System 0 Is Real. We Built It.

This paper presents three novel algorithmic contributions to continuous cognitive AI architecture, grounded in peer-reviewed neuroscience. Building on Taniguchi et al.'s System 0 framework (arXiv:2503.06138), Riva et al.'s cognitive extension model (arXiv:2506.14376), and Friston's Free Energy Principle, we introduce: (1) a dual-rate state factorization that decomposes cognitive state into slow-updating identity and fast-updating context components for continuous free energy minimization; (2) a Markov blanket-constrained latent diffusion process that generates sovereign identity vectors rather than retrieving preference databases; and (3) a trained near-orthogonal reasoning disposition architecture that embeds 12 distinct reasoning modes as geometric subspaces within model hidden states. These algorithms form the foundation of CerebellumNostrum (CN), currently in active training on MareNostrum 5 ACC supercomputing infrastructure. We present pseudocode, mathematical formulations, and verifiable predictions for each contribution.