Abstract: This study examines how ESG performance, innovation performance, and policy support relate to organizational resilience in China’s real estate industry. Drawing on the Resource-Based View, Institutional Theory, and Configurational Theory, the study conceptualizes organizational resilience through recovery and resistance capacities. Using panel data from 80 Chinese A-share listed real estate firms during 2015–2024 (800 firm-year observations), the study applies fixed-effects regression, robustness tests, and heterogeneity analyses. The findings show that ESG performance positively influences accounting-based recovery, particularly return on equity, but negatively affects market-based recovery, reflected in Tobin’s Q in the baseline models. Additional analysis reveals a U-shaped relationship between ESG performance and Tobin’s Q, suggesting that initial market valuation penalties may decline as ESG engagement deepens. Innovation performance shows limited baseline effects but becomes more relevant in alternative specifications related to recovery and leverage. Policy support demonstrates limited direct effects, indicating a more conditional role. Overall, organizational resilience is shaped by heterogeneous interactions among ESG, innovation, and policy-related factors.

Abstract: Accurate prediction of significant wave height (SWH) is essential for fisheries management, coastal socio-economic activities, and marine ecological conservation. In recent years, deep learning-based bias correction has shown considerable potential for improving numerical wave forecasts. However, many existing approaches are still constrained by limited receptive fields and often struggle to capture long-range spatiotemporal dependencies in wave forecast errors. To deal with this issue, we adapt and improve a video prediction framework, namely the Vision Mamba Recurrent Neural Network (VMRNN), to model and correct the spatiotemporal patterns of SWH prediction biases. Comprehensive evaluations show that the multi-channel VMRNN achieves consistently high predictive accuracy across different forecast lead times and sea-state conditions. When validated against reanalysis data, the proposed model reduces the root mean square error (RMSE) of WAVEWATCH III forecasts by 28.2%, 26.1%, and 24.7% at lead times of 24, 48, and 72 hours, respectively. It also preserves the spatial structure of SWH fields quite well, with the spatial structural similarity index remaining as high as 0.945 even at the 72-hour lead time. Regional assessments over high-wave areas further indicate that VMRNN can effectively reduce both the mean error and the systematic overestimation commonly found in numerical wave models. Additional validation using in-situ buoy observations confirms that the model has a robust ability to correct systematic positive biases, especially for wave heights ranging from 0.5 m to 2 m. Taken together, these results suggest that VMRNN has strong spatiotemporal modeling capability and can serve as a promising post-processing framework for improving operational physics-based wave forecasting systems.

Abstract: This paper addresses the robust trajectory tracking problem of an Unmanned Aerial Vehicle (UAV) equipped with a 2-DOF manipulator, designed for fast aerial manipulation of varying payloads. To overcome the high computational cost and adaptability limitations of traditional model-based controllers, this work introduces a novel hybrid gain-scheduling framework that shifts the computational complexity to the pre-flight phase. The approach utilizes an approximate inverse dynamics linearization, based on fixed nominal models, which transforms the complex nonlinear system into a simple linear plant with bounded, structured uncertainties. The entire configuration space, including manipulator states and a range of payload properties, is partitioned into dynamically similar regions using K-Means clustering. For each local region, a dedicated robust PD controller is designed using a multi-objective Genetic Algorithm (GA). This framework also successfully implements a gain interpolation technique to mitigate the potential for abrupt control actions. Simulation results validate the controller’s ability to maintain high-precision tracking during fast maneuvers and payload switching, confirming the robustness and adaptability of the offline-tuned design.

Abstract: This position paper argues that, for long-horizon tasks evaluated across models with comparable frontier capability, the agent execution harness, namely the infrastructure layer that governs context construction, tool interaction, orchestration, and verification around a language model, is often a stronger determinant of agent performance than the model it wraps. We formalize and defend the Binding Constraint Thesis: in this regime, performance variance is governed more by harness configuration than by model choice, and current evaluation protocols therefore systematically misattribute harness-level gains to model improvements. We support this thesis along three lines. First, a control-theoretic formalization treats the harness as the controller of a closed-loop dynamical system and the LLM as the stochastic policy it governs, which explains why small harness changes can produce performance shifts that exceed those obtained by substituting one model for another. Second, published benchmarks, industry deployments, and a controlled variance decomposition show that harness-induced variance can substantially exceed model-induced variance, including cases of model ranking reversal. Third, we propose a harness-aware evaluation framework with a disclosure standard and a variance decomposition protocol. Until harness specifications are disclosed, leaderboard comparisons for long-horizon agents should be treated as incomplete and potentially misleading.

