Search Results (32)

Background/Objectives: Large language models (LLMs) show promise for clinical decision support, yet their accuracy in interpreting specialized medical guidelines remains uncertain. Retrieval-augmented generation (RAG) may enhance performance by grounding responses in authoritative knowledge bases. This study aimed to compare the accuracy, comprehensiveness, and safety of RAG-enhanced versus standard LLMs for answering clinical questions derived from the German S3 guideline for oral cavity carcinoma. Methods: We conducted a prospective, single-blind benchmark study evaluating six LLMs: one RAG-enhanced model (Custom GPT with guideline access), one consensus-based model (ConsensusGPT), and four standard models (DeepSeek-V3.2, Mistral Small 3.2, Qwen3-Next-80B, GPT-OSS-120B). Fifty clinical questions covering 17 guideline domains were presented to each model three times, yielding 900 evaluations. Three expert reviewers assessed responses using 5-point Likert scales for accuracy, comprehensiveness, and clarity, under a single-blind procedure, the effectiveness of which was tested by a pre-specified manipulation check. We then ran a paired within-model experiment in which each base model was queried with and without guideline access through a transparent, openly released retrieval pipeline, and scored every response with a condition-blind automated judge alongside deterministic retrieval metrics computed from the logs. Secondary outcomes included hallucination rates and guideline citation behavior. Inter-rater reliability was assessed using intraclass correlation coefficients (ICCs). Results: In a paired within-model design that held each base model fixed, adding transparent guideline retrieval improved accuracy—significantly in the three weaker open-weight models (Mistral, Qwen3, and GPT-OSS) and directionally in the already-strong DeepSeek and GPT-5 bases. Because a pre-specified blinding check found that experts could still identify retrieval-augmented answers with 98.5% accuracy, we anchored causal interpretation on measures that do not depend on the human raters, ranked by their independence: deterministic, log-derived retrieval metrics first, and then an automated, condition-blind LLM judge, whose agreement with the experts (Spearman ρ = 0.81, 95.7% within-one agreement) establishes shared calibration rather than independence from their bias. Deterministically from the retrieval logs, citation groundedness rose from 0% to 51–89% and retrieval recall@5 was 92%. On the judge, content-level hallucination fell from 42% to 4% and accuracy rose by a pooled +0.64 points (95% CI 0.47–0.80); the accuracy gain persisted after adjustment for response length (+0.48, 95% CI 0.22–0.73), which retrieval shortened rather than lengthened. The accuracy gain was large for weaker base models and small or non-significant for already-strong ones, whereas the hallucination and auditability gains were consistent across all models. The human ratings reproduced the judge’s accuracy effect (+0.61, 95% CI 0.49–0.74), and GPT-5 run through the transparent pipeline showed no significant difference from the proprietary Custom GPT (judge accuracy 4.48 vs. 4.58). Conclusions: Guideline retrieval yields a reproducible, largely base-independent improvement in the safety and auditability of LLM answers to clinical guideline questions, with accuracy gains concentrated in weaker base models. Because retrieval-augmented answers are recognizable to experts, rigorous evaluation should rely on rater-independent measures, and residual hallucination continues to require human oversight. Full article

Efficient conservation prioritization of Huizhou historic buildings requires joint consideration of visible deterioration and value-bearing architectural elements. This study developed a deep learning-based decision-support framework using an original field dataset of 500 facade images and 2197 audited annotations from five towns in Shexian County, China. Separate YOLOv11n and YOLOv11n-seg models detected six decorative or typological element classes and segmented cracks, spalling, and stains, respectively. Model outputs were converted into a pathology severity index (PSI) based on mask-pixel-area ratios and a decorative value index (DVI) based on weighted element counts; the two indices were combined into a conservation priority index (CPI), whose weighting was examined through sensitivity analysis. On the validation set, decorative-element detection yielded a mAP@50 of 0.655, and pathology segmentation yielded a mask mAP@50 of 0.586. In a preliminary application to four held-out buildings, model-derived priority categories matched the blind ratings of three conservation experts in three cases. The framework offers interpretable evidence for preliminary screening and resource-allocation discussions, but it does not replace field diagnosis. Larger balanced datasets, external building-level validation, and metric calibration are required before regional deployment. Full article

