Computer Science and Mathematics

Abstract: Real-world agents extend language-model reasoning to digital operations and physical action, raising a central security question: how do digital-side attacks cross trust boundaries in robotic systems, and what physical safety conclusions can existing evidence support? This survey examines semantic hijacking, sensory manipulation, and middleware injection across software agents, modular robotic agents, and vision-language-action (VLA) systems. We compare how untrusted content acquires task authority, corrupted observations affect action generation, and middleware access reaches execution interfaces under different permission and control assumptions. An experiment-level analysis examines environment, trajectory, privilege and exposure, consequence, and assurance, distinguishing generated plans, software-state changes, simulated execution, and physical robot actions. The comparison reveals important evidence boundaries: RoboPAIR evaluates driving plans in Dolphins but API-mediated robot actions in Jackal, while Apollo point-cloud replay demonstrates altered braking decisions without measuring physical deceleration. These differences prevent planning or command-level success from being treated as direct evidence of physical safety violations. Across defenses, protection depends on the integrity of the state used for checking, coverage of executable command paths, and the timing of intervention. Reported attack suppression also need not imply useful protection: Safe Prompting reduces targeted attack success to zero while clean-task success also falls to zero in its LIBERO evaluation. Together, these findings identify the conditions under which digital compromise reaches physical execution, delimit the safety claims supported by current experiments, and motivate stronger assurance of action enforcement and recovery under compromised state.

Abstract: Multi-Agent Path Finding (MAPF) computes collision-free trajectories for robot teams. Conventional 4-connected cardinal-move MAPF restricts agents to axis-aligned moves, which produces unnecessarily long paths. Any-angle MAPF (AA-MAPF) can generate shorter line-segment trajectories, yet state-of-the-\art AA-MAPF solvers such as CCBS (Continuous-time Conflict-Based Search) suffer from heavy computational overhead: they rely on numerical iterative minimal safe waiting-time computation, enumerate similar constraints near current constraint for MC (Multi-Constraints) acceleration. This paper presents a bounded-suboptimal any-angle planner built upon the ECBS (Enhanced Conflict-Based Search) framework: AA-ECBS (Any-Angle ECBS). We redefine constraints as part of agent trajectories, avoid enumerate similar constraint for MC acceleration, derive analytic solutions for minimal safe waiting-time. The approach retains completeness and bounded suboptimality. Evaluated on standard benchmarks, our method achieves better runtime and success rate compared to existing solvers, and keep the same solution quality.

Abstract: Robotic process automation (RPA) is quick to deploy but difficult to govern: software robots hold privileged credentials and act unattended across production systems, yet the standards organisations rely on—COBIT 2019, ITIL 4, ISO/IEC 38500 and TOGAF—predate RPA and treat automation as an output of IT processes. This study builds and tests a standards-integrated RPA governance framework for the telecommunications sector of the United Arab Emirates. A four-phase deductive–confirmatory design combined standards analysis with a systematic review (34 of 543 records), content validation by five experts, a modified Delphi study (three rounds, 13 experts), and a goal–question–metric evaluation by nine practitioners in three operators. All 60 preliminary items were retained (S-CVI/Ave = 0.97; modified kappa ≥ 0.76). At full Delphi consensus (median 5, IQR 0) the framework grew to 18 factors and 80 items, up from 14 and 46. Practitioners judged it Accepted (composite 4.39 of 5; cross-case range 0.03). Three of the four expert-added factors matched coverage gaps recorded before the Delphi began, and all three concern the robot’s runtime behaviour. General IT standards therefore need supplementing at three identifiable points, and the framework provides an auditable baseline onto which AI-specific controls can later be layered.

Abstract: Efficient observation is essential for long-term autonomous robots because continuous and exhaustive sensing wastes computational resources while important or newly emerging information may still be missed. However, existing observation strategies often lack a unified mechanism for prioritizing impor-tant information, progressively covering initially neglected regions, adapting observation frequency to environmental change, and avoiding repeated processing of already understood information. This paper proposes Progressive Importance-Guided Observation (PIGO), which allocates limited detailed-observation capacity according to importance, uncertainty, novelty, and coverage demand. The method progressively shifts attention from urgent information toward neglected candidates, adapts observa-tion intervals according to environmental stability and change, and maintains viewpoint-conditioned observation memory to detect novelty, relation changes, removals, and true environmental changes. Stable low-impact information that has already been processed can be suppressed and reactivated when its state or relevance changes. Simulation and real-robot experiments show that PIGO achieves hidden-critical recall of 0.700 at a 35% observation budget, reduces repeated benign processing from 0.2964 to 0.0965, and decreases viewpoint false-change rate from 1.000 to 0.000.

