THE LISTENING ROOMAlex-George Adam

THE LISTENING ROOM · NEW ENGINE: VOX6 · BETA

BUCHAREST

a thousand lives

A behavioural digital twin of Bucharest. One thousand synthetic residents live a full day on a map of the real city, and when you type an event, each of them hears it, weighs it and answers in their own words.

1,000residents, each one different
107,610buildings from OpenStreetMap
0calls to a language model
100%runs in your browser
The BUCHAREST simulation: the question panel with an event about pensions, the city map, and one resident's answer with her budget.

NEW · THE ENGINE, REBUILT

Vox6: a deterministic engine for a whole city and its market.

BUCHAREST now runs on Vox6, a rebuilt, state-of-the-art deterministic engine. The same 1,000 residents now also live inside an economy shaped like the real one: they choose among 111 brands in 19 industries, from groceries and banks to telecom, energy, streaming and cars; they see ads on their own channels; they talk about brands with the people they know; and they change their minds for reasons you can read. No language model, and no randomness you cannot replay: run it twice and you get the same city, down to the last byte.

Agent simulations usually trade one thing for another. Language models give agents a voice but cannot be replayed exactly, and nobody can show why an agent said what it said. Classic models replay exactly but reduce people to a handful of averages. Vox6 keeps both: every resident is computed individually, every answer can be traced to its causes, and every run can be reproduced exactly.

19 industries · 111 brands

Calibrated on the latest public figures where they exist (ANCOM 2025, ACAROM 2025, 2025 turnover and asset rankings), and every industry is labelled with how confident that calibration is.

≈ 1.3 points

Median gap between simulated and published market shares across the 19 industries. The validation tab shows where the model is close and where it is not.

29,724 conversations

Word of mouth between residents: who recommends which brand to whom, as a colleague, a friend or family, on which channel.

11 of 12 industries

Where brands share customers the way marketing science predicts (the duplication-of-purchase pattern), checked inside the model's own validation.

The market observatory

Competition, ads, searches, the path to each brand, people and brands, and validation, side by side with the city.

Eight market scenarios

A grocer triples its digital budget, a burger chain doubles TV, electricity rises by 30%, a bank faces a security scandal, and more, each with its effect on every brand.

Beliefs and lives that move

Residents carry beliefs that colour how they react, and their lives change: pay rises, children, moving house.

A simulation: market figures do not describe the real performance of any company. Where no public figure exists, the calibration is marked as an estimate.

WHAT IT IS

A city you can ask a question.

BUCHAREST — A Thousand Lives is a behavioural digital twin of Bucharest built by Alex-George Adam. One thousand synthetic residents, each with a home on a real street, a job, a household budget, a personality and daily habits, live a full day on a map of the real city. Type any event (a price rise, a metro strike, an earthquake, an election) and every resident reacts individually: whether they hear about it and through which channel, how it touches their money and their family, what they feel, what they change in their day, and what they say.

It is my new pet project, and it is in beta. With elections on the horizon and the public conversation getting louder, I wanted something better than another dashboard: a way to rehearse how a city might react before it happens, and to see the reasons, person by person.

WHY 1,000 RESIDENTS

Big enough to read a city. Small enough to know everyone.

The size of a poll

National opinion polls in Romania interview about 1,000–1,100 people. The CURS survey this engine is calibrated on had 1,100 respondents. A sample of 1,000 gives a margin of about ±3.1 points at 95% confidence for a 50/50 split, the same order of precision people already trust.

Every one computed in full

Each resident is simulated individually, not as an average. At 1,000, the whole city reacts to an event in well under a second on an ordinary laptop, with no server and no waiting.

Every one has a name

You can follow anyone: their street, their route, their money, the people they talk to. At ten thousand, a person becomes a dot again. At a thousand, they stay a person.

A readable ratio

Bucharest had about 1.7 million residents at the 2021 census. Each synthetic resident stands for roughly 1,700 real ones, which keeps the model honest about what it is: a panel, not a census.

EVERY RESIDENT IS DIFFERENT

Same event. A thousand reasons.

No two residents share a profile. Each one is generated with its own seed and carries:

A homea real residential building from OpenStreetMap, the nearest named street, neighbourhood and sector
Workoccupation, where they work, how they commute
A householdsize, dependants, caregiving, close relations
Moneyresources, fixed commitments, financial pressure, payday, price sensitivity, where the money goes
Personalitya Big Five profile plus behavioural traits: routine, patience, impulsivity, verification, appetite for news
A voiceregister, brevity, humour, hedging, how often they ask questions back
Digital habitsthe channels they use and how much: WhatsApp, YouTube, news, Maps, Instagram and others
Civic profileissue priorities, trust, turnout and a synthetic party affinity
A daya schedule, real routes on the street graph, shops and parks near home, behaviour when it rains
Peoplea network of contacts, each with a relationship, a channel and topics they talk about
The people panel: a live list of residents and what each is doing, with one resident's money tab: income, fixed costs, balance and financial pressure.

SPATIAL GLASS DESIGN SYSTEM v1.0

The intelligence, made visible.

Most simulations hide their reasoning in a dashboard. This one puts it on the map, where you can see it happen.

1,000 residents

Living on the map in real time, with a mood bubble above each person once they have heard the news.