Resident-generated service requests are increasingly used in data-driven urban management, but complaint records may reflect uneven administrative visibility rather than objective service needs. This study develops a benchmark-based 311 visibility audit as a diagnostic tool for sustainability governance, examining whether local service conditions become sufficiently visible to support inclusive, anticipatory, and resilient urban management. Using New York City’s 2023 311 records for six complaint categories—Heat/Hot Water, Noise—Residential, Rodent, Water System, Sewer, and Air Quality—the analysis constructs a tract-by-month-by-complaint-type panel and compares tract-level complaint shares with population, housing-related exposure, and rodent inspection benchmarks. The final dataset contains 661,251 geocoded requests across 2243 positive-population census tracts. Results show substantial complaint-type differences in population-benchmark mismatch: Air Quality has the largest exact total variation distance and Noise—Residential the smallest, while under-visible coverage exceeds over-visible coverage in all six categories. Alternative housing benchmarks change selected Heat/Hot Water and Rodent interpretations, and the external rodent check identifies 46 primary blind-spot candidates, narrowed to 17 benchmark-confirmed candidates. The study concludes that 311 systems should be used as diagnostic infrastructures for identifying information blind spots, not as direct measures of urban need or sustainability performance. Full article

Clinical artificial-intelligence systems are starting to act across a course of care, not to answer one question at a time. Their safety is checked by methods that sample the input space: a test suite tries some inputs, a language-model reviewer reads some cases, a physician panel audits some cases. A sampling check can pass a safety rule and still miss the rare input that breaks it, such as a documented obligation dropped several encounters later. This study measures that gap and releases CIV-Bench, a public benchmark of 832 clinical rule sets with safety properties across eight whole-person domains, in single-encounter and longitudinal forms, plus a computational stress tier, each with independently established ground truth. We compare formal verification, which uses a satisfiability-modulo-theories (SMT) solver to check every possible input at once, against the methods used in practice: random unit testing, language-model judges, and a blinded physician panel. Formal verification detected all 612 violations, raised no false alarm, and returned no unsound verdict; for each item it returned either a proof that the rule holds over every input or one concrete input that breaks it. A frontier language-model judge matched this detection, but it returned a pass rate over sampled cases rather than a guarantee, at three orders of magnitude more compute per item. The general open-weights judge returned unsound verdicts on the computational stress tier; the medically fine-tuned judge was unsound far more widely, collapsing on the longitudinal properties despite strong single-encounter medical detection, so medical fine-tuning did not close the gap. Unit testing and the physician panel missed the deep, cross-encounter violations that hold a course of care together. Formal verification is set apart not by a higher detection rate but by the kind of evidence it returns: a proof over the whole input space, a replayable counterexample, or an explicit statement that it cannot decide. The guarantee holds for the decisions placed in this layer, and it depends on the safety rule being specified correctly. Full article

Architectural space is produced through decisions about form, material, construction, structure, program, and environment, and design education centers on it. In design studios, generative AI (GenAI) most often enters through visual representation, raising the risk we term the visualization trap: images can appear resolved before their tectonic implications are worked out. This exploratory study examines GenAI participation through protocol analysis of the documented process traces of 18 students in a single, AI-aware bioclimatic design studio, yielding 1107 protocols. The TRACE-SC methodology maps space-constituting components, GenAI use types, design phases, and cognitive breaking points; reliability was examined through a blind expert coding audit and cross-LLM comparison. GenAI appeared in 44.2% of protocols (489/1107), where visual generation and information gathering accounted for 81.2% (397/489). Suggestions were transformed before use in 68.1% (333/489) and adopted verbatim in 0.4% (2/489); interaction was designer-initiated. Cognitive breaking points appeared in 2.2% of protocols (24/1107), with indirect evidence in 70.8% (17/24). GenAI proves more than a visual production tool, but its engagement is limited, episodic, and designer-steered rather than a routine design partnership—partial support for the proposition. The transferable contribution is the TRACE-SC framework and codebook; its patterns describe this studio and invite comparison elsewhere. Full article