Abstract: Autonomous agents require the ability to continuously expand their competence while minimizing task-specific structures predefined by designers. However, most existing learning systems still assume fixed task sets, input representations, learning methods, retrieval mechanisms, or knowledge organizations. This paper proposes Minimal-Core Open Learning (MCOL), a framework that retains only a small generic core while moving reliably learnable components into an expandable part of the system. MCOL enables agents to discover what should be learned, expand useful inputs, acquire or switch learning methods, retrieve and reuse accumulated knowledge, discover new learning problems, and learn how heterogeneous knowledge should be organized. A Thinking-Guided Learning mechanism further uses learning outcomes to identify insufficient learning, acquire diagnostic evidence when needed, and revise subsequent learning under validation-based acceptance. Experiments show that MCOL achieves 0.8234 F1 with only one task-specific predefinition versus 0.8292 with eleven predefined components, improves temporal-task F1 from 0.532 to 0.955 through method expansion, and achieves 0.9098 F1 with 0.975 diagnosis accuracy in thinking-guided adaptation, approaching exhaustive adaptation at 0.9135 F1 with substantially fewer candidate evaluations.

Abstract: Widely used open-source tools for map-based LiDAR localization of mobile robots are still largely organized by sensor suite: a 2D scanner typically calls for an AMCL-style particle filter over an occupancy grid, a 3D LiDAR for an NDT-based tracker over a point-cloud map, and an inertial- or GNSS-aided platform for a Kalman fusion node, often with different topics, parameters, initialization, and recovery behavior. Existing libraries already share one particle-filter core across 2D and 3D sensor models, and other systems already relocalize a robot by branch-and-bound scan matching. We describe PRISM-Loc, an open-source ROS 2 localization stack that places three established estimator families behind one node contract: laser2d, a likelihood-field Monte Carlo localizer; ndt3d, a planar NDT-MCL backend (Saarinen et al., 2013) that reuses the same particle filter with a Gaussian-voxel measurement model; and fusion3d, a 15-state error-state Kalman filter fusing LiDAR, IMU, and RTK-GNSS. Every backend publishes the same map→odom transform and pose with covariance and accepts the same /initialpose seeding, all filter mathematics live in Eigen-only C++17 cores that are tested without a ROS runtime, and laser2d relocalizes a kidnapped robot on request (at startup or by a service call) by deterministic branch-and-bound scan matching over a map-sized search window restricted to cells not marked occupied, keeping the top-K candidate poses and committing one only after several odometry-chained scans concentrate a normalized heuristic evidence score on it, and otherwise reporting the place as ambiguous. We claim no new estimator theory and no new verification principle; the contribution is this particular integration, released as software. In seeded closed-loop experiments on synthetic data that drive the actual cores, the 2D tracker holds a steady-state mean position error of 0.105 m, branch-and-bound relocalization succeeds on 97 of 100 kidnapped-robot queries in a 6 m room with zero false accepts at its default threshold, on 40 m and 80 m maps with repeated structure the single-scan rule accepts a wrong pose on 6 and 15 of 120 queries, which multi-scan verification reduces to 0 and 1 of 120 (pooled 95% Wilson interval 0.1–2.3%) while committing as many correct poses (242 against 241 of 300), and the ESKF bounds a 10 s aiding dropout to 0.125 m RMSE, although a 100-run NEES test shows that its covariance is overconfident (full-state NEES above the 95% interval at 89 of 120 epochs, mainly in attitude; bias blocks consistent). No external baseline and no real or public data are used, and no field-accuracy claim is made. Code: https://github.com/kjungmo/prism_loc.

Abstract: Embodied AI aims to develop agents that perceive, reason, and act in physical environments. Beyond recognizing objects and scenes, agents must identify possible interactions, determine their relevance to the current task and embodiment, and translate them into executable behavior. Affordance learning supports this process by grounding perception in action possibilities and connecting it with task reasoning and control. Existing affordance surveys, however, provide only partial coverage of the field, often focusing on isolated aspects of perception or robotic interaction and lacking a comprehensive view of the full perception–reasoning–action pipeline in embodied intelligence. To bridge this gap, this survey presents a unified framework organized around three complementary roles: affordance perception identifies possible interactions and their locations; affordance reasoning selects opportu-nities relevant to the task, context, and agent state; and affordance-guided action uses affordance representations for planning and policy learning. Within this framework, we develop fine-grained taxonomies, compare representative methods, and trace the field’s progression from passive visual recognition toward action-aware embodied learning. We further organize representative datasets and evaluation protocols and identify open challenges and future directions. Overall, this survey clarifies how affordance learning connects perception, reasoning, and action to support generalizable and physically grounded embodied intelligence.

