Decentralized oracle networks pose significant security risks to blockchain systems due to transaction malleability, which can lead to double-spending and integrity issues. While existing solutions such as DAON, SegWit, and SecPLF improve specific aspects of security, they do not address Oracle-driven transaction malleability on a transaction level. DAON focuses on decentralized oracle consensus and reputation mechanisms, but it does not support the cryptographic binding of Oracle metadata to transactions. SegWit reduces signature malleability at the Bitcoin protocol level, but it does not protect the integrity of Oracle-fed data or require validation before transactions are added to the blockchain. SecPLF protects loanable-fund protocols from Oracle manipulation, but it lacks a comprehensive transaction-level solution to prevent Oracle-driven malleability. OracleTrust, on the other hand, uses a dual-layer scheme to bind Oracle metadata and signatures to transactions via provenance tracking and a smart contract validation layer. The first layer encodes transactions into verifiable provenance records, and the second layer dynamically verifies these records with salted Keccak hashing and ECDSA recovery to bind the Oracle signature. A time-constrained commit-reveal mechanism with penalty enforcement ensures that the data is tamper-resistant. OracleTrust outperforms existing solutions in detecting malleable transactions, reducing latency, and memory consumption. This demonstrates its superior robustness and efficiency in blockchain.
The identification of small molecule modulators of immune checkpoint proteins remains a significant challenge in drug discovery due to the flat, featureless nature of protein-protein interaction interfaces and the characteristically low hit rates observed in conventional high-throughput screening campaigns. Here we report OracleScreen-LILRB4 (HTS-Oracle v3), an ensemble machine learning framework trained on quantitative biophysical screening data from two structurally diverse compound libraries (19,800 compounds total) screened against the myeloid immune checkpoint leukocyte immunoglobulin-like receptor B4 (LILRB4/ILT3). By formulating binding prediction as a regression task targeting continuous ΔFnorm values rather than binary hit classifications, OracleScreen-LILRB4 achieved a mean Spearman R of 0.61 and ROC-AUC of 0.86 under scaffold-aware cross-validation. Prospective virtual screening of a 45,760-member compound library and experimental validation of the top 200 predictions yielded a 28.5% hit rate, representing a 15.0-fold enrichment over baseline, with 16 compounds demonstrating nanomolar-affinity LILRB4 (ILT3) engagement. Lead compounds ORS-22 and ORS-14 restored anti-tumor immune activity across patient-derived colorectal cancer and acute myeloid leukemia co-culture systems, reversing SCG2-mediated immunosuppression and recovering cytotoxic T-cell function. These findings establish OracleScreen-LILRB4 (HTS-Oracle v3) as an effective computational framework for accelerating small molecule discovery against non-enzymatic immune checkpoint targets.
High-throughput screening (HTS) remains the cornerstone of early phase small molecule discovery yet consistently underperforms against immunotherapy targets, yielding validated hit rates below 0.1%. Here we introduce HTS-Oracle v2, which features rigorous cross-validation that ensures honest performance estimates. HTS-Oracle v2 was trained and validated across four clinically significant immune checkpoint targets (CD28, ICOS, LAG-3, and TIGIT) achieving ROC-AUC values of 0.968, 0.969, 0.875, and 0.928, respectively, under rigorous cross-validation. For prospective experimental validation, HTS-Oracle v2 was applied to an 8960-compound Enamine Protein Mimetic Library, selecting only 25 compounds per target for experimental testing using temperature-related intensity change (TRIC) technology, a 99.7% reduction in screening burden. HTS-Oracle v2 identified 4, 5, 4, and 6 validated binders from 25 prospectively selected compounds per target, corresponding to validated hit rates of 16%, 20%, 16%, and 24%, respectively. Notably, 67-80% of all experimentally confirmed hits across the full 8960-compound library were captured within just 25 model-selected compounds per target. These results establish HTS-Oracle v2 as an efficient platform for AI-guided prospective hit discovery across immunotherapy targets.