Despite widespread adoption of cybersecurity awareness training (CSAT), a persistent knowledge–behaviour gap continues to undermine organisational security posture, particularly in resource-constrained and developing-country contexts. This 12-week quasi-experimental field study evaluated an AI-adaptive CSAT platform against traditional instructor-led training (ILT) across three Yemeni organisations (total N = 187; AI-Adaptive: n = 94; Control: n = 93). The system used a 4-parameter Bayesian Knowledge Tracing (BKT) engine—with interpretable guess and slip signals—as an auditable pedagogical decision layer that triggered Protection Motivation Theory (PMT) and Theory of Planned Behavior (TPB)-aligned interventions. Extreme baseline imbalances (Cohen’s d > 2.0), at which standard ANCOVA residual adjustment alone is known to be biased and which necessitated advanced causal-inference triangulation, were addressed via a four-method protocol (ANCOVA, Propensity Score Matching, Difference-in-Differences, mixed-effects). All four methods converged on consensus effect sizes of d = 0.66–0.89. IT-verified Tier 2–3 incidents declined by 48.9% (incidence-rate ratio [IRR] = 0.51, 95% CI [0.38, 0.68]); blinded phishing click-rates fell from 8.8% to 2.1% (χ²(1) = 8.74, p = 0.003). Bootstrapped mediation analysis (PROCESS Model 4; 5000 draws) indicated that coping self-efficacy and perceived behavioural control—but not threat appraisal—were jointly associated with 66.4% of the total compliance effect. Rosenbaum bounds Γ = 2.1; E-values ≥ 3.4. The findings are consistent with the hypothesis that AI-adaptive cybersecurity training produces robust, theoretically explicable benefits and that the coping-appraisal pathway, not threat salience, is the active psychological mechanism. The four-method triangulation framework offers a replicable standard for field evaluations with non-random assignment. the consensus envelope d = 0.66–0.89 is the observed range of point estimates across the four estimators; per-method 95% CIs are reported below indirect effect via coping self-efficacy = 0.843 [0.52, 1.19], via PBC = 0.524 [0.28, 0.81], via threat appraisal = 0.059 [−0.07, 0.21] (ns); direct effect c’ = 0.63 (p = 0.026); total effect c = 2.06 [1.58, 2.54]. Full article

Artificial intelligence (AI) agents increasingly need credentials such as application programming interface (API) keys and Secure Shell (SSH) credentials, but placing those secrets in the agent process exposes them to prompt injection, tool misuse, and exfiltration through ordinary agent outputs. We present CapSeal, a capability-based broker that replaces direct secret access with session-bound, non-exportable handles. Agents request policy-evaluated actions, while the broker performs credential-bearing Hypertext Transfer Protocol (HTTP) and SSH execution through typed executors with schema validation, replay protection, revocation epochs, and tamper-evident audit logging. We extend this design to hosted settings where the broker operator is not trusted with tenant secrets. Our main contribution is operator-blind secret mediation: a split-broker architecture in which a small trusted tenant gateway cooperates with an untrusted operator service that stores the master secret only as a fully homomorphic encryption (FHE) ciphertext and evaluates per-request derivations without decrypting it. We formalize the model and prove computational operator blindness from indistinguishability under chosen-plaintext attack (IND-CPA) security of the FHE scheme, together with conditional capability binding for any secure pseudorandom function/message authentication code (PRF/MAC) instantiation. We implement an end-to-end TFHE-rs prototype that exercises split-broker derivation, multi-tenant revocation and rate limiting, audit integration, and HTTP/SSH mediation. The prototype uses a non-cryptographic homomorphic stand-in and measures the cost of crossing the operator-untrusted boundary at about 9 s per request, roughly 17 million times slower than the plaintext path. We also give LowMC and Rasta transciphering designs and compare FHE with trusted execution environment (TEE)- and secure multiparty computation (MPC)-based alternatives, positioning each trust boundary by assurance and performance. Full article

