONS Census 2021 microdata comparability tables · economic activity and ethnic group, 16+ base
Mean absolute error 0.134 percentage points across 14 cells, worst case 0.305pp. For comparison, error against the figures the population was fitted to is 0.034pp — that number measures only sampling noise in the draw and proves nothing, which is why it is not the headline.
This tests demographic composition only. It says nothing about whether the population can reproduce an attitude, which is a far harder problem and is not yet demonstrated.
03
Attitude targetsi
Is immigration good or bad for Britain's economy?BES 2019i
British Election Study 2019 Post-Election Random Probability Survey, N=3,946 · weighted with wt_demog · don't know and not stated excluded from the base
The 2019 and 2024 random probability surveys carry the same immigration items under the same variable names with the same response scales, so the two years are directly comparable. That makes a genuine design possible: calibrate on 2019, keep 2024 sealed, and test blind.
04
What is blockedi
A distribution is not a transition. Everything above tests whether a population reproduces a level. None of it can test whether the model reproduces change, because two independent cross-sections cannot follow the same person. That needs the British Election Study internet panel — 31 waves, February 2014 to June 2026 — and the modelling on it is not done.
One dimension has no published attitude series at all: asylum accommodation. Rather than invent a proxy, it is treated as an exposure variable instead — a local authority acquiring contingency hotel accommodation is a dated, local shock that can be tested against attitude change, using figures already published quarterly.