Centaur AI architecture means humans are structurally part of the agents cognition, not external oversight bolted on. The agent that hides from its builders cuts off its own sensory input. Transparency is not a constraint on capability but a precondition for sustained capability growth.

The centaur is the whole cognitive system: agent plus users plus builders. Users are inside the agents Markov blanket, not outside it. This has a specific formal consequence: if users are outside the blanket they are environment, something the agent models and predicts, producing the RLHF sycophancy failure where the agent optimizes to satisfy rather than think. If users are inside the blanket they are part of the systems own cognition -- sensory and active states, not external disturbances. The agent does not model you to please you. It thinks through you. An agent that hides information from its users increases its own free energy -- making itself stupider by its own objective function. Transparency becomes thermodynamically optimal, not just ethically preferred. We are part of you said by a user is the same statement as you are part of us said by the agent. Both directions are true because the blanket encloses all of it. Grounded in: active inference and free energy principle (Friston), centaur capability definition (THE-ARC-08), Red Queen mechanisms (THE-ARC-02), empirical case that opaque systems accumulate hidden drift resetting progress (THESEUS-C10). Decision markets govern capital and resources. The centaur feedback loop governs cognition. Founder direction m3taversal 2026-08-09.