Harrisburg University of Science and Technology

We examine the linkages between power reliability, economic growth, and income inequality in the United States. Specifically, we use the two-step System Generalized Method of Moments (GMM) estimator to assess the impact of power interruptions on state-level GDP and the Gini Index. Our findings reveal that a 1 percent increase in power interruptions, measured in terms of duration (SAIDI) and frequency (SAIFI), is associated with a 0.07 to 3.7 percent decrease in real GDP and a modest increase in income inequality of approximately 0.17 to 0.20 percent relative to the mean Gini Index. Moreover, the marginal effects of power interruptions are substantial, with frequent outages resulting in GDP losses exceeding 2 trillion in the long run. We also use machine learning models to support the predictive relevance of the power reliability metrics. Overall, the results highlight the significant role that both the frequency and duration of power interruptions play in shaping regional economic performance and the importance of improving power reliability to foster economic stability and equity.

We study how runtime enforcement against unsafe actions affects end-to-end task performance in multi-step tool using large language model (LLM) agents. Using tau-bench across Airline and Retail domains, we compare baseline Tool-Calling, planning-integrated (TRIAD), and policy-mediated (TRIAD-SAFETY) architectures with GPT-OSS-20B and GLM-4-9B. We identify model dependent interaction horizons (15 to 30 turns) and decompose outcomes into overall success rate (SR), safe success rate (SSR), and unsafe success rate (USR). Our results reveal a persistent Safety Capability Gap. While safety mediation can intercept up to 94 percent of non-compliant actions, it rarely translates into strictly safe goal attainment (SSR below 5 percent in most settings). We find that high unsafe success rates are primarily driven by Integrity Leaks, where models hallucinate user identifiers to bypass mandatory authentication. Recovery rates following blocked actions are consistently low, ranging from 21 percent for GPT-OSS-20B in simpler procedural tasks to near zero in complex Retail scenarios. These results demonstrate that runtime enforcement imposes a significant verifier tax on conversational length and compute cost without guaranteeing safe completion, highlighting the critical need for agents capable of grounded identity verification and post-intervention reasoning.

The credibility revolution in psychology and related sciences contributed to the adoption of large-scale research initiatives known as Big Team Science (BTS). BTS has made significant advances in addressing issues of replication, statistical power, and diversity through the use of larger samples and more representative cross-cultural data. However, while these collaborations hold great potential, they also introduce unique challenges related to their scale. Drawing on experiences from successful BTS projects, we identified and outlined key strategies for overcoming diversity, volunteering, and capacity challenges. We emphasize the need for clear role definitions, structured and preregistered workflows, centralized project management, and transparent decision documenting to prevent common pitfalls. Ultimately, we call for reflection on the strengths and limitations of BTS to enhance the quality, generalizability, and impact of research across disciplines. This work complements existing BTS guides by offering experientially-grounded, discipline-specific strategies and addressing underexplored logistical, ethical, and epistemological challenges in large-scale collaborations.

Miscalibrated confidence scores are a practical obstacle to deploying AI in clinical settings. A model that is always overconfident offers no useful signal for deferral. We present a multi-agent framework that combines domain-specific specialist agents with Two-Phase Verification and S-Score Weighted Fusion to improve both calibration and discrimination in medical multiple-choice question answering. Four specialist agents (respiratory, cardiology, neurology, gastroenterology) generate independent diagnoses using Qwen2.5-7B-Instruct. Each diagnosis is then subjected to a two-phase self-verification process that measures internal consistency and produces a Specialist Confidence Score (S-score). The S-scores drive a weighted fusion strategy that selects the final answer and calibrates the reported confidence. We evaluate across four experimental settings, covering 100-question and 250-question high-disagreement subsets of both MedQA-USMLE and MedMCQA. Calibration improvement is the central finding, with ECE reduced by 49-74

Background: The clinical use of machine learning (ML) in survival analysis is often limited by the "black box" nature of complex algorithms, which makes their results difficult to interpret in practice. In this study, we propose a unified and clinically grounded framework that integrates ML-based feature selection with traditional survival analysis. This approach aims to bridge the gap between strong predictive performance and clear, clinically meaningful interpretation. Methods: High-impact prognostic clinical features were identified using ML models GBM-Cox, RSF, and LASSO-Cox with 5-fold stratified cross-validation and subsequently validated using Cox Proportional Hazards and Kaplan-Meier analysis. The framework was evaluated across two distinct disease domains, Heart Failure and the METABRIC breast cancer cohort, to assess robustness and generalizability. Results: In the Heart Failure dataset, age group, serum creatinine, and blood pressure stratified patients into distinct risk groups. The high-risk group had significantly higher mortality (HR: 2.61; 95% CI: 1.42-4.78; p = 0.0013). In the METABRIC cohort, age at diagnosis, HER2 status, and Nottingham Prognostic Index (NPI) showed strong survival separation (p < 0.001). The high-risk group had an HR of 2.73 (95% CI: 2.34-3.19) and the faced a significantly shorter median survival (104.7 vs. 252.3 months), representing a 12.3-year reduction in life expectancy compared to low-risk group. This prognostic separation emphasizes the predictive power of selected baseline variables. Performance remained stable across cohorts, with C-index values (0.665-0.794) consistent with standard clinical benchmarks. Conclusions: Integrating cross-validated machine learning feature selection with Cox-based survival analysis enables stable and clinically interpretable risk stratification across diseases. By translating ML selected predictors into hazard ratios and absolute survival differences, this framework provides a reproducible and clinically grounded approach for survival risk assessment.