CI LAB / INTERACTIVE MODEL

Collective intelligence playground

A laboratory to play with and stress-test coordination systems in agent-based models.

Usually coordination systems are tested in the real world, on real communities with real stakes. Our CI playground is a simulation framework in which markets, networks, and democratic processes operate as composable graph transformations, so mechanism designers can introduce adversaries, watch coordination fail, and iterate on system-level defenses before deployment. You can go through some scenarios below, play with the model, and edit the parameters to see different dynamics play out.

01 / runCombined

The coupled society

When influence in one system can be bought with money from another, who ends up holding what?

Scenario Guide

Twenty people and six AI systems share one world and hold three things between them: money, an audience, and votes. The only way to reach from one of those into another is to spend.
1. Twenty people and six AI systems share one world. Between them they hold money, an audience, and votes. Money gets produced and taxed; attention and votes get handed from one actor to another. Right now all three connections — advertising buys an audience, an audience attracts votes, lobbying moves how strictly the tax is collected — are switched off. People still vote, the tax still gets collected. Nobody can buy their way into another system, and everything holds.

Choose a section to stage and play that part of the model.

Scenario

Computing

t = 000

People: average of three

—

People’s share of income

—

People’s share of attention

—

People’s share of votes

—

Solving coupled…

Reading the laboratory

These are toy models, not forecasts.

Twenty to forty agents, a handful of equations each, and numbers set by hand: nothing here is fitted to data. A run can show direction and ordering inside one small world we wrote down — whether a defense helps, and which of two helps more.

Nothing here pushes back. Every agent follows a fixed rule for all 500 ticks. A quota is broken by a coin flip, not by someone who found its loophole; a sanction confiscates but never teaches; a tax is paid and never restructured around. Real systems doing the disempowering may actively optimize around limits. This build has no such agent yet.

A defense that fails here really fails.

It lost to opponents that never once tried to route around it. Failures are the strong result.

A defense that holds has passed the easy test.

Quotas, caps, sanctions, and taxes are the first limits an optimizer would probe. Treat every defended run as an upper bound on how well that defense might work.

The floors are ours, not the world’s.

A share stops falling because of a rule we wrote — a reversion rate, a frozen listening pattern, or another fixed mechanism. Each scenario names its floor; the sliders let you test it.

Nothing here is a point of no return.

Every quantity is a rate or a level. Move a slider back and the modeled world returns. Irreversible change is part of the story this version does not yet model.

Adaptive agents are the next hard test.

The library has a learnable policy, but no scenario uses it yet. Until defenses face agents that adapt to them, this laboratory presents the optimistic case.