Autonomous inspection of discrete obstacles (e.g., tree trunks in orchards and forests) requires UAVs to visit every target with proper observation distance and heading, while simultaneously exploring the unknown environment. Existing space-guided exploration methods focus on eliminating unknown space and are inherently agnostic to the inspection targets themselves, leading to incomplete coverage and redundant traversal. We observe that the obstacles themselves encode the spatial topology of the environment and can serve as natural planning anchors. Based on this insight, we propose ORACLE, an Object-centric Autonomous Coverage Exploration framework that shifts the planning paradigm from space-guided to target-guided exploration. ORACLE integrates: (1) an online target detection and persistent identification module via occupied-voxel connected component labelling, (2) a density-aware global coverage planner that modulates ATSP costs to prioritize target-dense regions, and (3) a target-guided local planner that replaces frontier viewpoints with direct obstacle observation points in a Sequential Ordering Problem formulation. Experiments in two point-cloud environments reconstructed from real-world forests with contrasting tree densities (Environment I: 50 trunks, n¯=1.56; Environment II: 70 trunks, n¯=2.19; both with non-uniform spacing) show that ORACLE achieves 98.8% and 99.7% target coverage compared to 22.7% and 25.1% for the space-guided baseline, while reducing the mission overhead ratio from 202.9% to 129.2% (Environment I) and from 176.8% to 126.6% (Environment II). Ablation studies confirm that zone reactivation is the decisive factor for coverage completeness (-18.8 and -17.2 percentage points when disabled in Environments I and II, respectively) and that density weighting improves path efficiency.
For patients with complex health needs, the periods between clinical encounters are times of significant vulnerability, during which unobserved risks can escalate into acute events. The Waymark community-based care management program was designed to reduce this vulnerability for the patients it serves who are enrolled in Medicaid programs, but the organization faced the challenge of processing thousands of unstructured daily encounter notes from its field-based teams. To meet this challenge in a scalable way, the authors developed and implemented an artificial intelligence (AI) oracle, a system that continuously analyzes these notes. The oracle has three functions: (1) to identify urgent safety red flags; (2) to surface opportunities for preventive care; and (3) to generate monthly, strengths-based feedback to foster a learning health system. Following a phased implementation that began with a silent validation period, the system now delivers real-time, actionable alerts directly into the care teams' messaging system. Iterative refinement of the AI prompts, driven by clinical leadership review, improved the actionable true-positive rate of safety alerts from 68% of 147 to over 95% of 203 over 12 weeks. The system's value lies in its ability to synthesize disparate data streams - such as community health worker observations; pharmacy data; rising risk scores; and hospital admission, discharge, and transfer feeds - to reveal complex risks that could be missed by manual, siloed review. The AI oracle has become an integrated, "over-the-shoulder" digital coach, transforming a reactive care model into a proactive, learning health system and offering a scalable solution to improve outpatient safety for vulnerable populations.
Contract management and quoting processes are mission-critical operations in modern enterprises, yet they remain prone to inefficiencies, compliance risks, and human error-particularly in IT and SaaS sectors where contract complexity is high. This paper introduces the Artificial Intelligence Contract Risk and Intelligence Model (AICRIM), a simulation-validated, prototype-level framework that integrates generative AI (GPT-4) and BERT-based semantic matching into Oracle Configure-Price-Quote (CPQ) systems to perform real-time compliance and risk assessment during SaaS and cloud deal negotiations. AICRIM autonomously identifies anomalies in data privacy clauses, service-level agreements (SLAs), and pricing terms by interpreting unstructured contract text and benchmarking it against GDPR, HIPAA, and ISO 27,001 regulatory standards. The framework is evaluated through structured simulation experiments over a corpus of 500 synthetic SaaS contract documents, using manual legal review and rule-based NLP as baselines. Simulation experiments demonstrate a 27-32% reduction in contract errors, a 38.2% reduction in deal cycle time, and a compliance detection accuracy of 92.1%, with statistically significant improvements over both baselines (p < 0.01). These results are simulation-derived and serve as indicative upper-bound estimates pending real-world validation. AICRIM also embeds AI governance controls-AES-256 encryption, bias monitoring, and structured audit trails-to ensure ethical and auditable deployment. By extending Oracle CPQ via REST API event hooks, this model provides enterprises with a scalable, reproducible approach to contract risk automation.
A prevailing approach for learning visuomotor policies is to employ reinforcement learning to map high-dimensional visual observations directly to action commands. However, the combination of high-dimensional visual inputs and agile maneuver outputs leads to long-standing challenges, including low sample efficiency and significant sim-to-real gaps. To address these issues, we propose Oracle-Guided Masked Contrastive Reinforcement Learning (OMC-RL), a novel framework designed to improve the sample efficiency and asymptotic performance of visuomotor policy learning. OMC-RL explicitly decouples the learning process into two stages: an upstream representation learning stage and a downstream policy learning stage. In the upstream stage, a masked Transformer module is trained with temporal modeling and contrastive learning to extract temporally-aware and task-relevant representations from sequential visual inputs. After training, the learned encoder is frozen and used to extract visual representations from consecutive frames, while the Transformer module is discarded. In the downstream stage, an oracle teacher policy with privileged access to global state information supervises the agent during early training to provide informative guidance and accelerate early policy learning. This guidance is gradually reduced to allow independent exploration as training progresses. Extensive experiments in simulated and real-world environments demonstrate that OMC-RL achieves superior sample efficiency and asymptotic policy performance, while also improving generalization across diverse and perceptually complex scenarios.
