From Portals to Case Graphs: A Reference Architecture and Benchmark for Safety Investigation Operations with Agentic Orchestration

Abstract

National safety investigation organizations operate a continuous lifecycle spanning public occurrence intake, operations-center triage, multi-source situational context, investigation workflows, evidence handling, and controlled publication. These functions typically exist as separate portals lacking unified, case-centric back-ends that support cross-channel consolidation, event-time enrichment, auditability, and policy-governed automation. This paper presents N-iSOP, a reproducible reference architecture and benchmark specification for multimodal safety investigation operations using the Nigerian Safety Investigation Bureau online service as a motivating case. N-iSOP retargets an established three-layer functional model (integration, processing/intelligence, and control/service) and cloud-edge-device implementation patterns into a bureau-grade "case graph" platform with event-driven microservices, stream processing, and zero-trust security controls (Adepoju & Segun, 2025). The core extension is agentic orchestration: a policy-bound, tool-using agent that consumes case events and generates auditable proposals for deduplication, triage, follow-up tasks, incident-window context enrichment, and publication metadata. Reliability is enforced through knowledge-grounded prompt chaining and structured outputs, aligned with KG+LLM emergency decision support and disaster management LLM governance (Chen et al., 2024; Xu et al., 2025). We define a benchmark suite with minimal datasets measuring (i) cross-channel entity resolution quality (Binette & Reiter, 2023; Koumarelas et al., 2020), (ii) triage and enrichment latency under cloud-edge placement (Veith et al., 2023; Xhafa et al., 2020; Zeuch et al., 2022), (iii) authorization/audit overhead under Zero Trust (Rose et al., 2020; Sengupta & Lakshminarayanan, 2021), and (iv) custody trace completeness for digital evidence (Nath et al., 2024; Malik et al., 2023).