Abstract: Service robots in crowded pedestrian environments re-decide at every control cycle which side to pass a person on; when this re-decision is greedy, the passing direction can oscillate across cycles in ambiguous configurations, or the robot can freeze. Topology-aware planners bias selection toward the previously chosen class through soft costs, selection hysteresis, or consistency weights; opinion-dynamics methods make the passing side an explicit, stateful decision with deadlock-breaking guarantees; and minimum-dwell or fixed-rate selectors bound the switching rate by construction. We introduce COMPASS (COMmitment-based PASSing for Social navigation), a decision-layer planner whose commitment rule is a leaky evidence accumulator — a rectified, leaky variant of Page’s one-sided CUSUM statistic with a progress-dependent threshold — under a lexicographic safety override of the standard priority-arbitration kind. Passing classes are bound to person track IDs, a design contract that the current implementation only partly realizes. We state five properties (P1–P5); they apply standard arguments and are not new mathematics. The load-bearing one, P2, is a deterministic, pathwise lower bound on the number of control cycles between discretionary switches that follows from the accumulator rather than from a fixed dwell timer; at the default knobs its worst case is 3 cycles (0.15 s). P2 is paired with a conditional response-time bound, and P1 refines it to an exponentially small per-cycle switching probability under an i.i.d. hypothesis. In an offline harness that drives the actual decision core (5 scenarios × 50 seeds), an immediate-argmin comparator produces about 98 decision changes per encounter under ambiguous ties, whereas COMPASS produces at most one decision change in each tested encounter, as does a 0.6 s persistence (dwell) timer in that scenario; the two rules differ under transient, reversing and intermittent advantage. In a historical decision-core-only measurement, latency at K = 3 (8 classes) is p99 58.6µs and max 651.3µs (the max is about 1.30% of the 50 ms budget at 20 Hz). All comparators are internal. Closed-loop evaluation, external baselines (including a previous-class consistency-weight baseline), and user studies are not performed; they are specified as a planned-evaluation protocol. Code: https://github.com/kjungmo/compass.

Abstract: Sand scorpions are a useful model for deriving principles of navigation by familiarity using a local sensor. Immediately after digging its burrow, a sand scorpion performs a series of learning walks that may facilitate the acquisition of home-directed tastes and touches via its midventral pectines. Later, during subsequent homebound journeys, the scorpion repeatedly compares current tastes and textures to the memories acquired via learning walks, using this information to step in the direction of the most familiar memory until it reaches its home. Previously, we have created computer simulations to show proof-of-concept of this navigation process. My goal here is to illustrate the idea with a simple, easily replicated robotic toy “scorp-car” that navigates based on the familiarity algorithm. I expect this robot to not only shed light on scorpion sensory biology but also inspire translational applications involving autonomous navigation within human-built environments.

Abstract: Robotic systems have become critical assets across manufacturing, healthcare, logistics, and field operations, yet their maintenance remains fragmented across a literature that treats sensing, actuation, mechanical wear, and system-level failure as unrelated topics rather than facets of a single problem. This survey addresses that fragmentation by organizing robot fault detection, diagnosis, prognosis, and fault-tolerant control around a unified four-category taxonomy, sensor, actuator, mechanical, and system-level faults, applied consistently across eight robot platforms: industrial robotic arms, wheeled mobile robots, legged robots, collaborative robots, aerial robots, surgical robots, underwater robots, and humanoid robots. Following a PRISMA-inspired screening process across four databases (IEEE Xplore, Google Scholar, Scopus, and ScienceDirect) covering publications from 2022 to 2026, an initial pool of approximately 1,800 identified records was narrowed through duplicate removal, title and abstract screening, and full-text eligibility assessment to 115 studies meeting this survey's inclusion criteria. Within this literature, industrial robotic arms account for the largest share of coverage, while acoustic emission sensing, optical and fiber Bragg grating sensing, communication-layer system faults, and humanoid-specific fault detection are addressed by a comparatively small number of studies despite the same search effort applied uniformly across every category. Methods are reviewed independently of fault category, spanning classical machine learning, deep learning architectures including recurrent, convolutional, state-space, and generative models, graph neural networks, federated learning, hybrid approaches, and model-based probabilistic methods. Recurring challenges are identified across the reviewed literature, including data scarcity relative to normal operating hours, heterogeneous data across robot fleets, domain shift between laboratory and deployment conditions, sensor corruption under real-world noise, real-time feasibility on constrained onboard hardware, and platform-specific limitations most pronounced for humanoid, surgical, and underwater robots. This survey does not propose new fault-detection algorithms; its contribution is the taxonomy itself and the structured synthesis of existing methods against it, together with a research agenda mapping identified challenges to near-term, mid-term, and long-term directions for the field.

