The Andromeda architecture was designed by Art Code Outdoors, with the architecture first demonstrated to me in 2019.
These documents are my personal interpretation of that work. I don't fully understand it, and I may have gotten things wrong.
The architecture diagram and GraphViz source below are the designer's own creations. The written documents are mine, and any errors in them are mine. This is a step on the path, not the destination. Free to copy and share.
Andromeda is not a proposal for a larger model. It is a documented cybernetic architecture for adaptive machine intelligence outside the neural-network paradigm.
These documents are provided for careful readers, informed operators, and future builders.
If you translate the terms back into weights, rewards, benchmarks, or training runs, you are no longer reading the architecture described here.
New readers: begin with the Layman's Guide.