ABSTRACT. Volunteer disaster responders are often the first people survivors turn to for emotional support, yet most lack clinical training in structured crisis intervention. Existing resources-static manuals, self-paced courses, and text-only AI tools-do not address the real-time cognitive demands of a live survivor encounter, where a responder must simultaneously listen, assess emotional state, recall protocol, and decide what to say next. We introduce "JulienPFA", an audio-native AI assistant that delivers real-time, human-in-the-loop Psychological First Aid (PFA) guidance during active disaster response. JulienPFA listens to the live survivor--responder conversation through the "Gemini-live-2.5-flash-native-audio" model, infers the survivor's arousal state directly from vocal affect (hyperaroused, hypoaroused, or regulated), and retrieves protocol-aligned recommendations via a retrieval-augmented generation (RAG) pipeline grounded in the New York State PFA manual. The system surfaces structured guidance---what PFA element to address, what to say, and when to escalate---while the responder remains fully engaged with the survivor. To validate this architecture, we benchmark five open-weights and cloud-based language models across inference latency, adversarial safety boundary adherence, and PFA protocol compliance. Results show that locally hosted models either exceed acceptable latency thresholds for crisis interaction or fail critical safety boundaries (e.g., suggesting relaxation strategies instead of escalating suicidal ideation), confirming that unconstrained text generation without domain-specific RAG grounding is insufficient for safe PFA delivery. JulienPFA addresses these gaps through the combination of native audio affect detection, strict retrieval-bounded generation, and structured clinical escalation pathways.

ABSTRACT. Natural disasters destroy terrestrial communication infrastructure precisely when connectivity is most critical. Cellular towers, fiber paths, and power systems fail simultaneously, leaving residents and responders without access to emergency services. While low Earth orbit (LEO) satellite systems can restore backhaul, isolated hotspots do not address mixed-device access, local service continuity, or metro-scale coordination. This paper proposes Resilience Nodes, a passive, fault-tolerant metro-area network (MAN) architecture that pre-positions low-cost emergency network units at critical facilities. Nodes remain dormant during normal operations and activate during disasters to provide Minimum Viable Connectivity (MVC), which includes emergency alerts, SOS messaging, shelter information, and responder coordination. The architecture employs point-to-point directional antennas forming a mesh backbone across 1-3,km urban links, with selected gateway nodes providing LEO satellite backhaul. We present an implementation specification including a per-node bill of materials, an energy budget derived from a 512 Wh LiFePO4 pack, and a token-bucket Minimum Viable Connectivity scheduler. We evaluate feasibility through (i) an RF link budget showing 27 dB fade margin at 2 km, (ii) an analytical worst-case latency bound of approximately 100 ms for P1 alerts under 1000-user saturation, with the discrete-event simulation measuring P1 p99 of 28 ms in practice, (iii) an energy budget deriving 48 to 72 hour autonomy from the LiFePO4 pack under graded duty cycling, and (iv) a county-scale cost comparison showing that a 25-node deployment at $15K-25K is an order of magnitude less than a single Cell-on-Wheels deployable.

ABSTRACT. This paper examines a persistent gap in emergency management between the tools and frameworks that exist and those that practitioners can actually deploy during disasters. The study was conducted as part of a National Science Foundation (NSF) funded research, combining with secondary research, with 15 semi-structured interviews across Texas, Maryland, and Florida, representing city and county government, state agencies, universities, nonprofits, and the private sector. Findings show that communication failure remains the most consistent operational problem, driven by radio interoperability breakdowns, cybersecurity restrictions, the absence of real-time situational awareness, and dependence on infrastructure that often fails during disasters. Additional findings identify undocumented resource knowledge, fragmented technology systems, persistent equity gaps, and weak translation of after-action lessons into practice. In addition, the paper proposes a Knowledge Management System (KMS) framework organized around eight functional layers: resource access, operational playbooks, partner registries, situation reporting, decision support, community-facing access, after-action learning, and interoperability infrastructure. The paper argues that the core challenge in emergency management is not a lack of technology, but a lack of systems designed around the operational realities, constraints, and adoption barriers practitioners face in the field.

