The rapid expansion of autonomous Large Language Model (LLM) agents introduces critical security risks, particularly the confused deputy problem caused by orchestrator compromise or indirect prompt injection (IPI). Traditional defenses rely on in-band filtering, which fails to prevent a subverted orchestrator from bypassing security checks. We propose Intent-Execution Integrity Binding (I-EIB), a reference architecture designed to enforce complete mediation by isolating execution capabilities within a Trusted Execution Environment (TEE). By leveraging Amazon Web Services (AWS) Nitro Enclaves and Key Management Service (KMS)-based capability sealing, I-EIB ensures that sensitive API keys and session tokens remain inaccessible to the host. These credentials are only released when a TEE-based verifier confirms that the processing provenance matches a user-signed domain-specific language (DSL) policy and a predefined chain layout. We formalize the system’s security invariants and provide a structured proof sketch demonstrating how I-EIB maintains intent-execution integrity even under full host compromise. Compared with a host-mediated baseline that performs policy checking without enclave isolation or attestation-gated capability release, the proposed architecture introduces a mean additional overhead of 14.2 ms. Experimental results show a 100% detection rate for bounded step-omission attacks and a false positive rate below 0.1% after normalization. A workflow-oriented case study and computational-path analysis further indicate that the architecture is feasible for high-stakes domains such as finance and healthcare, while still carrying deployment constraints associated with TEE availability, policy precision, and wrapper management.

This paper presents a new four-node strain-based finite element (SBQ12) formulated within the framework of Reissner–Mindlin plate theory for the dynamic and stability analysis of isotropic plates resting on elastic foundations. The independent approximation of the bending and transverse shear strains in the proposed element is efficient in eliminating the shear locking and enhancing the numerical accuracy and stability for thin and thick plates. In this study, the strain-based finite element is used to investigate the free vibration and linear buckling of plates on Winkler, Pasternak and Kerr elastic foundations. The SBQ12 element is extensively validated for square and rectangular plates with different boundary conditions (all edges simply supported (SSSS), all edges clamped (CCCC), two opposite edges simply supported, two opposite edges clamped (SCSC), and two opposite edges simply supported, two opposite edges free (SFSF), etc.), thickness ratios ($a/h$ ranging from 5 to 1000), and aspect ratios. The numerical results show excellent agreement with the analytical and reference solutions, with maximum relative errors generally less than 3.5% for free vibration analyses and 4% for buckling analyses in the majority of cases tested. Particular attention is given to the influence of foundation stiffness parameters on natural frequencies and critical buckling loads. The obtained results confirm the accuracy, reliability, and computational efficiency of the proposed element. Overall, the developed SBQ12 element proves to be a robust and highly accurate tool for the analysis of isotropic plate structures resting on elastic foundations, offering a valuable contribution to computational mechanics.

The evolution of prosthetic devices has been influenced by the advent of additive manufacturing, particularly 3D printing, to produce prosthetic devices with reduced weight and customized designs. However, the mechanical properties of 3D printed prosthetic shanks have been affected by the printing orientation, considering the anisotropic nature of the fused deposition modeling (FDM) process. This article presents the effects of layer orientation on the mechanical and fatigue properties of prosthetic components made of polylactic acid (PLA) using 3D printing and FDM. Standard tensile and fatigue test samples were prepared using PLA and printed using FDM. Three printing orientations were used to prepare the samples. The results of the tensile test showed the anisotropic nature of the printed samples, as the yield strength of the samples printed horizontally was greater (61.3 and 60.7 MPa) than that of the samples printed vertically (24.7 MPa), representing a reduction of approximately 60%. Fatigue life analysis of the samples showed that the fatigue life of the samples printed in orientations A and B was greater than that of the samples printed in orientation C, as the bonding between the filaments of the printed samples in these two orientations was greater. Analysis of the ground reaction force (GRF) showed that the highest force occurred during the toe-off phase of the gait cycle. These results were used to evaluate the structural safety of prosthetic shanks using finite element analysis (FEA). From the results, it is evident that the printing orientation significantly affects the stress and failure of the prosthetic shanks.

The early and accurate diagnosis of neurodegenerative diseases presents a significant clinical challenge, particularly in distinguishing between conditions with overlapping symptoms. Much of the existing research has focused on binary classification, which inadequately addresses the multi-class nature of real-world differential diagnosis. This study’s objective is to conduct a comprehensive evaluation of multi-class machine learning classifiers for the early detection of neurodegenerative diseases using gait signal data. Furthermore, we propose and implement a novel decision support system to automate the selection of the most effective classifier based on defined clinical priorities. We utilised a public gait dynamics dataset from Physionet, comprising data from healthy individuals and patients diagnosed with Parkinson’s Disease, Huntington’s Disease, and Amyotrophic Lateral Sclerosis (ALS), forming a four-class classification problem. A feature set including gait signals and demographic variables such as age and body mass index was established. Eleven classifiers, categorised as density-based, linear, and non-linear, were trained and evaluated. To automate the selection of the optimal model, a decision-making framework was employed to assign weights to evaluation metrics and rank the classifiers. The classifiers demonstrated varied performance across multiple evaluation metrics. The Bayes Normal-U (UDC) classifier achieved the highest accuracy at 65.0%, with a precision of 86.4%, sensitivity of 63.0%, and specificity of 70.0%. The Bayes Normal-L (LDC) classifier yielded an accuracy of 62.5%, with 85.7% precision, 60.0% sensitivity, and 70.0% specificity. The implemented decision support system ranked the UDC classifier as the optimal choice. Notably, the system ranked Fisher’s classifier third, ahead of others with higher accuracy, by prioritising its superior sensitivity (57.5%) and lower Type II error rate, which are critical for reducing missed diagnoses in a clinical setting. Simple accuracy is an insufficient metric for evaluating classifiers in complex, multi-class medical diagnostic scenarios. Our proposed decision support framework provides a robust and automated methodology for selecting the most clinically relevant classifier by systematically balancing multiple performance indicators. This approach enhances the transparency and reliability of machine learning in clinical decision-making and contributes to the development of more effective, deployable diagnostic tools for neurodegenerative diseases.

