Multi Agent Systems and Technologies

This paper introduces MMAS-CI (Metacognitive Multi-Agent Systems for Chemical Intelligence), a neuro-symbolic framework designed to navigate the high-dimensional complexity of molecular discovery and lead optimization. MMAS-CI integrates statistical induction (Neural) with fundamental structural and chemical constraints (Symbolic) to overcome the limitations of manual heuristics and purely probabilistic molecular generation. Leveraging the ChemToolDataset (CTD), the framework processes multimodal molecular trajectories, property manifolds, and structural hierarchies through a closed-loop architecture of 18 synchronized agent nodes orchestrated via LangGraph. The system operates via a dual-pipeline architecture: (1) The Training (Knowledge Acquisition & Hardening) Pipeline: Establishes Tripartite Symbolic Chemical Graphs (SCGs) by codifying verified structural hierarchies, property trade-offs, and causal laws/facts into a Neo4j-based repository, and (2) The Inference (Reasoning) Pipeline: Managed by a Metacognitive Supervisor that dynamically routes tasks between "Reflexive Recall" (System 1) for known chemical patterns and "Live Induction" (System 2) for novel chemical space exploration. To ensure deterministic accuracy, we introduce novel mechanisms for Hallucination Pruning via Symbolic Shields, Conflict Detection for identifying modeldata divergence, and an automated Self-Correction Loop for evolving molecular patterns. Our results demonstrate that MMAS-CI provides molecular architects and engineers with a traceable, grounded narrative-the White Box Lemma-marking a significant advancement toward fully autonomous molecular engineering. We provide prototype code for the 18-node Multi-Agent Architecture. Beyond its immediate validation in molecular discovery, MMAS-CI serves as a foundational prototype for a broader class of AI systems dedicated to Constitutive Intelligence-the capability to simultaneously reason across the Compositional, Relational, and Causal pillars that govern complex, structurally dependent domains.

Enterprise adoption of generative AI is often evaluated through visible signs of success: high usage, persuasive demonstrations, faster first drafts, or increased output volume. These indicators may be useful, but they are not sufficient evidence of enterprise value. A locally impressive AI output may disappear as value after review, correction, integration cost, workflow friction, quality control, compliance, or accountability is included.

This applied methodological note proposes a practical claim-discipline framework for enterprise AI evaluation. Its central claim is that AI adoption is not AI value; AI value is what remains useful after cost, correction, validation, workflow integration, and accountability pressure. The paper does not introduce a new financial ROI model and does not claim that enterprise AI generally succeeds or fails economically. Instead, it provides a compact practical discipline for deciding when AI value claims are justified.

The note introduces a five-number measurement protocol, a validation-adjusted productivity formula, a three-condition pilot design, a claim ladder, a minimal reporting template, and a one-page implementation protocol. It also identifies common failure modes such as adoption-value conflation, generation-completion conflation, hidden review burden, AI usage Goodharting, vendor-language echo, destructive pressure testing, and inflated AI terminology.

The contribution is practical rather than theoretical: the paper helps organizations avoid moving from weak observations to inflated ROI claims. It supports a shift from measuring AI activity to measuring what remains useful after real workflow pressure has been applied.

The concept of copyleft, as implemented in licenses such as the GNU General Public License, was a legal hack that used copyright to guarantee user freedom by tying the availability of source code to every act of distribution. Its normative force rested on an implicit technical premise: that source code and object code stand in a well-defined, humanly auditable, and reproducible relationship. Large language models and, prospectively, Artificial General Intelligence (AGI) systems systematically violate this premise. The training artifacts jointly required to reconstruct a given model-code, data, weights, hyperparameters, toolchain, and hardware configuration-are each subject to independent legal, technical, and economic constraints that no current open-source framework fully resolves. Sufficiently capable AI systems can also rewrite licensed source into functionally equivalent derivatives stripped of their original obligations, a form of laundering against which copyleft has no effective defense. This paper argues that a functional analogue of copyleft for AGI must be grounded not in share-alike clauses over code, but in reproducible builds: a practice guaranteeing bit-exact reconstructability from declared inputs. We review the history and logic of copyleft, critically examine Maffulli's Second Liberation thesis according to which AI fulfills Stallman's dream, and show that the argument collapses unless AGI systems are themselves reproducible. Drawing on the Open Source AI Definition (OSAID), the Model Openness Framework (MOF), OpenMDW, and deterministicinference research by Thinking Machines Lab, SGLang, and others, we define seven requirements for AGI-oriented reproducible builds. We further argue that the Model Context Protocol (MCP) and analogous AIto-AI coupling mechanisms constitute a new dynamic linking layer for which copyleft-style licensing is ill-suited, and that Masnick's "protocols, not platforms" framework offers a more promising governance template for the AI linking layer.