Abstract: The automation of vehicle refueling and charging stations is an emerging enabler of sustainable transportation, but the robotic platforms able to perform this task remain fragmented across disconnected research communities. This review examines two evidence-based platform families - fixed and mobile manipulator arms - and two emerging candidate families - humanoid and quadruped robots, retained as future research directions - across four refueling modalities: EV conductive charging, EV wireless charging, hydrogen refueling, and battery swapping. A PRISMA-inspired, criterion-driven search retained 84 primary sources from an initial pool of 194 records. Quantitative evidence is reported wherever available: docking success rates of 90--100\%, positional accuracy of 2-10 mm, cycle times from tens of seconds to a few minutes, and Technology Readiness Levels. Every comparative table entry carries an explicit Evidence Class (A-C), separating evidence-based findings from the extrapolation required for humanoid- and quadruped-based refueling, for which no direct study exists. Dedicated analyses address hydrogen functional safety and industrial battery-swapping architectures (NIO, CATL, Ample, Gogoro). Six technological and normative gaps, including cybersecurity, feed a five-direction, priority-ranked research roadmap. This review offers the first cross-platform, evidence-graded comparative framework in this domain, making explicit where current knowledge ends and extrapolation begins.

Abstract: We ported EASA (Embedded Action Selection Architecture) [1], based on the basal ganglia model of [2], originally validated on a Khepera I with a gripper in a foraging task, to a modern rigid-body physics engine, with actual geometry extracted from the same hardware's robot definition files. The decision circuitry remained unchanged in its constants. The actual contact physics revealed several behaviors that previous platforms did not demonstrate in the same way, including gripping failures due to an off-center grasp, an arm that could remain indefinitely in a lowered position, and a periodic navigation orbit that avoided the center of the environment. A systematic dopamine sweep showed that the model's original calibration point, 0.2, corresponds to the observed minimum total task failure among five tested levels. We observed a specific mechanism behind a motor distortion effect already described in previous work with the same model: motor activity of behaviors that lose selection competence by filtering into the final robot command. Our results indicate that the effect does not appear to depend on simultaneous channel selection at the thalamic level but rather on a specific threshold at the motor activity combination stage [3]. A preliminary execution mode inspired by the procedure reported for real hardware showed worse performance at all dopamine levels tested, with no classical form of dithering explaining the effect [4]. This work suggests that porting a historical neuro-robotic model to a modern physics engine requires adjustments beyond a simple code translation.

Abstract: Autonomous robots should continuously extend what they can learn rather than remain limited to a fixed set of predefined tasks and skills. However, existing learning systems typically optimize within predetermined task, data, or capability spaces and provide limited mechanisms for making new learning opportunities themselves reachable. This paper proposes an autonomous reachable learning space expansion model in which newly acquired knowledge and skills progressively enlarge the space of future learnable knowledge. The model maintains persistent learning beyond individual tasks and expands learning along environmental, temporal, and thinking dimensions. Environmental interaction exposes previously inaccessible observations and opportunities, temporal accumulation enables delayed and long-term relations to be learned, and thinking generates new candidate problems through abstraction, transfer, combination, and variation. Newly acquired capabilities are recursively reused to unlock further learning opportunities, while an online learning process adapts which materials should receive priority. Experiments show that the full model increases structural diversity from 0.740 to 0.947, raises reachable scope from 19.70 to 29.60 over random exploration, and improves cumulative learning utility from 124.99 to 148.38 over a fixed priority strategy.