Remote tele-rehabilitation requires objective pain assessment, but existing approaches fail in two distinct ways. Self-report scales such as the Visual Analog Scale and the Numeric Pain Rating Scale are easy to falsify, opening a special case of the Oracle problem in blockchain-based insurance. Cloud-based computer vision handles falsification but transmits raw biometric video off the patient's device, violating privacy requirements. A decentralized Edge AI-Oracle architecture is proposed that combines MediaPipe Face Mesh landmark extraction with a recurrent classifier mapping Action-Unit feature sequences to a learned pain score aligned with the Prkachin and Solomon Pain Intensity scale. The recurrent cell is selected empirically across short-context (T = 2) and long-context (T = 120 frames at 24 fps) regimes, with a two-layer Long Short-Term Memory (LSTM) network adopted for deployment. Inference and Elliptic Curve Digital Signature Algorithm (ECDSA) signing run inside an ARM TrustZone Trusted Execution Environment (TEE). Biometric logs are stored off-chain on the InterPlanetary File System (IPFS). Smart contracts anchor results on-chain and open a 24 h optimistic verification window for an off-chain Watchtower auditor. On SynPAIN the LSTM reaches F1 = 0.683 on T = 120 video (leave-one-stratum-out), with a directional but non-significant advantage over Gated Recurrent Unit (GRU) (Wilcoxon p = 0.167). Cross-dataset validation on BioVid Heat Pain Database Part A (87 subjects, 174 paired observations, leave-one-subject-out) yields F1 = 0.519 for LSTM and 0.499 for GRU (Wilcoxon p = 0.549). A processor-only TEE surrogate benchmark estimates 1.96 ms (FP32) and 0.45 ms (INT8) inference latency at T = 120 with a 0.34 MB footprint and 707 µs ECDSA signing latency, leaving the INT8 inference latency more than an order of magnitude below the 33 ms per-frame budget. The dual-layer storage reduces gas costs by a factor of 23.4 (160,261 vs. 3,744,872 gas), corresponding to an illustrative mainnet cost of approximately 0.53 USD per submission at 1 gwei, rising to roughly 16 USD at a busier 30 gwei, and falling to approximately 0.005 USD on Arbitrum One (April 2026 reference parameters), so that continuous monitoring is economically practical on Layer-2. An adaptive-adversary analysis of the Watchtower shows that gross score tampering is detected at every usable operating threshold, whereas a rational adversary who inflates by less than the dispute threshold, or who shapes the injected score to fall just inside it, evades detection. Because the false-positive rate reaches zero only for δ≳0.15, the protocol bounds rather than eliminates patient-side fraud and motivates a zero-knowledge proof-of-inference successor. The framework is architecturally and economically feasible as a cryptographically verifiable, privacy-preserving tele-rehabilitation substrate aligned with General Data Protection Regulation (GDPR) and Health Insurance Portability and Accountability Act (HIPAA) requirements through the Zero-Video Transmission principle, while remaining economically viable under post-Dencun mainnet and Layer-2 conditions. Recognition accuracy on real-world data and robustness to small-magnitude tampering remain limitations that the interchangeable recognition and audit components must improve before clinical deployment.
