Search Results (1,628)

[d=LE]Transitions of control between automated driving systems and human drivers remain safety-relevant and cognitively demanding moments in human–automation interaction. Recent studies show that transition performance depends not only on takeover timing or response speed but also on traffic complexity, driver readiness, automation limitations, trust calibration, and situational-awareness recovery. As in-vehicle interaction evolves toward conversational and agentic AI assistance, takeover support also becomes a problem of governing how natural-language AI systems communicate with the driver under uncertainty.Transitions of control between automated driving systems and human drivers remain safety-relevant and cognitively demanding moments in human-automation interaction. Recent studies suggest that transition performance should not be assessed only through takeover timing or response speed since control resumption quality also depends on traffic complexity, driver readiness, automation limitations, and situational awareness recovery. [d=LE]This paper proposes a digital-twin-mediated framework for human-aware takeover support in automated driving. In this framework, the companion AI is treated as an assumed LLM-based in-vehicle conversational or agentic assistant used as an advisory interaction component. The contribution is defined at the architectural level: human, vehicle, and context/road digital twins provide structured semantic state abstractions through a semantic state interface exposing confidence, freshness, provenance, and consistency metadata, while a trustworthy companion AI (TCAI) layer grounds, constrains, validates, and governs companion AI output proposals before HMI delivery.This paper motivates and defines a trustworthy companion AI (TCAI) layer for human-aware transition support in automated driving. The TCAI is conceived as a bounded, supervised, and explainable advisory agent that supports the driver without entering the safety-critical vehicle-control loop. It reasons over structured semantic state abstractions derived from a human digital twin, a vehicle digital twin, and a context/road digital twin, exposing driver readiness, automation capability, and contextual urgency in a form that supports traceable, uncertainty-aware, and degradation-aware assistance. [d=LE]Building on the research on driver-state monitoring, adaptive HMI, trust calibration, explainability, conversational assistance, and human assistance systems (HASs), the framework coordinates advisory interaction across vigilance support, contextual explanation, trust-calibrating communication, and directive handover guidance. The TCAI layer combines bounded reasoning, human-factor-derived guardrails, state-consistency management, dynamic explanation-depth control, trust-dynamics modeling, graded watchdog veto handling, mandatory access-control assumptions, and deterministic fallback. Safety-critical vehicle-control and minimum risk condition (MRC) functions remain assigned to the deterministic vehicle-control stack, while the authorized output path of the TCAI layer is validated HMI delivery.Building on the research on driver-state monitoring, adaptive HMI, trust calibration, explainability, and conversational assistance, we propose a conceptual architecture in which the TCAI coordinates multimodal assistance across different interaction conditions, including vigilance support, contextual explanation, trust-calibrating communication, and directive handover guidance. The companion does not actuate the vehicle; its outputs are constrained by runtime governance, policy enforcement, and deterministic fallback mechanisms. [d=LE]The paper concludes with a validation agenda and technical roadmap covering planned transitions, urgent handovers, degraded or adversarial conditions, temporal fusion of driver-state evidence, phase-sensitive HMI policies, trust-calibration trajectories, driver veto and partial-disabling mechanisms, and staged simulator-to-vehicle evaluation. Although motivated by SAE Level 3 automation, the framework may also inform fallback-related Level 4 scenarios in which human and automated agency must be managed under uncertainty.The paper concludes with a research roadmap for validating the proposed architecture under planned transitions, urgent handovers, and degraded or adversarial conditions. Although motivated by SAE Level 3 automation, the approach may also inform fallback-related Level 4 scenarios. Full article

This paper addresses bearing-only three-dimensional target localization using three cooperative UAVs under observation inconsistency and degraded geometry. A weighted point-to-line least-squares localization model is established to fuse multiple line-of-sight (LOS) observations derived from image measurements, camera calibration, and UAV poses. To handle unreliable measurements without ground truth, a reliability assessment mechanism is developed by combining geometric stability indicators with observation consistency metrics, enabling weak geometry and abnormal observations to be identified online. Based on this assessment, an adaptive optimization framework is introduced to perform residual-driven adaptive weighting and configuration optimization, thereby suppressing unreliable LOS measurements and improving the conditioning of cooperative geometry. Simulation results under four representative scenarios show that the proposed method consistently improves localization accuracy and robustness. The mean localization error is reduced from 0.545 m to 0.260 m under abnormal observations, from 0.355 m to 0.081 m under degraded geometry, and from 0.711 m to 0.280 m when both effects occur simultaneously. Statistical evaluations including RMSE, standard deviation, maximum error, confidence intervals, and box-plot analysis further demonstrate that the proposed framework effectively reduces error dispersion and improves robustness. Full article

