The agent maintains a generative model—an internal representation of how the world produces sensations. Initially empty, this model grows with experience.
In Piaget's terms, these are schemas: organised patterns of knowledge that guide interpretation and prediction.
Before encountering an object, the agent generates predictions based on its model. The difference between prediction and reality is prediction error.
High prediction error signals disequilibrium—the uncomfortable state that drives learning.
When a new experience matches an existing schema, assimilation occurs. The prediction succeeds, free energy stays low, and the model remains stable.
This is the default mode—organisms prefer to confirm rather than revise their models of the world.
When prediction fails significantly, the agent must accommodate—restructure its generative model to reduce future surprise.
In Active Inference, this means updating beliefs μ to minimise variational free energy F(μ).
As more objects are encountered, the agent forms abstract categories—schemas that group objects by shared properties.
This emergence of hierarchical structure mirrors how children develop increasingly sophisticated conceptual frameworks.
The agent doesn't passively receive sensations—it actively explores to reduce uncertainty. Clicking shapes simulates the agent's curiosity-driven sampling of its environment.
This is epistemic action: acting to gain information that refines the generative model.