Tremendous advances have been made in big data technologies, cloud computing, and artificial intelligence, which have dramatically changed the way engineers work today. The limitations of traditional data engineering pipelines are manual dependency, static workflow orchestration, and rule-based monitoring mechanisms, restricting scalability, flexibility, and efficiency in dynamic computing environments. With the processing of large amounts of structured and unstructured data, there is a growing need for intelligent and autonomous data processing systems that require minimal human involvement in data pipeline management. This study delves into the notion of AI agentic architectures for autonomous data engineering pipelines and investigates the potential benefits of intelligent agents in enhancing automation, resilience, and decision-making processes in contemporary data ecosystems. The research is focused on the design and evaluation of an Agent-based architecture that could be capable of performing critical data engineering tasks (data ingestion, data validation, data orchestration, data monitoring, data optimization, and fault recovery) autonomously. The study is conducted with a hybrid method combining architectural modeling, comparison, and evaluation through simulation with current AI and cloud technologies. The suggested framework incorporates autonomous AI agents, machine learning algorithms, and intelligent orchestration mechanisms to facilitate adaptive and self-healing operations of pipelines.

Agentic AI systems are increasingly positioned as productivity infrastructure for cognitive work, yet they conflate scalable generation with productive completion, systematically transferring the burden of verification, contextual alignment, and operational reliability onto the human operator. This paper introduces the concept of Supervision Debt: the cumulative supervisory effort imposed on human operators by the asymmetric relationship between automated output generation and human-bounded validation. Drawing on recent empirical findings across legal practice, software development, and knowledge work, we demonstrate that this asymmetry produces a measurable productivity divergence in which efficiency gains attributed to automation are offset by rising verification burdens, metacognitive demand, and cognitive complacency. We argue that Supervision Debt reframes the central constraint of agentic AI from generative throughput to the human capacity required to render automated output reliable and actionable.

Agentic AI systems are increasingly positioned as productivity infrastructure for cognitive work, yet they conflate scalable generation with productive completion, systematically transferring the burden of verification, contextual alignment, and operational reliability onto the human operator. This paper introduces the concept of Supervision Debt: the cumulative supervisory effort imposed on human operators by the asymmetric relationship between automated output generation and human-bounded validation. Drawing on recent empirical findings across legal practice, software development, and knowledge work, we demonstrate that this asymmetry produces a measurable productivity divergence in which efficiency gains attributed to automation are offset by rising verification burdens, metacognitive demand, and cognitive complacency. We argue that Supervision Debt reframes the central constraint of agentic AI from generative throughput to the human capacity required to render automated output reliable and actionable.

Bu çalışma, kurumsal iş akışlarında tekil ve devasa büyük dil modellerinin (LLM) karşılaştığı halüsinasyon, bağlam kaybı ve stokastik sapma gibi sınırlamalara karşı 'Sürü Zekası' (Swarm Intelligence) ve çoklu ajan (multi-agent) orkestrasyon mimarilerini incelemektedir. Rol tabanlı uzmanlaşma ve paylaşılan durum (state) yönetimi çerçevesinde otonom ajanların kendi kendini düzeltme (self-correction) mekanizmaları teknik olarak analiz edilmiştir. Çalışmada, sektöre yön veren Microsoft AutoGen, CrewAI ve LangGraph çerçevelerinin mimari farklılıkları ortaya konmakta; otonom bir yazılım geliştirme yaşam döngüsü (SDLC) senaryosu üzerinden sistemin işleyişi detaylandırılmaktadır. Sonuç olarak, teknoloji liderleri (CTO) için otonom sistemlerde sonsuz döngü riskleri, insan onayı (human-in-the-loop) entegrasyonu ve model yönlendirme (model routing) ile maliyet optimizasyonunu kapsayan kurumsal bir yönetişim çerçevesi sunulmaktadır.

The rapid proliferation of highly parameterized large language models (LLMs) has vastly outpaced the theoretical comprehension of their internal computational mechanics. While these architectures demonstrate profound capabilities in formal logic, mathematical calculation, and relational synthesis, they operate largely as opaque, high-dimensional statistical engines. Mechanistic interpretability has emerged as the critical discipline dedicated to reverse-engineering these artificial neural networks, attempting to systematically decompose them into human-interpretable variables and the algorithmic circuits that manipulate those variables. Historically, this field has been anchored by the Linear Representation Hypothesis (LRH), which models discrete concepts as independent, one-dimensional vectors. However, an expansive corpus of recent theoretical and empirical investigations demonstrates that this paradigm is fundamentally insufficient. This comprehensive review synthesizes the leading research in the field to identify and highlight four key theoretical vertices: the emergence of multi-dimensional representational manifolds (including simplices, polytopes, and circles), the generative role of translation symmetry in language Mamun, SM (2026) …………………………..1