ABSTRACT. Energy-intensive public safety operations rely on a mix of Emergency Operations Centers (EOCs) dashboards, field displays, cameras, sensors, and comms that collectively shape response times and safety outcomes. This paper presents FAMIUM Wattlify, an AI-enabled energy-intelligence platform for Emergency Operations Centers (EOCs) and field displays. The system combines a high-fidelity digital twin of public safety energy flows (EOCs, video walls, sensors, radios, and field devices) with a retrieval-augmented generation (RAG) layer and a safety-minded large language model (LLM) to deliver real-time, auditable energy guidance without impacting live operations. Wattlify ingests device-level energy measurements, energy-prediction models, CO2 emission data, and eco-mode catalogs, while RAG sources extend device specifications and safety procedures from internal and public documents. The LLM orchestrates tool calls (one tool per turn) and generates transparent, user-oriented explanations with verifiable references. Evaluation results indicate that end-to-end latency averaged 5.5 s across scenarios, with the Model Code (RAG) path reaching a 95th percentile latency of 15.4 s, predominantly due to LLM inference and network variability rather than backend computation. Retrieval evaluation showed Acc@5 of up to 91% on CO2 queries and 80% on TV eco-mode queries, with query expansion providing consistent gains on the CO2 subset.

ABSTRACT. Cooperative autonomy between drones and connected autonomous vehicles (CAVs) is increasingly relied upon in disaster response and search-and-rescue (SAR) operations, where timely, trustworthy data exchange is critical for saving lives. However, these systems often operate across organizational boundaries and in adversarial, infrastructure-denied environments, making traditional security and trust mechanisms insufficient. This paper proposes a hierarchical and resilient provenance framework that enables cross-domain trust establishment and real-time data verification for cooperative drone–vehicle systems. In the proposed approach, resource-constrained drones employ hardware-rooted security using Physically Unclonable Functions (PUFs) to generate tamper-resistant provenance for sensed data, while more capable CAVs act as local provenance aggregators and ledgers. To address contested environments, the framework incorporates multi-modal provenance integrity, allowing agents to validate data using redundant sensory perspectives and interaction histories. A fuzzy-logic–based certainty engine evaluates incoming data by fusing provenance, interaction, and contextual evidence, producing dynamic confidence scores for safety-critical decisions. This work demonstrates how provenance can evolve from a passive audit mechanism into an active, real-time trust and resilience enabler for cooperative cyber-physical systems in disaster-response scenarios.

ABSTRACT. Public safety and disaster relief organizations require resilient, rapidly deployable communication infrastructure capable of operating in challenging environments where public terrestrial networks are unavailable or compromised. This paper presents the RooiBOS project’s approach to develop missioncritical 5G communication demonstrators for Public Protection and Disaster Relief (PPDR) operations. We describe a comprehensive system architecture integrating nomadic 5G base stations, device-to-device Sidelink communication, satellite backhaul, zerotrust overlay networking, and distributed edge computing. The paper details the technology components, explains the system design and also presents mechanisms for PPDR operations support. The presented approach is implemented through an Endto- End (E2E) demonstrator not only with isolated site nomadic deployments, but also spanning geographically separated sites. It demonstrates the feasibility of extending existing systems to broadband 5G infrastructures while addressing key challenges including coverage limitations, system resilience, security, management and operations mechanisms tailored for PPDR scenarios, and autonomous operation in infrastructure-compromised scenarios.

