Toward Safe Attractor Architecture: Geometry, Prediction, and Control of Structural Safety

抄録

We propose a unified theoretical framework for Safe Attractor Architecture (SAA),
integrating geometric stability, predictive dynamics, and control-theoretic safety. The
framework is built upon minimal attractor geometry and reversibility conditions in
a prediction-error landscape, extended by information acceleration as a pre-rupture
indicator, and completed by an angular geometric formulation incorporating observer
dynamics and small-gain control. We demonstrate that structural instability arises
not merely from magnitude divergence but from directional fixation combined with
sustained positive information acceleration. The resulting theory establishes a minimal
and unified architecture for structural safety in dynamical systems, providing a unified
foundation for predictive monitoring and controlled stability in complex adaptive and
artificial intelligence systems.