This working paper proposes a filtration-gated control architecture for building trustworthy artificial intelligence systems. Rather than relying on post-hoc safeguards, the framework introduces deterministic regulatory layers that constrain probabilistic reasoning before output and action occur. The architecture integrates filtration, reasoning permission, evidence validation, bias regulation, authority gating, memory control, and auditability into a bounded system design. The central claim is that trust in AI should not be derived from capability alone, but from the presence of structured, testable control mechanisms that regulate when and how intelligence is permitted to operate. The paper presents the architecture, explores failure modes, and outlines testable criteria for trustworthy AI. An applied prototype (Professor Santi) is referenced as a development environment for exploring these principles, though implementation details remain outside the scope of this work. This document is a working paper intended for feedback, discussion, and further development.