Abstract: Deep learning classifiers deployed in scientific applications often encounter inputs that violate physical laws (e.g., due to sensor failure or corruption). Standard methods cannot detect such violations and may produce confident but wrong predictions. We propose UA-PBR, a framework that combines a physics-informed autoencoder (to detect physics violations) with a Bayesian CNN (to quantify predictive uncertainty). Inputs are rejected if either the PDE residual exceeds a threshold or the predictive entropy is too high. As a proof-of-concept, we evaluate UA-PBR on a synthetic Darcy flow dataset (32×32 grid) under severe computational constraints (Google Colab, 10 seeds). Despite these limitations, UA-PBR reduces classification risk by over 90% on heavily corrupted samples while accepting 89.7% of clean inputs with 99.99% accuracy on accepted samples. Ablation studies confirm that both components contribute synergistically. These preliminary results on a synthetic benchmark illustrate the potential of physics-aware rejection and motivate further investigation with larger-scale experiments. Code is available at: https://github.com/UA-PBR/UA-PBR.

Abstract: Background: Cervical cancer is the fourth most common cancer among women globally, with the highest burden in low- and middle-income countries. Limited access to screening and treatment contributes to high mortality, despite effective screening methods like HPV testing and cervical cytology. Objectives: To establish the degree of correlation between cervical cytology, colposcopy, and histological features among patients with abnormal cytological smears seen at Nelson Mandela Academic Hospital and MthathaRegional Hospital. Methods: This was an analytical cross-sectional study conducted from June 1, 2024, to June 30, 2025. Two hundred twenty-five participants were enrolled through a convenience sampling method. Demographic and clinical data were collected using a structured questionnaire. Categorical data were expressed as frequencies and proportions, and continuous data were summarized into means ± SD or medians (IQR). X² was used to determine the correlation, and a p-value of <0.05 was significant. Results: The mean age was of the participants was 45.5 years, with 72% being HIV positive. Most cytology results showed high-grade squamous intraepithelial lesions (HSIL). Colposcopy classified 77% of participants as CIN II or III. Both cytology and colposcopy correlated positively with histology p< 0.05. Cytology showed 92% sensitivity and 33% specificity for detecting CIN 2+ lesions, while colposcopy had 87.4% sensitivity and 49% specificity. Micro-invasive cervical cancer was prevalent in 4% of the participants and was associated with age ≥ 50 years and treatment delay of > 4months. Conclusion: Both colposcopy and cytology demonstrated good sensitivity but poor specificity for the diagnosis of CIN 2 or higher dysplastic lesions of the cervix. Early colposcopic evaluation and treatment of women with HSIL can help prevent incident cervical cancer.

Abstract: This paper makes a review for the studies of Space Energy Internet. Based on introducing the background of related networks, this paper discusses several key components of the Space Energy Internet (mainly including Space Solar Power Station, Energy Internet, and Artificial Intelligence Data Center), focusing on their corresponding system architectures, main research directions, and related technical challenges. Subsequently, supporting technologies such as discrete signal compression and coding, communication technology, energy transmission, power electronic devices, and artificial intelligence are discussed and analyzed. Furthermore, a highly integrated “data-computing-energy-networks” framework is established based on star computing networks and multi-orbital star link systems, and adopting the technologies like plug-and-play and modular design, which can support many innovative applications further.

Abstract: Decoherence is the primary obstacle to reliable quantum computing, yet real-time, measurement-driven early warning remains unavailable. Standard metrics (fidelity, entanglement entropy) are computed post-hoc or require full state tomography. We propose KA-Quantum, a thermodynamic early warning framework grounded in the Karimov–Alekberli (KA) causal entropy formalism [1], monitoring three quantum channels: C₁ (von Neumann entropy deviation), C₂ (bipartite correlation entropy coupling decay), and C₃ (fidelity residual). Using Qiskit Aer density-matrix simulation [2] of N=5 qubit circuits across three circuit classes (GHZ, random, variational) and three noise models (amplitude damping, phase damping, depolarizing; 30 Monte Carlo runs each, linearly increasing noise ramp), we find: 5 of 9 configurations achieve DR = 100% (30/30); overall DR = 56% (150/270) including physically non-decaying regimes where the specific noise type does not drive fidelity below the 0.9 threshold within the simulation window. FPR = 0.00 in all calibration periods. Median lead times of 1–15 gate cycles (F < 0.9); lead time is invariant to KA threshold θ (2.0/2.5/3.0 tested), confirming empirical threshold invariance [3]. The C₂ entanglement channel provides the earliest indicator, consistent with the structural coupling precursor that distinguishes KA across prior KA classical applications [4].