The alluvial clay deposits at Al-Fao, Southern Iraq, with deep soft clay, offer a great foundation challenge due to low bearing capacity and high risk of settlements. To address these issues, this study evaluated the performance mechanism of floating geogrid-encased stone columns (GESCs) through three-dimensional finite element analysis using PLAXIS 3D with a hardening soil (HS) constitutive model. A parametric study was conducted based on column diameter (0.4–0.8 m), a slenderness ratio (L/D = 3–30), and encasement lengths of (1/3 L, 2/3 L, and Full L). The results demonstrated that increasing the column diameter is the most effective strategy, achieving a maximum bearing capacity ratio (BCR) of 1.75 compared to unimproved soil. Notably, the findings revealed that a 2/3 L partial encasement provides performance nearly identical to full-length encasement (with a difference of less than 0.5%) while significantly reducing material costs by 33%. The geogrid encasement provided an improvement factor (IF) of 1.09 over ordinary stone columns (OSCs). This efficiency is attributed to the encasement’s ability to restrain bulging failure within the upper active zone. The study concluded that 2/3 L partial encasement offers superior technical and economic benefits for floating systems in deep soft clay deposits.

This research investigates the aerodynamic performance and dynamic response of high-speed elevators. The study was conducted using a numerical model based on a two-way air-structure coupling. This is achieved by integrating computational fluid dynamics (CFD) and finite element analysis (FEA) techniques. Three different elevator cabin designs (flat, elliptical, and dome) were analyzed at different operating speeds (6, 8, 10, and 12 m/s) to evaluate the effect of geometry on flow and vibration characteristics. The results showed that the dome cabin shape achieved the best overall performance, contributing to reductions of approximately 41% in acceleration, 35% in deformation, 28% in stress, and the vibration frequency by approximately 50–60% compared to the flat shape. It also exhibited a significant reduction in vibration amplitude. Furthermore, a critical dynamic amplification region was identified at approximately 10 m/s, where the response reaches its peak. This region should be considered when designing damping systems. This improvement is attributed to the streamlined properties of the cabin’s dome shape, which reduce flow decoupling and pressure fluctuations. The results show that improving the streamlined shape may reduce air resistance, thereby positively impacting the required operating power.

Multi-aspect sentiment analysis aims to identify different aspects and associated sentiments within user-generated reviews. In recent years, bidirectional encoder representations from transformer (BERT) have been widely used for sentiment analysis due to its strong ability to capture contextual information. However, BERT has limitations in explicitly identifying aspect boundaries and aligning sentiments, especially when multiple aspects with different sentiments appear in the same review. To address this issue, we propose a combination of rule-based and bidirectional encoder representations from transformer (RB-BERT). The main idea of RB-BERT is to utilize domain-specific linguistic rules to automatically generate weak labels for aspect and sentiment pairs, which are then used to fine-tune the pretrained BERT model. A key contribution of this study is addressing BERT’s limitations in aspect-based sentiment analysis (ABSA) by enhancing aspect identification and sentiment assignment. The dataset consists of 3811 user reviews about Sarangan Lake, a popular tourist attraction in East Java, Indonesia. We collected the dataset from Google Maps. The aspects used in this study are scenery view, price, and local environment. The sentiment polarities are positive and negative. We applied four rule levels to enhance the BERT model. The first rule handles aspect extraction, the second addresses sentiment extraction, and the third determines the dominant sentiment based on the frequency of positive and negative words. The fourth rule combines aspects and sentiment in each review to produce a label. BERT tokenization and BERT embeddings are used for feature extraction, with a fully connected linear layer serving as the classification head. RB-BERT performs best with a precision value of 0.9218, a recall of 0.9748, a Micro-F1 of 0.9476, and a Hamming Loss value of 0.0132. Thus, RB-BERT can be used as an approach to perform automatic labeling in multilabel classification by offering speed, low cost, and good performance.

Reliable predictions of high temperature events are of great significance to enhance urban resilience in arid regions, especially for cities such as Baghdad which lie at the southern end of the jet stream with summer temperatures frequently exceeding 50 °C. However, linear models such as the autoregressive integrated moving average (ARIMA) are limited; they have difficulties in modeling nonlinear patterns. Deep learning techniques (e.g., long short-term memory (LSTM) networks) pose yet another difficulty as they are sensitive to overfitting and they demand large amounts of data to be trained on. In this paper, introduce a hybrid ARIMA-LSTM based on residual decomposition is proposed. This method takes the best of statistical and deep learning methods. The time series of temperature is decomposed into two parts: the linear part which is modeled by ARIMA and the residual nonlinear part which is modeled by LSTM. Based on the daily temperature information during 2000–2023, this hybrid model outperformed the ARIMA and LSTM models individually. For example, it obtained a mean absolute error (MAE) of 1.56 °C, root mean square error (RMSE) of 2.11 °C and $R^2$ of 0.92. Note that the model remained highly accurate during extreme heat events over 45 °C (producing an MAE of 2.01 °C). These findings point to the model’s potential for early warning and climate adaptation, particularly in dry urban districts confronted with escalating heat stress.