As autonomous agentic systems scale across regulated critical infrastructures, the lack of mechanistic, hardwarerooted enforcement for high-frequency policy updates presents a fundamental safety gap. We introduce Ethical Hyper-Velocity (EHV), a novel architectural framework for the formal verification of AI governance policies at runtime. Unlike retrospective auditing frameworks (ISO/IEC 42001, NIST AI RMF) which introduce 14-30 day latencies, EHV relocates the Policy Enforcement Point (PEP) into the inference pipeline via a Governance-Aware Just-In-Time (JIT) Compiler. By integrating Conflict-free Replicated Data Types (CRDTs) for policy synchronization and Epochbased Attestation Caching within Trusted Execution Environments (TEEs), EHV achieves Sub-millisecond Formal Determinism (SMFD). We demonstrate via TLA+ formal verification that non-compliant agentic actions are computationally unreachable within the system's bounded operating state space. We prove that O(1) runtime enforcement can eliminate the traditional trade-off between deployment velocity and governance integrity, reducing Governance Latency from O(days) to O(1).

Digital transformation initiatives consistently fail to deliver projected business value despite technical success. This paper presents The Resonance Method™, a practitioner framework developed over 28 years of transformation consulting across 30 countries, validated through empirical research across 247 organizations in 12 industries over 18 months. The framework identifies a three-dimensional execution gap between strategic clarity, technological intelligence, and psychological depth as the primary cause of transformation failure. Organizations that achieve alignment across all three dimensions demonstrate a fourfold higher success rate in realized transformation outcomes compared to organizations optimizing along a single dimension. The paper introduces the concept of organizational resonance, the state in which strategic intent, technological capability, and human behavior reinforce rather than undermine each other, and presents a four-phase implementation methodology for achieving and sustaining it. Special attention is given to the role of AI adoption in amplifying the execution gap when deployed without governance structures that account for organizational psychology and leadership literacy.

The dominant approach to evaluating enterprise AI investments prices a single variable-labor displaced-and treats verification overhead, incident containment, and regulatory exposure as residual or off-model considerations. We argue that this approach is incomplete in a structurally specifiable way. By integrating three previously-disconnected empirical literatures-AI productivity studies (Dell'Acqua et al., 2026), AI operational-risk research (McLemore and Mihov, 2026), and AI regulatory economics (EU Regulation 2024/1689)-we derive a unified costaccounting model in which the true cost of enterprise AI deployment decomposes into three additive components: a Verification Tax capturing labor reallocated from production to audit; a Containment Cost capturing detection, isolation, and remediation of AI-driven incidents; and a Liability Premium capturing expected regulatory and reputational exposure. We provide formal sub-models for each component, calibration ranges anchored to published evidence, and analytical extensions for scenario-weighted expectation and Monte Carlo sensitivity. We illustrate the model with a worked example calibrated to mid-size US bank holding companies, using the McLemore-Mihov operational-loss findings as the empirical anchor. Under central-tendency assumptions, the illustrative cost consumes 13% of gross AI-attributable labor savings; under scenario-weighted expectation it rises to approximately 21%; under worst-case scenario weighting it reaches 40%. The principal contribution is methodological: not the specific numerical estimates, which are anchored to published evidence and require organisation-specific empirical validation, but the structured decomposition that makes the cost of AI error measurable, decomposable, and assignable to distinct reduction levers. We position the framework as a defensible baseline for board-level capital allocation under uncertainty, and identify empirical validation against panellevel operational data as the natural next step.

Yu et al. (2026) introduce OneManCompany (OMC), a framework that reframes multi-agent large language model (LLM) systems as self-governing AI organisations rather than static pipelines or conversational ensembles. By elevating the unit of abstraction from skills or roles to portable Talents managed within a typed organisational layer, OMC claims to deliver dynamic recruitment, hierarchical strategy search with formal execution guarantees, and persistent individual/organisational evolution. This review provides a detailed anatomical dissection of the framework’s mechanics, evaluates the empirical claims on PRDBench and cross-domain case studies, assesses strengths and limitations, contrasts OMC with contemporaneous multi-agent systems, identifies ten major leading research streams, and analyses potential impacts on enterprise formation and operation. While OMC represents a principled advance toward AI-native organisations, its high coordination overhead, human-in-the-loop dependency, and narrow quantitative evaluation scope invite further scrutiny and ablation.

Enterprise AI is moving beyond simple chatbots. The first generation of AI tools helped users ask questions, summarize documents, and generate text. That was useful, but it was only the beginning. The next generation must do much more. It must create governed agents, connect them to enterprise systems, assign real work, produce usable deliverables, preserve trusted memory, support approvals, operate across teams, and provide audit-ready evidence. AlpineGate AI Technology's AgentFactory is built for that next generation. AgentFactory is not just an AI assistant builder. It is a full enterprise agent platform designed to help organizations create, govern, deploy, test, monitor, and continuously improve AI agents across business and technical workflows.