The decarbonization of urban energy communities increasingly requires coordinated integration of hydrogen, electricity, heat, and mobility under market-regulated environments. This study develops a hydrogen-driven digital transactions market embedded within a clustered, integrated energy hub architecture, where digital transactions markets, such as carbon emission trading (CET) and green certificate trading (GCT) mechanisms, are endogenously incorporated into operational scheduling. The framework coordinates hydrogen-diversified utilization, dual electric-hydrogen transportation systems, multi-vector storage, and renewable generation under carbon accounting constraints and social multi-stakeholder interactions. A decentralized multi-carrier optimization model is formulated to minimize system-wide scheduling cost while integrating CET/GCT revenues directly into dispatch decisions. Uncertainties in renewable generation, demand, and electricity prices are modeled using an inexact probabilistic stochastic programming approach with scenario generation and reduction. To extend evaluation beyond economic performance, a hydrogen-centric eco-social welfare layer comprising ten normalized indicators is introduced, quantifying emission mitigation, accessibility, equity, cost relief, and public acceptance. The model is validated on a four-hub clustered configuration under baseline and stress-test scenarios, including demand surges, renewable shortfalls, hydrogen price shocks, and market price fluctuations. Results demonstrate effective coordination between hydrogen production, storage, and mobility demand, with demand-side flexibility reducing operational costs by more than 16% in selected hubs. Carbon and certificate oracles market participation improves financial performance while enhancing emission compliance. Sensitivity analysis confirms robustness under combined worst-case disturbances. The proposed framework establishes a unified operational market structure that links hydrogen diversification, digital carbon-regulated transactions, and measurable eco-social welfare within sustainable urban energy systems.
暂无摘要（点击查看详情）
暂无摘要（点击查看详情）
The orbitofrontal cortex (OFC) is recognized as being responsible for constructing and representing internal models of the causal structure of the world. Current theories task the OFC with filling gaps in our experiential knowledge of this structure, recognizing hidden causes, making inferences, and even creating so-called belief states. Most accounts emphasize the importance of the functions assigned to the OFC in driving decisions. However, mediating such critical aspects of our internal representation of the world should also be important for learning, as adjusting our knowledge depends in part on our expectations and their violation, and constructing and revising a model of the world involves learning in the first place. Here, we will review data consistent with the proposal that the OFC is critical for performance and learning, focusing on neural correlates and OFC-dependent deficits. We will then discuss other actors comprising this circuit, including, most prominently, midbrain dopamine neurons.
暂无摘要（点击查看详情）
The exponential growth of publicly available genomic data has created unprecedented opportunities for sequence-based discovery. Locating specific k-mers is fundamental to diverse applications, including metagenomic classification, pathogen and cancer detection, and variant calling yet efficient identification of multiple k-mer patterns across large sequencing data and massive databases remains a significant computational challenge. We implement two quantum algorithms for DNA multi-pattern string matching for k-mer detection, leveraging Grover's amplitude amplification under the idealized quantum random access memory (QRAM) framework. The first algorithm uses an enumerate-m oracle that sequentially checks a loaded text substring against all m patterns achieving O (√S) query complexity for S text positions but requiring O (m · L) work per oracle call. The second algorithm employs nested Grover search with an outer loop over text positions and an inner loop over pattern space, reducing oracle complexity to O(L) while performing O (√S · √m) in total. These asymptotic gains highlight the potential advantages that could be unlocked by future large-scale, low-noise QRAM architectures, positioning our results as a promising proof-of-concept foundation. This work introduces two quantum implementations of multi-pattern string matching tailored for k-mer detection. Leveraging quantum parallelism and Grover-inspired search primitives, our methods accelerate dictionary-based pattern matching, particularly in contexts involving large sequences, such as genomic data, and extensive pattern sets. While implementation challenges such as QRAM overhead remain, this study demonstrates both the promise and current limitations of quantum-enhanced string matching, establishing a foundational step toward quantum readiness in bioinformatics. To maximize accessibility and practical use, we provide our methodology at: https://github.com/Georgakopoulos-Soares-lab/quantum-multi-motif-finder.
Accurate fault diagnosis of analog circuits is critical for ensuring the reliability of modern electronic systems. Two practical challenges hinder data-driven methods: data imbalance, where certain fault modes are rare, and variable operating conditions, where models trained under one fault severity fail to generalize to different severity levels. Although both challenges have been studied in isolation, their co-occurrence in analog-circuit diagnosis has received little systematic attention. This paper proposes DiffDA-Net, a unified two-stage framework that addresses both challenges simultaneously, and reports a systematic benchmark of generative augmentation and domain-adaptation components under this joint setting. In the first stage, a conditional Denoising Diffusion Probabilistic Model generates representative class-conditioned fault signals to balance the training dataset. In the second stage, a domain adaptive network couples adversarial training with Maximum Mean Discrepancy regularization to transfer diagnostic knowledge across operating conditions. On a challenging 13-class cross-severity transfer task (50%→25% parametric deviation), the dual-alignment domain-adaptation module reaches [Formula: see text] target accuracy under an oracle target-model-selection protocol, exceeding the strongest single-mechanism baseline by 21.79 percentage points; under a fully label-free source-validation protocol it still attains roughly twice the accuracy of the no-adaptation baseline. In the joint imbalanced and cross-domain setting ([Formula: see text], 10 seeds), the full framework attains [Formula: see text] accuracy. Controlled experiments isolate the contribution of each module: removing domain alignment causes the largest degradation ([Formula: see text] pp, [Formula: see text]), identifying dual-alignment domain adaptation as the dominant performance driver, while diffusion augmentation contributes a further [Formula: see text] pp ([Formula: see text]) and performs on par with SMOTE and ADASYN yet markedly better than GAN-based generation. All reported accuracies under the oracle protocol are explicitly labelled as upper-bound estimates.