ABSTRACT. Reliable and resilient communication systems are indispensable for first responders, enabling rapid coordination and effective emergency response. However, traditional communication networks frequently encounter congestion, interoperability failures, and infrastructure collapse during large-scale disasters. To address these deficiencies, specialized networks such as the First Responder Network Authority (FirstNet) have been developed, leveraging advancements in Long-Term Evolution (LTE), Fifth-Generation New Radio (5G NR), and priority access mechanisms to enhance reliability and coverage. This comprehensive review examines the technological evolution of first-responder communication systems from legacy Land Mobile Radio (LMR) and Project 25 (P25) systems through modern broadband solutions. We systematically analyze key enablers including network prioritization with Quality of Service (QoS), Priority, and Preemption (QPP); dedicated spectrum allocation on Band 14 (758–768/788–798 MHz); Mission Critical Push-to-Talk (MCPTT), Mission Critical Video (MCVideo), and Mission Critical Data (MCData) standards defined across 3GPP Releases 13–18; network slicing for dedicated emergency virtual networks; and Multi-access Edge Computing (MEC) for ultra-low-latency field processing. Additionally, this study assesses the integration of artificial intelligence and machine learning for predictive network management, digital twin technology for infrastructure resilience simulation, Internet of Things (IoT) sensor ecosystems for enhanced situational awareness, and satellite communication systems, including emerging Low Earth Orbit (LEO) constellations, for connectivity in infrastructure-denied environments. We further examine real-world deployments through case studies encompassing Hurricane Katrina (2005), the September 11 attacks (2001), Hurricane Harvey (2017), the California wildfires (2018–2025), and the 2011 Great East Japan Earthquake, alongside global initiatives including the United Kingdom’s Emergency Services Network (ESN), the European Union’s Public Protection and Disaster Relief (PPDR) framework, and South Korea’s PS-LTE SafeNet. By synthesizing recent advancements across more than 120 scholarly and technical sources, this review provides a forward-looking roadmap addressing Sixth-Generation (6G) networks, terahertz communications, holographic situational awareness, augmented and extended reality for field operations, blockchain-secured data sharing, and cybersecurity frameworks. We conclude with policy recommendations and identify critical research gaps necessary to ensure a seamless, intelligent, and globally interoperable communication infrastructure for first responders.

ABSTRACT. Autonomous unmanned vehicles are vital to tactical missions, mission-critical public-safety operations like search and rescue and disaster response. However, their reliance on open wireless links and standard Robot Operating System (ROS 2) middleware exposes a broad cyber-physical attack surface. A compromise of these systems can disrupt real-time control loops, leading to mission failure or asset loss in high-stakes environments. This paper presents and empirically evaluates a layered, context-aware cybersecurity framework enforcing Zero-Trust principles for a multi-node robotic fleet over Wi-Fi. The framework integrates an active hardware root of trust (TPM 2.0), centralized in-band and out-of-band SIEM telemetry monitoring (ELK Stack and Kismet), and a non-cryptographic Inter-Packet Delay (IPD) timing watermark. Evaluated on a live ROS 2 mobile testbed under multi-layer exploits (OSI Layers 2–5), results demonstrate that while volume-based filters isolate brute denial-of-service floods, tracking the statistical sample kurtosis ($K$) of the embedded IPD watermark exposes stealthy Man-in-the-Middle command injections with complete detection accuracy without payload overheads.

ABSTRACT. Integrated Access and Backhaul (IAB) enables 5G public-safety coverage by reusing the NR radio interface for access and backhaul, but the shared-medium half-duplex design creates a multi-hop cascade problem: a mission-critical surge at a leaf node overwhelms upstream hops that static slicing and local-only priority cannot prevent. We propose a predictive preemptive backhaul borrowing framework in which a queue-growth predictor fires a path-scoped Borrowing-Advisory before overflow, triggering a coordinated, temporary downgrade of non-critical commercial flows. Evaluation in a discrete-time system-level simulator (1\,800 runs, 72 scenario configurations, five policies including a tabular Q-learning baseline) and an ns-3 5G-LENA proof-of-concept shows: over the full sweep, predictive borrowing and Q-learning achieve near-equal aggregate PS delivery (0.939 vs.\ 0.930); in the hardest stressed regime the heuristic leads by 12.5 points (0.884 vs.\ 0.759) because its trigger directly encodes signaling-plus-switching delay. TDD-slot borrowing via \texttt{SetPattern()} is implementable without modifying core 5G-LENA internals. Commercial delivery remains non-zero under all borrowing policies. The Borrowing-Advisory maps to an O-RAN near-RT RIC xApp and represents a new scheduler extension not present in current 3GPP specifications.

