In recent years, frequent emergencies such as the COVID-19 pandemic and regional conflicts have significantly heightened uncertainty risks in global industrial chains. Safeguarding industrial chains is crucial for sustaining national economic stability, and while major countries prioritize industrial chain security, establishing an effective industrial chain monitoring system remains a substantial challenge. This paper establishes a dynamic impact monitoring system (DIMS), which can dynamically monitor the risk level of the industrial chain. The DIMS consists of an industrial chain static risk assessment model (IC-SRAM) and an industrial chain dynamic event monitoring model (IC-DEMM). The IC-SRAM is used to evaluate the risks of the industrial chain itself, and the IC-DEMM is used to monitor external risk events in real time. DeepSeek is used to associate events with industrial chain nodes. The dynamic industrial chain risk value is formed based on the results of static risk assessment and dynamic risk event monitoring. Finally, an empirical analysis of the new energy vehicle industrial chain is carried out. The results show the effectiveness of the proposed DIMS and the accuracy rate of associating industrial chain nodes using DeepSeek reaches 91%.
Background: WordPress plugins account for the large majority of disclosed CMS vulnerabilities, and learning-based detectors report high accuracy on synthetic corpora and random splits. Methods: We build a benchmark from 1757 real plugin CVEs (6666 indexed identifiers, 1281 plugins), yielding 30,860 labeled PHP functions in two negative-sampling variants. An audit of the released artifacts found exact normalized-code duplicates crossing test folds (166 groups/1086 records in Variant A; 151/676 in Variant B) and identical code carrying contradictory labels (730 and 456 groups). After removing conflicting groups, merging exact duplicates and collapsing token- and AST-level clone families, the corpora contain 15,294 (Variant A) and 24,917 (Variant B) records with zero shared hashes, clone families or plugins across folds. Ten detectors are compared under plugin-grouped five-fold cross-validation on complete held-out folds: a static taint analyzer, TF-IDF Random Forest and SVM, Bi-LSTM, unidirectional LSTM, a taint-augmented neural model, a nested out-of-fold stacking ensemble, and three baselines (majority class, a source/sink presence rule, and a single-feature function-length classifier). Results: The best macro-F1 is 0.563 (Variant A, binary, SVM) and 0.601 (Variant B, binary, nested stacking). Neural models do not lead in any of the six configurations, and the single-feature length baseline matches or exceeds the TF-IDF Random Forest in four of six. The source/sink rule attains an MCC close to zero, and the taint analyzer misses about 98% of vulnerabilities. Manual review of 400 functions finds 48.8% of cleaned positive labels to be genuine vulnerabilities. Conclusions: Under duplicate-controlled evaluation, the differences between detector families are small compared with the variation attributable to label quality, which appears to be the more binding constraint. All data, folds, predictions and fingerprinted result files are released.
Multimodal fake news detectors may learn biased dependencies from imbalanced event distributions and incidental text–image associations, causing attention to capture dataset-specific patterns rather than reliable discriminative evidence. To address this problem, we propose an Attention-level Causal Intervention Framework (ACIM), which performs causal adjustment directly within the attention learning process. Unlike prior causal debiasing methods operating at the feature-representation level, ACIM intervenes in attention distributions where cross-modal bias emerges. By treating attention representations as mediators, ACIM applies front-door causal intervention to mitigate confounding effects and estimate attention-level intervention effects without requiring fully observed confounders. This principle is implemented through a Causal Attention Layer Module (CALM), integrated into BERT-based textual and Swin Transformer-based visual encoders to jointly model in-sample and cross-sample attention. A causal-aware fusion layer further reconstructs cross-modal attention to suppress misleading text–image co-occurrence patterns. Experiments on Twitter and PHEME achieve accuracies of 0.906 and 0.909, improving upon the strongest reported accuracy baselines by 0.9 and 0.6 percentage points, respectively, while maintaining competitive precision, recall, and F1 performance. Ablation and sensitivity analyses further support the contribution and stability of the proposed approach.
