Lighthouse Ethical Architecture: Coherence as a System Property for AI Governance (Working Paper)

Abstract This paper presents the Lighthouse Ethical Architecture, a governance architecture intended to make coherence an operational governance object for AI systems and the institutions that design, deploy, and govern them. It is a working paper that presents a conceptual governance architecture and public evaluation protocol; empirical validation and implementation artifacts are out of scope for this release. Existing efforts in AI governance frequently stall at principle lists and high-level frameworks that do not translate into enforceable controls or operational mechanisms, a gap noted across regulatory and research bodies (European Union, 2024; Mittelstadt, 2019; NIST, 2023; OECD, 2019; UNESCO, 2021). The Architecture responds by anchoring system and institutional behavior to the Coherence Code Layer, a foundational representation that encodes explicit normative commitments as invariants, constraints, incentives, and boundary conditions, and by structuring how those commitments are monitored, tested, and revised over time. The primary unit of analysis is a system–institution configuration, for example, a high-risk or high-impact AI system deployed within a particular institutional context. Within this configuration, coherence is defined as the relation, over time, between three elements: the declared invariants and encoded constraints in the Coherence Code Layer, the incentive structures and reward mechanisms that shape behavior, and the observable behavior of the system–institution configuration in practice. Explicit normative commitments include internal policies, institutional values, regulatory obligations, and other declared constraints that an institution elects to encode and treat as binding for the purposes of governance. These commitments are understood to be contestable and revisable rather than uniquely correct. The Architecture includes mechanisms for documenting, reviewing, and updating them as evidence, context, and regulatory expectations change. Drift is defined as governance-relevant deviation of observed behavior from these declared coherence commitments across technical, institutional, and sociotechnical dimensions (Selbst et al., 2019; Raji et al., 2020). A divergence function is introduced as a conceptual measurement hook. It links encoded commitments to observed behavior within specified tolerances and produces structured divergence signals that can be interpreted against documented commitments. Drift is treated as present when these divergence signals cross documented governance thresholds or fall into predefined drift categories. The Architecture adopts a governance posture that treats drift as a default risk trajectory in high-risk or high-impact AI settings unless actively monitored and addressed through controls, escalation paths, and periodic review, while explicitly not claiming that drift is a universal empirical law for all systems in all contexts. Structurally, the Architecture is organized around three interlocking components: The Coherence Code Layer provides the foundational representation of declared commitments and associated constraints, instantiated as policy-as-code and institutional policy. Next, a five-layer Coherence Stack organizes how these commitments are connected to measurement, decision-making, and deployed AI services in the surrounding institutional environment, in alignment with lifecycle concepts such as transparency, traceability, quality management, and meaningful oversight in high-risk systems (European Union, 2024; OECD, 2019). Traversing these layers, the Coherence Stack Decision Loop routes signals of divergence through detection, evaluation, decision, and institutional learning into a continuous governance cycle with role-assigned responsibilities and traceable decision logging. These components are supported by two framing constructs: The Lighthouse Coherence Principle (Coherence Principle), which defines coherence as bounded alignment under tolerances, and The Fractal Coherence Model and associated vertical maturity levels are used as analytic lenses to examine coherence and drift at individual, team, institutional, and ecosystem levels and across technical lifecycle stages, rather than as additional governance layers. The public version of the Architecture presented in this paper is limited to the conceptual architecture, governance logic, and artifacts required for academic, regulatory, and institutional review. Proprietary scoring logic, numerical thresholds, test suites, and implementation mechanisms are intentionally excluded and are handled separately under commercial terms; they are expected to undergo independent legal, ethical, and technical review wherever they are developed. The paper advances a conceptual governance architecture, not an implementation blueprint or legal instrument. It positions coherence and drift as explicit operational governance objects, coupled to divergence functions, thresholds, and decision processes that can be documented, interrogated, and revised. The paper argues that, under appropriate institutional conditions, this explicit treatment of coherence and drift can contribute to improved governance performance. Demonstrating the magnitude, limits, and sector-specific implications of such improvements requires formalization, empirical evaluation, and piloting by regulators, model developers, and deploying institutions, with particular attention to validation, limitations, and potential failure modes of the Architecture itself.