Reconfigurable intelligent surfaces (RISs) are emerging as a key enabling technology to engineer the wireless propagation environment in 5G/6G systems. This paper presents the design, characterization, and control of a high-resolution RIS targeting slowly time-varying scenarios, in which fine-grained and stable phase control is more valuable than fast reconfiguration. Starting from the OpenRIS unit-cell layout, the cell is re-optimized for single-polarization operation and continuous phase tuning through a single varactor diode, exploiting the full tuning range enabled by a high-resolution DAC infrastructure rather than multi-bit quantization. The unit cell is analyzed via full-wave 3D FEM simulation in COMSOL Multiphysics and optimized to maximize the reflection-phase excursion while limiting amplitude modulation across the 5G N78 band (3.60–3.78 GHz). The design is experimentally validated in a WR-284 waveguide fixture, exhibiting a phase excursion approaching the full 360∘ near resonance, with an amplitude variation below 1 dB over the bias sweep at any given frequency, while the average reflection level decreases by about 2 dB from the center to the upper band edge. A fifth-order polynomial phase–voltage calibration feeds a lookup table driving a layered control system based on a Python HMI, a server, and STM32-driven 16-bit DACs. Experimental measurements confirm that the control chain delivers the commanded bias voltages to the addressed unit cells within measurement uncertainty; the array-level beamforming is assessed at simulation level under idealized (unit-magnitude) assumptions, while the experimental characterization of the assembled surface is left to future work.
In SRAM-based compute-in-memory (CIM), read-bitline (RBL) charging and discharging depend on the physical bit pattern stored in the memory array, so the energy-relevant code statistic should be defined with respect to the actual read-port polarity. This paper presents a read-polarity-aware row-wise offset-encoding method for W4A8 INT4 weights. In the evaluated Q-sensed 8T topology, the stored logical one is the discharge-active state; hence, the topology-specific read-active density equals the stored-one fraction. Under the exact whole-row INT4-feasibility protocol, a nonzero row offset is accepted only when every translated valid signed-INT4 code remains within [−8, 7]; no clipping, saturation, wraparound, or remapping is permitted, and zero offset remains the fallback. The complete software evaluation covers 286 quantized modules, 579,464 physical 16 × 16 tile positions, and 147,156,296 quantized weights across ResNet-18, MobileNetV3-Small, and SmolLM2-135M. Circuit re-validation uses 300 independent tile-policy samples, 1200 matched baseline-selected bitplane pairs, and 2400 successfully completed transistor-level Spectre simulations. The balanced circuit population yields an aggregate local SRAM readout-energy reduction of 5.03%, with a sample-cluster bootstrap 95% confidence interval of 4.12–6.02%. After four-bitplane aggregation, relative read-active-density reduction and local SRAM readout-energy reduction exhibit Pearson r = 0.81 and Spearman ρ = 0.75, indicating a substantial but imperfect relationship. The directly validated no-offset, positive-offset, and signed-offset policies preserve the corresponding model-level Top-1 accuracy or perplexity. Proposal-specific digital overheads and metadata-storage capacity are quantified separately, whereas representative physical SRAM/ROM metadata-access energy remains uncharacterized. Accordingly, the measured energy benefit is limited to local SRAM readout; a net energy reduction at the complete CIM-macro or system level, robustness across all evaluated PVT conditions, and robustness to process mismatch are not established by the present evidence.
Millimeter-wave (mmWave) radar-based gesture recognition has attracted increasing attention for real-time human–computer interaction owing to its robustness to illumination changes, privacy-preserving sensing capability, and suitability for embedded deployment. However, existing single-stream models often couple heterogeneous point-cloud and temporal statistical features, which may limit their ability to capture fine-grained motion patterns and key action frames. To address this problem, this paper proposes a dual-stream long short-term memory network (LSTM) and a bidirectional gated recurrent unit (BiGRU) with an attention mechanism (Attention-BiGRU) network for mmWave radar-based hand gesture recognition, termed as DSTG-Net. In the proposed DSTG-Net framework, an LSTM branch is used to process radar point-cloud sequences and extract fine-grained spatio-temporal features, while an Attention-BiGRU branch models global motion trends from statistical and temporal-difference features. The attention mechanism in the Attention-BiGRU branch is introduced to emphasize discriminative frames during gesture transitions, and the complementary features from the two branches are fused through feature concatenation for final classification. The proposed method is evaluated on a public mmWave radar gesture dataset to verify its recognition performance, and an additional self-built near-field dataset is used to test its effectiveness under a constrained acquisition scene. The proposed method achieves recognition accuracies of 97.40% and 98.75% on the two datasets, respectively, outperforming several baseline models. The Raspberry Pi-based implementation with a TI IWR1642 radar confirmed the functional feasibility of the proposed pipeline.