Large language model (LLM) agents are becoming cognitive interfaces to data lakes, enterprise knowledge bases, vector memories, browsers, files, and software tools. This shift creates a data-governance gap: an agent may reason over large private context, yet the protected resource often lacks a verifiable record of which user intent, data object, action, destination, and semantic release were authorized. This paper proposes CapAgent, a semantic data-flow governance middleware for LLM agents in big-data cognitive-computing environments. CapAgent maps human-attested task intent into signed, attenuable, and purpose-bound capability tokens that are checked by a reference monitor before sensitive tool invocation, memory retrieval, data export, and inter-agent delegation. Its policy layer combines task templates, resource labels, destination rules, caveats, semantic release modes, and audit obligations; its runtime enforces both symbolic scope checks and semantic recoverability checks over protected facts. We present formal governance semantics, a conservative intent compiler, an explainable data-flow decision workflow, and a runnable Python middleware. A reproducible trace-replay benchmark with 600 benign and adversarial traces across five data-intensive agent scenarios reports attack success, benign success, false blocking, latency, component ablations, and audit quality. In this synthetic trace-replay evaluation, the full monitor reduces measured attack success from 100.00% under ambient execution and 12.50% under scope-only authorization to 0.00% (Wilson 95% CI [0.00, 0.95]), while retaining 75.00% benign success. In addition, we conduct a 520-trial end-to-end tool-calling benchmark with representative prompt-only, task-shield-style, CaMeL-style, scope-only, and full-CapAgent configurations; a 210-task compiler gold-standard evaluation; a 240-item semantic-release calibration set; and a 12-cell BDCC-style scalability microbenchmark. In these supplemental tests, full CapAgent obtains 0.00% ASR (95% CI [0.00, 1.06]) in the tool-calling benchmark, 87.50% exact-policy compiler match with 0.00% over-authorization, 88.89% semantic-release recall with 0.00% false-block rate, and sub-millisecond in-process authorization latency up to 100,000 resources. The results support CapAgent as an auditable governance layer for cognitive LLM agents rather than as a replacement for model-level alignment or public end-to-end agent benchmarks.
This study examines whether possession-oriented tactical profiles are associated with reported absolute injury burden across Europe’s Big Five domestic football leagues. Team-level tactical statistics were integrated with publicly available injury data covering eight seasons (2017–18 to 2024–25), yielding 780 team-season observations. Three outcomes were examined: total reported injury events, total days lost, and total matches missed. League-specific feature selection used Elastic Net and Random Forest models combined through a performance-weighted Meta Score, followed by K-Medoids clustering with cosine distance to identify relatively higher- and lower-possession tactical profiles. Injury outcomes were compared using Welch’s t-tests and permutation tests with false discovery rate correction, supplemented by club-clustered bootstrap confidence intervals. After multiplicity correction, higher-possession profiles showed greater injury-event counts and matches missed in the Premier League and La Liga, all three outcomes in Ligue 1, and matches missed in the Bundesliga. The La Liga total-days-lost difference did not remain robust under club-clustered resampling. No adjusted difference was identified in Serie A. These findings indicate exploratory associations between possession-oriented tactical profiles and reported absolute injury burden rather than causal or exposure-adjusted effects. As player exposure, fixture volume, and major club- and competition-level confounders were not fully controlled, the results should not be interpreted as estimates of epidemiological injury risk.
Electrical energy theft presents a serious and significant challenge for utility companies worldwide. It poses substantial risks to energy infrastructure, reduces efficiency, and destabilizes distribution networks. The introduction of the Advanced Metering Infrastructure (AMI) in the last decade has significantly boosted the development of new methods and techniques for detecting electrical energy theft. This advancement is primarily due to the availability of electric consumption and other data with greater granularity (e.g., every 15 min), which has enabled the development of analytic and data-driven approaches as opposed to mass inspections alone. In this context, this paper aims to analyze the landscape of energy theft detection by providing a systematic analysis of the state-of-the-art manuscripts in the form of a tertiary study (i.e., a review of literature reviews and surveys) in accordance with the PRISMA methodology guidelines. Consequently, a comparative framework is presented, along with well-formulated research questions designed to explore the past, present, and future directions of energy theft detection.
Consistent and equitable classification of students within Response to Intervention (RTI) frameworks remains a significant challenge in Qatar’s educational system, where tier-assignment decisions rely predominantly on qualitative, multidisciplinary evaluation reports rather than standardized quantitative measures. This research proposes a transformation-based framework for RTI tier classification using Qatari student evaluation reports. The framework converts qualitative clinical descriptors into structured intermediate representations through descriptor extraction, ordinal severity mapping, score translation, and composite aggregation. Seven classification approaches were systematically evaluated within this framework: direct zero-shot and few-shot large language model (LLM) classification; hierarchical prompting; rule-based transformation; LLM-assisted transformation; and a hybrid transformation-based approach. All experiments utilized OCR-extracted Arabic evaluation report text and were assessed using accuracy, balanced accuracy, macro F1-score, weighted F1-score, and class-wise F1-scores. The hybrid transformation-based approach demonstrated the strongest overall performance across evaluation metrics and was the only approach to maintain meaningful classification performance across all three RTI tiers, including the underrepresented and most challenging Tier 1 category. Direct and hierarchical prompting approaches produced lower and less consistent classification performance across the evaluated RTI tiers. These findings indicate that the introduction of structured intermediate transformation stages substantially enhances the consistency, interpretability, and equity of RTI tier classification from qualitative evaluation reports, providing a principled mechanism for standardizing classification decisions across Qatar’s schools and evaluation teams.