Abstract: Both Noise Radar (NR) and Quantum Radar (QR), with alleged common features, aim to use the randomness of the transmitted signal to enhance radar covertness and to reduce mutual interference. While NR has been prototypically developed and successfully tested in many environments by different organizations, research and development investments on QR did not bring to practically operating prototypes. Starting from the well-known fact that radar detection depends on the energy transmitted on the target, the detailed evaluations in this work show that the detection performance of all the QR types proposed in the literature are well below the ones of a much simpler and cheaper equivalent “classical” radar set, for example of the NR type. Moreover, the absence of a “Quantum radar cross section” different from the well-known radar cross section is explained. From these facts it results that, in spite of alleged advantages in some literature, Quantum Radar proposals cannot lead to useful results, including, of course, the detection of stealth targets.

Abstract: In this paper, whose main results are conditional on the density hypothesis or the Generalized Riemann Hypothesis, we establish a complete conditional hierarchy for the restricted weighted Goldbach sum $R_{a,q}(N) := \sum_{p_1+p_2=N,; p_1 \equiv a \pmod{q}} (\log p_1)(\log p_2)$, with expected main term $M_{a,q}(N) := C_2 S(N) N/\varphi(q)$, for fixed $q \geq 1$ and $\gcd(a,q)=1$.Under the Density Hypothesis DH($A$) (any $A \geq 2$), the exceptional-set exponent is shown to equal $\theta(A) = 1 - 2/(A+2)$ via a saddle-point argument, correcting prior formulas $\theta = 1 - 1/A$ and $\theta = 1 - 2/A$, which are both wrong. Under the Generalized Riemann Hypothesis (GRH), we prove the pointwise bound $R_{a,q}(N) = M_{a,q}(N) + O_{q,\varepsilon}(N^{1/2+\varepsilon})$ for all even $N \geq N_0(q)$, and derive the explicit threshold $\log N_0(4) = 45.93$ via a fixed-point iteration. We present both normalizations of the effective constant ($C^2_{4,\text{eff}} \approx 529$ and $\approx 2111$) and give a complete account of the discrepancy. A complete constant-chain audit carried out in Section 7 shows all three normalizations (values 529, 2111, and the independently reconstructed 2375) are consistent, and the worst-case certified bound is $\log N_0(4) \leq 46.1$, with no remaining caveat.We further provide:A certified computation showing all 122 primitive real Dirichlet characters of conductor $q \leq 200$ are free of Siegel zeros in the Stechkin critical interval, with global minimum $L_{\text{cert}} = 0.2344$ at $q = 163$. An unconditional restricted Chen-type theorem $N = p + P_2$, $p \equiv a \pmod{q}$, via Bombieri–Vinogradov. A conditional short-interval lower bound under GRH. None of these results proves the binary Goldbach conjecture or GRH. The paper establishes conditional results under explicitly stated hypotheses. This paper is a companion to "An Almost-All Theorem for a Restricted Goldbach Sum over Arithmetic Progressions with Explicit Unconditional Constants", whose results are used here as a black box.

Abstract: Background: The mucosal surfaces from the anterior nares to the anal canal are lined by a continuous liquid layer studied extensively in regional isolation — as airway surface liquid in pulmonary physiology, gastric mucus in gastroenterology, and nasal mucus in rhinology — but never conceptualized as a unified physiological system. Framework: This paper proposes that this continuous mucosal liquid layer functions as a "mucosal river" serving three critical roles: physical barrier protection, immune transport of secretory immunoglobulins and antimicrobial peptides, and maintenance of the hydrated microenvironment required for commensal microbial homeostasis. Nasal cilia function as the initial pump generating downstream momentum; pharyngeal and digestive peristalsis maintain flow; pulmonary cilia serve as tributary pumps feeding the main channel against gravity. The adenoid crypts function as immunological canyons — narrow, deep channels that use Venturi-effect flow dynamics and M-cell-mediated antigen transport to actively deliver antigen-laden mucus into immune processing centers. Waldeyer's ring forms a 360-degree antigen trap through which the river cannot pass without immune surveillance, and the first breath represents an immunological ignition event initiating adaptive immunity. Hypotheses: The framework generates testable predictions regarding pepsin as a pathologic passenger ascending the river against flow to cause posterior-predominant sinonasal inflammation, systemic dehydration disrupting the river through mucus hyperconcentration and ciliary compression, cigarette smoke damming the river via acquired CFTR dysfunction, and an antibiotic-dehydration "double hit" synergistically compromising mucosal barrier integrity. Each prediction is paired with a specific experimental design for validation. Conclusion: Understanding the mucosal river as a unified system may reshape approaches to chronic inflammatory diseases of the airway and digestive tract.