The frequency-dependent impedance of electrical connectors, switchgear loops, and related contact structures can indicate degradation at electrical contact interfaces. Conventional contact measurements with a vector network analyzer require dedicated adapters and impedance matching, and the object under test usually has to be removed from the operating system. To simplify field measurements and avoid direct electrical contact, this study employs an established two-probe inductive-coupling/ABCD de-embedding framework in conjunction with a purpose-designed broadband current probe and calibration fixture. Commercial injection and receiving probes were first evaluated to identify practical limitations associated with probe resonance, magnetic-path air gaps, cable clamping, and probe-to-probe coupling. A parametric CST study on ferrite material, winding turns, and core geometry was then used to select a 3W800 NiZn ferrite core with six winding turns and dimensions of 5 mm inner diameter, 10 mm outer diameter, and 5 mm height. The prototype exhibited relatively flat S12/S21 responses from 1 to 400 MHz, with S21 changing from approximately −13.5 dB at 1 MHz to −17.5 dB at 400 MHz and without a pronounced resonance. Quantitative impedance accuracy was validated only over 1–100 MHz: validation against an impedance analyzer using a 108 nH inductor and an 82 pF capacitor yielded mean relative magnitude errors of 4.00% and 4.5%, respectively. Field measurements on a 10 kV switchgear circuit were additionally conducted over 1–30 MHz to demonstrate non-invasive broadband impedance acquisition under practical contact conditions. Accordingly, the 1–400 MHz range in this work refers to probe-transfer characterization rather than experimentally validated impedance accuracy over the full band.
In recent years, the fusion of millimeter-wave radar and vision has emerged as a prominent research hotspot and a mainstream solution for autonomous driving perception. This integration spans multiple hierarchical levels, and the evolution of each level is not an isolated technological advancement, but rather a synergistic outcome driven by technological maturity, computational constraints, and mass-production requirements. Despite the inherent information loss associated with decision-level fusion, it remains the predominant engineering approach in the industry due to its superior functional safety and cost-effectiveness. Conversely, feature-level fusion has developed rapidly, propelled by a positive feedback loop of deep learning, bird’s-eye view (BEV) representations, and cross-modal attention mechanisms, moving beyond exclusive reliance on the Transformer architecture. Meanwhile, data-level fusion directly integrates raw radar point clouds and image pixels, a strategy that theoretically minimizes information loss. However, its large-scale deployment in practical engineering applications is hindered by critical bottlenecks, including poor interpretability, vulnerability to cross-sensor fault propagation, and severe challenges in safety isolation. From an engineering perspective, this paper systematically analyzes the evolutionary trajectory of millimeter-wave radar and vision fusion technologies, clarifying the parallel coexistence and adaptive deployment of these three fusion levels in practical autonomous driving scenarios.
With the rapid global expansion of wind energy and increasing deployment of large offshore turbines, achieving reliable sensorless control of wind energy conversion systems (WECSs) has become increasingly important. This paper presents a comprehensive review of modern sensorless control techniques for the doubly fed induction generator (DFIG), which is a dominant technology in variable-speed WECSs. This review focuses on control strategies and rotor speed and position estimation techniques, covering their theoretical foundations, operational characteristics, and emerging research trends. In addition, sensorless operations of the DFIG beyond power generation, including grid synchronisation and stand-alone operation, are presented, providing a foundation for future research in the field.
Assessing the quality of super-resolved images is important for comparing reconstruction algorithms, but pixel fidelity, perceptual appearance, and structural preservation do not always agree. We investigate whether keypoint detector response maps and detected keypoints can act as trainable structural indicators for aligned full-reference super-resolution image quality assessment (SR-IQA). A contrastive Multi-Scale Index Proposal (MSIP) objective specializes Key.Net toward SR-like resolution loss, candidate checkpoints are screened without subjective labels, and six response- and keypoint-based measures are evaluated on four subjective benchmarks. The results show a redistribution rather than a uniform improvement: MSIP correlations increase on three of the four benchmarks, both repeatability variants decrease for every trained family on SISAR and RealSRQ, and general keypoint performance on HPatches decreases for trained checkpoints. At the benchmark level, established comparators such as TOPIQ-FR, RQI, and DISQ remain stronger, and on three of the four datasets the best keypoint-based result is still obtained with the pretrained detector. The measures nevertheless retain quality-related variation after conditioning on ten IQA controls in 88 of 120 tested hypotheses, with markedly weaker evidence on RealSRQ. Their practical value is therefore diagnostic: the response maps localize the structures behind an HR–SR discrepancy, complementing rather than replacing established SR-IQA metrics.