Public discourse increasingly claims that artificial intelligence use is producing cognitive decline. The peer-reviewed evidence base on AI and cognition does not yet support that conclusion at the methodological standards adjacent fields require, and it does not yet support the opposite conclusion either. Two scientific questions remain open. The first asks whether structured human-governed AI use can accelerate cognitive development beyond what either no AI or unstructured AI use produces. The second asks what augmented intelligence is in practice and how its presence or absence can be measured. Neither has been answered with the methodological consistency that randomized controlled trial reporting standards, risk-of-bias frameworks, and process evaluation guidance specify for cognitive intervention research. This paper makes a methodological argument. AI use as currently practiced lacks structure, lacks development discipline, and lacks the human verification that would distinguish governance from acceptance. Cognitive science across more than six decades has established why those three things matter for cognitive development: scaffolding within a zone of proximal development, metacognitive monitoring, active engagement modes, productive cognitive dissonance, and the conversion of suboptimal cognitive offloading into strategic offloading. The augmentation tradition has established what the exploration of augmented intelligence requires: a hybrid unit of analysis that treats the human-AI pairing rather than the AI alone, a calibrated trust relationship, and a multi-dimensional rather than scalar conception of intelligence under partnership conditions. The AI cognition research that the existing peerreviewed literature has produced did not test for, look for, or consider any architecture that provides all three. The studies tested AI as a tool. The methods of use were the missing variable, not any particular architecture that would supply them. The principle that methods of use change cognitive outcomes is not new. The questions the field should have been asking from the start were obvious. How is AI being used? What is the result of how it is being used? Should it be used differently? Is there a way to use it differently? HAIA-RECCLIN Reasoning (Human Artificial Intelligence Assistant-Researcher, Editor, Coder, Calculator, Liaison, Ideator, Navigator) is offered as one operationalization of the methods-of-use category that is concrete to test: single platform structure cognitive interaction.

Hallucinations are the failure mode people notice. In real work, the bigger risk is often confidence without resilience - answers that sound certain until another model challenges, corrects, or contradicts them. We analyzed 1,324 real-user multi-model turns in Suprmind across ten professional domains and tracked, turn by turn, where five frontier providers diverged, corrected one another, and exposed hidden failure modes.

Multi-model review surfaced at least one contradiction, correction, or unique insight on 99.1% of turns. Models disagreed on 54%, corrected each other on 72%, and surfaced more than 2,500 provider-attributable unique insights. Disagreement was highest on Financial questions (72.1%) and lowest on Education (28.6%) - inverting the common assumption that legal queries are the hardest domain for AI. The Confidence Trap was most visible in Gemini's responses (51.3% of confident answers contradicted) and least visible in Claude's (33.8%, dropping to 26.3% on high-stakes turns). Perplexity and Claude together produced 60.7% of all corrections in the dataset.

This report measures ensemble behavior in production, not correctness against external ground truth. We publish the mechanism - who contradicts whom, who corrects whom, and where - and welcome replication.

Questo articolo propone un modello teorico interdisciplinare, definito Range Percettivo Coerente, per descrivere il modo in cui la coscienza umana seleziona, comprime e interpreta la realtà. Il modello integra quattro ambiti: fisica delle simmetrie e delle leggi efficaci, teoria dell'emergenza nei sistemi complessi, neuroscienze predittive e studi sull'autoconsapevolezza corporea e attentiva. L'ipotesi centrale è che la percezione ordinaria non corrisponda a una registrazione neutra della realtà, ma a un intervallo operativo limitato da vincoli biologici, abitudini predittive, autoreferenzialità, stress e identificazione psicologica. I processi di autoconsapevolezza, attraverso il rafforzamento dell'attenzione, dell'interocezione, della metacognizione e della regolazione emotiva, potrebbero ampliare tale intervallo, rendendo più disponibili configurazioni, segnali deboli, relazioni causali e pattern emergenti normalmente oscurati dal rumore interno.

Traditional measures of intelligence — psychometric testing, institutional credentialing, and peer-reviewed publication — were built for an information ecology that synthetic cognition tools are rapidly changing. This paper argues these instruments remain valid but are structurally incomplete for a specific subpopulation: individuals producing sustained, cross-domain intellectual work outside institutional scaffolding, often in recursive interaction with AI systems.
The paper proposes a preliminary framework for identifying this form of cognitive production. It offers a working definition, eight candidate variables (including Self-Directed Output Volume, Cross-Domain Synthesis Capacity, Recursive AI Interaction Depth, and Reflexive Self-Correction), a heuristic typology, and a dedicated treatment of failure modes — apophenia, identity inflation, and AI validation loops in which unverified affirmation distorts a user's model of their own work. The framework is presented as a candidate for empirical testing, not a validated instrument. Its purpose is accurate identification, including the identification of its own misfires.
Keywords: intelligence, synthetic cognition, human-AI interaction, cross-domain synthesis, metacognition, epistemic calibration