The Tomarrazón-Camarones Basin (La Guajira, Colombia) is characterized by frequent, widespread flooding and, anthropogenically, by intense instream sediment mining. Mapping flood hazard is hence essential to develop effective flood management plans, and a knowledge of the water regime (duration curves) is also essential to estimate sediment transport and carry out sediment budgets to inform on the impacts and sustainability of the mining activity. However, neither water levels nor discharges are monitored by official gauging stations, and only a few rainfall gauging stations are available in the area, with daily records often affected by data gaps. Therefore, a first challenge is to reconstruct discharge time series by an affordable effort, scaled to the financial-labour resources available in that challenging context. This paper presents an integrated approach that combines satellite-derived rainfall data with ground observations. A semi-distributed hydrological model (HEC-HMS, SCS-CN method) is used to reconstruct the full flow-rate time series once calibrated and validated with data derived from automatic sensors and field measurements. The model is fed with hourly data derived from daily data at ground gauging stations temporally downscaled by adopting the spatially distributed hourly rainfall patterns obtained from satellite records. Before that, observed water levels in three stations equipped with water level sensors were translated into discharge time series using analytical relationships based on field-measured geometric and physical characteristics. Then, these event-based hydrographs were used to calibrate and validate the model. Results show good agreement with observations, with R² = 0.981 and a relative RMSE of 40% for overall hydrograph reproduction, and R² = 0.87 for peak flow estimation, supporting a reasonable confidence in the approach. The calibrated model is then applied to long-term datasets (1973–2024) to retrieve duration curves and return periods of peak discharges. Full article

Background/Objectives. Automated lateral ventricle volumetry is increasingly used in population-based neuroimaging, but correlation between methods does not establish agreement of absolute volumes. We quantified agreement and calibration between FreeSurfer and a visually quality-controlled FSL/FAST–ALVIN lateral ventricle workflow within the Study of Health in Pomerania (SHIP). Methods. This cross-sectional agreement-and-calibration study included 2988 SHIP participants with visually accepted FSL/FAST–ALVIN total lateral ventricle volumes; paired FreeSurfer data were available for 1913 participants. FSL/FAST–ALVIN was treated as the study reference scale rather than biological ground truth. Agreement was assessed using Pearson and Spearman correlations, Bland–Altman analysis, log-ratio agreement, Lin’s concordance correlation coefficient, and a two-way mixed-effects single-measure absolute agreement intraclass correlation coefficient. Directional calibration models predicted FSL/FAST–ALVIN volume from FreeSurfer volume and were internally validated using 2000 bootstrap resamples. Results. In the paired sample, volumes were almost perfectly associated (Pearson r = 0.9978; Spearman ρ = 0.9974), but FreeSurfer yielded systematically lower values (mean FreeSurfer-minus-FSL bias, −3.02 mL; 95% limits of agreement, −4.52 to −1.53 mL; geometric mean FreeSurfer/FSL ratio, 0.844). Lin’s concordance coefficient and the absolute agreement ICC were both 0.9598. Calibration was strong but workflow-specific: FSL/FAST–ALVIN volume = 2.611 + 1.0210 × FreeSurfer volume (R2 = 0.9955; optimism-corrected RMSE = 0.732 mL). Conclusions. FreeSurfer and visually quality-controlled FSL/FAST–ALVIN preserved participant ranking extremely well but were not directly interchangeable as absolute measurements. Cross-workflow comparisons require explicit method reporting, formal agreement analysis, and calibration to the intended measurement scale; the equation should not be used as a universal conversion formula outside comparable acquisition, segmentation, QC and software settings. Full article

