Civilization Simulator

What happens if you change the rules? The civilization simulator lets you model how different economic structures, governance models, and social policies produce different outcomes over time.

What it does

This is a research-oriented simulation of societal dynamics — economic systems, governance structures, social trust, environmental impact, wealth distribution, and more. You can set up scenarios based on real countries and see how different variables interact over time.

What it is not

It is not a prediction engine. No model can predict the future of complex human systems with precision. What a model can do is reveal structural tendencies — the directions a system is likely to move given its internal dynamics. This is a tool for thinking, not a crystal ball.

No setup required

The simulator runs entirely in your browser using a built-in rule engine. No AI service, API key, or external connection is needed. On first launch, the simulator will offer to configure optional AI-powered interviews. To use the complete built-in simulation without an AI service, choose "Skip — Play Without AI."

How it differs from other simulators

Most policy and macroeconomic simulators are built to reproduce known outcomes as accurately as possible — they are calibrated tightly to historical data so that their hindcasts match the record. That makes them powerful for projecting within known structures, but it also means they cannot model structures that have never existed. They reproduce the world as it is; they cannot ask "what if the rules were fundamentally different?"

This simulator is built for a different purpose. It models the mechanisms — the structural dynamics that produce outcomes — rather than targeting precise reproduction of any particular country's trajectory. It captures how wealth concentrates, how trust erodes, how governance structures interact with economic systems, how environmental pressures feed back into social stability. This means it can model speculative scenarios — cooperative economies, radically different governance models, transitions between paradigms — that no historically-calibrated model can represent, because there is no historical data to calibrate against.

The trade-off is explicit: the simulator approximates real-world dynamics rather than reproducing them exactly. A hindcast of Industrial Britain will show wealth concentration rising, environmental pressure increasing, and institutional capacity growing — the right structural story — without hitting the precise GDP figures for 1850. That is by design, not a limitation to be apologized for.

Validation and testing

The simulator has been tested using the same methodological standards applied to research-grade social simulation models; it has not yet undergone external peer review. Full diagnostic results are documented in the project's MODEL_DIAGNOSTICS.md file (available in the source repository).

Hindcasting

The simulator is tested against historical scenarios — configurations modeled on real societies (Classical Athens, the Roman Republic, Industrial Britain, Tokugawa Japan, the Mughal Empire, Soviet-type economies, post-colonial states, and others). Each scenario has expectations derived from the historical record before running the model, not after. The simulator's output is compared against these expectations to verify that structural dynamics are correctly represented.

Held-out validation

Scenarios are divided into two sets: a tuning set (used during development to identify and fix defects) and a held-out set (never used for tuning — its results are the only genuine measure of generalization). This is standard practice in model validation: if you tune against all your test cases, you have no way to know whether the model is capturing real dynamics or just fitting the specific cases it was shown.

Current graded scores: 93.8% on the tuning set, 75.7% on the held-out set. The gap between these numbers is itself informative — it shows where the model captures known dynamics well and where further structural work is needed.

Invariant suite

Thirteen permanent automated checks guard against specific defects found during development — including numerical instabilities, unreachable code paths, discontinuities in state transitions, and subsystems that appear active but produce no measurable effect. All 13 pass. Each invariant names the exact defect it guards against.

Sensitivity analysis

Every tunable parameter has been tested for its influence on outcomes. Parameters whose values were interpolated rather than derived from empirical data are flagged, and the two most influential of these have been re-anchored to published research (Tamburini et al., Science Advances, 41,946 comparisons on ecosystem-yield relationships; measured transit-reach ratios for active travel). Interaction analysis confirmed that parameter effects are substantially redundant through a shared wellbeing channel, meaning the model's behavior is more stable than single-parameter testing suggests.

What is honestly not yet done

Two areas remain as known limitations, documented rather than hidden:

• Scenario thresholds are currently authored by the developer from historical knowledge. A planned upgrade will derive them from the Seshat Global History Databank (414 societies, 51 variables, 10,000 years of coded data) — converting validation from "does it match my expectations" to "does it match the coded historical record."
• Two subsystems (schism dynamics and wealth concentration pathways) have documented structural defects that are partially repaired but not fully resolved. These are flagged in the source code and the diagnostic log rather than silently shipped.

Design philosophy

When a structural repair made the model demonstrably more correct but lowered the numerical score, the repair was kept and the scoring instrument was examined. The project treats honesty about what the model gets wrong as more important than optimizing a headline number. A model that hides its failures behind favorable scoring is less trustworthy than one that documents them.

Current status

The simulator is functional and actively maintained. It is a research tool under continued development, not a finished product. Structural improvements are made through a documented diagnostic process, and each change is tested against both the tuning and held-out scenario sets before it ships.

The simulator works best on desktop. On mobile devices, the interface will be small but functional.