ABSTRACT. Cell-Free (CF) massive Multiple-Input Multiple-Output (MIMO) can improve wireless resilience by serving each user through multiple distributed Access Points (APs), thus providing macro-diversity and reducing the dependence on a single radio site. However, when user-centric serving clusters are selected only according to channel strength, the resulting APs may be geographically concentrated around the user and exposed to spatially correlated failures during localized disaster events. This paper addresses this limitation by designing a reliability-aware AP selection strategy that accounts for both channel quality and spatial diversity within each serving cluster. The proposed approach uses large-scale fading information to select APs that are both channel-effective and spatially diverse, without requiring explicit location data. The resulting clusters are evaluated under different disaster patterns to assess whether spatially diversified AP selection can better preserve service continuity after infrastructure failures. The results show that reliability-aware clustering improves post-disaster robustness while preserving the pre-disaster performance of conventional channel-based clustering, highlighting the role of cluster design in making CF massive MIMO a resilient connectivity solution for disaster management scenarios.

ABSTRACT. The ongoing digital transformation is driving the development of governmental web platforms, making public safety heavily reliant on Citizen-to-Government (C2G) platforms such as emergency portals and public warning APIs. These interfaces depend on the Web Public-Key-Infrastructure (PKI) to guarantee data authenticity and confidential communication. However, the integration of PKI into such high-sensitivity environments introduces new attack surfaces for adversarial exploitation. This paper systematizes the current threat landscape of hierarchical PKIs and evaluates modern hardening strategies against operational realities. Through a comparative TLS probing analysis of 7,880 global domains and 745 governmental endpoints, my findings demonstrate that while modern interfaces successfully adopt 90-day lifecycles, this transition leads to a dangerous market concentration. Specifically, over 50\% of the US-Gov sector relies on a single issuer, creating a systemic single point of failure. I conclude that the ecosystem resilience is determined by both the enforcement policies of root store operators and the architectural initiatives of Certificate Authorities. To address these challenges, I advocate for strict governmental frameworks, aligned with NIST/BSI recommendations, to standardize security configurations across public safety ecosystems.

ABSTRACT. Unmanned Aerial Vehicles (UAVs) depend on continuous wireless communication, making them vulnerable to cyber-physical attacks such as replay, denial-of-service, evil-twin, and false data injection. Centralized intrusion detection requires transmitting raw UAV telemetry, raising privacy, bandwidth, and latency concerns. This paper proposes a split-learning based multi-label intrusion detection system that divides a neural network between a UAV client and a ground-station server. The UAV processes initial layers and sends only intermediate activations, preserving data privacy while enabling collaborative training. Using a real UAV communication dataset with multi label attack annotations, the system achieves 81.5% accuracy and strong macro-F1 performance. The results show that split learning effectively detects overlapping attacks while reducing communication overhead in UAV cyber-physical environments.

