Jianlong Zhou

Concept bottleneck models are only as useful as the concept space they expose. Existing concept-annotation-free CBMs often treat this space as text-pool retrieval: an image is explained by whichever prompt words score highly, so visual evidence that is fine-grained, compositional, or hard to name can be missed. We introduce SAND-CBM, which shifts the bottleneck from prompt retrieval to image-side basis reconstruction in a frozen vision–language space. Text-similarity channels are retained as named anchors, but they are complemented by learned visual-factor channels that can surface candidate evidence beyond the vocabulary. The core design is a signed–nonnegative–calibrated–shared (SNCS) representation contract: signed dictionary rows preserve CLIP-space directions, nonnegative activations give additive channel strengths, training-set calibration makes channels rankable, and a class-conditional selector reuses one shared dictionary across classes. SAND-CBM is concept-annotation-free rather than label-free: it uses no concept names, concept annotations, or concept-level targets for the learned image-side basis, while task labels train the final readout and an explicit low-bandwidth class gate. This design provides a more diagnostic interface: important channels should be removable, low-contribution or random removals should have weaker effects, discovered activations should align above null controls, and class gates should reuse rather than duplicate evidence. Across four visual benchmarks, SAND-CBM remains competitive in accuracy while showing a more concentrated and intervention-consistent concept-candidate interface than matched controls. Core code is available at https://github.com/LMBTough/SAND.

Language-model agents act through tools and credentials. A score from one tested configuration should not keep authorizing actions after the system, permissions, or threat model change. We argue that agentic AI is the first high-stakes domain in which assurance-invalidating change routinely outpaces evidence renewal. Continuous assurance is therefore not an improvement on benchmark evaluation, but a precondition for an authorization claim being true when it is relied upon. The machinery exists in nuclear, aviation, and medical-device regulation. Three requirements do not: a configuration baseline that makes change detectable, a change taxonomy calibrated to hazard relevance, and qualification criteria for evaluation environments. We propose the Agentic Assurance Stack (Stack), a versioned assurance profile linking each action path to a bounded claim, five evidence obligations, and a renewal rule.

Cardiovascular disease is the leading cause of death among women worldwide. However, women remain underdiagnosed, undertreated, and underrepresented in the clinical studies, health records, images, and device data used to build artificial intelligence (AI). Machine learning, large language models, and consumer digital health tools can therefore inherit and amplify a century of male-default medicine. When designed and deployed with equity in mind, the same tools can help identify and reduce these disparities.This survey connects three literatures that are usually reviewed separately: women's cardiovascular health, algorithmic fairness in medical AI, and digital health and “FemTech.” We organize the evidence around four questions. Where does inequity enter the AI lifecycle? Which AI applications address women's cardiovascular needs? How do digital health and FemTech change access and risk? Which technical and governance methods can improve fairness?Across imaging, risk prediction, and language models, current studies document disadvantages for women. Imbalanced imaging data reduce diagnostic accuracy. Chest-radiograph classifiers underdiagnose female patients. Language models overgenerate male cardiac cases and can downplay women's needs. We also review sex-specific risk models, electrocardiographic and imaging systems, multimodal foundation models, simulation, telehealth, mobile health, and wearables. The methods literature covers bias measurement, auditing, mitigation, causal and mechanistic analysis, regulation, open science, and World Health Organization ethics guidance. We close with open problems in intersectionality, sex-disaggregated data, the reproductive life course, real-world fairness, and benchmarks designed for women's cardiovascular AI. Overall, equity must be incorporated throughout system design and deployment rather than treated as a post-deployment correction.

Language-model agents can plan multiple steps, call tools, browse the web, execute code, and act over many turns. As these systems move from demonstrations into deployment, measuring what can go wrong has become urgent. The relevant literature has grown from a handful of papers into several hundred in roughly three years. It is scattered across benchmark releases, red-teaming studies, frontier-lab capability reports, and governance documents that rarely cite one another. This survey organizes that literature. We assembled and screened a corpus of agentic-risk evaluations from a large paper index. The core covers general capability benchmarks, safety-behavior testbeds, and adversarial or prompt-injection evaluations. It also covers dangerous capabilities, deception and scheming, multiagent risk, runtime control, benchmark validity, and governance-facing testing regimes. Three key findings emerge consistently across the evidence. Direct paired studies show that safety alignment measured on a chat model can transfer poorly to the same model acting as an agent. Refusal rates that look strong in conversation can collapse once the model is given tools and a goal. Benchmark validity is also fragile. Underspecified tasks, weak outcome checking, contamination, and evaluation gaming can distort reported results. Finally, capability has standardized, executiongrounded suites, whereas propensity evidence remains largely scenario-based. We can increasingly say what an agent can do but not reliably what it will do under deployment pressure. We use these findings to propose a taxonomy and compare what the leading benchmarks actually measure. We then identify five open problems: construct validity, held-out difficulty, sandbagging-resistant elicitation, evaluation under adaptive adversaries, and system-level measurement.

