Author: Matthew Andrews
Date: September 15, 2026
Status: Working Paper
This working paper develops the current foundation of Generative Adaptive Dynamical Systems. Beginning from a commitment against central control of representational development, it derives the architectural roles required for a search process to create and regulate its own representational language without collapsing under the consequences of doing so.
The paper defines the Layer 0 substrate, develops six failure-driven chains leading to the full GADS architecture, describes its coupled timescales and compositional geometry, and defines variational free energy as passive telemetry rather than an objective optimized by the system.
The derivation establishes why the identified architectural roles are necessary under its foundational commitments. It does not claim that their presence alone guarantees stable dynamics, adaptation, or biological equivalence. Those remain empirical questions addressed through implementation and experiment.
For an introductory walkthrough, start with What Is GADS? and What kind of thing is MEGA?
The architecture is also available to study in code. GADS_single implements one system developing its own representations. GADS_network connects independent systems that exchange representations through bounded local interfaces. Both repositories include instructions for running the implementations and descriptions of their fitness environments.