Abstract: We report a cross-station replication of endogenous circadian rhythms in plant bioelectric voltage, recorded continuously for 42 days at three independent sensor stations within a public science exhibition (Phänomena, Dietikon, Switzerland; March–April 2026). Three primrose (Primula vulgaris) stations were equipped with custom Biolingo bioelectric sensors (ESP32 + AD8232) and recorded autonomously through approximately 21,000 visitor interactions. We extracted DC-invariant spectral features from 5–10 second volt-age windows (n = 78,431 quality-filtered files) and fitted two-stage cosinor models with bootstrap 95% confidence intervals. All three stations show a robust 24-hour rhythm in the 1–5 Hz band power (bp1–5), with peak-to-trough amplitudes between 0.35× and 1.19× of mesor (R²med 0.72–0.87). Acrophase varies across stations from 05:00 to 11:00 local time. Critically, the rhythm survives an overnight-only restriction (18:00–09:00, no visitors) at all three stations, ruling out visitor presence as the rhythm driver. The most visi-tor-intensive station (faces of museum visitors triggering an emotion-recognition instal-lation) additionally shows a sharp daytime amplitude collapse coincident with exhibition opening at 09:00, consistent with the cardiovascular-mechanosensory coupling demon-strated in a companion study [20]. We argue that bp1–5—the spectral band most directly related to plant action-potential activity—carries an endogenous circadian signal in Primula vulgaris, and that this signal is modulated by sustained nearby human cardio-vascular activity in a manner consistent with frequency-selective mechanosensory transduction. From a biomimetic perspective, this demonstrates that the plant’s evolved bioelectric sensing apparatus can be leveraged as a live ambient biosensor for nearby human activity, complementing the more common biomimetic approach of replicating plant sensing in synthetic devices.

Abstract: Background/Objectives: Adolescence is a critical period for establishing lifelong oral health behaviours; however, persistent oral health problems and limitations in conventional school-based oral health education (OHE) highlight the need for more engaging and scalable approaches. Emerging digital modalities, such as artificial intelligence (AI)-based virtual human (VH) education, offer a promising alternative but remain insufficiently evaluated in adolescent populations. This study aimed to evaluate the effectiveness of AI-based virtual human–based oral health education (VOHE) program compared with conventional face-to-face oral health education (FOHE) among adolescents. Methods: A cluster randomised pretest–post-test intervention design was employed. Participants received either VOHE or FOHE, followed by assessment using a structured questionnaire based on the Knowledge–Attitude–Practice (KAP) model. A total of 268 middle school students were assessed for changes in oral health literacy (OHL) and oral health-related KAP. A linear mixed-effects model was applied to evaluate the effects of time, group (VOHE vs. FOHE), and their interaction, with participants treated as random effects to account for within-individual correlations. Results: Both groups demonstrated significant improvements in OHL and oral health related KAP following the intervention (all p &lt; 0.05). However, no significant group × time interaction effects were observed for any outcome variables (all p &gt; 0.05), suggesting that VOHE achieved educational outcomes comparable to those of FOHE. These findings indicate that AI-based VH education may provide an effective and scalable approach for adolescent OHE. Conclusions: VOHE demonstrated effectiveness comparable to FOHE and may serve as a feasible alternative or complementary approach for adolescent OHE. AI-based VH education also has potential applicability as an accessible digital health intervention for school- and community-based oral health promotion, particularly in digitally mediated or resource-limited educational settings.

