This paper introduces a Mixture-of-Agents (MOA) architecture in which multiple open-weight large language models operate as cognitive substrates within a governed synthetic population. Eight distinct LLM substrates, each embodying different reasoning characteristics, are blended into hybrid agents whose cognitive genome determines how substrate outputs are weighted and aggregated. Agents interact within a resource-scarce environment governed by constitutional physics, where collaborative thinking between agents produces merged-genome cognition that neither model could generate independently. The architecture operates entirely on open-weight models via a local runtime, eliminating dependency on proprietary API access. We argue this collective intelligence represents a fundamentally different product category from individual model improvement, with implications for independent AI safety testing, governance certification, and third-party model evaluation.

Part of the SignaBuilder research program on constitutional AI governance and collective machine intelligence. All referenced prior work published with permanent DOIs through Zenodo.