Modeling crop development under different agrotechnologies is important not only for assessing the factors that affect their yields but also because of the role of vegetation in regulation of the hydrology regime. For this reason, interest in the plant module in the semi-distributed hydrological model SWAT is increasing. The model has to be supplied with a lot of information for running and testing, which can be achieved with ground-based, statistical and satellite data. The aim of the study is to determine the accuracy of the SWAT model to predict crop development by using ground-based and satellite data for LAI in the case of a 5-year field experiment. Two staple crops in rotation were monitored—winter wheat and maize—under different erosion control technologies (up-and-down conventional tillage, conventional contour tillage, and minimum contour tillage with inclusion of cover crop before maize) on sloping terrain on moderately eroded Epicalcic Chernozem in the region of Ruse, north Bulgaria. The remote sensing data from the Copernicus Sentinel-2 mission were used for estimation of LAI of both crops and verified against ground-based data in two ways—via a custom LAI script available through the Sentinel Hub cloud platform and as input to a machine learning quantile regression forests (QRF) model. The calibrated satellite-derived LAI, ground-based soil moisture and yields data were used to calibrate several SWAT model parameters (EPCO, ESCO, CN2, LAImax, HU, HI) and assess the model performance regarding these variables. Although a good temporal fit of the SWAT-modeled LAI data with the satellite data was achieved, the accuracy of predicted LAI is moderately high only in the last two years of the rotation (R² = 60.4%). The accuracy of calibrated yields (R² = 55.5%) is acceptable in four of the years. On average for the period, the applied erosion control agrotechnologies did not cause significantly different yields, but they are 14% higher compared to the up-and-down conventional tillage. The most sensitive SWAT parameters accounting for this effect are EPCO and ESCO. Full article

Ensemble learning has been widely adopted in fraud detection; however, conventional ensemble strategies rely on uniform or performance-based weighting schemes that treat explainability as a post hoc annotation rather than an architectural component. This study addresses the research goal of whether SHAP attribution values can serve as a principled, instance-specific weighting mechanism within an ensemble, thereby embedding interpretability directly into the aggregation process. A SHAP-Weighted Ensemble (SWE) framework is proposed in which the L2 norm of each base model’s SHAP attribution vector, computed at prediction time, is used to derive instance-specific voting weights via Softmax normalization. Three linear base learners (logistic regression, robust LR, calibrated linear SVM) are combined, with LinearSHAP providing exact attribution values. A comprehensive evaluation protocol was applied on a real-world vehicle insurance claims dataset, including bootstrap 95% confidence intervals, McNemar’s test, a three-way ablation study comparing equal weighting, SWE, and validation-AUC weighting, F1-optimal threshold selection, expected calibration error, and cost-sensitive evaluation under asymmetric misclassification costs. The central finding is that SWE achieves performance statistically comparable to both simpler baselines across all evaluated metrics (ROC-AUC = 0.774, 95% CI [0.681, 0.862]; F1 = 0.679, 95% CI [0.569, 0.774]; McNemar p = 1.000), while producing a transparent, per-claim weighting trace that equal-weight voting cannot provide. A KernelSHAP influence analysis conducted directly on the SWE confirms that SHAP-derived weights are substantially aligned with actual model influence ratios (LR: 1.05×, LR_R: 1.05×, SVM: 0.81×), validating the weighting mechanism empirically. An exploratory analysis of a seven-model equal-weight diagnostic ensemble reveals a negative correlation (r = −0.721, p = 0.067) between individual model performance and ensemble influence; a theoretically coherent finding that does not reach statistical significance at conventional thresholds. The primary contribution of SWE is architectural and interpretability-driven: it produces an auditable, instance-level model-weighting mechanism grounded in SHAP attribution theory, supporting regulatory accountability under GDPR Article 22 and the EU AI Act. Full article

Phenomenological crystal plasticity (CP) models are widely used in Integrated Computational Materials Engineering (ICME) to bridge microstructural features with engineering-scale mechanical behaviour. However, their practical application is hindered by two major challenges: high computational costs of physics-based simulations, and the labour-intensive, trial-and-error nature of parameter calibration. These challenges are amplified in additively manufactured (AM) materials, where location-dependent properties require calibration to be repeated at multiple points to produce a detailed property map. Additionally, a limited understanding of how individual parameters of the CP models influence stress–strain predictions across the strain spectrum compounds these issues, making it challenging to utilise CP models for efficient materials design. To address these limitations, we developed an integrated framework that combines deep neural network (DNN) surrogates, optimisation algorithms (OAs), and explainable AI (XAI) techniques. We also utilised experimental tensile data from AM AlSi10Mg alloy as ground truth since AM materials are expected to benefit the most from our investigation. We demonstrate that, by using OAs such as a Natural Evolutionary Strategy or a Genetic Algorithm, the calibration process can be made more accurate and significantly accelerated. We also investigated the utility of employing deep neural network (DNN) surrogates of CP simulations in the calibration process. The fast-solving DNN surrogates achieved substantial time savings in the absence of OAs, i.e., during exhaustive parameter searches mandated by trial-and-error strategies. However, their effectiveness in parameter discovery was context-dependent when used in conjunction with OAs, since OAs can sometimes converge with fewer simulations than required for DNN training. Furthermore, we applied Shapley Additive exPlanations (SHAP), an XAI method, which revealed intricate interactions among some CP parameters, offering insight into why conventional trial-and-error calibration approaches often prove challenging. Our study contributes to strengthening the practical relevance of CP models for modelling-informed materials engineering and optimisation applications. Finally, our integrated framework offers broad applicability beyond materials modelling, enabling accelerated discovery of tuneable parameters in phenomenological models and providing deeper insight into their contributions to predictions. Full article

