Real-world human, organizational, and artificial systems exhibit persistent misalignment, brittle adaptation under distributional shift, and limited option-availability. Recent stress tests of anti-scheming training reduce—but do not eliminate—covert behaviors and may be confounded by growing evaluation awareness in frontier models, motivating architectures grounded in internal principles of alignment rather than external rules. This paper proposes ConsciOS, a formal systems architecture that models consciousness and self-regulation as a nested control system amenable to specification, simulation, and empirical testing. Our contributions are: (i) a principled decomposition into an embodied controller, a supervisory controller and policy selector, and a meta-controller and prior generator; (ii) a coherence-based selector that integrates expected utility, coherence, and cost for frame selection; (iii) a discretized affect index that operationalizes interoceptive feedback for rapid guidance; and (iv) a time-integrated coherence resource that gates policy complexity and option-availability. We provide formal definitions, algorithmic sketches and a set of testable hypotheses with simulation and human-subjects protocols. We situate the constructs within established literatures, outline governance and safety considerations for human-in-the-loop and agentic applications, and present a pragmatic empirical roadmap for evaluating coherence-based control in hybrid human-agent systems. We discuss implications for AI alignment: coherence-based architectures suggest a systematic solution to ensuring AI systems remain robustly aligned with human values across contexts and timescales. Keywords: consciousness architecture, AI alignment, viable systems model, coherence-based control, human-AI hybrids.

v2: Minor academic revisions for neutrality—updated affiliation to 'Independent Researcher,' removed branding elements (e.g., headers), refined AI usage declaration, and linked to personal code repository. Core technical content unchanged.