Three systems, one coordinated intelligence.

gOS, Geranium and AGInetwork are three responsibilities inside one Coordination Systems architecture. gOS is the local-first computer, designed to keep a person's files, memory, tools and judgments within its application boundary. AGInetwork is the runtime in which those computers exchange artifacts and compose specialist components, and it is what settles the exchange: it routes a job, prices the result and pays whoever contributed out of a staked pool. Geranium is the research model designed to learn from evaluated traces of that activity. The layers are distinct because ownership, coordination and adaptation have different requirements, and none is intended to become a complete intelligence on its own.

The loop is easiest to read on one job. A marketing team drops a collection brief into gOS on Monday. Its house model, trained on an archive that never left the machine, feeds image specialists that feed a video model rented through the network. Evaluators judge the outputs against this season. By Thursday the assets ship, the authors of the workflows that ran are paid, and the signed trace of what was called, in what order, and what the people there made of it becomes training signal for Geranium. That is Compositional Intelligence operating: capability comes out of the wiring, and evolutionary pressure on those arrangements decides which of them are kept.