From Prediction to Compilation: A Manifesto for Intrinsically Reliable AI

从预测到编译：本质可靠 AI 的公理化宣言 From Prediction to Compilation: An Axiomatic Manifesto for Intrinsically Reliable AI ________________________________________ 定义｜Definitions 定义 1（预测系统） Definition 1 (Predictive System) 以概率方式输出未来状态或动作分布的系统。 A system that outputs future states or actions in probabilistic form. 定义 2（执行系统） Definition 2 (Executable System) 其输出将直接驱动物理世界状态变化的系统。 A system whose outputs directly cause physical state changes. 定义 3（执行合法性） Definition 3 (Execution Legitimacy) 一个输出在物理上存在唯一、确定、可验证执行路径的性质。 The property that an output admits a unique, deterministic, and verifiable physical execution path. ________________________________________ 核心命题｜Core Proposition 命题 1 Proposition 1 任何缺乏执行合法性的系统，不得被视为可靠的执行系统。 Any system lacking execution legitimacy cannot be considered a reliable executable system. ________________________________________ 公理体系｜Axiom System 公理一：非臆想公理 Axiom I: Non-Hallucination Axiom 系统的任何输出，若不存在唯一的物理执行映射，则该输出在执行层面是非法的。 Any system output that lacks a unique physical execution mapping is illegal at the execution level. ________________________________________ 公理二：预测–执行分离公理 Axiom II: Prediction–Execution Separation Axiom 概率系统仅允许生成目标、约束与假设，不得直接生成可执行动作。 Probabilistic systems may generate goals, constraints, and hypotheses, but must not generate executable actions directly. ________________________________________ 公理三：编译优先公理 Axiom III: Compilation Primacy Axiom 所有可执行动作，必须由确定性物理模型与约束条件编译生成。 All executable actions must be compiled from deterministic physical models and constraints. ________________________________________ 公理四：拒绝合法性公理 Axiom IV: Refusal Legitimacy Axiom 在约束冲突或无可行解时，系统拒绝执行构成合法且必要的输出。 When constraints conflict or no feasible solution exists, refusal to act is a valid and necessary output. ________________________________________ 公理五：能效合理性公理 Axiom V: Energy Rationality Axiom 在确定性问题中使用概率搜索构成不必要的能量与算力浪费。 Using probabilistic search to solve deterministic problems constitutes unnecessary energy and computational waste. ________________________________________ 推论｜Corollaries 推论 1 Corollary 1 生成式模型在未经编译层约束的情况下，不能直接接管物理执行权。 Generative models must not assume physical execution authority without a compilation layer. 推论 2 Corollary 2 世界模型适用于认知与规划层，但不构成执行充分条件。 World models are sufficient for cognition and planning, but not for execution. 推论 3 Corollary 3 模型预测控制（MPC）及其等价方法在执行层中具有结构上的必然性。 Model Predictive Control (MPC) and equivalent methods are structurally necessary at the execution layer. ________________________________________ 结论｜Conclusion 结论 Conclusion AI 系统若要在高风险、不可逆的现实环境中运行， 其核心能力不应被定义为预测准确性， 而应被定义为执行合法性。 An AI system intended to operate in high-risk and irreversible environments must be evaluated not by predictive accuracy, but by execution legitimacy. 从预测到编译，不是实现路线之争， 而是可靠智能的必要条件。 From prediction to compilation is not an implementation preference, but a necessary condition for reliable intelligence.