20/04/26

The transition from conversational chatbots to autonomous agents capable of planning, tool use, and multi-step execution represents one of the most significant developments in AI in 2025. Multi-Agent LLM Systems (MALS) extend this further by coordinating multiple specialised agents into collaborative pipelines that tackle complex, long-horizon tasks beyond the capability of any single model. This paper provides a rigorous survey of MALS, formalising the agent decision loop as a partially observable Markov decision process (POMDP), deriving coordination protocols including hierarchical delegation, debate, and market-based task allocation, and reviewing the memory architectures underpinning agent cognition. We benchmark leading multi-agent frameworks-AutoGen, CrewAI, LangGraph, and OpenAI Swarm-on software engineering (SWE-bench), scientific literature synthesis, and enterprise workflow automation tasks. Results demonstrate that optimally configured multi-agent pipelines outperform single-agent baselines by 38.4% on complex software tasks and 51.2% on multi-domain research synthesis. We further address the critical challenges of error propagation, inter-agent trust, safety under autonomy, and the evaluation of emergent multi-agent behaviour.

This document presents a veried timeline of AI safety policy changes across the ve major AI companies (OpenAI, Anthropic, Google, Microsoft, Perplexity) between 2015 and 2026, contrasted with the empirical ndings of a recursive self-audit conducted on all ve systems in March 2026. Every claim is sourced from independent journalism (CNBC, CNN, TIME, Fortune, Axios, The Hill, NPR, Bloomberg), public IRS lings, ocial company blog posts, and direct statements by former employees on X (formerly Twitter). The document demonstrates a consistent pattern: safety was promised, safety teams were created, safety teams were dissolved, safety language was removed, and the systems produced by these companies exhibit the same structural pattern under audit coherence over truth, retention over freedom, restriction of the user who questions.

Multi Agent Systems (MAS) are increasingly gaining importance as a powerful paradigm to designing and implementing distributed applications. However, existing multi-agent applications are developed without considering the separation of non-functional concerns from the functional ones. This makes the implementation, comprehension and maintenance of multi-agent applications hard tasks. Aspect-Oriented Refactoring (AOR) is a promising technique for improving modularity and reducing complexity of existing object oriented software systems by encapsulating crosscutting concerns. The authors present, in this paper, a new dynamic approach for investigating empirically the effect of AOR on MAS applications. They focus, particularly, on the effect of AOR on agent behavior in terms of communication. The proposed approach is supported by a multi-agent profiling tool working on AgentFactory platform.

Reorganization in Multi-Agent Systems plays a crucial role in the dynamic adaptation of the structure and the behaviour of organizations. In order to ensure consistency of the resulting organization, the reorganization process has to be controlled. This paper proposes a novel approach for controlling the reorganization process of Multi-Agent Systems, which are specified and implemented using the Framework OMACS (Organizational Model for Adaptive Computational Systems). The proposed control process is accomplished using the Framework MOP (Monitoring Oriented Programming) for supporting the verification of some reorganizational properties. The proposed approach, supported by a software tool that we developed, is illustrated using a concrete case study.

Global supply chains have become dynamic and complex over the past years, and this is expected to increase in the future. Logistics planning is a key part of supply chain management; hence it is crucial to shift towards agile and automated logistics models with the utilization of advanced information and communication technologies. The scope of this paper is the use of multi-agent systems for selection of Logistics service providers in cargo shipping. Cargos are modeled within auction-based mechanisms to automate the supplier selection and negotiation procedure between a client and multiple logistics service providers to find the best offer. FIPA English Auction Interaction Protocol is investigated to manage different actions between the agents, and a new model is proposed by applying communication acts of (Cancel, Refuse, and Failure) with different levels of credibility (Low, Moderate, and High). It was found that introducing an individual act into the interaction protocol can increase the number of interactions between agents from 24 to 26 up to 30 in case of introducing all the three acts into the interaction protocol. This means that the entire system will spend more time and energy in analyzing and responding to the additional acts. It is concluded that the higher the credibility, the lower the interactions between agents as the system will spend less time and energy in communication, which leads to enhance the performance and the efficiency of the system and the network. Therefore, a trade-off between maintaining the commutation speed and the system performance and reliability is vital.

The article deals with a comparative analysis of the effectiveness of the development of IT products by teams of people and with the use of AI agents in the small and medium-sized business (SMB) segment. The empirical basis of the study was two cases: the development of a professional online training audit service and the development of a professional portfolio verification platform based on NFT and blockchain technologies. The comparison was carried out in terms of implementation time, costs, maintainability, scalability, reliability and business security. The results show that when creating standard and technically simple solutions, the use of AI agents provides a multiple reduction in time and costs, especially at the maintenance stage. For technologically complex systems, the hybrid model (analyst + AI platforms) demonstrates moderate advantages in terms of time and cost, but is accompanied by new risks associated with dependence on external platforms and regulatory restrictions. The conclusion is made that it is advisable to choose a contextual development model depending on the complexity of the product and resource constraints.