Abstract: Common middleware topics and message types do not ensure that control commands have the same meaning across robotic platforms. We audit version-pinned source code, configurations and specifications of Collaborative Robotics Toolkit (CRTK) interfaces in surgical robotics, identifying differences in Cartesian reference frames, physical units, state preconditions and stop behaviour. We derive task-space error bounds, including the exact maximum positional error of a frame mismatch over a command ball. Frame and uniform-unit mismatches can remain invisible when command and feedback share the same binding; identifying them requires an independent spatial relation or dimensional anchor. We release crtk-conformance, a black-box prototype that compares observable bindings with a client’s declared expectations and returns undetermined when evidence is insufficient. Against a configurable reference node, spatial error intervals contained truth in 240 of 240 replicated runs under Gaussian noise. Anchored scale estimates recovered injected scales from 0.1 to 10; one false divergence occurred at the fixed 1mm boundary. Liveness intervals contained the injected timeout in all 23 formed estimates, 22 conditional on observed run-maximum latency allowances. A heuristic source age interval contained event-log truth in 98 of 150 windows, with failures at both ends; rate conformance is therefore not decided. Runs against two released Surgical Robotics Challenge interfaces on AMBF corroborate topic inventory and state behaviour while leaving frame, unit and rate conformance undetermined. These results support an artifact audit and identifiability study with prototype verification, not hardware certification. We recommend explicit, discoverable frame, unit and timing contracts.

Abstract: Heterogeneous multi-robot task allocation (MRTA) requires the efficient coordination of robots with different capabilities while balancing task benefits, travel costs, and completion time. To address the challenges of complex constraints and premature convergence in conventional optimization methods, this paper proposes a heterogeneous MRTA algorithm based on chaotic adaptive dung beetle optimization algorithm, termed t-SDBO. First, a task allocation model is established by incorporating time-dependent task benefits, robot resource requirements, and task execution constraints into a weighted combinatorial optimization formulation. A real-valued encoding-decoding mechanism and constraint-repair strategy are then employed to transform continuous search positions into feasible discrete task assignments and execution sequences. To improve the optimization performance of the dung beetle optimizer, Sine chaotic mapping is introduced to enhance initial population diversity, while adaptive t-distribution mutation is used to strengthen exploration and alleviate premature convergence. Furthermore, a dynamic selection strategy adjusts the application of mutation to balance solution quality and computational overhead. Comparative experiments demonstrate that t-SDBO achieves better solution quality and convergence performance than particle swarm optimization, the sparrow search algorithm, and standard DBO. Experiments across different problem scales, together with indoor and outdoor ROS/Gazebo simulations, further validate the effectiveness and feasibility of the proposed method.

Abstract: World models have become a key foundation for artificial general intelligence and autonomous agents, enabling perception, understanding, and reasoning via internal representations. Diffusion models, empowered by high-fidelity generation and flexible conditional modeling, have become an important avenue for building representation-based world models. Despite rapid progress, the community still lacks a dedicated and systematic survey of diffusion-based world models that systematically analyzes their distinctive strengths and inherent limitations and clearly formulates the key scientific problems that must be addressed going forward. To bridge this gap, we present a comprehensive and structured review of diffusion-based world models. We first summarize the principles that connect diffusion processes to world modeling, and organize mainstream methodological frameworks and representative systems across diverse application scenarios. We then synthesize research progress, intrinsic characteristics, and commonly used datasets, providing a unified taxonomy to position existing works. Importantly, we offer an in-depth assessment of diffusion models for world imitation, highlighting their advantages (e.g., multimodal hypothesis coverage, controllable conditional generation, and high-fidelity generation) and limitations (e.g., computational inefficiency, long-horizon consistency challenges, error accumulation, and evaluation bottlenecks). Building on this analysis, we identify critical open problems and future research directions. Overall, this survey provides a consolidated reference framework and a clear research roadmap for diffusion-based world models, facilitating rapid understanding and guiding sustained progress in this emerging area. https://github.com/energy588/Diffusion-based-World-Models.

Abstract: Autonomous agents require not only deliberate reasoning for unfamiliar situations but also the ability to convert repeatedly useful reasoning into efficient automatic behavior. However, existing LLM-based agents and knowledge-distillation approaches generally couple reasoning and learning through predefined training objectives or direct teacher--student supervision, leaving unclear how an agent can autonomously determine what should become learned behavior. This paper proposes Thought-Mediated Learning (TML), a framework that separates LLM-based deliberative thinking from an independent ML/DL background learning process. Thinking cannot directly update the automatic model; instead, it produces thought materials whose content and frequency determine what becomes available for learning. Repeated thinking implicitly strengthens important experiences, attention determines which factors are represented, jointly considered events may create spurious associations, and imagined experiences can provide learning materials before corresponding real events occur. Observed outcomes further generate corrective materials that revise previously learned automatic behavior. Experimental results show that TML achieves 96.67% hidden-pattern acquisition and 99.93% post-learning automatic accuracy, reaches 97.83% OOD (out-of-distribution) accuracy versus 9.50% for raw learning, and improves zero-real rare-case success to 59.38% while corrective learning raises final hidden-case success to 96.82%.