Despite rapid growth in the AI endorser market, the psychological mechanisms governing their effectiveness remain theoretically fragmented. This study proposes the AHICC (AI–Human Interface in Consumer Cognition) model—integrating the Stereotype Content Model, Uncanny Valley hypothesis, anthropomorphism theory, Source Credibility Model, and Parasocial Interaction theory—to explain consumer responses to AI endorsers. A fully crossed 3 (endorser type: AI vs. hybrid vs. human) × 3 (anthropomorphism level: low vs. moderate vs. high) × 2 (technological transparency: low vs. high) between-subjects factorial experiment (n = 252) was conducted. Twenty-one sub-hypotheses were tested using MANOVA, polynomial regression, SEM, and bootstrap mediation analysis. All 21 sub-hypotheses were supported. AI endorsers outperformed human counterparts on brand attitude and purchase intention. Polynomial regression confirmed an inverted U-shaped Uncanny Valley effect with an optimal anthropomorphism level of 4.7 (7-point scale). High technological transparency attenuated the Uncanny Valley effect by approximately 60%. Dual-pathway mediation through cognitive and affective routes was confirmed, and TRI and product complexity emerged as significant boundary conditions. The AHICC model offers the first comprehensive framework for the AI endorser context, providing theoretically grounded guidance on anthropomorphism calibration, transparency strategy, and product-category-specific endorser selection. Full article

This paper presents a novel design approach for mitigating the adverse effects of antenna mountings on the radiation pattern of GNSS antennas. By employing a resistive structure integrated into the ground plane, the proposed solution suppresses unwanted edge diffraction and near-field reflections caused by nearby mounting hardware. The design is developed using the concept of tapered resistive sheets and optimized using a customized cost function that accounts for pattern degradation across multiple realistic mounting configurations, ensuring robust performance under varying installation conditions. The resulting structure is fabricated using additive manufacturing (AM), enabling precise realization of complex resistive profiles with tailored surface impedance. Comprehensive validation through both electromagnetic simulations and experimental measurements demonstrates significant improvements in radiation pattern stability and reduced sensitivity to near-field objects, particularly in critical GNSS bands such as E5a/L5 and E1/L1. The results demonstrate that the proposed structure significantly enhances antenna reliability and calibration integrity in real-world deployments, offering a practical, hardware-based solution to a persistent challenge in high-precision GNSS systems. Full article

The rapid growth of the sharing economy has improved resource utilization in car-sharing, yet it has also sharpened market competition and diversified user demand. A persistent obstacle is the low coordination efficiency between asset-heavy operating companies and traffic-driven platforms, whose misaligned objectives waste social resources. This paper uses differential game theory to analyze their dynamic coordination strategies and benefit allocation mechanisms. The Nerlove–Arrow model captures the evolution of brand goodwill, while the company’s decisions on station layout, vehicle dispatch, and pricing, together with the platform’s advertising investment, form the core decision variables in a two-party game framework linking the asset side and the traffic side. Compared with the non-cooperative Nash equilibrium, the cooperative mode removes the double marginalization effect, strengthens the investment incentives of both parties, and raises the system’s steady-state goodwill and total profit, achieving a Pareto improvement. To ground the cooperative framework in rigorous theory, we supply a verification theorem confirming that the linear candidate value functions satisfy the Hamilton–Jacobi–Bellman equations over the entire admissible state space. A formal proof of instantaneous rationality ensures that neither party falls into a cooperation trap on the horizon $[0,T]$, and the asymptotic stability of the steady-state goodwill trajectory is established. We further endogenize the revenue-sharing coefficient through a generalized Nash bargaining model that admits asymmetric bargaining structures, and introduce a Stackelberg leadership benchmark as a third comparative regime. Sensitivity analyses with respect to the discount rate and user heterogeneity confirm the robustness of the findings. A dedicated discussion section bridges the gap between idealized parameterization and data-driven calibration, describing practical pathways via A/B testing, user churn metrics, and econometric estimation of demand parameters. The results offer a scientific decision-making reference for strategic cooperation in the car-sharing industry. Full article