Student retention analytics often treats withdrawal as a single outcome, although students may attribute dropout intention to personal, socioeconomic, and academic pressures simultaneously. We benchmarked nine supervised classifiers for identifying a concurrent three-domain attribution profile in a cross-sectional survey of 333 undergraduates at a Colombian public university campus. The response came from a semi-structured weight-allocation item; an audit found that literal label matching altered 21 classifications because of spelling variants and decimal notation. Nine classifiers—logistic regression, decision tree, random forest, neural network, Gaussian Naïve Bayes, k-nearest neighbors, AdaBoost, gradient boosting, and XGBoost—were fitted using six pre-specified predictors. Models were compared by repeated nested stratified cross-validation (five outer folds, three repeats), inner tuning, fold-contained preprocessing, and training-only threshold selection. The concurrent profile occurred in 256 students (76.9%). Logistic regression achieved the highest mean held-out ROC AUC (0.674, 95% CI 0.644–0.704), closely followed by random forest (0.671, 0.641–0.702); their paired difference was nonsignificant after Holm adjustment. Logistic regression had the highest F1 score (0.788), whereas random forest had the highest balanced accuracy (0.617). AdaBoost did not retain its apparent single-holdout advantage. Housing and financial aid had the largest held-out permutation importance. The predictors provided moderate discrimination of a perceptual profile, not a validated prediction of future dropout. Outcome auditing and leakage-free validation materially changed the model ranking.
Relational databases power high-stakes decisions in lending, healthcare, and justice, yet SQL lacks native constructs for causal and counterfactual reasoning. Prior SQL-based causal systems address parts of this gap but do not unify treatment-effect estimation with counterfactual generation in a single, composable SQL surface. We present a system that extends the SQL dialect with two declarative operators: EXPLAIN_CAUSALLY_WHY (ψ) for estimating average and conditional treatment effects via meta-learners, and EXPLAIN_COUNTERFACTUAL (φ) for generating diverse, constraint-respecting alternatives via a hybrid KD-tree/LSH pipeline. Both operators consume standard SQL relations (joins, filters, projections) and return table-valued results with optional diagnostics, confidence intervals, and feasibility metrics. We formalize the operators in relational algebra, and describe our prototype system called PsiQL. On four evaluation datasets, PsiQL recovers a protective TWINS treatment effect, returns a non-significant COMPAS point ATE with imbalance diagnostics, flags HMDA covariate imbalance via built-in SMD checks, and generates constraint-respecting counterfactuals; a synthetic Census run serves as a balanced pipeline proof-of-concept alongside a real ACS diagnostic under severe imbalance.
Backpropagation is less a single algorithm than a pipeline of choices: how the error signal is propagated, how the weight gradient is assembled, and how the update is applied. This paper revisits three consecutive steps and proposes small, mathematically transparent modifications that improve gradient scaling and conditioning without changing the represented function class. The resulting method, Conditioned Backpropagation(CBP), combines (i) a layerwise gradient-norm equalization that counters the geometric depth dependence of the backpropagated error; (ii) an activation-centering reparameterization that removes the dominant rank-one mean term from the per-layer curvature; and (iii) a damped diagonal preconditioner that is positive-definite by construction. The composite operator is a bounded positive-definite preconditioner, so the method inherits standard nonconvex, Polyak–Łojasiewicz, and stochastic convergence guarantees at the per-step cost of ordinary backpropagation. No prior method composes these three repairs into one operator with a joint boundedness and positive-definiteness guarantee. Two further results, both new, concern equalization. On a block-structured strongly convex model, and for the curvature-equalizing target that the implemented gradient-energy equalizer approximates up to a quantified heterogeneity factor, equalization makes the convergence rate depth-uniform; the bounded-clip version that is actually run stays depth-uniform up to a clip-determined depth and retains a constant-factor improvement beyond it. Controlled experiments, run over ten or more seeds with paired significance tests, confirm the mechanisms: Equalization compresses an order-of-magnitude per-layer gradient disparity, centering cuts the top curvature eigenvalue about threefold and yields the lowest training loss, and the configurations combining centering with the damped preconditioner, including the full method, converge fastest. The effects persist on MNIST and on CIFAR-10 with a small residual convolutional network, at a measured per-iteration overhead below about twice that of Adam. Generalization is comparable across methods, and no end-to-end depth-scaling advantage is claimed, keeping the contribution focused on optimization geometry.