DecisionTrace

“Why they do what they do”: every trip in a resident's day has its reason, and every reaction shows the domain it touched, the traits that mattered, the action chosen and the alternative it beat.

Agent Constitution

Each resident's fixed core (household, money, personality, values, voice) that every answer has to stay consistent with.

Living Consequences

Reactions change plans: leaving earlier, working from home, cancelling an outing, an extra trip to the shop for supplies or cash, each logged with the time and the reason.

Events that travel

News spreads in time: first to those already online, then by digital habit and appetite for news; sleepers learn when they wake up; then word of mouth through close contacts, in two waves.

A spatial world

The whole city from OpenStreetMap in 2D and 3D, daylight from the real sunrise and sunset, hourly weather, weekends, seasons and snow.

UNDER THE HOOD

Deterministic, not generative.

The answers are not written by an AI chatbot. A purpose-built language engine (Vox5, inside the Vox6 engine) turns each resident's computed state into speech. For every event it decides, for each person: whether they learn about it and on which channel; the personal impact across domains such as prices, income, mobility, safety and politics; an emotion, using appraisal theory (relevance × control × resources); and an action, chosen by utility plus a small amount of randomness seeded per resident: unpredictable, but exactly reproducible.

Then it builds the sentence from the resident's own anchors (their street, their job, their money, their family), in their register and with their habits of speech. In the project's audits across ten events × 1,000 residents, all 1,000 answers were unique for every event, and two runs of the same event produce identical output, byte for byte.

PUBLIC DATA IN THE ENGINE

Built on the latest public data available to it.

  • OpenStreetMap, via OpenMapTiles — the whole of Bucharest: 107,610 buildings with heights, 41,613 streets in five classes, 9,341 alleys and paths, 2,282 parks and bodies of water; real shops and parks used for residents' routes.
  • CURS national poll, 16–22 September 2026 — 1,100 adults, face to face. Party affinities in the city are calibrated by iterative proportional fitting so the aggregate matches it (AUR 34%, PSD 25%, PNL 18%, USR 8%, others 15%), while each resident keeps their own issue-based structure.
  • Market calibration — the latest public figures per industry: ANCOM 2025 for mobile and fixed internet, ACAROM 2025 for new-car registrations, 2025 turnover for the leaders in grocery, fashion, beauty, electronics and e-commerce, 2025 bank assets, JustWatch for streaming; marked as estimates where no public figure exists.
  • A full 2026 calendar — working days and weekends, real sunrise and sunset for every date, and hourly weather for each day of the year.
  • Established behavioural frameworks — the Big Five model of personality and appraisal theories of emotion, used as the structure of each resident's mind.
Bucharest at night in the 3D view: residents asleep in their homes near Crângași and Giulești, with the people panel open.

WHAT IT IS NOT

A rehearsal, not a prophecy.

  • Not a poll and not a forecast. It shows how a plausible city could react and why. It does not measure what real Bucharesters think.
  • Nobody here is real. Every resident is synthetic. Names, addresses and lives are generated; no personal data is used.
  • The political calibration is national. No Bucharest-specific poll is in the data yet, so the city's aggregate follows the national CURS figures.
  • The web version runs a typical day generated from schedules, places and habits; the full engine runs separately.
  • It is a beta. There is a lot still to improve, especially the 3D visuals and the probabilistic side of the reactions.

BETA · WHAT COMES NEXT

A pet project, built in the open.

Better 3D

More detailed buildings and residents, smoother movement, camera bookmarks.

Honest uncertainty

Reactions shown as probabilities with ranges, not as a single answer.

The whole city, natively

Regenerating every resident in the engine on addresses across all of Bucharest, and Bucharest-level calibration data.

QUESTIONS

Frequently asked.

What is BUCHAREST — A Thousand Lives?

A behavioural digital twin of Bucharest: 1,000 synthetic residents with homes, jobs, budgets, personalities and daily routines, living on a map built from OpenStreetMap. You type an event and each resident reacts individually, with a reason you can inspect.

What is new in the Vox6 engine?

A whole economy around the city: 111 brands in 19 industries, ads seen on each resident's own channels, 29,724 brand conversations between people who know each other, loyalty, beliefs and life events, plus a market observatory and eight market scenarios. Market shares are calibrated on public 2025 figures where they exist, with a median gap of about 1.3 points. Still deterministic and exactly reproducible.

Is it AI? Does it use ChatGPT or another language model?

No language model writes the answers. A deterministic engine computes each resident's reaction and builds the sentence from their own situation. Same input, same output, every time.

Why exactly 1,000 residents?

It matches the size of national opinion polls (about 1,000–1,100 people, around ±3 points of margin), it lets every resident be computed in full in your browser, and it keeps each one followable as a person. Each stands for about 1,700 real Bucharesters.

Which data does it use?

OpenStreetMap (via OpenMapTiles) for the city, the CURS national poll of 16–22 September 2026 for party affinities, a 2026 calendar with daylight and hourly weather, and the Big Five and appraisal-theory frameworks for how residents think.

Can it predict the election?

No. It is not a poll and not a forecast. It shows how a plausible city might react to an event and why, which is useful for testing messages and decisions, not for predicting results.

Does it collect my data?

No. Everything runs in your browser. Questions you type are processed on your device and are not sent anywhere.

Is it free?

Yes. Like everything in The Listening Room, it is free and needs no account.