gOS and AGInetwork can produce traces as a by-product of use: which specialists were called, how they were wired, what context they received, and what the human made of the outcome. Geranium is designed to learn from that record through reinforcement learning, a way of updating a policy from evaluated consequences. In this training design, RL links prior traces to later routing decisions. That mechanism specifies how the research model would adapt; any improvement still has to be shown against outcomes and evaluations from the surrounding system.

That places Geranium at the centre rather than the top of the architecture. The design assigns it no privileged view, canonical memory or authority of its own. What it could learn is bounded by the traces and evaluations supplied by gOS and AGInetwork, and its training signal depends on those layers producing grounded feedback.

Compositional Intelligence is the program Geranium answers to. If capability lives in the wiring, then the wiring is the relevant search space: routing, control flow, memory, evaluators and the models themselves. Geranium is designed to search candidate assemblies under evolutionary pressure: arrangements compete on real jobs, evaluators with their own stake in the outcome score them, and the structures that survive are retained for reuse. Where appropriate, some of that structure may be fused back into model weights.