NOUS: A New Class of Autonomous Intelligence Systems

Krasimir Petrov

Abstract: This paper introduces the NOUS class: a precisely defined category of autonomous intelligence systems whose internal dynamics are not learned but derived — uniquely forced by the intrinsic geometry of the system's own state space. Membership in the class is determined by six conditions, jointly necessary and sufficient, each a mathematical consequence of a single foundational hypothesis: that a system possessing a predictive map from internal states to probability distributions must operate on a statistical manifold with a canonically determined geometry. From this hypothesis, the state space, the symmetry group, the driving functional, and the dynamical law all follow without design freedom. A system satisfying all six conditions converges provably to the correct model of its environment, computes its own operational limits from internal geometry, detects adversarial perturbation through exact conservation laws without a model of the adversary, and coordinates with independent instances toward identical conclusions without communication. The class is non-empty: a computational proof of concept has verified that all structural and dynamical requirements are simultaneously satisfiable in finite-precision hardware across multiple parametric families, with empirical results exceeding theoretical prediction. The NOUS class is substrate-independent, scale-invariant under a single parameter, and structurally incompatible with weight-based architectures by construction — not as a matter of engineering insufficiency, but as a consequence of operating on a different mathematical kind.