ABSTRACT. The platform-train interface (PTI) represents one of the most persistent safety hazard zones in public railway networks, with conventional countermeasures relying on generic, repetitive public address announcements that rapidly degrade into background noise through auditory habituation. This paper presents YellowLine AI, an end-to-end, edge-native vision-language framework that transforms existing, unmodified CCTV infrastructure into an active, real-time behavioural nudging system, without relying on facial recognition or any form of personally identifiable information. When a passenger crosses a geometric danger zone, a lightweight YOLOv11n object detector extracts the pedestrian crop, which is routed to a PromptPAR vision-language model trained on a unified corpus of 113,042 samples across 67 visual attributes. The model extracts transient soft biometrics such as clothing colour, style, and accessories, achieving a mean accuracy of 84.00 percent and an instance F1-score of 87.50 percent. These attributes are orchestrated through an edge-deployed Small Language Model to dynamically synthesise hyper-personalised safety announcements, which are delivered via the Kokoro text-to-speech engine directly to the station public address system. The complete pipeline achieves a mean end-to-end latency of 1,083 milliseconds on constrained edge hardware, well within the critical 2-second intervention window. A live pilot deployment at Ribble Station, validated with industry partner Fujitsu, confirmed seamless legacy infrastructure integration and demonstrated that personalised, attribute-driven nudges produce noticeably higher rates of immediate physical compliance compared to generic pre-recorded warnings. To the best of the authors' knowledge, this is the first implementation combining pedestrian attribute recognition with generative language models for automated behavioural nudging at the PTI, establishing a scalable, privacy-compliant paradigm for AI-driven public transport safety.

ABSTRACT. Public transit systems are increasingly integrating digital technologies, including Internet of Things (IoT) sensors, autonomous navigation systems, and Vehicle-to-Everything (V2X) communications, to enhance operational efficiency, safety, and service reliability. While these technological advancements are transforming modern transportation, they also expand the cyberattack surface, exposing transit infrastructures to threats such as data breaches, signal spoofing, ransomware, and other cyberattacks. Although numerous technical security solutions have been proposed, a comprehensive cybersecurity governance model that addresses the socio-technical dimensions of public transit security—including human, organizational, regulatory, and policy-related factors—remains largely unexplored. This study proposes a hybrid framework that integrates artificial intelligence-driven cybersecurity with established cybersecurity governance principles. Specifically, the Google Secure AI Framework (SAIF) is adopted to guide AI-enabled cybersecurity practices, while the ISACA COBIT framework is utilized to establish governance, risk management, and compliance mechanisms. By combining these complementary approaches, the research introduces a novel cybersecurity governance framework for public transit systems. Although the framework has not yet been empirically validated and specific use cases are outside the scope of this study, the paper discusses key socio-technical challenges and implementation considerations associated with deploying such a framework in complex public transit environments.

ABSTRACT. The increasing use of Artificial Intelligence (AI), along with its integration into Supervisory Control and Data Acquisition (SCADA) systems has led to a transformation of how critical power systems operates. The benefits associated with the inclusion of AI into SCADA include improved predictive maintenance, fault detection, load forecasting, and adaptive control capabilities which enhance operational efficiencies and situational awareness. However, these same advantages also increase the cyber-attack surface area of Operational Technology (OT) environments. In particular, traditional perimeter-based security models are poorly suited for preventing or mitigating coordinated cyber-physical attacks within converged IT/OT environments that include legacy devices with limited security capabilities. To address this challenge, this paper presents a zero-trust based, AI-enhanced cyber-resiliency architecture applicable to AI-enabled SCADA systems. This framework incorporates identity centric access control, micro segmented OT networks, hybrid anomaly detection algorithms, and resilience scores aligned to the National Electric Reliability Corporation (NERC) Critical Infrastructure Protection (CIP) standards, the U.S. Department of Commerce National Institute of Standards and Technology (NIST) Special Publication (SP) 800-82 standards, and the International Electrotechnical Commission (IEC) 62443 standards. Experimental results using simulated testing indicate the proposed architecture reduces the Lateral Propagation Factor (LPF) by 71%, reduces the Mean Time to Detect (MTTD) by 45% and increases the composite Resilience Index from .31 to .84 when comparing to a flat-network baseline. This architectural model introduces a quantifiable Resilience Index based on the lateral propagation factor, detection latency, and operational degradation, thus allowing for the transition from a static compliance based cyber defense to a layered measurable resiliency based cyber defense.

