SafeMachine: An evidence-centric architecture for traceable and context-aware machine safety assessment

Machine safety assessment requires the systematic evaluation of regulatory requirements, heterogeneous evidence sources, and expert judgment. In practice, these assessments are often conducted through manual procedures that lack explicit traceability between regulatory obligations, inspection evidence, and resulting compliance decisions. These limitations reduce transparency, reproducibility, and scalability of safety assessment workflows. This paper introduces SafeMachine, an evidence-centric artificial intelligence architecture designed to support traceable machine safety assessment in engineering environments. The proposed system integrates structured regulatory knowledge ingestion, retrieval-augmented reasoning, and a multi-agent orchestration layer that supports requirement-oriented safety evaluation workflows. Rather than treating large language models as autonomous decision-makers, the architecture constrains the model reasoning through explicit evidence retrieval, authority-aware regulatory filtering, and mandatory source attribution. A central contribution of the system is the introduction of a traceability model based on structured artifact dependencies linking regulatory requirements, authoritative document fragments, audit evidence, intermediate assessment artifacts, and final compliance-oriented outputs. This representation enables machine-verifiable audit artifacts and systematic validation of traceability and reproducibility properties. The system is evaluated through ten controlled synthetic industrial scenarios designed to reflect varying levels of contextual completeness, documentation availability, and safety relevance. The evaluation focuses on context-aware evaluability, conservative abstention under insufficient information, and structured traceability of requirements, evidence, findings, and actions. The results show that SafeMachine progressively increases its evaluability as contextual evidence improves, while avoiding unsupported compliance judgments in underspecified scenarios. These findings demonstrate that generative AI can be safely integrated into knowledge-driven engineering decision-support systems when constrained by evidence-centric architectures and explicit traceability mechanisms.