The rapid development of renewable energy technologies and the proliferation of microgrids have led to increasingly complex power optimization and scheduling challenges in microgrid clusters. These challenges primarily stem from the heterogeneity and scale of the internal units within these clusters. To address these critical optimization issues effectively, we propose the Dual Exploration Mechanism Enhanced Diffusion Soft Actor-Critic (DEDSAC), a novel centralized, diffusion-based Soft Actor-Critic (SAC) framework. This framework features dual probabilistic exploration mechanisms, consisting of diffusion-based latent action generation and probability-triggered selective noise injection, together with a SAC-style squashed Gaussian reparameterization for actor optimization. It is designed to generate globally coordinated, high-quality decisions through direct modeling of the multi-peaked joint action space. The diffusion model serves as a robust policy network, enabling effective exploration of the action space. Extensive experiments across diverse conditions demonstrate that, upon convergence, DEDSAC effectively explores and converges to higher-quality scheduling strategies compared to standard SAC and other existing baseline algorithms. Based on the final converged evaluation reward, DEDSAC improves over SAC by 29.5% and 13.2% in small-scale and large-scale microgrids, respectively. Furthermore, DEDSAC approaches the Oracle Model Predictive Control (MPC) upper bound, achieving final evaluation rewards of -0.79 and -10.75 compared to Oracle MPC rewards of -0.34 and -8.55 in small-scale and large-scale scenarios, respectively. Results confirm DEDSAC's efficacy for optimizing power scheduling in large-scale microgrids, enabling efficient and sustainable energy management amid increasing system complexity and operational uncertainty.
Wearable ultrasound bladder monitoring often assumes static body positions, which is unrealistic for all-day usage. This work aims to enable reliable, non-invasive bladder-fullness classification by mitigating performance degradation caused by sleeping position variations. A miniature wearable ultrasound system for bladder monitoring has been developed. It weighs 32.2 gram (inc. battery) and draws an average current of 8.59 mA, supporting long-term operation. To ensure robustness to sleeping position variations, three modeling strategies were investigated: (i) Pooled training on multi-sleeping-position data, (ii) Oracle routing to sleeping-position-specific models based on dual-modal data, and (iii) Adversarial training using a confounder-free network (CF-Net) to achieve sleeping-position-independent feature extraction. The system has been validated on 12 participants across diverse sleeping positions. The baseline model trained on specific sleeping positions achieved only 75.95% average accuracy when tested on unseen position samples. Relative to this baseline, after applying the Pooled, Oracle and CF-Net strategies, classification accuracy improved by 16.48%, 20.03% and 18.56%, respectively. The proposed system enabled wearable ultrasound bladder monitoring with substantially improved robustness to sleeping-position variations. This study is the first to address the challenge of sleeping-position variation in wearable bladder monitoring, demonstrating its feasibility of all-day, home-based monitoring with clinically meaningful pre-void alerts.
Migrating legacy on-premise electronic health record (EHR) systems in tertiary hospitals to modern cloud-native platforms presents technical and strategic challenges. We aimed to establish an optimized roadmap for transitioning legacy monolithic systems to a microservice architecture-based cloud-native EHR (MCEHR). We conducted a 3-month strategic assessment and case study based on the modernization requirements of a global healthcare provider. The methodology incorporated semi-structured interviews with key stakeholders, including clinical informatics officers and system architects, to identify critical pain points such as .NET 4.0 end-of-support risks and performance bottlenecks during peak clinical hours. Phased hybrid migration was adopted, analyzing over 20 TB of Oracle-based legacy data and evaluating the technical feasibility of transitioning to .NET 9 and RESTful APIs. To ensure clinical safety, a proof-of-concept (PoC) environment was developed to simulate high-concurrency clinical workloads, emphasizing system resilience and transaction integrity during intensive order-entry periods. The transition to .NET 9 and MCEHR demonstrated 100% transaction integrity across 1,012 complex clinical test cases. Frontend and backend modernization showed high feasibility; however, migration of business logic embedded within legacy Oracle views represented a primary technical bottleneck, necessitating targeted decoupling. The PoC confirmed that RESTful API-based services maintained stable throughput under heavy concurrent loads, significantly reducing the risk of system-induced delays in clinical workflows. Transitioning to an MCEHR architecture is complex but strategically essential. The proposed task force team roadmap outlines staged upgrades incorporating core technology modernization (.NET 9 and RESTful APIs), selective business component migration, and parallel DevOps adoption.