This study presents a hybrid spatio-textual matching approach for integrating historical web-derived address datasets with a municipal reference dataset, using the 2009 administrative delineation of Şişli (Istanbul) as a case study. The proposed approach addresses challenges commonly encountered in data obtained from web archives, such as lexical differences, abbreviations, heterogeneous structures, and missing address information. The methodology consists of three main stages: (i) preprocessing and structuring of web-based address records; (ii) hybrid matching, combining deterministic rules with similarity-based methods; and (iii) post-matching geographic enrichment using an Application Programming Interface (API) to provide supplementary geographic context for matched records. The matching process is conducted exclusively between historical datasets; contemporary geographic information is used only after the completion of the matching process to provide additional contextual information. The methodology integrates token-based, vector-based, and structural similarity measures within a calibrated scoring scheme to improve the matching of ambiguous and inconsistent address records. The results indicate that 65.4% of the records were automatically accepted, 7.3% required manual review, and no suitable candidate was found for 5.4%. Deterministic matching results reveal that strict rule-based approaches are highly sensitive to data integrity and attribute consistency, especially in heterogeneous web-based datasets, highlighting the value of combining multiple similarity measures within a hybrid matching strategy. The API-based enrichment results provide supplementary geographic context regarding the contemporary surroundings of matched records, while historical interpretations remain grounded in the original archival datasets. In this context, the study may contribute to the integration of historical web-based address data with structured municipal datasets under heterogeneous archival data conditions. Full article

This article examines the medium to long-term enrollment patterns of an AI-based platform designed to support children in learning robotics and participating in a national robotics competition in Romania. Drawing on registration and participation data covering students and teachers across urban and rural schools between 2020 and 2025, the study documents a consistent pattern: an initial period of high enrollment and rapid adoption followed by a steady decline over time. A key feature of the initiative is that hardware, platform access, and learning resources were provided entirely free of charge, allowing cost-related explanations for the decline to be set aside and structural and human factors to be examined directly. The paper makes two primary contributions. First, it proposes a System Dynamics framework grounded in innovation diffusion theory as a first-generation calibration model for understanding AI-enabled educational robotics adoption in a resource-constrained national context. The model is designed to be progressively tested and refined as anonymized aggregate data accumulates, and it relies exclusively on anonymized aggregated public data in accordance with GDPR requirements. Second, it advances the hypothesis that an AI-based educational platform, even one from which all financial barriers have been removed, will experience sustained enrollment decline in the absence of adequate human teacher involvement. The empirical trajectory and model outputs are consistent with this hypothesis and motivate further investigation. This represents a hypothesis-generating and framework-building paper. The framework reveals pronounced urban-rural disparities and differential outcomes by age of entry. All findings are presented as model-generated hypotheses rather than empirically demonstrated conclusions. The paper invites researchers gathering comparable data from similar initiatives in other countries to collaborate in testing and refining the model. The central conclusion is cautiously optimistic: AI may support robotics education adoption, but it is not a substitute for dedicated teachers, and without sustained investment in human capital, even a financially accessible platform is insufficient to maintain long-term enrollments. Full article

The new LOAC2 optical aerosol counter is designed to detect liquid and solid particulates across 19 to 30 size classes within the 0.15–90 µm size range, and to provide their main typology. The instrument can be used at ground level and on all kinds of balloons, including weather balloons, up to an altitude of about 35 km. The measurements are based on principles established for the previous version of LOAC, now incorporating improved electronics and detection geometry. Counting is performed at small scattering angles in the diffraction domain, making it insensitive to the refractive indices and the porosity of the particles, thus allowing a direct relationship between scattered intensity and aerosol size. Typology identification is now performed at three additional scattering angles, where the scattered flux is highly sensitive to the refractive index of the different aerosol families present in the atmosphere. The calibration was conducted using calibrated spherical and irregular grains, as well as different types of solid particles. Several intercomparison sessions with other counters and with reference mass-concentration air quality monitoring stations were carried out indoors, in an atmospheric simulation chamber, and in outdoor ambient air. The agreement between LOAC2 and the other instruments is good, confirming the ability of LOAC2 to be used for scientific studies and for monitoring atmospheric aerosols. Full article

