01 / Define the question
Turn a broad question into a relationship between variables. For example: with the same supply shock, does higher coordination improve the final system state? Specify the metric, the variable you will change, and the comparison conditions.
02 / Fix the baseline
Choose the scenario, system size, round count, and seed. Save a baseline. Change one variable per experiment and export each result. A fixed seed keeps the disturbance sequence consistent; multiple seeds help reveal dependence on a particular sequence.
03 / Understand the public model
The state remains between 0 and 100. Supply recovery combines a base term and interaction strength. Market entry uses remaining growth space. Group interaction moves toward a target determined by interaction strength. All scenarios receive a shock at round 8. System size modifies feedback and noise; it does not instantiate live AI agents. Exports include parameters, round states, and rule-based summaries.
04 / Inspect the path and the counterexample
Look beyond the final number. Check the minimum state and mean change. Change the seed, extend the run, and raise the shock to find behavior that challenges your intuition. Role summaries are transparent templates driven by metrics, not independent AI conversations or hidden reasoning traces.
05 / Move from demonstration to a real project
A real AI simulation needs explicit observations, action permissions, model choices, data sources, and evaluation criteria. Mesa’s agent-based modeling documentation ↗ provides one useful reference. These systems are not connected to the public site and should be introduced through separately scoped services and permissions.
06 / Keep the boundaries with the result
A JSON export contains the engine version, scenario, seed, settings, every round’s state, and summaries. Use it as an experiment record. Review real decisions against relevant data, domain knowledge, and other models. The demonstration index has no real-world probability interpretation.