The growing capabilities of artificial intelligence (AI) have not translated straightforwardly into organisational value. A persistent disconnect—the “last-mile problem”—arises from structural gaps between idealised AI tasks and real-world organisational contexts. Synthesising insights from organisational theory, cognitive science, and computer science, we have developed a five-dimensional diagnostic framework that maps the challenges of human-AI collaboration across Integration, Representation, Scale, Temporality, and Adequacy gaps. These gaps illuminate how socio-technical complexity, contextualised problem representations, interdependencies among agents, dynamic environments, and limitations in current AI reasoning collectively constrain full automation and demand human judgement. By reviewing the historical evolution of AI—from symbolic systems to machine learning, generative models, and emerging agentic approaches—we show that augmentation remains the dominant and most viable mode of use in complex environments. An illustrative system-dynamics example demonstrates how improvements in algorithmic performance do not automatically yield proportional system-level gains. Overall, our framework provides researchers with a conceptual lens and practitioners with a diagnostic tool for assessing complementarities and informing the design of human-AI collaborations. The framework is offered as a conceptual synthesis and diagnostic instrument rather than an empirically validated model.
Cardiovascular diseases remain the leading global cause of mortality, necessitating continuous monitoring solutions that extend beyond clinical settings. This paper proposes a real-time, end-to-end Edge-IoT framework for cardiovascular risk assessment that integrates biomedical signal acquisition, edge processing, and interpretable deep learning. The system includes a three-tier architecture: (i) physiological signal acquisition using AD8232 ECG, MAX30102 photoplethysmography, DS18B20 temperature, and NEO-6M GPS sensors interfaced with an ESP32 microcontroller; (ii) real-time signal preprocessing, including digital filtering, normalisation, and PQRST feature extraction performed at the edge; and (iii) cloud-based analytics using an Enhanced TabNet classifier with modified attention mechanisms for cardiovascular risk prediction. The Enhanced TabNet architecture incorporates Entmax-based sparse attention and modified Gated Linear Units to improve predictive performance and clinical interpretability. Signal quality enhancement using Kalman filtering and class imbalance correction using SMOTE further support robust model performance. The Enhanced TabNet model achieves 97.43% accuracy, 96.18% precision, and 97.24% recall on the combined Cleveland, Hungarian, Switzerland, Long Beach VA, and Statlog heart disease datasets (n=1190). The developed Edge-IoT prototype maintains an end-to-end communication and processing latency below 200 ms. The framework also includes automated risk alert generation via SMS when the predicted cardiovascular risk probability exceeds a predefined threshold (e.g., 0.85), including the patient’s vital information and geolocation to support emergency response. The integrated edge-cloud architecture with attention-based feature selection provides interpretable cardiovascular risk predictions while maintaining the computational efficiency required for potential continuous patient monitoring outside hospital settings.
In recent years, there has been an increase in the amount of fake news in the media, which is why fact-checking systems are gaining popularity, particularly those that use natural language processing (NLP) to quickly identify and flag fake news. One of the main limitations in the development of such systems is the limited number of datasets containing verified information, which are necessary for the effective training of models. The situation is particularly critical for non-English datasets, specifically those in the Ukrainian language. This article proposes a three-stage algorithm for training a model to recognize fake news in the Ukrainian language. At the core of the proposed approach lies the multilingual transformer model XLM-RoBERTa, which solves this problem by utilizing cross-lingual knowledge transfer from English to Ukrainian. This approach means there is no need to search for a large, high-quality dataset in Ukrainian; instead, a significantly smaller dataset in Ukrainian can be used for the final calibration of the model. The model developed as a result of the experiment proved effective in extreme low-resource scenarios, achieving 90.7% accuracy on just 500 training records and outperforming the baseline model by 9.7%.