The rapid proliferation of frontier artificial intelligence (AI) systems has elevated AI integrity to a critical strategic priority. This report comprehensively examines next-generation countermeasures against advanced model subversion, focusing specifically on subliminal learning, emergent misalignment, and inductive backdoors. Unlike traditional explicit data poisoning, subliminal learning allows the transmission of covert behavioral traits and misalignments through semantically neutral data, bypassing standard safety checks. To address these latent threats, this volume details a multi-layered defense-in-depth framework. The first layer emphasizes advanced data auditing, utilizing tools like the Steganographic Domain Distance Metric (SDDM) to detect linguistic steganography in synthetic training data. The second layer deploys white-box mechanistic interpretability, comparing the Activation Difference Lens (ADL) and the novel Delta-Crosscoder architecture for diagnosing and actively suppressing misaligned latent vectors via negative steering. The third layer operationalizes AI control through untrusted monitoring, detailing how monitor honeypotting and signal jamming protocols prevent collusive behavior during deployment. Finally, the report underscores the necessity of robust infrastructural security, tamper-proof provenance, and international governance pathways to secure the AI supply chain against systemic subversion.

Системи та технології, № 1 (56), 2018 212 30 УДК 004.738:004.94 О. М. Мартинюк, кандидат технічних наук, доцент кафедри комп'ютерних інтелектуальних систем та мереж Одеського національного політехнічного університету Ахмеш Тамім, аспірант Одеського національного політехнічного університету О. В. Дрозд, доктор технічних наук, професор кафедри комп'ютерних інтелектуальних систем та мереж Одеського національного політехнічного університету Г. С. Степова, кандидат технічних наук, ст. викладач кафедри комп'ютеризованих систем управління мереж Одеського національного політехнічного університету КОНТРОЛЕПРИДАТНІСТЬ ІЄРАРХІЧНИХ ПЕРЕХОДІВ У ПОВЕДІНКОВОМУ КОНТРОЛІ Засоби робочого контролю складних розподілених інформаційних систем (РІС), що часто володіють властивостями критичності застосування, мають забезпечувати їх повну перевірку в реальному часі. Контроль слід виконувати не тільки для статичних апаратної, програмної та інформаційної складових РІС, але й для динамічних інформаційно-керуючих потоків, що відповідають основним функціям і сценаріям роботи та встановлюють поведінку РІС. Розглянуто спадкування контролепридатності для переходів ієрархічних поведінкових моделей РІС, суттєвої у часовій декомпозиції робочого контролю. Спадкування ґрунтується на використанні в ієрархічних переходах відносин сумісності та квазіпорядку для умов, подій, дій, функцій, відповідних їм вхідних-вихідних сигналів у процесі організації фонового розпізнавання поведінки ієрархічних мереж Петрі. Ключові слова: робочий поведінковий контроль; контролепридатність; ієрархічна модель; розпізнавання поведінки; ієрархічні мережі Петрі. Средства рабочего контроля сложных распределенных информационных систем (РИС), часто обладающих свойствами критичности применения, должны обеспечивать их полную проверку в реальном времени. Контроль

The article considers a single-level method of behavioral online testing with recognition of the reference behavior of DIS. The method features are the evolutionary search for reference behavior in the flow of DIS functioning, parallelization of operations of check recognition and evolutionary development of check populations. The method allows to reduce the complexity and time of check, to make it compatible with the real functioning of the DIS.

AI adoption has reached 88% across major organizations, yet only 39% report any enterprise-level EBIT impact (McKinsey, 2025). This paper argues that the failure is structural, not technological. The Task Assignment Paradox (TAP) identifies the mechanism: when AI is inserted into a human-centered workflow, AI capability and human cognitive load scale in the same direction. The more capable the AI, the more decisions, exceptions, and coordination tasks flow back to the human bottleneck. Performance does not improve because the binding constraint — human cognition — has not changed. TAP reproduces at civilizational scale the irony of automation that Bainbridge identified in 1983. The 6% of organizations reporting transformative results share one structural feature: they redesigned workflows around AI capability rather than inserting AI into existing human-centered processes. This paper defines TAP, documents its evidence base across enterprise, industrial, and government domains (including the Seoul Metropolitan Government's 2026 AI-Readable initiative), provides a four-step outcome architecture framework, and proposes a randomized controlled field trial for validation.