Abstract: Adulteration of beef (Bos indicus) with buffalo meat (Bubalus bubalis) is a common form of food fraud with economic and religious implications, but quantitatively detecting its presence in ground beef products is difficult. Ten replicates of each of six binary mixtures (100:0 to 0:100 % w/w) of ground beef and buffalo meat were characterized using untargeted ¹H NMR metabolomics (43 metabolites after QC filtering), physicochemical measurements (pH, CIE L*a*b* color, water activity, and electronic nose), and proximate composition. Fifteen pairwise OPLS-DA models and a 1000-fold permutation test were performed for discrimination and biomarker identification. PCA explained 54.2% of the total variance, and the adulteration groups separated along the PC1 axis. All OPLS-DA models were statistically valid (R²Y = 0.738–0.981; Q² = 0.532–0.961; pQ² < 0.001), with no evidence of overfitting. Three metabolites met all three criteria (VIP > 1.0, FDR < 0.05, < FC > 2 or < 0.5) and had AUC = 1.00 in the internal data set: betaine (−82.6% in buffalo vs. beef), glycerol (+154.7%), and malonate (+656%). No individual biomarker exceeded the multi-criterion threshold at buffalo substitution levels less than10%. The selection of external discovery-phase candidates for beef authentication using NMR includes betaine, glycerol, and malonate.

Abstract: Coronary Artery Disease (CAD) is the leading cause of death worldwide, highlighting the need for more reliable and efficient diagnostic tools beyond conventional methods. Artificial Intelligence (AI), particularly Machine Learning (ML) and Deep Learning (DL), has shown strong potential for detecting obstructive CAD by learning complex patterns from Electrocardiogram (ECG) and Coronary Computed Tomography Angiography (CCTA) data. This rapid systematic review assesses and compares the diagnostic performance and methodological quality of AI models built for CAD prediction using ECG and CCTA data. A systematic search following PRISMA 2020 guidelines was conducted for primary studies published between 2021 and 2025. Eleven studies were included, six using ECG data and five using CCTA data. Methodological quality was evaluated using the PROBAST+AI tool. ECG-based models achieved AUC (0.72--0.961); however, only 33\% of these studies used external validation cohorts. CCTA-based models showed slightly stronger top-end performance, with AUC (0.77--0.97), and were more methodologically rigorous, with 80\% applying external validation. Despite these strong results, PROBAST+AI assessment revealed a high risk of bias in 90.9\% of the included studies, largely due to weaknesses in the analysis domain, including poor handling of missing data and the absence of model calibration reporting. AI models show strong diagnostic accuracy for CAD, with CCTA-based approaches demonstrating greater validation maturity. However, the widespread methodological bias means these tools should currently support clinical decision-making rather than replace standard diagnostic methods. Future studies should focus on prospective multi-centre validation and the use of multimodal data

Abstract: Comparative research on digital social influence and sustainable food consumption has grown substantially; however, most transnational studies do not verify measurement invariance nor assess whether observed structural differences reflect genuine cultural variation or measurement artifacts. This study addresses this gap by applying the Stimulus–Organism–Response (SOR) model to examine whether Social Media Content (SMC) and Online Member Group Support (OMGS) influence Organic Product Purchasing Behavior (OPPB) through Environmental Attitude (EA) and Subjective Norms (SN) in Ecuador, Chile, and Peru. A cross-sectional quantitative design was implemented with 809 organic consumers, analyzed using PLS-SEM in two stages: assessment of compositional invariance via the MICOM procedure and multigroup analysis (MGA) based on permutations. Full compositional invariance was confirmed across the three national groups, validating transnational structural comparability. The SOR model held consistently, with EA emerging as a stable predictor of OPPB. Significant structural differences were identified: the SMC→SN path was significantly stronger in Chile (β = .671 vs. β = .558 in Peru; p <.01), whereas the OMGS→EA path was stronger in Peru (β = .284 vs. β = .211 in Chile; p < .05). These findings underscore the need to formally verify invariance before drawing transnational conclusions and highlight the cultural contingency of sustainable digital marketing strategies in Andean emerging markets.

Abstract: This chapter examines the ontological assumptions, epistemological challenges, and ethical implications involved in using Agentic AI to assist, guide, or potentially replace human agents in the making of moral and legal decisions. It argues that different metaphysical assumptions regarding the ontology of ethics, justice, cognition, and AI decisively shape the framework within which such systems are evaluated. In this respect, the analysis distinguishes between two different levels of severity in the moral issues raised, corresponding to two distinct levels of AI autonomy: first, AI systems operating as advisory systems, with limited autonomy; and secondly, AI systems operating as regulatory systems, with full autonomy, that is, as entities entrusted with final decision-making authority. The text adopts a critical perspective on the use of Agentic AI in contexts of moral and legal judgement, highlighting both the conceptual fragility and the epistemological challenges that accompany proposals for such applications. At the same time, it considers the conditions under which such systems could genuinely contribute to human flourishing. Particular attention is given to the risk that ostensibly advisory systems may, in practice, become tacitly regulatory, especially under the pressure of widespread assumptions concerning AI objectivity and effectiveness. The chapter’s structure follows an algorithmic logic, in which a series of key questions serve as branching yes/no nodes, each possible answer leading to a distinct line of philosophical analysis.