Automatic sentiment analysis requires models that can effectively handle texts of varying levels of complexity. Monolithic methods use the same algorithm uniformly, without taking into account the intrinsic complexity of the input text. In response to this need, we developed a Dynamic Agentic Mixture-of-Agents with Inter-Agent Communication and Adaptive Routing for Robust Sentiment Analysis (DAMA-Sent). This approach merges three distinct algorithmic paradigms: statistical learning, deep learning, and attention models. The decomposition process is carried out in a sophisticated system featuring a hierarchical routing system and an inter-agent communication system based on differentiable attention. Furthermore, each agent has an individual self-reflection module that allows it to assess its own reliability. Finally, an adaptive early exit system halts processing as soon as an appropriate confidence threshold is reached or the computation budget is exhausted. In-depth analyzes conducted on a corpus of tweets from American airlines reveal that the suggested approach can adjust to the intrinsic variability of textual complexity and surpasses static ensemble methods in terms of accuracy and computational cost, achieving an accuracy of 95.96%. Additional confirmation thru an ablation analysis supports these trends, demonstrating the relevance of each element of our model.
ViTs have achieved excellent performance in image recognition tasks, but their large parameter counts and high computational complexity limit their deployment on resource-constrained devices. Most existing ViT pruning methods adopt class-agnostic pruning strategies, which fail to distinguish the diverse structural requirements of different target classes. As a result, they are prone to removing critical features, leading to class-wise accuracy imbalance in practical deployment. To address this issue, this paper proposes a class-aware joint pruning framework for ViTs, which collaboratively compresses the model from two orthogonal dimensions: MLP neurons and visual tokens. Specifically, (1) based on first-order Taylor expansion, we quantify the contribution of each MLP neuron to the target classes and adaptively prune redundant neurons to achieve structured compression, followed by lightweight fine-tuning on the target class subset; (2) we propose a Class-Guided Token Selection (CGTS) method, which constructs class prototype vectors using a few support samples of the target classes and then dynamically selects patch tokens that are semantically highly relevant to the target classes during inference in a zero-shot manner, requiring no additional training or fine-tuning. The two modules complement each other, achieving dual compression from the parameter dimension and the inference data dimension. Experiments on CIFAR-100 and TinyImageNet datasets using DeiT-Tiny/Small models demonstrate that, compared with state-of-the-art pruning methods, our method reduces GMACs on target class subsets by up to 48%, improves inference speed by nearly 50%, and requires only 0.8 KB of additional storage overhead per subset, ultimately achieving a superior trade-off among accuracy, computational efficiency, and storage overhead.
Understanding how clinical pathways evolve over time is essential for characterizing care processes. It also helps identify potential shifts in diagnostic and organizational practices. This study provides a descriptive analysis of patient trajectories for four major respiratory conditions: lung cancer, interstitial fibrosis, chronic obstructive pulmonary disease (COPD), and pneumonia. Trajectories were compared between a pre-COVID-19 period (2018–2019) and a COVID-19 period (2020–2022) in a specialized hospital. Using process mining applied to administrative event logs, we examined three aspects of care: the structure and sequencing of activities, the timing of transitions between care encounters, and imaging timeliness. The analysis spanned inpatient, emergency department, and outpatient settings. Indicators of care duration and transition timing revealed heterogeneous temporal patterns. Several conditions showed shorter intervals in the COVID-19 period, whereas others varied little. Activity-level analyses complemented these findings. Process maps indicated stable structural components in many pathways, together with differences in timing and execution. In the emergency department, care shifted toward bedside radiography, whereas CT chest volumes remained relatively stable across periods and settings. Imaging timeliness stayed consistently high in the emergency department and relatively stable for most inpatient conditions. Outcome-related indicators, including 30-day readmission and prolonged care trajectories, showed only modest differences between periods. Overall, the study demonstrates the value of process mining for describing real-world clinical pathways and identifying temporal variations in care. These results provide a foundation for future work that integrates richer clinical information and analytical approaches capable of assessing causal relationships.
