Hybrid CRE for AI governance

Abstract Current AI governance approaches predominantly frame alignment as a problem of behavioral optimization, relying on reward shaping, constraints, or post-hoc oversight. This paper advances a different paradigm: a constitutional computational architecture in which artificial decision-making is institutionally structured rather than behaviorally steered. Inspired by Montesquieu’s republican separation-of-powers and Hans Kelsen’s normative hierarchy, the framework decouples intelligence from sovereignty by embedding systemically consequential AI systems within a layered jurisdictional order composed of an AI Core, an Alignment Traceability Code (ATC), and a Computational Reflective Equilibrium Core (CRE). The framework is explicitly non-universal: constitutional review is activated only when decisions exceed calibrated systemic impact thresholds, preserving full statistical autonomy for low-stakes AI systems. Systemic risk is operationalized through the Systemic Impact Score (SIS), a real-time structural impact function that determines when decisions exit the domain of ordinary automation and enter computational constitutional review. SIS does not encode values; it encodes jurisdictional thresholds, introducing a reproducible procedural calibration protocol. Legitimacy thus resides not in fixed numbers but in traceable calibration processes satisfying procedural invariants: auditability, contestability, versioning, and accountability. By introducing Habeas Log as a projected fourth-generation Human Right, the architecture ensures that AI remains a tool of institutional stability, guaranteeing that ultimate decision-making authority resides strictly within the sovereign's jurisdictional boundaries. Beyond instantaneous impact detection, the architecture integrates Normative Entropy (Sₙorm) as a dynamic indicator of institutional stability. Entropy accumulation modulates jurisdictional tolerance, tightens action thresholds, shifts burdens of proof, and triggers phase transitions from operational autonomy to cautionary or restrictive regimes, culminating in constitutional paralysis when coherence collapses. This transforms safety from a behavioral property into a form of institutional homeostasis. The framework is designed to be politically agnostic yet procedurally constrained. Through the concept of Hybrid Flow, it demonstrates how constitutional AI governance can operate across divergent political systems, while preserving invariant requirements of traceability, reviewability, and accountability. The ATC further functions as an integrity jurisdiction, detecting deception, telemetry manipulation, or systemic compromise independently of impact evaluation. Comparative simulations against Hierarchical Reinforcement Learning (HRL) baselines show that optimization-based systems cannot represent jurisdictional limits, exhibit monotonic normative drift, and fail to recognize structural incoherence. Hybrid CRE, by contrast, displays stable escalation behavior, explicit limit recognition, and resistance to gaming through multi-layer verification and stochastic threshold defenses. These results suggest that robust AGI and high-impact AI governance may emerge not from better reward models but from machine-speed constitutional orders capable of enforcing jurisdictional boundaries, preserving human sovereignty, and dynamically regulating institutional legitimacy. Principio del formulario 中文摘要 (修订版) 当前的人工智能治理方法主要将对齐问题框定为一种行为优化问题, 依赖奖励塑形、约束机制或事后监督。本论文提出一种不同的范式: 一种“宪制型计算架构”, 在其中, 人工决策并非通过行为引导实现, 而是通过制度结构加以组织。该框架受到孟德斯鸠共和制分权理论与凯尔森规范层级理论的启发, 通过将具有系统性影响的人工智能系统嵌入分层的司法管辖秩序之中, 实现智能与主权的分离。该秩序由 AI 核心 (AI Core) 、对齐可追溯性代码 (Alignment Traceability Code, ATC) 以及计算反思均衡核心 (Computational Reflective Equilibrium Core, CRE) 构成。 该框架明确并非普遍适用于所有人工智能系统: 只有当决策超过经校准的系统性影响阈值时, 才会触发宪制审查机制;对于低风险系统, 则保持完全的统计自治。系统性风险通过“系统影响评分” (Systemic Impact Score, SIS) 进行操作化, 该评分是一种实时结构性影响函数, 用以判定何时决策行为超出常规自动化范围并进入计算宪制审查领域。SIS 并不编码价值, 而是编码司法管辖阈值, 并引入可复现的程序化校准协议。因此, 合法性并不建立在固定数值之上, 而存在于满足程序不变性的可追溯校准过程中, 包括可审计性、可争议性、版本化以及责任追溯。通过引入“日志人身保护权” (Habeas Log) 作为一种拟议中的第四代人权, 该架构确保人工智能始终作为制度稳定性的工具, 保证最终决策权严格保留在主权司法边界之内。 除即时影响检测外, 该架构还整合了“规范熵” (Normative Entropy, Sₙorm) 作为制度稳定性的动态指标。熵的累积调节司法容忍度、收紧行动阈值、转移举证责任, 并触发从运行自治向审慎或限制性制度的阶段转变, 最终在规范一致性崩溃时进入“宪制瘫痪”。这一机制将安全从一种行为属性转化为制度稳态的形式。 该框架在政治层面保持中立, 但在程序层面受到约束。通过“混合流” (Hybrid Flow) 的概念, 本文展示了宪制型人工智能治理如何在不同政治体系中运行, 同时保持可追溯性、可复审性与问责性的程序不变要求。ATC 还作为一种完整性司法域运作, 能够独立于影响评估检测欺骗行为、遥测操控或系统性破坏。 与分层强化学习 (Hierarchical Reinforcement Learning, HRL) 基线模型的对比模拟表明, 基于优化的系统无法表示司法管辖界限, 表现出单调的规范漂移, 并且无法识别结构性不一致。相比之下, Hybrid CRE 展现出稳定的升级行为、明确的边界识别能力, 以及通过多层验证与随机阈值防御机制实现的抗操纵能力。 这些结果表明, 稳健的 AGI 与高影响人工智能治理 或许并非源于更优的奖励模型, 而是源于能够在机器速度下执行宪制秩序的系统, 这些秩序能够强制执行司法边界、维护人类主权, 并动态调节制度合法性。 Final del formulario Keywords: Constitutional AI, Hybrid CRE, Systemic Impact Score (SIS), Normative Entropy, Alignment Traceability Code (ATC), Jurisdictional Logic, Institutional Homeostasis, AGI Governance. 关键词: 宪制型人工智能 (Constitutional AI), 混合 CRE (Hybrid CRE), 系统影响评分 (Systemic Impact Score, SIS), 规范熵 (Normative Entropy), 对齐可追溯性代码 (Alignment Traceability Code, ATC), 司法管辖逻辑 (Jurisdictional Logic), 制度稳态 (Institutional Homeostasis), 通用人工智能治理 (AGI Governance) Contact: gastonrey76@gmail. com