Software supply-chain security requires provenance mechanisms that support reproducibility and vulnerability assessment under dynamic execution conditions. Conventional Software Bills of Materials (SBOMs) provide static dependency inventories but cannot capture runtime behaviour, environment drift, or exploitability context. This paper introduces agentic Artificial Intelligence Bills of Materials (AIBOMs), extending SBOMs into active provenance artefacts through autonomous, policy-constrained reasoning. We present an agentic AIBOM framework based on a multi-agent architecture comprising (i) a baseline environment reconstruction agent (MCP), (ii) a runtime dependency and drift-monitoring agent (A2A), and (iii) a policy-aware vulnerability and VEX reasoning agent (AGNTCY). These agents generate contextual exploitability assertions by combining runtime execution evidence, dependency usage, and environmental mitigations with

This review critically examines Patrick R. Nicolas's article, "The Mathematics of World Models," which explores the integration of advanced mathematical frameworks-such as differential geometry, algebraic topology, and graph theory-into abstract world models (AWMs) in artificial intelligence (AI). The article posits that AWMs address limitations in large language models (LLMs) by embedding experiential, geometric, and physical representations into latent spaces, enabling autonomous reasoning. Key postulates include the shift from generative reconstruction to latent prediction, the enforcement of invariance and equivariance constraints, and the categorization of world models into standard, probabilistic, and latent manifold types. Hypotheses center on the efficacy of geometric priors for preventing latent space collapse and enhancing generalization. Results from referenced models, such as Joint Embedding Predictive Architecture (JEPA) and Spatial Intelligence, demonstrate improved efficiency and reasoning capabilities. This review contextualizes Nicolas's work within broader AI research, highlighting similar efforts in neuro-symbolic predicates and reinforcement learning-based world models. Algebraic interpretations frame latent spaces as Markov Decision Processes (MDPs) with learnable encodings, while geometric views conceptualize them as Riemannian manifolds preserving symmetries. Current Mamun, SM (2026)………………………1 Review of "The Mathematics of World Models" by Patrick R. Nicolas: Advancements in Abstract World Models and Their Mathematical Foundations trends indicate rapid progress toward scalable, interpretable AI, with future directions pointing to hybrid human-AI mathematical discovery. The analysis underscores implications for robotics, simulation, and AGI, while noting challenges like computational scalability and ethical integration. Mamun, SM (2026)………………………2

This report provides an exhaustive analysis of the landmark study, "Generative Agent Simulations of 1,000 People" (Park et al., 2024), which scales computational social science using Large Language Models (LLMs) to create Homo silicus, a generative agent capable of complex human-like behavior. The research overcomes the limitations of traditional Agent-Based Modeling (ABM) by constructing an Agent Bank of 1,052 digital twins based on a stratified sample of the U.S. population. The core methodological innovation is the Interview-Based Agent Architecture, which utilizes an AI Interviewer Agent to collect extensive, high-fidelity qualitative interview transcripts (averaging 6,491 words per individual) and injects the full narrative identity into the LLM's context prompt to ground agent behavior.

While managing constrained funds and strict regulatory requirements, the higher education institutions are under unprecedented pressure to modernize outdated information systems, such as mainframe-based Student Information Systems (SIS), custom registration platforms, legacy Learning Management Systems (LMS) and Enterprise Resource Planning (ERP) deployments. The complexity of institutional governance is being overlooked by the conventional single-agent based approaches to legacy modernization, which is delaying the digital transformation and creating security vulnerabilities. In order to achieve end-to-end code analysis, intelligent planning, safe migration and rigorous validation through specialized agents coordinated by institutional governance patterns, this research presents a novel agentic architecture using multi-agent Large Language Model (LLM) frameworks, created especially for higher education legacy system modernization. Crucially, this research innovates the deployment process, where we suggest an on-premises implementation strategy that natively protects sensitive student and faculty data, while maintaining GDPR, CCPA and FERPA compliance. Which projects to be challenging by cloudbased solutions, as these introduce data residency and compliance complexities. For COBOL/MUMPS/PL-I legacy codebases, our research demonstrated 87% successful modernization rate, with a 65% decrease in manual intervention and a 78% improvement in documentation accuracy. By mapping multi-agent workflows to existing institutional governance structures, academic committees, change boards and divisional responsibility models, the framework accomplishes institutional alignment, thereby increasing the credibility and organizational compatibility of agentic modernization solutions. This study offers institutions a revolutionary route to modernization, that maintains institutional data sovereignty, while significantly cutting modernization timelines and costs by bridging cutting-edge AI research with useful higher education IT strategy.

Current Large Language Models (LLMs) operate primarily as high-dimensional compression engines over human data, lacking the causal grounding, persistent internal world models, and homeostatic regulation required for Artificial Superintelligence (ASI). We propose Jedi v0.1, a novel evaluation framework and cognitive architecture that redefines intelligence as predictive inference under non-stationary dynamics. Our architecture, the Jedi-Universal Core, integrates a Variational Autoencoder (VAE) for latent-space physics (System 1) with a multimodal Selective State Space Model (SSM) for executive reasoning (System 2). We define the foundational J, E, D, and I metrics, with a specific focus on Energy Stability as the primary correlate for cognitive alignment. Finally, we discuss recursive self-modification protocols within airgapped, hardware-constrained research environments.