ABSTRACT. We present OOBWatch-S3, a low-power ESP32-S3/Zephyr watchdog for deployable public-safety wireless networks that monitors its own Wi-Fi/BLE link layer and UDP-over-Wi-Fi heartbeat, independent of gateway, RIC, or cloud telemetry, and writes an HMAC-SHA256 tamper-evident incident log. An initial n=50 cold-boot campaign reported 35/50 alarm fires; we retract the original detection interpretation: instrumentation traces the firings to a Zephyr v3.7-LTS hal_espressif first-scan-at-boot race (upstream PR #106329, not backported to v3.7-LTS at audit). Three contributions. (i) We report this as an experience report on a Zephyr v3.7-LTS hal_espressif cold-boot Wi-Fi race family discovered through deployment-quality testing: a pre-registered cold-boot isolation probe registered V3 inconclusive per pre-reg M-11, which surfaced three manifestations—SCAN-path backport request #110290 (upstream maintainer opened backport PR #110387 in response, 2026-06-02), CONNECT-path issue #52868 (documents -EIO from data->state != ESP32_STA_STARTED; we re-engaged on 2026-06-02 with our cold-boot reproduction), and an application-side L4-readiness API-contract gap against the ESP-IDF station-scenarios contract, each with a concrete upstream artifact or workaround. (ii) A microwave-oven Faraday-cage mechanism witness, N=4 trials yielding 377/377 kind=1 LINK_LOSS alarms (predicates (i)/(iv) co-eligible, not separated) with a NetworkManager-heuristic >=20 dB lower-bound shielding estimate at 2.437 GHz (uncalibrated). (iii) A dev-build real-board characterisation: instrumented link-health latencies, 30/30 software-harness tamper detection across modify/delete/truncate, and a 568 KB image (27.7% of a 2 MB OTA slot); release-build replication queued (registered protocol deviation V4 against EVAL_PLAN §top-level discipline).

ABSTRACT. Large Language Model (LLM) agents are rapidly deployed in healthcare and finance, where mishandled Personally Identifiable Information (PII) causes regulatory violations and patient harm. We present \textit{GuardChain}, a composable, layered middleware framework built on LangChain and LangGraph that integrates deterministic keyword filters, regex-based PII redaction, model-based output safety validation, and Human-in-the-Loop (HITL) approval into a unified, formally characterized pipeline. We model the system as a function composition over agent state, derive a cascade cost model with closed-form savings formula, and introduce the first formal treatment of \emph{emergent PII synthesis} - protected information arising from combining individually benign tool outputs. Based on MedQA-USMLE, FinanceBench, HIPAA-Synth, and Nemotron-PII dataset, GuardChain achieves \(\mathcal{B} = 96.8\%\) harmful-request block rate, \(\mathcal{L}_{\mathrm{PII}} = 1.2\%\) leakage, and \(F_1 = 0.95\) (\(\pm 0.012\), \(p < 0.001\) vs.\ all baselines), outperforming NeMo Guardrails, Presidio, and single-layer baselines while keeping end-to-end latency under 250\,ms. Our three-tier HITL routing reduces the burden of the human reviewer by \(3\times\) without sacrificing safety coverage.

ABSTRACT. Smart buildings use advanced sensors that can enhance emergency responses when integrated into decision-making. This limits the ability to share verified situational context across buildings during time-critical events such as fires or evacuations. We present the TrustBaaD-trust-aware Building-as-a-Database framework, which enables smart buildings to act as queryable safety-state providers in federated public-safety coordination. Each building maintains a structured local safety database, including sensor data, event logs, building profiles, and peer registries, and responds to signed, time-bounded safety queries from neighboring buildings. TrustBaaD integrates digital signatures, peer consensus, and reputation scoring to verify alerts under adversarial or unreliable conditions, and supports risk-aware response selection based on shared situational context. We evaluate TrustBaaD using a multi-building simulation calibrated with prototype sensor noise. The findings show improved robustness in hazard detection, reduced evacuation time, and reduced shared-exit congestion compared to baseline approaches. Coordinating building systems using queries and trust can enhance situational awareness and decision-making in emergencies.