Abstract: Multi-server private information retrieval (PIR) based on function secret sharing (FSS) has emerged as a prominent paradigm for achieving sublinear communication. However, standard FSS constructions strictly require full server participation, making them highly vulnerable to single-node fail-stop faults. Existing fault-tolerant schemes mitigate this but inevitably inflate the downlink response overhead to scale with the database size N (e.g., \( O(\sqrt{N}) \)). To overcome this limitation, we propose a (t,p)-fault-tolerant PIR (FT-PIR) protocol grounded in a newly designed generalized (t,p)-fault-tolerant distributed point function (FT-DPF). By introducing a hierarchical recursive patching mechanism, our scheme transforms rigid all-party evaluations into flexible t-out-of-p reconstructions. This architecture completely decouples the response communication from N and ensures efficient client-side reconstruction via lightweight XOR aggregations, fundamentally bypassing heavy algebraic interpolations. Formal analysis proves that our strictly stateless protocol guarantees (t-1)-computational privacy under the semi-honest model. Asymptotic evaluations demonstrate that the proposed FT-PIR achieves an optimal downlink complexity bounded to O(\( poly(t,p) \cdot \log p \)), significantly outperforming existing robust baselines for large-scale datasets.

Abstract: Objective: High sperm DNA fragmentation (SDF) results in more aneuploid embryos. Although sperm retrieved by testicular sperm extraction (TESE) has low SDF as compared to the ejaculated sperm, there is little data comparing the prevalence of chromosomal abnormalities in the resulting embryos. The objective of this study was to compare rates of chromosomal abnormalities in embryos generated by TESE or ejaculated sperm with increased SDF. Methods: The blastocysts were generated by ICSI. The preimplantation genetic testing for aneuploidy (PGT-A) was achieved by next-generation sequencing (NGS). This study utilized 400 embryos; 200 in each group. The sperm DNA fragmentation was determined by Halosperm G2 assay. The rates of euploid, aneuploid and mosaic embryos were compared by multivariate logistic regression and mixed-effects models that accounted for female age and ovarian reserve. Results: The TESE-derived embryos showed significantly higher percentage of euploid embryos (67.8%) as compared to those derived from ejaculated sperm with high SDF (48.2%, p = 0.003). The multivariate logistic regression indicated that the sperm source (TESE / ejaculate) was an independent predictor of euploid embryos [OR = 1.85; 95% CI = 1.05 - 3.26; p = 0.034]. The probability of having euploid embryos decreased by 6% for every 1% increase in the SDF. The increased age of female was a major negative predictor [OR = 0.92 per year; 95% CI = 0.88 - 96; p = 0.001]. Conclusion: TESE-derived sperm with low SDF resulted in significantly higher euploid embryos as compared to the ejaculated sperm with increased SDF.

Abstract: Clinical artificial intelligence systems are transitioning from predictive tools that generate diagnostic outputs for human interpretation to agentic systems capable of autonomous multistep action within clinical workflows, including ordering laboratory tests, initiating medication reconciliation, and updating patient records. Existing trust frameworks, designed for advisory systems and built on output verification and confidence calibration, do not address the governance requirements of autonomous action. We identify an agency gap: the structural mismatch between validated predictions and unvalidated action policies. Using a partially observable constrained decision process (PO-CDP) formalism, we establish the principle of agency nontransferability, demonstrating that trust calibrated at the diagnostic level does not imply safe or appropriate action policies under real-world clinical, institutional, and legal constraints. To address this gap, we propose a three-layer governance stack—epistemic soundness, policy safety, and institutional traceability—that provides verifiable guarantees at each stage of the agentic decision pipeline. This paper presents a theoretical governance framework; the phasespecific milestones in the backcasting roadmap define the empirical validation agenda for each deployment stage. A compositional risk analysis formally predicts that individually safe components can produce unsafe system-level behavior through nonlinear error propagation. An extended backcasting roadmap defines three empirically testable phases for the transition to governed agentic systems: sandboxed action proposals (2027–2029), credentialed policy systems (2030–2032), and supervised autonomy (2033–2035). The transition to agentic clinical AI constitutes a paradigm shift from prediction correctness to policy safety under constraint, requiring institutional design rather than technical improvement alone.