Tool-calling LLM agents are vulnerable to indirect prompt injection: externally retrieved data can redirect tool calls without system-prompt access, and prompt-level defences leave three harm classes undefended (path traversal, user-guided exfiltration, high-frequency tool abuse). We present a Policy Enforcement Point (PEP) that intercepts at the tool-call boundary with declarative rules over a cross-step information-flow label system (source integrity, data sensitivity) and a synchronous SHA-256 hash-chained audit log. On a controlled dataset across four attack classes, the full system cuts the attack success rate (ASR) from 40.0% to 5.0% (deepseek-v4-pro, five repeats) versus 35.0% for the strongest prompt-only baseline; disabling cross-step label propagation raises the call-level false-negative rate by 26.4 points. The 30.0% task-level false-positive rate is dominated by by-design least-privilege capability-token denials, not rule false positives—an expanded 30-task benign set yields 0/30 rule false positives under scripted isolation. A conservative-DS mitigation (intent-taint) closes the constructed denied-read reconstruction blind-spot variant (ASR 100% to 0%) at no cost on standard workflows. The audit log detects all three tested tamper classes; the in-process enforcement overhead is sub-millisecond per call. Across four further backends, ASR drops under the full system, though LLaMA-3.3-70B retains 16.7% (a rule-coverage gap). A preliminary run over a real MCP stdio transport (an official filesystem server) shows the mechanism operates at a real boundary with a sub-millisecond execution-path increment. We frame these as mechanism-coverage evidence on a controlled benchmark, not a deployability claim for production MCP workloads. Code, data, and metrics are openly available in the replication repository. Full article

Electronic voting requires the simultaneous admission of only legitimate participants, ballot uniqueness, vote confidentiality, storage integrity, and result verifiability. Blockchain alone does not solve these problems, since ledger immutability does not guarantee anonymity, ballot correctness, or reduced trust concentration. The purpose of this work is to develop a parameterizable research framework for electronic voting scenarios with enhanced cryptographic protection, allowing the security level to be varied according to the requirements of a voting scenario. The main contribution of the work is a parameterizable research architecture for composing and experimentally comparing electronic voting configurations with different security and computational profiles. The cryptographic and audit mechanisms integrated into this architecture include blind-signature-based anonymous authorization, encrypted ballot submission, blockchain-style audit, receipt verification, homomorphic tally publication, and threshold-supported tally artifacts. These mechanisms are not proposed as new cryptographic primitives; rather, they are integrated into a reproducible prototype to study how their combination affects verifiability, privacy support, auditability, and computational cost. Compared with basic blockchain-based voting prototypes, this architecture explicitly separates security, privacy, and verifiability profiles and makes their computational cost observable. The implemented prototype is used as an experimental platform for analyzing supported security properties, threat modeling, and computational cost estimation. The results show that authentication, anonymous token issuance, and receipt verification maintain an almost constant cost at the studied scale, while the main cryptographic burden is associated with encrypted ballot submission and threshold-supported tally publication. The scientific novelty of the work lies in constructing a parameterizable architecture that integrates several cryptographic mechanisms and a blockchain audit layer into one reproducible research prototype. At the same time, the proposed approach retains prototype-level limitations associated with the absence of a full zero-knowledge proof stack, independently deployed threshold authorities, and coercion-resistance mechanisms. Full article

This paper presents a Systematization of Knowledge (SoK) on the evaluation methodologies and capability boundaries of Large Language Models (LLMs) in cybersecurity. We propose a Three-Dimensional Taxonomy Matrix to systematize existing metrics across offensive domains, defensive applications, and inherent architectural flaws. Beyond categorization, this matrix functions as a predictive framework to expose structural evaluation blind spots. Specifically, by intersecting target domains with failure attributions, it identifies a critical, unresolved frontier: measuring cross-architecture semantic equivalence in low-level reverse engineering. Empirically, synthesizing 39 frontier benchmarks reveals a systemic evaluation gap: static metric success rarely translates into end-to-end adversarial efficacy. In offensive domains, high penetration rates correlate strongly with pre-training data contamination. When subjected to semantics-preserving code obfuscation as a stress test, zero-shot, tool-free exploit success rates collapse to near 0%. In defensive contexts, cross-procedural code auditing struggles, yielding a peak F1-score of only 23.83%. Furthermore, models suffer from over-alignment-induced functional degradation, with joint-testing frameworks recording up to a 77% functional loss in automated program repair. Our analysis strongly suggests that purely autoregressive mechanisms drive severe technical hallucinations, evidenced by a 19.7% package dependency fabrication rate. Evaluations also expose significant attack surfaces and a significant safety-utility tradeoff: models succumb to prompt leakage attacks at rates up to 86.2%, while heavily aligned versions simultaneously exhibit excessively high False Refusal Rates (FRR) for benign, borderline security queries. Finally, we delineate a theoretical neuro-symbolic roadmap—integrating LLM heuristics with deterministic formal methods—to structurally mitigate the limitations of the autoregressive paradigm. Full article