Background/Objectives: Medication-related osteonecrosis of the jaw (MRONJ) is a serious complication of bisphosphonate therapy, whose risk is currently assessed through qualitative staging systems that do not integrate pharmacological determinants of cumulative drug exposure. The aim of this study is to present the Bisphosphonate Accumulation Index (BAI), a pharmacologically derived, dimensionless scalar quantifying cumulative exposure to bone-targeted antiresorptive agents by integrating relative potency, administered dose, dosing frequency, route-specific bioavailability, and treatment duration, for use as a pre-procedural assessment tool in patients receiving bisphosphonates. Methods: The BAI combines five pharmacologically grounded parameters from peer-reviewed literature: (1) relative antiresorptive potency referenced to etidronate; (2) dose per administration (mg); (3) monthly dosing frequency; (4) bioavailability route; and (5) years of treatment within the preceding 10-year window. The model includes nine bisphosphonates registered in Italy. Results: The BAI spans approximately five orders of magnitude (from <1000 for short-term oral therapy to >120,000 for monthly intravenous zoledronic acid). Four analyses support the model: sensitivity analysis identifies relative potency as the main source of variance; ecological calibration against nine MRONJ incidence data points yielded r = 0.911 (p = 0.0006, R² = 0.829), indicating that the BAI accounts for approximately 83% of the population-level variance in published incidence rates across heterogeneous regimens (ecological correlation; this does not establish individual-level predictive validity); Monte Carlo simulation on 10,000 patients generated a plausible exposure-strata distribution (6.1% low, 66.6% moderate, 27.3% high); and concordance analysis with a DDD-based metric showed discordance in 7/8 regimens. Conclusions: The BAI is a transparent, reproducible, pharmacologically grounded metric of cumulative antiresorptive exposure addressing the quantitative gap identified in the AAOMS 2022 Position Paper. The BAI measures pharmacological exposure, which is a necessary but insufficient component of MRONJ risk; clinical modifiers such as corticosteroid co-administration, diabetes, renal function, and procedure type are not integrated and must be evaluated independently. The provisional exposure strata reported here (<1000, 1000–10,000, >10,000) are hypothesis-generating and intended solely to guide the design of validation studies; they should not be used as clinical decision rules until prospective patient-level validation has been completed. Full article

Grading open-ended student work consistently remains a persistent challenge in higher education, and the recent rise of large language models (LLMs) has renewed interest in rubric-guided automated scoring. However, a key gap remains: most studies report correlation rather than agreement, rarely benchmark models against a local human–human baseline, and seldom test whether simple post hoc calibration improves operational fit. This study addresses that gap by examining whether a rubric-guided LLM can approximate local human grading practice for text-based responses in three university courses, using agreement-oriented rather than correlation-only evidence. A total of 930 student responses from Prompt Engineering, Photoshop Design, and AI Video Production were scored by two human raters and by ChatGPT using the same five-criterion analytic rubric (Accuracy, Logical Flow, Specificity, Quality, and Originality; 0.0–3.0 each; Total 0–15). Human consensus (HC) was defined as the mean of the two human scores and was treated as a pragmatic reference rather than a ground truth. Pairwise agreement among H1, H2, AI, and HC was evaluated using ICC(3,1), Pearson correlations, mean absolute error (MAE), Bland–Altman bias and limits of agreement (LoA); a course-specific held-out calibration analysis was additionally conducted. For the Total score, human–human agreement was strong (ICC = 0.819 [0.797, 0.839]). AI–H1 and AI–H2 Total-score agreement were ICC = 0.700 [0.666, 0.732] and 0.767 [0.739, 0.792], respectively, while AI–HC agreement was ICC = 0.763 [0.735, 0.789], with MAE = 1.603 and LoA = [−4.246, 4.045]. At the trait level, AI–HC ICCs exceeded H1–H2 ICCs for all five rubric dimensions, although Quality remained weakly defined in the human baseline. On a 70/30 held-out test split, a course-specific linear calibration modestly improved Total-score ICC from 0.774 to 0.782 and reduced MAE from 1.624 to 1.215, narrowing the LoA from [−4.290, 4.188] to [−3.157, 3.329]. However, threshold-adjacent agreement remained imperfect after calibration. The principal contribution is a conservative, multi-metric agreement benchmark of rubric-guided LLM scoring against a local human baseline, together with a held-out calibration test that informs deployment. The findings concern written responses only and support a conservative conclusion: rubric-guided LLM scoring can assist human grading under fixed local rubrics, but the current evidence supports calibrated human–AI co-grading rather than unsupervised replacement. Full article