On resource-constrained edge hardware, fixed-point Transformer inference is constrained by GEMM kernels and non-GEMM normalization paths. In LayerNorm and Softmax, division, inverse square root, and denominator reciprocal operations complicate fixed-point datapath design. We propose DIFT (Division and Inverse-Square-root-Free Transformer normalization), a fixed-point scheme that removes division and inverse square root from the targeted normalization paths. Starting from a trained checkpoint, we keep the weights, matrix multiplications, and Softmax exponentials unchanged, replacing only the LayerNorm inverse-standard-deviation path and Softmax denominator reciprocal path with Q7I2 fixed-point approximations. Calibration remains separate from final evaluation. We evaluate fine-tuned BERT-base on four GLUE tasks with three seeds and fixed-length-block perplexity (PPL) of GPT-2 small on WikiText-2. LayerNorm-only, Softmax-only, and Joint (DIFT) settings isolate the error sources. For BERT, the mean accuracy change under Joint (DIFT) is approximately −0.24 to −0.08 percentage points. GPT-2 test PPL increases from 29.009 to 31.092, a ΔPPL of +2.083. Under this protocol, the same approximation affects token-level-likelihood PPL more strongly than classification accuracy. The results provide an algorithm-level assessment before hardware implementation on edge NPUs, FPGAs, or custom ASICs. Synthesis and hardware measurements of latency, area, power, and energy are outside the scope of this study.
This paper proposes a Koopman-based stochastic model predictive control (SMPC) approach for unknown nonlinear systems by exploiting partial probabilistic information. Unlike existing Koopman-based SMPC methods that primarily rely on the first- and second-order moments of stochastic Koopman modeling error, the proposed approach further incorporates available support information into the controller design. By jointly utilizing the mean, covariance, and support information of the resulting uncertainty, the chance-constrained optimization problem is reformulated into a tractable deterministic optimization problem with reduced conservatism. Moreover, the proposed approach is theoretically shown to be no more conservative than the corresponding RMPC in terms of constraint tightening under identical constraint requirements. Furthermore, recursive feasibility and closed-loop quadratic stability of the proposed control scheme are established theoretically. Simulation studies on spacecraft attitude control demonstrate that the proposed method reduces the performance index by 6.7% and 17.0% compared with a Koopman-based SMPC using mean and covariance information and RMPC, respectively, while maintaining a comparable average computation time of approximately 0.010 s.
Timely and accurate classification of lung diseases from chest X-ray images remains an important healthcare challenge. Machine-learning and deep-learning methods can support automated classification, but their evaluation may be affected by class imbalance, feature redundancy, dataset leakage, and computational cost. This paper evaluates an artificial immune system (AIS)-evolved decision-tree ensemble using fold-specific ResNet18 features. All within-dataset experiments use duplicate-family-aware five-fold splits. Within each fold, standardization and adaptive principal component analysis (PCA) are fitted to the training features, and the Synthetic Minority Over-sampling Technique (SMOTE) is applied only to the reduced training data. Each candidate tree is assigned an affinity based on out-of-bag macro-F1. In the primary run, mean within-dataset macro-F1 was 98.07%, 99.48%, and 98.26% for Datasets 1–3, respectively, and 96.66% for the exploratory Dataset 4. Because of extensive cross-dataset image reuse and conflicting labels, Dataset 4 does not provide independent evidence of clinical lung-cancer detection. A five-seed repeated-initialization analysis repeated the complete fold-specific feature and classification pipeline while preserving the same folds. Mean macro-F1 differences between AIS and the prespecified static comparator for each dataset, calculated as AIS minus the comparator, were −0.18, 0.00, −0.35, and −0.13 percentage points for Datasets 1–4, respectively. Using the same sign convention, mean differences between AIS and the fixed random tree ensemble ranged from −0.05 to +0.05 percentage points. Population diagnostics showed that evolution improved individual-tree macro-F1 but reduced pairwise disagreement, without a consistent majority-vote gain. Median latency from an already decoded image to prediction ranged from 38.86 to 61.20 ms on one CPU thread and from 3.43 to 6.36 ms on an RTX 4090. The results do not establish a practically important or consistent predictive advantage from AIS evolution. The study provides a reproducible and duplicate-controlled framework for evaluating AIS-based tree ensembles.