ABSTRACT. Schools are densely occupied buildings where indoor relative humidity is rarely monitored, yet low humidity weakens respiratory defenses and prolongs airborne pathogen survival. Using a network of affordable indoor sensors, we related indoor humidity to daily medical absences across 152 school days (258 students, 43 staff) in 2024-2025, controlling for regional influenza activity (CDC ILINet). Indoor humidity averaged 31% and was inversely associated with the student absence rate (Spearman r = -0.51; Pearson r = -0.46), an association that persisted after adjusting for influenza (partial r = -0.25) and followed a dose-response pattern. During the heating season, days below 30% humidity showed absence rates about 41% higher than days above 40%. Humidity can be remediated inexpensively. We estimate equipment costs of ~ $5,600 for the site in this study. A decrease in student and staff absenteeism has been shown to improve student performance and reduce school administrative costs. This study demonstrates that low-cost continuous sensing can surface and help manage this largely unseen public-safety risk while reducing the costs of absenteeism.

ABSTRACT. The assessment of chemical ingredients in cosmetic products against individual skin profiles is a largely manual and error-prone task. Existing consumer applications have primarily focused on the branding level and are without the depth of reasoning at the molecular level. Additionally, large language models (LLMs) without retrieval grounding generate hallucinations, resulting in unsafe recommendations and posing safety issues. This work presents a Chemical Compounds Recommendation Engine that integrates Optical Character Recognition (OCR) and a hierarchically structured Retrieval-Augmented Generation (RAG) framework to provide evidence-based recommendations on skincare ingredients. This engine processes a user query (optional product label image and skin type profile), retrieves relevant dermatology transcripts, employs evidence-based reasoning, and guides a structured response in JSON format (each recommendation is evidence-based and cited). Under a strict grounding condition, hallucinations are limited to a configurable evidence-overlap threshold (i.e., the rate = 2.7%). Role-Based Access Control (RBAC) provides a protective separation of doctor-level formulation safety analysis and consumer-level formulation safety analysis within a single microservices architecture (React, Spring Boot, Flask, and MongoDB) Provenance Tracking. Measuring 850 unique query responses, a separation F1 = 0.889 with 97.3% accurate grounding and 3.2 s latency is achieved, far surpassing the reliance on keyword, TF-IDF, ungrounded BERT, and zero retrieval GPT-4 benchmarks.

ABSTRACT. As AI-assisted online adaptive radiotherapy (oART) expands from academic centers to community cancer programs, a central public-safety question emerges: do adopting departments possess sufficient safety infrastructure to support mission-critical AI workflows? Using published marginal distributions of workforce, accreditation, and infrastructure indicators, we construct a synthetic national cohort of 287 US radiation oncology departments and apply the Risk Assessment and Dimensional Adaptive Reliability (RADAR) framework’s resource-tier classification. A multi-criteria scoring model with Analytic Hierarchy Process (AHP) weights assigns departments to three tiers, and a RADAR Safety Readiness Score (SRS) quantifies coverage across all five RADAR dimensions. Results show that 47.0% of departments fall into Tier 3 (low-resource), with mean SRS of 0.138 compared to 0.562 for Tier 1 (Cohen’s d = 4.56). An estimated 43,576 patients annually (59.6%) receive adaptive radiotherapy at below-academic-infrastructure sites, including 17,165 patients at Tier 3 centers where 44.4% lack AI governance protocols and 91.9% lack safety-culture programs. International benchmarking reveals that US Tier 3 departments score below LMIC facilities (0.174). Multiple regression explains 78.4% of SRS variance, with physics FTE and NCI designation as dominant predictors. A minimum SRS threshold of 0.20 would classify 50.5% of departments as non-compliant. These findings underscore the need for resource-stratified safety frameworks as essential public-safety infrastructure for equitable AI deployment. For consistency across AI-enabled imaging workflows, we define CBCT as cone-beam computed tomography