Abstract: Autonomous robots require not only reliable task execution but also the ability to identify what new local knowledge should be learned, when explicit thinking is necessary, and when learned knowledge can be reused automatically. However, existing methods often bind learned knowledge to predefined tasks or fixed reasoning procedures, making it difficult to autonomously discover reusable local concerns and reduce repeated reasoning after adaptation. This paper proposes a thinking-triggered aspect learning framework for autonomous knowledge acquisition and transfer. Routine behavior is handled by automatic models, while novelty, uncertainty, or periodic heartbeat inspection triggers thinking to localize unresolved concerns. The robot then discovers task-independent aspects containing relevant factors, applicability conditions, handling principles, and local automatic models. Validated aspects are reused across tasks and domains, while aspect-specific calibration and checkpoint decay progressively convert explicit thinking into automatic processing; environmental changes can reactivate thinking and initiate new aspect acquisition. Experiments show aspect and factor F1 scores of 0.975 and 0.965, perfect zero-shot cross-task and cross-domain transfer in all successfully validated source runs, and a reduction in thinking rate from 0.508 to 0.100 after familiarization; the proposed method also reduces LLM calls by 76.4% compared with Always Think while achieving 0.976 post-validation detection and handling success.

Abstract: Autonomous robots should be able to initiate meaningful activities even when no explicit task is currently assigned. However, most existing systems remain task-centered or rely on predefined internal mechanisms, limiting their ability to preserve and acquire cognitive influences that remain relevant after current reasoning has ended. This paper proposes a self-initiated activity-triggering framework in which robot activities can arise from external perception, autonomous internal processes, or prolonged inactivity. Internal processes maintain independent states and provide non-commanding messages whose effects are interpreted contextually by the main thinking process. An independent process manager further learns acquired internal processes from previous thinking experience. Process formation is gated by future influence rather than recurrence alone and supports both recurrent delayed needs and rare but severe future vigilance needs. Learned activation signatures determine what future evidence should reactivate acquired processes. Experiments show that the proposed external triggering mechanism achieves 96.00% strict accuracy while reducing LLM tokens by 76.1% relative to centralized reasoning. The learned process manager achieves 98.00% acquisition F1, while activation signatures improve strict balanced future correctness from 62.67% to 82.33%. Compared with repeated history-based LLM reasoning, process internalization reduces LLM calls by 91.80% and token consumption by 94.07%.

Abstract: Diagnostic uncertainty in neurological rehabilitation motivates robotic systems that can select informative sensing actions adaptively rather than rely on fixed assessment protocols or static clinical records. This work introduces the Embodied Evidence Acquisition and Reasoning Loop (EARL), a typed architecture in which five role-specialized critics propose and assess sequential sensing actions, a probabilistic simulation sandbox estimates their expected information value, and a deterministic safety governor retains exclusive execution authority. Observed, derived, and simulated evidence remain provenance-distinct in a replayable, hash-linked ledger. EARL was evaluated retrospectively on 64 subjects from the PhysioNet Gait in Neurodegenerative Disease Database using repeated subject-level cross-validation, 11 predeclared conditions, and 3520 replay-verified runs. The primary endpoint was area under cumulative posterior-entropy reduction. EARL achieved 7.689 (95% CI 7.535–7.840), exceeding fixed-order and random selection by 0.459 and 0.763, respectively; the EARL-minus-EIG difference was -0.321. EARL nevertheless achieved higher macro accuracy (55.8% versus 49.5%) and a lower Brier score (0.781 versus 0.813) than pure expected-information-gain selection, demonstrating a trade-off among uncertainty reduction, discrimination, calibration, and safety-constrained decision making. All 10 deterministic safety-conformance scenarios produced their expected outcomes. These results establish reproducible software behavior for bounded sequential retrospective sensing; they do not establish clinical diagnostic performance, treatment benefit, physical-robot safety, or patient efficacy.