Drug–target interaction (DTI) prediction is a critical step in drug discovery, and accurate prediction of potential interactions can significantly accelerate the drug-development process. Although deep-learning approaches have achieved promising performance in DTI prediction, two challenges remain: single models often fail to comprehensively capture heterogeneous sequence information, resulting in limited stability and generalization, while insufficient integration of local and global features restricts interaction representation. To address these limitations, we propose HFEDTI, a DTI prediction model that integrates hierarchical feature fusion and weighted ensemble learning. Specifically, a residual convolutional neural network (ResCNN) is employed to extract local structural features of drugs and targets, while a self-attention-based hierarchical bidirectional long short-term memory network (SAHBiLSTM) captures global contextual dependencies. Furthermore, a hierarchical heterogeneous attention mechanism is introduced to align and fuse multi-level cross-modal representations, and a weighted ensemble strategy based on validation performance ranking is developed to enhance model robustness and generalization. Experimental results on three benchmark datasets demonstrate the effectiveness of HFEDTI. On the DrugBank dataset, HFEDTI achieves an AUC of 0.9238 and an AUPR of 0.9327, improving the best-performing baseline by 0.90 and 1.40 percentage points, respectively. Moreover, HFEDTI consistently achieves strong performance on the C. elegans and Human datasets, further validating its effectiveness and generalization capability for DTI prediction.
EASE-CloudNet is a two-phase safety-alignment framework for generative small language models (SLMs) deployed on resource-constrained edge nodes. Its input is a natural-language user query and its output is a safe, helpful natural-language response or refusal; network-traffic classification and resource-scheduling actions are outside the task evaluated in this study. In Phase 1, a cloud teacher uses a security policy graph to generate structured safety rationales and response targets, which are distilled into Qwen2.5-1.5B/3B and Llama3.2-3B students. In Phase 2, an offline heterogeneous graph and a two-layer GraphSAGE model identify vulnerable semantic regions; these vulnerability targets supervise a lightweight edge-side router. We formulate deployment cost as a differentiable gate-conditioned expectation, so measured latency and energy constants affect the router through its reasoning probability. In the Qwen2.5-1.5B ablation experiments, the full model obtains 3.9% StrongREJECT ASR, 54.7% MMLU accuracy, and 70 average generated tokens; an A100 reference profile reports 18.8 ms/query and 2.37 J/query, or 3.3% latency and 2.6% measured GPU-energy overhead over the unaligned model. Physical edge runs measured a direct/reasoning end-to-end latency of 41.2/68.7 ms on Jetson Orin NX and 62.5/105.3 ms on Snapdragon 8 Gen 3, with a direct/reasoning energy of 0.48/0.79 and 0.71/1.18 J/query, respectively. Equal-seed Holm–Bonferroni-corrected tests confirm lower ASR than EASE on StrongREJECT and WildJailbreak for all three base models (p<0.01).
Production scheduling in industrial settings requires the simultaneous coordination of tasks, machines, operators, and precedence relations under limited resource availability. This study proposes a Constraint-Aware Cuckoo Search Algorithm (CACSA) for resource-constrained production scheduling and makespan minimization. The modification combines CSA exploration with a constraint-aware and operator-aware repair layer that converts continuous Levy-flight perturbations into feasible discrete schedules by selecting admissible machine–operator pairs, respecting precedence relations, and applying local repair-guided refinement within the same computational budget. The experimental section combines instance-level schedule documentation with comparative statistical evaluation. The documented instance set includes Scenarios I-VI and two 35-task variants, while the repeated comparison uses 30 independent runs per method and an equal computational budget for CACSA, baseline CSA, Genetic Algorithm, Particle Swarm Optimization, Differential Evolution, Simulated Annealing, and Ant Colony Optimization. The evaluation reports best, mean, standard deviation, worst makespan, runtime, feasibility rate, lower-bound gaps, and parameter sensitivity. The results show that CACSA preserves feasibility in all tested scenarios and generally improves baseline CSA performance, with the clearest gains observed in the larger 35-task instances. The study therefore positions the proposed CACSA as a practical and reproducible scheduling approach, while distinguishing feasibility, robustness, and comparative dominance as separate empirical claims.