The National Register of Historic Places (NR) is a structured artifact of meaning-making that encodes disciplinary values linking architectural and cultural significance to wealth and stylistic distinction. In doing so, it systematically underrepresents vernacular, working-class, and the built environments of racially and ethnically marginalized communities. This paper uses artificial intelligence (AI) to examine how that meaning is constructed. We analyze the preservation record across three scales: a national dataset of 100,117 NR listings (1966–2025), a state-level profile of Illinois’s 1997 NR listings, and a close analysis of Lake Forest, Illinois, a community whose exceptional concentration of NR-listed estate architecture makes it an ideal site for examining how preservation significance has been defined and what it excludes. Two parallel AI methods are applied to eighteen Lake Forest nomination documents and their associated photographs. Natural Language Processing (NLP) analyzes nomination text to trace how preservation professionals connect buildings to cultural value; blind AI image analysis examines the same properties to assess how a model trained on cultural imagery constructs visual meaning independently. NLP analysis reveals a corpus dominated by architectural description, with social history, landscape, and labor systematically underrepresented. The visual analysis confirms and amplifies the nomination record’s class-based assumptions while reproducing the same omissions regarding labor, diversity, and community context. These findings inform debates about AI’s potential to audit existing listings and support nominations for underrepresented property types, while showing that without deliberate corrective design and policy reform, such tools are as likely to replicate the preservation system’s inequities as to repair them. Full article

Mandatory Environmental, Social, and Governance (ESG) disclosure requirements—anchored in Corporate Sustainability Reporting Directive (CSRD), International Sustainability Standards Board (ISSB), and Task Force on Climate-related Financial Disclosures (TCFD)—have placed unprecedented demands on supply chain data quality and auditability. Blockchain technology, combining immutability, decentralised governance, and smart contract automation, has emerged as a candidate infrastructure for addressing verification deficits across multi-tier supply chains. To our knowledge, no prior systematic review has simultaneously examined the blockchain specifically for formal ESG transparency and sustainability reporting across all three ESG dimensions within the post-CSRD mandatory reporting landscape. This study presents a systematic literature review (PRISMA 2020). Scopus and Web of Science searches identified 1166 records (2016–2026); after deduplication, 761 unique records were screened, and after blinded screening (κ = 0.84), 96 studies were included. Five blockchain application typologies are identified (T1–T5), spanning provenance tracing, smart contract compliance, carbon accounting, supplier data aggregation, and ESG disclosure systems. A structural asymmetry is identified: governance is addressed in 96% of studies (77.1% under the strictest G-CONFIRMED recoding; 95.8% under the moderate interpretation, including borderline cases), the environmental pillar in 49%, and the social dimension in 21%, explained through institutional theory, with significant implications for CSRD and Corporate Sustainability Due Diligence Directive (CSDDD). Key barriers include scalability, interoperability, and the blockchain–GDPR (General Data Protection Regulation) tension. Three principal contributions are made: (i) a systematic typology of blockchain for ESG transparency; (ii) institutional-theory explanation of ESG dimension asymmetry; and (iii) a research agenda centred on AI–blockchain convergence and post-CSRD empirical studies. The review is limited to English-language peer-reviewed literature. Full article

At racing speeds above 300 km/h (≈83 m/s), hazard awareness becomes a vehicular-communications problem: 100 ms already correspond to about 8.3 m of blind travel before an alert can influence braking, line choice, or torque delivery. Cloud-only telemetry is therefore insufficient under intermittent coverage and variable round-trip delay, while conventional trackside and pit-wall links do not provide direct inter-bike hazard dissemination. We propose Hybrid Epistemic Offloading (HEO), an edge–mesh–cloud architecture for high-mobility V2V/V2X hazard dissemination that explicitly separates an ephemeral safety plane from a durable cloud-analytics plane. On-bike edge nodes ingest high-rate ECU/IMU signals over CAN and persist full-fidelity traces into standardized ASAM MDF containers, enabling loss-tolerant buffering, deterministic replay, and post hoc auditability across coverage gaps. For real-time safety, motorcycles form a local V2V mesh that disseminates compact hazard digests using latency-bounded gossip with adaptive fanout, TTL-based suppression, and redundancy-aware forwarding over sidelink-capable V2X links. The hazard channel is formulated as uncertainty-aware to account for localization error and propagation delay at race pace. We evaluate the system in two stages: (i) a reproducible mobility-coupled simulation/emulation campaign for mesh dissemination and durable edge → gateway → cloud delivery; and (ii) an MDF4 replay-based Jerez pilot for stability-oriented co-design analysis. Under the tested conditions, the durable MQTT path achieved an 83.4 ms median, 175.9 ms p95, and 303.74 ms maximum end-to-end latency with no observed event loss. In the Jerez pilot, the co-design workflow reduced mean wheel slip from 6.26% to 3.75% (−40.10%) and a control-volatility proxy from 0.1290 to 0.0212 (−83.58%). Full article