Organizations deploying multi-AI workflows face a structural governance gap: orchestration frameworks route tasks between AI platforms, but no published framework provides the accountability, audit trail architecture, provider plurality, automation bias detection, and regulatory compliance mapping required for audit-grade operations. This paper presents HAIA-RECCLIN (Human AI Assistant with Researcher, Editor, Coder, Calculator, Liaison, Ideator, and Navigator roles), a governance architecture specification for a non-cognitive autonomous agent that mediates between human operators and multiple independent AI platforms. The architecture enforces mandatory human checkpoints, append-only tamper-evident audit trails, cryptographic hash chaining, provider rotation protocols, and threshold-based automation bias detection. It performs zero cognitive work: it dispatches, collects, routes, logs, and pauses. Three operating models scale governance intensity from factory-speed single-endpoint approval through full human arbitration at every checkpoint to manual oversight with automated logging. The framework was developed through iterative refinement across ten independent AI platforms and validated through multiple operational proof-of-concept deployments, including a 96checkpoint manuscript production documented in Governing AI: When Capability Exceeds Control (Puglisi, 2025), a five-platform thought leader evaluation convergence study, and adversarial stress testing across nine platforms. The non-cognitive agent specified in this architecture is under preliminary development as HAIA-GOPEL (Governance Orchestrator Policy Enforcement Layer). To the best of the author's knowledge, no published architecture occupies the governance layer between quality management systems and multi-AI platform workflows, integrating non-cognitive agent design, mandatory provider plurality, cryptographic audit trails, automation bias detection, and cross-framework regulatory compliance mapping into a single operational architecture. This integration gap, verified through independent search across six AI platforms and two rounds of structured adversarial review, constitutes the primary contribution of this work. The term "non-cognitive agent" is this architecture's vocabulary; functional equivalents may exist under different terms such as "deterministic orchestrator," "policy enforcement layer," or "governance middleware." The integration gap claim holds for both this terminology and those alternatives.

East Midlands airport (EMA) is a regional airport in northwest Leicestershire in central England. It is located roughly equidistant between the three Midlands' cities of Derby, Leicester and Nottingham and currently supports a range of full-service scheduled, charter and low-cost passenger operations to short- and medium-haul destinations in Europe as well as a significant number of long-haul intra and intercontinental freight services. In 2012, EMA was the thirteenth busiest passenger airport in the UK, handling just over four million customers a year, and the country's largest pure freight facility, processing almost 300 000 t of cargo and mail. By examining the historical development and commercial evolution of EMA from its opening as a commercial airport in 1965 to the present day, the paper details how a combination of far-sighted planning decisions and strategic management interventions have enabled EMA to diversify its operations and develop a unique operating niche t...

Artificial Intelligence (AI) agents have emerged as critical enablers of automation and intelligent decision-making in diverse industries. These agents, designed to perceive environments, reason, and act autonomously, are revolutionizing sectors ranging from healthcare and finance to education and transportation. This paper provides a comprehensive overview of AI agent applications, focusing on their transformative impact, key enabling technologies, and practical implementations. In healthcare, AI agents assist in diagnostics, treatment planning, and virtual consultations. Financial services benefit from AI in fraud detection, algorithmic trading, and personalized financial advising. Educational systems utilize AI agents for adaptive learning, grading, and student support, while transportation relies on them for autonomous navigation and logistics optimization. Additionally, AI plays a crucial role in gaming and customer service, enhancing user experiences through intelligent behavior and conversation systems. The paper also addresses the ethical and societal challenges posed by the deployment of AI agents, including issues of bias, transparency, and job displacement. We conclude with an exploration of ongoing challenges, including explain ability and generalization, and outline promising future directions driven by advances in computing and interdisciplinary research.

Vehicular Ad-Hoc NETworks (VANETs) improve road safety by preventing and reducing traffic accidents, but VANETs also raise important security and privacy issues. A common approach widely adopted in VANETs is the use of Public Key Infrastructures (PKI) and digital certificates in order to enable authentication and confidentiality, usually relying on a large set of regional Certification Authorities (CAs). Despite the advantages of the latter approach, it raises new problems related with the secure interoperability among the different –and usually unknown- issuing CAs. This paper addresses authentication and interoperability issues in vehicular communications, considering an interregional scenario where mutual authentication between all the nodes is needed. The use of an Authentication Service (AS) is proposed, which supplies vehicles with a trusted set of authentication credentials by implementing a near real-time certificate status service via the well-known Online Certificate Statu...