Residual vibration of industrial manipulators can limit positioning efficiency and dynamic accuracy during high-speed motion. This study develops an integrated vibration-suppression framework for a rigid-link manipulator with flexible-joint dynamics. A controller-oriented rigid–flexible model with lumped disturbances is established, and a disturbance observer (DOB) is employed as the inner-loop compensation layer under a small-gain robustness constraint. On the compensated nominal model, partial eigenstructure assignment (PESA) selectively increases the damping of the retained flexible modes while preserving the rigid-body eigenstructure associated with trajectory tracking. A pose-dependent gain-scheduling mechanism further updates the PESA feedback gain to accommodate configuration-dependent modal-frequency variation. Numerical comparisons with conventional PID, standalone DOB, and standalone PESA demonstrate improved residual-vibration attenuation and settling behavior. Hardware tests on an Aubo i5 manipulator, with 16-channel responses directly acquired under the respective control configurations, further show an approximately 80% reduction in the representative low-frequency vibration amplitude relative to the PID baseline under the considered operating condition.
To prevent DC-bus voltage sag and overvoltage during power-supply takeover after a public-grid outage in an industrial microgrid, this study proposes a continuous power-supply control strategy for critical loads that combines dynamic source-side-deficit feedforward with DC-bus energy compensation. Direct feedforward of real-time load power can overlap with residual injection from the grid-side converter, repeatedly charging the DC-bus capacitor and causing an overvoltage. Based on the three-port power-balance relationship among the grid-side interface, bidirectional DC–DC converter, and load-side grid-forming interface, the handover transient is attributed to residual source-side supply, delayed DC–DC power buildup, and persistent load-side consumption. The source-side power deficit, defined as the difference between the load-side DC-power demand and actual source-side injected power, is filtered to generate a dynamic feedforward signal. Thus, storage power is established adaptively as source-side supply withdraws, avoiding repeated compensation. An energy-compensation branch derived from the relation between DC-bus capacitor energy and squared bus voltage corrects residual mismatch caused by filtering, converter losses, and DC–DC dynamic lag. Mathematical models and a small-signal three-port DC-bus model are developed. Simulation and hardware-in-the-loop results verify the effectiveness of the proposed strategy.
Mobile industrial devices increasingly cross administrative domains. A maintenance terminal certified in one factory can be dispatched to another, while autonomous vehicles cross fog domains while reaching cloud digital twins. Existing solutions force a trade-off. Long-lived certificates expose a stable identity that every visited domain can track and that is clonable once a device is captured, while single-gateway token services concentrate issuing power in one trusted node and asynchronous revocation leaves an unquantified window in which a revoked device is still accepted. Here, we present HRCred, which converts hardware identities rooted in physical unclonable functions (PUFs) into domain-bound, epoch-bound, and threshold-issued short-lived pseudonymous credentials. A device proves possession of its reconstructed root key to its home domain using only symmetric primitives. A set of fog issuers jointly signs each credential with a t-of-n threshold BLS signature, so no coalition of fewer than t issuers can mint one. Finally, a monotonic revocation-epoch mechanism yields a configurable upper bound on how long a revoked credential can still be accepted, which we prove and validate. On constrained hardware, the device side costs 3.6 mJ (18.8 ms authentication and 5.4 ms verification, on par with the lightest single-gateway token) while resisting up to t−1 compromised issuers, cutting cross-domain linkage AUC to 0.55, and keeping all 12,000 measured post-revocation acceptances below the analytical bound.
Binary third-party library detection is a fundamental task in software composition analysis, vulnerability tracing, and software supply chain security. Existing approaches mainly rely on function-level matching or global similarity computation, while paying insufficient attention to region-level reuse structures formed by function call relationships. Moreover, their detection results generally lack interpretability. To address these limitations, this paper proposes LibXSub, an explainable subgraph-based framework for binary third-party library reuse region detection. Based on function call graphs, LibXSub integrates instruction operation features and control-flow structural features of function nodes, and employs a Siamese graph neural network to compute the structural-semantic similarity between target candidate regions and candidate library regions, enabling region-level third-party library reuse detection. To improve interpretability, LibXSub introduces an explainable subgraph generation mechanism that identifies critical call relationships through edge masks and characterizes the contribution of different function feature dimensions to region similarity prediction through node feature masks. Experimental results show that LibXSub achieves F1-scores of 0.900 and 0.877 on Dataset_1 and Dataset_2, respectively, outperforming existing methods. Furthermore, it generates compact explanatory subgraphs while maintaining high explanation fidelity, demonstrating the effectiveness of the proposed framework for both region-level third-party library reuse detection and result interpretation.