Background/Objectives: Acute pancreatitis (AP) lacks disease-modifying pharmacotherapy. Neuroimmune, serotonergic, and redox-regulated pathways may modulate inflammatory amplification and acinar injury, although pharmacovigilance data link some psychotropic drug classes to AP risk. This review synthesized controlled rodent studies evaluating neuromodulatory interventions with serotonergic, stress-axis, or ferroptosis-linked targets in experimental AP. Methods: PubMed, Scopus, eLIBRARY.ru, and Elicit were searched in January 2026, supplemented by Google Scholar audit and citation chasing. Eligible studies were controlled in vivo rodent experiments using validated AP models with quantitative outcomes. Intervention timing was classified a priori as a primary analytic variable. Risk of bias was assessed with SYRCLE. A prespecified audit showed that no subset met the criteria for quantitative pooling because of heterogeneity in model class, compounds, timing, outcome definitions, units, and sampling timepoints. Mechanism-stratified qualitative synthesis was therefore performed. The protocol was registered on OSF (doi: 10.17605/OSF.IO/CZXDJ). Results: Nine studies (1992–2023) yielded 410 outcome rows across three mechanistic strands. Serotonergic modulation (5-HT₂/5-HT₂A-focused; six studies) reduced serum amylase/lipase (−37% to −65% vs. disease controls) and histological injury, with receptor-selectivity data supporting 5-HT₂A-mediated mechanisms. Stress-axis modulation with thiadiazine L-17 reduced 7-day mortality in two severe models (from 50–70% to 30%). Olanzapine attenuated ferroptosis-linked injury via off-target antioxidant activity independent of serotonergic receptors. All interventions were prophylactic, peri-induction, or very early post-induction; no delayed therapeutic-window studies were identified. Most SYRCLE domains were unclear, particularly allocation concealment and blinding-related procedures. Conclusions: Neuromodulatory pathways modulate experimental AP in rodents, but evidentiary strength differs across mechanistic strands. Inference is constrained by absent therapeutic-window testing, heterogeneous endpoints, and reporting deficits. The findings support mechanism-level target prioritization rather than clinical repurposing. Full article

The proliferation of digital assets has catalyzed a profound decoupling between intangible property and traditional inheritance jurisprudence. Under the existing legal framework in Taiwan, practitioners must rely on the testamentary forms prescribed in Article 1189 of the Civil Code, which are fundamentally ill equipped to handle cryptographic assets. Specifically, Notarized Wills (Article 1191) necessitate full disclosure to a notary, creating a “Privacy–Security Paradox” where revealing private keys exposes assets to misappropriation. Conversely, while Sealed Wills (Article 1192) offer confidentiality, they are plagued by risks of physical degradation and technical non-executability. This study proposes zkWill, an EVM-compatible decentralized testamentary framework designed to bridge these structural gaps. By leveraging Zero-Knowledge Proofs (ZKPs), zkWill achieves a state of “blind compliance,” verifying that a sealed will meets the statutory requirements of the Civil Code without disclosing its underlying content. The system integrates the Permit2 protocol for secure asset migration and combines AES-256 encryption with IPFS to immunize testaments against centralized storage failures. Unlike conventional services that demand custodial trust, zkWill employs decentralized oracles to trigger automated execution, ensuring legacy distribution without compromising wallet private keys. Empirical data from the Arbitrum Sepolia testnet confirms that the framework maintains constant verification efficiency and a judicially resilient audit trail, providing a paradigm that harmonizes legal pragmatism with cryptographic security for digital inheritance. Full article