Closed-set heterogeneous domain adaptation (HDA) for Internet of Things (IoT) intrusion detection aims to transfer detection capabilities across environments that differ in devices, telemetry, feature schemas, attack implementations, label taxonomies, and target supervision availability. Although recent HDA methods report strong performance, their deployment meaning is often unclear because improvements over a weak source-only baseline do not show how much target supervision headroom has been recovered or whether adaptation is preferable to direct target-side labelling under the same budget. This paper presents a controlled, anchor-based benchmark for closed-set HDA in IoT intrusion detection. Edge-IIoTset is used as the main fixed target dataset, with transfer from CICIDS2017, UNSW-NB15, CICIDS2017 + UNSW-NB15, and CICIDS2017 + NSL-KDD under single-source and multi-source settings. The benchmark defines fixed resolved contexts, Intersection and Union representation contracts, a five-class closed-set label contract, leakage-safe preprocessing, and an anchor ladder consisting of source-only, correlation alignment (CORAL), matched-budget target-only, and oracle target-only references. Geometric Graph Alignment (GGA) and the Joint Semantic Transfer Network (JSTN) are evaluated as the primary selected native single-source semi-supervised HDA (SS-HDA) and multi-source semi-supervised HDA (MS-HDA) exemplars, while the Prototype-Matching Graph Network (PMGN) and Conditional Weighting Adversarial Network (CWAN) provide 1:10 method coverage checks. Each method-context-ratio configuration is evaluated across twenty fixed seeds, and DA-versus-target-only differences are tested using paired seed-level statistical evidence. A compact second-target confirmatory experiment using ToN-IoT assesses whether the qualitative headroom recovery and same-budget deployment patterns remain visible under a different IoT/IIoT target. The results show that primary native HDA can recover substantial source-only-to-oracle headroom, but not uniformly. At the 1:10 labelled target ratio, GGA recovers 0.633-0.835 of the available headroom across C1-C4, while JSTN recovers 0.776-0.897 in the contemporary-source MS-HDA family and 0.872-0.926 in the mixed-vintage family. Same-budget comparisons show that DA is deployment-competitive only in some contexts; in others, direct target-side supervised learning is stronger. The benchmark therefore shows that closed-set HDA should be evaluated as target-conditioned, context-resolved evidence rather than as a pooled method leaderboard.
Accurate segmentation of cardiac substructures on computed tomography (CT) scans is essential for radiotherapy planning. This study evaluated whether pretrained transformers enabled data-efficient training using a fixed architecture with balanced curriculum learning while achieving robust generalization to imaging and patient variations. A hybrid pretrained transformer-convolutional network, self-distilled masked image transformer (SMIT), was fine-tuned using lung cancer patient scans (Cohort I, training N = 180) and tested on held-out Cohort I lung cancer scans (testing N = 60) and breast cancer scans (Cohort II, N = 65). Two configurations were evaluated: SMIT-Balanced (32 contrast-enhanced CTs, 32 non-contrast CTs) and SMIT-Oracle (180 CTs). Performance was compared with nnU-Net and TotalSegmentator. Segmentation accuracy was assessed primarily using the 95th percentile Hausdorff distance (HD95), along with radiation dose and overlap-based metrics as secondary endpoints. SMIT-Balanced approached SMIT-Oracle performance despite using 64% fewer training scans, with mean HD95 of 6.6 versus 5.4 mm in Cohort I and 10.0 versus 9.4 mm in Cohort II. On the Cohort I held-out test set, SMIT-Balanced mean HD95 was within 1.0 mm of nnU-Net. Cross-cohort testing showed larger accuracy degradation with nnU-Net than SMIT-Balanced (62% versus 50%, absolute change 4.5 mm versus 3.4 mm). Dose metrics derived from SMIT-Balanced were equivalent to manual delineations. Balanced curriculum training reduced labeled data requirements within the SMIT architecture. SMIT-Balanced was comparable to nnU-Net on Cohort I held-out data and showed smaller cross-cohort HD95 degradation.