Generating a virtual anchor from speech has to satisfy three demands at once: the face must stay recognizable as the same person, the motion has to be temporally coherent, and the lips must follow the audio. Existing methods often fall short on the first two points because identity and expression share one entangled representation, and temporal dynamics are modeled only implicitly. We propose DynaID-VAE to address these problems. At its core is an identity–expression disentangled conditional VAE (DC-VAE) that splits the latent space into a time-varying expression subspace and a static identity subspace, held apart by mutual-information minimization and orthogonality regularization. A temporal memory module (TMM) then regularizes the expression trajectory: a GRU propagates sequential state, attention retrieves from a learnable key–value prototype memory, and residual fusion combines the two. Multiscale adversarial supervision and lip–audio synchronization losses complete the training objective. We evaluate on VirtualAnchor-100, a benchmark we recorded ourselves (100 h, 10 anchors), under two complementary protocols. Cross-identity driving is scored only with non-paired measures, namely lip synchronization, distributional video quality, and identity preservation; full-reference image metrics are confined to a self-reenactment protocol, where a genuine paired ground truth exists. DynaID-VAE outperforms the one-reference baselines Wav2Lip, PC-AVS, SadTalker, and DiffTalk under both protocols and on unseen VoxCeleb2 identities. The margins are stable across five identity-disjoint, nested cross-validation folds and are confirmed by an external SyncNet evaluator that never takes part in training, while the model runs at 41.2 FPS with 14.3 M parameters. Ablations separate the contribution of each regularizer and each TMM component. Linear and capacity-matched non-linear probes quantify the factorization as a large reduction of decodable reference identity; full independence is not claimed. A user study confirms the perceptual gains.
Concept drift can degrade encrypted-traffic classifiers deployed at the network edge as applications, protocols, and usage patterns evolve. This paper formulates federated continual learning under asynchronous real- and virtual drift and proposes DriftGuard, a framework combining a two-level per-node detector, selective adapter-based adaptation with Fisher importance masking and class-balanced replay, and drift-aware server aggregation. Level 1 detects distributional changes in learned representations, while Level 2 monitors supervised prediction errors to provide evidence consistent with decision-relevant drift before selective adaptation is activated. We further derive a convergence bound under stated assumptions that explicitly incorporates environmental variation, detection delay, and false alarms. DriftGuard is evaluated in a controlled simulation using reproducible synthetic traffic-like features under sudden, gradual, virtual-only, and recurring drift. Across five independent runs, results are reported with standard deviations, 95% confidence intervals, and paired statistical comparisons. DriftGuard maintains competitive classification accuracy while limiting forgetting of stable classes, with its clearest advantage observed under gradual asynchronous drift. Results also show that immediate adaptation using ground-truth drift states does not necessarily improve performance under the evaluated adaptation policy. The findings provide controlled methodological validation rather than evidence of production-scale deployment performance.
Despite recent advances in dialogue topic segmentation, existing work provides limited evidence about why individual utterances are predicted as boundaries and how local explanations depend on the selection strategy and perturbation protocol. Using fixed checkpoints of 3LHSeg, a hierarchical dialogue topic segmentation model, we evaluate a boundary-centered framework that compares five local utterance-selection strategies through complementary perturbation diagnostics. The audit also examines configuration sensitivity and explanation stability and uses Random-Baseline Gain (RBG) as a diagnostic of relative local distinctiveness by contextualizing comprehensiveness against matched random subsets from the same local candidate window. Experiments on TIAGE, QMSum, and Friends show that Leave-One-Out ranks most favorably under the adopted perturbation-based diagnostics, although this result is protocol-specific. The diagnostics provide partially overlapping information and vary with the local configuration. Raw deletion area-under-the-curve values are consistently positively associated with initial boundary confidence, but a centered control substantially attenuates this association in most dataset–strategy combinations. Overall, explanation behavior, probability quality, confidence, stability, and boundary correctness should be treated as distinct dimensions, supporting multi-diagnostic and configuration-explicit auditing rather than reliance on a single explanation score.
