What We Know We Do Not Know: Systematic Risk Identification from Contradictory Safety Documentation

Ensuring accountability in human-centered safety-critical systems, such as autonomous vehicles and medical devices, requires analysis methods that expose latent hazards implied by incomplete documentation. These systems must maintain safety, satisfy normative and ethical requirements, comply with regulations, and manage complex human–machine interactions. Existing methods fall short in verification of documented properties and attribution of responsibility; a key cause is the presence of undocumented safety-relevant predicates (known unknowns) that remain undetected. We propose a Kripke-based approach that connects the Open World Assumption of ontological knowledge graphs with Closed World Assumption verification by exhaustive enumeration of feasible completions. The approach exposes latent hazards that remain implicit under Open World Assumption validation of incomplete documentation. Our methodology addresses the lack of precision in the extraction of knowledge graphs from informal documentation, which prevents reliable automated validation and formal correctness guarantees. We construct Kripke structures whose states are models that satisfy knowledge graph configurations derived from potentially contradictory documentation. The framework incorporates expert feedback to align models with domain knowledge and regulatory expectations. A case study of a simplified automotive seatbelt warning system demonstrates how the framework compares documented safety properties with latent-hazard criteria. The case reveals a 40.9% coverage gap in safety classification: many states classified as safe by the specification nevertheless harbor latent hazards resulting from sensor–actuator conflation and unobservable safety properties.