90,000

A scenario engine has no people in it, and policies land on people who support, reject or adapt to them.

Does making agents more human make a synthetic population behave more like the real one?

A synthetic population at national scale: 90,000 agents for a country of about 87,000 people.

A language model writes 123 archetypes ($9.55 in total); deterministic code expands them into 90,000 people (76,321 adults, 13,679 children) in 40,520 households, a four-layer social network of 575,405 ties, and a day of trips routed on the street network (313,527 trips, 312,736 of them routed). Every generated field carries one of four evidence licenses. Two JSON contracts couple the population to the scenario engine.

Household structure matches the official register (band TVD 0.0043).

81% of my first synthetic population did. 1% of real residents did. I wrote it up as a fail.

The population was demographically faithful and attitudinally wrong.

The second test was pre-specified and hash-locked in my repository before a single run, an internal pre-registration. I ran it once, on 8 September 2026. Housing tenure matched (TVD 0.003). Economic self-assessment missed on all six items.

Across all items, 12 of 24 missed.

Results moved by less than 0.11 standard deviations. A twin population with Andorra’s names removed reproduced the results, so the fit and the misfit travel with the generated profile, not with the model’s knowledge of Andorra.

Only four constructs in the whole schema can be checked against real surveys at all. The synthetic unemployment rate is 9.63%; the official one is 2.0%.

Mobility is unvalidated: a telecom-based validation protocol is drafted, the data request has not been sent, and the comparison has not been run. It is the subject of a paper in preparation.

The honest answer to the question at the top is still partly no – structure transfers and dispositions do not yet.

Synthetic Agent Generation Pipeline (working title): the methodology and architecture for generating realistic synthetic populations.

• IN PREPARATION

Mobility Validation Against Telecom Data (working title): validating simulated agent movements against real-world telecommunications data. The comparison has not been run yet.

• IN PREPARATION
• COMPARISON NOT YET RUN

Public Opinion Validation Against Surveys (working title): comparing synthetic agents’ opinions and responses against actual survey data. The Round 1 and Round 2 findings reported on this site are its evidence.

• IN PREPARATION

I designed the pipeline, the validation harness, the pre-specifications and the coupling contracts.

City Science team (TechRxiv preprint: Mora, Atchade-Adelomou, Alonso-Pastor et al.). My part: the coupling, identification checks and outcome memory.

Survey data: Observatori d’Andorra, aggregates only. Supervision: Kent Larson, Luis Alonso.