Language-model agents can plan multiple steps, call tools, browse the web, execute code, and act over many turns. As these systems move from demonstrations into deployment, measuring what can go wrong has become urgent. The relevant literature has grown from a handful of papers into several hundred in roughly three years. It is scattered across benchmark releases, red-teaming studies, frontier-lab capability reports, and governance documents that rarely cite one another. This survey organizes that literature. We assembled and screened a corpus of agentic-risk evaluations from a large paper index. The core covers general capability benchmarks, safety-behavior testbeds, and adversarial or prompt-injection evaluations. It also covers dangerous capabilities, deception and scheming, multiagent risk, runtime control, benchmark validity, and governance-facing testing regimes. Three key findings emerge consistently across the evidence. Direct paired studies show that safety alignment measured on a chat model can transfer poorly to the same model acting as an agent. Refusal rates that look strong in conversation can collapse once the model is given tools and a goal. Benchmark validity is also fragile. Underspecified tasks, weak outcome checking, contamination, and evaluation gaming can distort reported results. Finally, capability has standardized, executiongrounded suites, whereas propensity evidence remains largely scenario-based. We can increasingly say what an agent can do but not reliably what it will do under deployment pressure. We use these findings to propose a taxonomy and compare what the leading benchmarks actually measure. We then identify five open problems: construct validity, held-out difficulty, sandbagging-resistant elicitation, evaluation under adaptive adversaries, and system-level measurement.

Artificial intelligence (AI) is now a credible translational research platform across the Alzheimer’s disease (AD) pathway, but its clinical value is often difficult to judge from performance metrics alone. This scoping-informed methodological review maps AD-AI evidence across prevention and risk stratification, diagnosis and staging, treatment and therapeutic monitoring, and care and safety support, drawing on PubMed/MEDLINE, IEEE Xplore, ACM Digital Library, ScienceDirect/Scopus-indexed venues, Google Scholar, and citation chaining for literature from 2000 to May 2026. We contribute a decision-centered evidence-maturity framework and apply it to a diagnostic-anchored sample of 31 stage-specific appraisal entries from 29 unique reports, with a PROBAST+AI-informed diagnostic concern screen. Diagnostic AI provides the strongest current evidence base, while prevention, treatment, and care and safety applications show how AI can support earlier risk recognition, anti-amyloid workflow planning, longitudinal monitoring, and caregiver-facing safety systems. In the purposive appraisal sample, explicit reporting of calibration, decision-curve or net-benefit analysis, subgroup fairness, prospective workflow testing, and patient- or caregiver-centered endpoints was uncommon and, for several stage-specific items, absent; exact counts are reported in the appraisal table. The review argues that AD-AI is most persuasive when each model is linked to a user, decision, actionable threshold, uncertainty communication strategy, and outcome, supported by leakage-resistant validation, calibration, decision-curve assessment, subgroup testing, workflow evaluation, and lifecycle governance.

The emergence of agentic AI systems — software that autonomously executes consequential decisions without human-in-loop approval — has exposed a critical governance failure in Australian and global corporate practice. Boards have relied on probabilistic assurance mechanisms: annual audits, accuracy benchmarks, and post-hoc explainability reports. A system operating at 95% accuracy is wrong one decision in twenty. At agentic velocity and scale, that error rate constitutes a systemic risk, not a tolerable variance.Three regulatory instruments have converged to render this posture legally untenable. ETSI TS 104 008 V1.1.1 (CABCA, January 2026) establishes Continuous Auditing as the technical baseline for AI governance, eliminating point-in-time compliance as a credible control. NIST AI 800-4 (March 2026) formally designates post-deployment monitoring gaps as a named control failure. The Cyber Risk Institute Financial Services AI Risk Management Framework V1.0 (February 2026) enumerates 230 discrete control objectives that no human compliance team can satisfy through manual, periodic attestation.Against these standards, the annual AI assurance report remains the primary governance artefact in most Australian boardrooms. Under Corporations Act s.180 — a conduct standard, not an outcomes standard — directors who cannot demonstrate continuous, enforceable oversight of agentic AI systems now face a personal liability question that annual audits cannot resolve.This paper presents the Deterministic Risk Architecture (DRA), developed through the UTS Human-Centred AI Lab and UTS Data Science Institute. DRA wraps probabilistic AI models in a deterministic enforcement shell comprising three layers: a Parameter Boundary Layer, a Causal Measurement Layer grounded in Professor Zhou's Causal Quantification Framework, and a Hard Logic Gate that prevents execution of any action deviating from board-approved risk parameters. The architecture shifts the governance posture from Post-Hoc Observation — detecting failure after it occurs — to Deterministic Constraint: mathematically preventing non-compliant execution before it occurs. Speed limiter, not speed camera.