Assurance-Centered Agentic AIOps (ACAA)

Industrial cyber operations in local-cloud IIoT/OT environments face a persistent systems-level gap: analytical components for detection, triage, and response exist in isolation, yet decision quality degrades at handoff boundaries where competing objectives -- security assurance, operational continuity, governance compliance, and human accountability -- must be reconciled under uncertainty. This thesis addresses that gap through the design, implementation, and proof-of-architecture evaluation of Assurance-Centered Agentic AIOps (ACAA), a layered decision-support architecture that composes statistical inference, machine learning, deep learning, generative AI, and agentic orchestration under deterministic policy controls, explicit uncertainty handling, and human-in-the-loop authority. The architecture is developed across seven progressive project chapters. A reproducible data workflow (P1) establishes ingestion contracts and provenance discipline over OpenRCA telecom telemetry. Statistical inference (P2) extracts governance-relevant structure from heterogeneous cyber observability data. Leakage-aware machine learning (P3) produces vulnerability prioritization priors over NVD/CISA KEV metadata under extreme class imbalance. Deep learning (P4) applies sequence and representation models to LANL cybersecurity telemetry with controlled ablations and guardrails. A generative RCA layer (P5) synthesizes bounded, confidence-labeled narrative hypotheses for analyst scaffolding. Policy-gated multi-agent orchestration (P6) operationalizes deterministic safety boundaries around adaptive reasoning. Finally, an integrative synthesis (P7) introduces parallel contested orchestration, where dual-branch reasoning (assurance versus continuity) is adjudicated by a meta-orchestrator with explicit human-in-the-loop escalation. The central finding is that decision quality in safety-critical cyber AI systems is a property of composed workflows rather than any single model. Contested orchestration produced meaningful branch differentiation, policy-gate invariants held without violation, and governance traceability was maintained end-to-end across all analytical layers.