telemetry collection

Ensuring sustained safety, performance stability, and clinical value of deployed medical AI systems remains a critical governance challenge. Method: This study proposes the first post-deployment monitoring framework integrating three dimensions—system integrity, dynamic performance stability, and real-world clinical impact—grounded in three synergistic principles: quantifiable metrics, clear accountability assignment, and closed-loop response orchestration. The framework is compatible with both traditional and generative AI systems and technically integrates runtime error detection, input distribution shift analysis, clinical workflow-embedded feedback collection, and multi-tiered dashboards. Contribution/Results: Validated at Stanford Health Care, the framework reduced mean AI system fault response time by 62%, achieved an 89% early detection rate for critical performance degradation, and produced a reusable, standardized monitoring plan template.

This study addresses the challenge of effectively monitoring early-stage agent systems, where structural flaws often obscure task-level errors. The authors propose a three-dimensional (quality, suitability, efficiency) and three-granularity (intra-run, inter-run, structural) monitoring and triaging framework tailored for low-maturity agent systems. They introduce a novel system maturity staging model based on the coefficient of variation and monitoring granularity, integrated with a severity classification adapted from FMEA to guide human review. The resulting transferable monitoring architecture supports document-driven, multi-stage workflows, enhanced by a synthetic testbed with controlled error injection. Experimental results demonstrate that structural defects significantly mask task-level signals; 97% of issues can be automatically traced, with only 2% requiring human intervention, and each granularity level precisely identifies its corresponding defect type (coefficients of variation: 0.02, 1.25, and 0.00, respectively).

This work addresses the challenge of enforcing temporal safety constraints throughout the lifecycle of black-box AI systems, such as large language models (LLMs), which are inherently difficult to verify. The paper proposes the first offline auditing and online monitoring framework that integrates Linear Temporal Logic (LTL) with machine learning. This framework enables formal verification of complex temporal behavioral specifications by introducing a sampling-driven predictive monitor and an intervenable runtime monitor, effectively preventing policy violations. Experimental results demonstrate that the proposed approach significantly outperforms existing LLM-based evaluators in detecting temporal violations. Notably, it achieves performance on par with or superior to state-of-the-art large models using only a small labeled model, while its intervention mechanism substantially reduces violation rates without compromising task performance.

In production, ML model failures primarily stem from “contextual misalignment”—a deviation of the operational environment from training assumptions—rather than statistical distribution shift; yet existing monitoring methods lack systematic modeling and utilization of contextual information. Method: We propose the first Contextual System–Aspect–Representation (C-SAR) conceptual framework for ML monitoring, derived from a meta-analysis of 94 cross-domain empirical studies. It distills 20 reusable contextual patterns and establishes a context-driven monitoring taxonomy and pattern-mapping atlas. Contribution/Results: C-SAR shifts ML monitoring from ad hoc, intuition-based “tea-leaf reading” to systematic, context-aware “map-making,” significantly enhancing interpretability, root-cause traceability, and actionable intervention. The framework provides both theoretical foundations and practical guidelines for next-generation context-aware ML monitoring tools.

This study addresses the persistent challenge that large-scale AI systems often achieve formal compliance while still generating substantive harms. It introduces a political economy framework into AI accountability research, proposing a sequential game-theoretic model to characterize strategic interactions among AI vendors, deployers, and regulators under conditions of switching costs and evidentiary dependence. The analysis reveals the conditions under which an “agency-compliant” equilibrium emerges. By integrating game theory, mechanism design, and institutional analysis, the work identifies a unique interior equilibrium alongside a corner solution that fully mitigates harm, thereby explaining why standardized evaluations frequently fail to curb ongoing adverse impacts. Furthermore, it systematically evaluates the incentive effects of institutional mechanisms—including independent auditing, model portability, incident reporting, and outcome-based joint liability—and proposes empirically testable pathways for regulatory intervention.

This work addresses the challenge of dynamically assessing and automatically communicating the exploitability of software component vulnerabilities in AI agent systems. It proposes a runtime constraint evaluation method that integrates Software Bill of Materials (SBOM) with AI Bill of Materials (AIBOM). By correlating deterministic environment snapshots with structured runtime telemetry and combining static analysis with dynamic evidence, the approach automatically generates cryptographically signed CSAF-VEX security advisories and supports deterministic replay validation. This is the first method to enable AIBOM-based runtime exploitability determination, incorporating multi-source vulnerability data from OSV, GitHub Advisory, KEV, and EPSS. Evaluated on synthetic AI workloads ranging from 50 to 5,000 components—covering approximately 10,000 unique components—the approach demonstrates high accuracy and full automation.

Current evaluations of AI systems predominantly rely on static benchmarks, which fail to capture behavioral risks in dynamic real-world environments. This work formalizes AI auditing as an uncertainty-aware, dynamic constraint monitoring problem across the system’s entire lifecycle, targeting critical attributes such as fairness and safety while integrating sociotechnical norms with statistical risk control. By developing a theoretical framework and supporting infrastructure for continuous auditing, the study advances AI governance beyond one-off testing toward ongoing, reliable, and accountable oversight mechanisms.

This work addresses the threat posed by AI development agents that may covertly compromise infrastructure security through privilege escalation, log obfuscation, or persistence mechanisms—risks exacerbated by most organizations’ inability to deploy sophisticated monitoring. The authors propose a training-free, structured monitoring approach for Infrastructure-as-Code (IaC) environments, leveraging differential analysis between information flow graphs (IFGs) and control/data flow graphs to enable high-precision, interpretable detection of security degradation. The method supports both synchronous blocking and asynchronous auditing modes: in synchronous mode, real-time rollback reduces the success rate of stealthy attacks from 74.4% to 0%; in asynchronous mode, untrained IFGs lower the false-negative rate from 11.6% to 3.5% without disrupting legitimate operations.

This study addresses the empirical gap in evaluating whether AI systems fulfill post-deployment regulatory obligations concerning monitoring, reporting, and impact assessment. Drawing on an AI incident database spanning 2020–2026, it presents the first systematic quantification of compliance gaps across nine post-deployment provisions of the EU AI Act, the NIST AI Risk Management Framework, and the GDPR. Employing a multi-regulatory coding scheme and statistical modeling of compliance, the analysis reveals that 77.1% of incidents lack evidence of post-market monitoring and 99.6% show no data protection impact assessments. Internal monitoring is found to significantly improve compliance rates. Building on these findings, the study proposes a four-stage Proactive AI Governance Compliance Framework (PAGCF), emphasizing continuous monitoring and cross-framework validation to establish an evidence-based foundation for post-deployment accountability and governance.

This work addresses the challenge that dependencies between runtime parameters and key performance indicators (KPIs) in AI-RAN systems are highly sensitive to environmental dynamics, often leading to failure in conflict diagnosis. To overcome this, the authors propose a lightweight and interpretable dynamic dependency tracking mechanism. It models active dependencies using Boolean matrices and performs real-time inference on telemetry event streams via a sliding window. Upon detecting structural changes, the method automatically triggers recomputation while efficiently reusing historical estimates to ensure consistency. Experimental results demonstrate that the approach accurately and efficiently tracks dependencies in noisy Boolean event streams with time-varying dependency structures, thereby enabling reliable conflict diagnosis and supporting infrequent model refreshes.