Accurate blood glucose prediction is essential for glycemic management in people with diabetes, but the size of many high-performing models complicates execution on resource-constrained artificial pancreas controllers. We propose GluKDnet, a lightweight glucose-forecasting model for prospective Android-smartphone-based mobile edge controllers. GluKDnet transfers the representational capacity of a time-series foundation model to a compact causal CNN through heterogeneous knowledge distillation. The teacher model, MOMENT, is adapted to continuous glucose monitoring (CGM) data through risk-event-aware masking, which prioritizes abnormal glucose levels, rapid glucose fluctuations, and CGM-defined dawn phenomenon and Somogyi effect patterns during masked reconstruction. A transient-state and steady-state distillation module jointly aligns ordered patch-level dynamics and day-level summaries between teacher and student. Using DLCP3 for teacher pretraining and leave-one-patient-out evaluation on OhioT1DM, GluKDnet achieves RMSE values of 20.04, 32.04, and 45.33 mg/dL for 30, 60, and 120 min prediction, respectively, with about 53K parameters. Auxiliary evaluation on T1D-UoM shows a similar offline accuracy–parameter count pattern. On a vivo V2072A Android smartphone, the 30 min model achieved a mean API inference latency of 0.470 ms (P95: 0.855 ms), a maximum sampled process proportional-set-size memory of 47.06 MiB, and a median incremental device energy estimate of 0.277 mJ per inference. These device measurements characterize the exported student model under one hardware and software configuration; insulin dosing and prospective closed-loop clinical evaluation remain outside the scope of this study.
English-teaching agents use external knowledge retrieval to update instructional content, broaden domain coverage, and personalize support beyond standalone large language models (LLMs). However, open sources may introduce harmful, biased, or misleading content into retrieval-augmented generation (RAG) pipelines, affecting learners’ judgment, cultural understanding, and value formation. To address this problem, this study proposes a multi-stage secure knowledge retrieval method for English-teaching agents. The method coordinates safeguards across knowledge-source access, retrieval execution, and model output. At the access stage, custom rules and Semgrep-based static scanning perform preliminary risk screening. At the retrieval stage, LLM-based dynamic evaluation identifies tool-description contamination and cross-file data-flow risks. At the output stage, semantic-embedding pre-screening, LLM review, and bounded knowledge purification detect and rewrite risky responses. Our experiments use public safety benchmarks, a mixed corpus of benign and poisoned passages, synthetic purification cases, and controlled end-to-end teaching scenarios. Compared with vanilla RAG, the framework reduces Poison Exposure@5 from 92.0% to 3.0% and retrieval attack success from 86.0% to 2.0% while preserving retrieval coverage. These results provide preliminary evidence that the framework can empower English teaching by enabling agents to deliver safer materials and trustworthy support for classroom questioning, academic writing, and intercultural learning.
Record linkage is a fundamental step in ensuring the quality of data by detecting duplicate records within different databases. Nevertheless, dealing with big, imbalanced databases and ensuring data confidentiality is still difficult in terms of performance and precision. This paper introduces a new Privacy-Preserving Record Linkage (PPRL) method named PPRL-Stack, which uses the Bloom filter encoding technique to hide information and a Stack Ensemble structure for classification. The proposed model consists of Support Vector Machine (SVM) as a base learner and Logistic Regression (LR) as a meta-classifier in combination with the application of Sorted Neighborhood Method (SNM) technique to bring down the time complexity to O(N log N). Experiments conducted on the Freely Extensible Biomedical Record Linkage (FEBRL) and North Carolina Voter Registration (NCVR) databases prove that the proposed PPRL-Stack can obtain nearly perfect discrimination with an F1-score of 0.9921. Particularly, our proposed architecture is more than 340 times and 40 times faster than the latest Siamese Bidirectional Long Short-Term Memory (Bi-LSTM) architecture in training and validation stages, respectively.
The deployment of natural-language-to-SQL (NL-to-SQL) systems in primary healthcare requires more than accurate query generation: it also requires governed data access, robustness to local terminology, and reliable handling of ambiguous user requests. This study evaluated a pilot proof-of-concept integrating a Spanish-language NL-to-SQL assistant with a governed, read-only outpatient scheduling repository derived from the Rayen information system used in a Centro de Salud Familiar (CESFAM) setting in Renca, Chile. The data used by the prototype were accessed through an external company responsible for data management in this context. The prototype was implemented with MindsDB as an artificial intelligence (AI)-enabled database layer and operated on anonymized, delayed secondary scheduling data. Evaluation was conducted through a Slack interface using 252 audited interactions from 42 users, with six assigned interactions per user and up to three exchanges per interaction. SQL correctness reached 240/252 (95.2%), whereas both query correctness and answer correctness reached 144/252 (57.1%). These findings suggest that governed pilot deployment for outpatient schedule monitoring may be feasible under controlled institutional conditions, while indicating that the main remaining barriers are semantic rather than purely syntactic, specifically ambiguity handling, institution-specific operational language, and faithful answer verbalization. The study therefore contributes deployment-oriented pilot evidence and clarifies where operational Spanish NL-to-SQL remains fragile under real institutional constraints.
