Modern manufacturing systems operate under high variability, uncertainty, and reduced decision time, which limits the applicability of classical optimisation and static heuristic scheduling approaches. This paper proposes a knowledge-driven scheduling architecture integrated with a digital twin to support adaptive decision-making in dynamic production environments. The framework combines context-sensitive strategy restriction, multi-criteria evaluation, and simulation-based validation within a closed-loop structure. Scheduling strategies are dynamically selected and ranked based on real-time system conditions, and validated through high-fidelity digital twin simulations prior to deployment. A formal mathematical model of the architecture is presented. The proposed architecture establishes a foundation for future empirical validation in industrial environments.