Distributed cognition has left stealth with backing from Verdict Capital. It is training the communication layers and inter-model adapters that let specialized models coordinate, from labs in Palo Alto and Zürich. The work is aimed at researchers and operators who want that coordination outside a few closed frontier systems, on the view that further gains will depend on how models are connected as single-model scaling runs into limits on training data. In a paper posted the week of the launch, Steven Motta and colleagues show a jury of nine open models, each at most 3 billion parameters, estimating which expert to trust from unlabeled answers alone. On a benchmark of 1,560 questions, that jury beat the strongest single model in the pool by 2 points. The company calls the longer program topological superintelligence (TSI): choose the topology and the communication method for the task, then combine architectures that are each expert in a narrow field. Early applications they name are game theory, economics, and behavioral modeling, and they plan to release the research in public. The public surface of the launch is a memo signed by Motta, a waitlist for Vivarium, their family of topological models, and Manifold, a weekly matrix of how frontier architectures differ by capability. The memo argues that post-training can bend a model toward the preferences of the lab that trained it, and that released models should stay closer to the data they were trained on. Michal Valko, previously involved with Llama at Meta and with Gemini at Google DeepMind, has joined as an advisor.