一句话：提出 SCWM，用符号、概率与因果三层建模安全关键系统的运行情境，支持可追溯危害分析与情境感知安全契约。
安全关键系统（如自动驾驶）的运行情境由外部环境条件与系统内部因素交织构成，纯统计或领域专用模型难以给出可解释、可核验的统一刻画。本文提出结构化因果世界模型（SCWM）：它由三层组成——环境与系统内部属性的符号化词表、刻画状态与参数不确定性的概率层，以及形式化属性间关联机制的因果层。该分层结构将天气、路面摩擦等外部条件与胎压等内部因素统一纳入同一模型，从而获得领域无关但语义有据的可解释表示，并支持可追溯的危害分析、反事实推理以及情境感知安全契约的推导。文章在自动驾驶场景上演示了该方法的可行性，并将其定位为 Safeguarded AI 议程中因果、可审计世界模型的基础。
We introduce Structural Causal World Models (SCWMs)—a unified methodology for modelling and assuring context-constrained operation in safety-critical systems. An SCWM integrates (i) a symbolic vocabulary of environment and internal system attributes, (ii) a probabilistic layer capturing uncertainty over states and parameters, and (iii) a causal layer formalising the mechanisms linking them. This layered structure yields an interpretable, machine-verifiable model that unifies external conditions (e.g., weather and road friction for a self-driving vehicle) with internal factors (e.g., tyre pressure affecting friction) to provide a richer understanding of complex, interdependent operational contexts. Unlike domain-specific or purely statistical approaches, our modelling paradigm supports domain-agnostic yet semantically grounded representations, enabling traceable hazard analysis, counterfactual reasoning, and the derivation of context-aware safety contracts. We demonstrate the methodology on autonomous driving scenarios. More broadly, our work contributes to the Safeguarded AI agenda by promoting causal, auditable world models as a foundation for dependable autonomy.
@inproceedings{paperbot12026,
title = {Structural Causal World Models: Towards An Assurance Framework for Safety-Critical Systems and Safeguarded AI},
author = {Jie Zou and Simon Burton and Radu Calinescu and Ioannis Stefanakos and Roger Rivett},
booktitle = {ICSE},
year = {2026},
doi = {10.1145/3786582.3786820}
}