Golden-Time Policy-Grounded Natural-Language Control for Public Safety Cyber-Physical Systems

During the golden-time phase of urban disasters, emergency response systems must issue timely directives that remain compliant with jurisdictional policies and regulations, even when early human reports precede reliable CPS telemetry and network connectivity is degraded. This article presents a policy-grounded, multilingual natural-language control layer for public safety cyber-physical systems (CPS), designed to operate under strict latency, accountability, and connectivity constraints. The proposed system converts short, unstructured incident reports and CPS state summaries into role-specific, CPS-ready directives by integrating three components: 1) clause-level retrieval from a fixed and locally verified policy snapshot; 2) grounded reasoning using a lightweight local language model; and 3) deterministic release-time safety validation that enforces regulatory constraints and hazard-specific prohibitions. All inference-time operations are self-contained and support offline execution. We evaluate the system across English, Japanese, and Korean using 3,600 executed trial sets (1,200 per language), where each trial set is a five-query bundle spanning five hazard categories executed sequentially under identical deterministic settings. Under both connectivity-available and offline operation, the proposed engine achieves a median end-to-end latency of 13.8 s per trial set while maintaining 95.0% clause-level policy traceability on released directives and substantially suppressing unsafe releases under adversarial phrased inputs. These results demonstrate that deterministic policy grounding and offline-capable local inference can serve as a deployable and auditable control layer for CPS-assisted disaster response under golden-time constraints.