| ai.viXra.org > Artificial Intelligence > 2603 |
|
 |
| ai.viXra.org > Artificial Intelligence > 2603 |
|
|
[10] ai.viXra.org:2603.0101 [pdf] submitted on 2026-03-29 11:52:42
Authors: Stephane H Maes
Comments: 36 Pages. All related details of the projects (and updates) can be found and followed at https://shmaes.wordpress.com/
Information Technology Service Management (ITSM) and the execution of ITIL 4 frameworks are currently bottlenecked by legacy, deterministic software platforms that rely heavily on manual workflows and human operators. They also come with lock-in, and heavy TCO. This paper argues for replacing these monolithic, or composable, applications with a fully autonomous Agentic AI application, and predicts that it will happen soon. It is organized first with typical practices for agentic AI platforms. Accordingly, by leveraging hierarchical agent topologies, standardized open protocols (MCP and A2A), and dynamic temporal graph memory, enterprises can transition from manual orchestration to the algorithmic execution of ITIL 4 practices, e.g., achieving zero-touch incident resolution and predictive problem management. This approach extends far beyond IT operations, nowadays often part of ITSM offerings. It is already being done by some vendors, but with time and efforts to develop, that still often precludes enterprises to do it themselves, while the innovators dilemma limits what incumbent vendors are willing to transform to agentic AI, i.e., definitively not the core processes of their software, sticking instead to side-car copilots and agents extensions, associated to end user task, which already provide significant ROI. New vendors should not have such qualms.The shift toward "Do-It-For-Me" (DIFM) autonomous execution fundamentally collapses the seat-based licensing models of traditional software, triggering the obsolescence of massive enterprise suites like ITMS/ESM/ITOM and including ERP and CRM platforms. We detail how the "Agentic Strangler Fig" pattern, catalyzed by Application-aware (agentic) AI with real-time discovery and coding (RTDC) mechanisms, defined in the paper, allows organizations to bypass multi months or year migrations, by surrounding, with AI agentic processes that extend then replace and decommissioning the legacy enterprise applications, while saving in TCO, and achieving fully autonomous outcome. This way, enterprises are able to surround their ITSM/ESM/ITOM software, as well as other enterprise software, until left only with agentic AI and a database system of record, leading to ROI from significant cost reductions, autonomous automation, and the ability to customize their preferred processes, without the usual risks that this will lead to difficult upgrades in the future, or other problems.We also argue that the market of enterprise application market is about to be disrupted, but contrary to many recent discussion, this does not necessarily mean the dead of SaaS.
Category: Artificial Intelligence
[9] ai.viXra.org:2603.0094 [pdf] submitted on 2026-03-23 05:02:21
Authors: Sayali Patil
Comments: 10 pages, IEEE two-column format, 15 references, 4 figures, 2 tables. Includes formal soundness proof and 300-episode agentic simulation. AI-assisted article.
The kill switch problem for autonomous AI agents is conventionally treated as a design property: either an agent has a shutdown mechanism or it does not. This binary framing obscures the more consequential engineering question, which is not whether a kill switch exists but when it should trigger. This paper introduces the Halt Condition from Operating Envelopes (HCOE) framework, a formal architecture for deriving kill-switch criteria from intent-based operating envelopes, grounded in the chaos-level engine paradigm of U.S. Patent No.u202012,242,370u2009B2. HCOE treats halt conditions not as static design properties but as computable, real-time threshold functions over a continuously measured Operating Envelope Deviation (OED) signal. The framework introduces five formally defined halt classes, a graduated response architecture mapping OED zones to halt disposition (continue, pause, halt), and a proof of soundness establishing that HCOE activation prevents irreversible agent actions outside the declared operating envelope with probability approaching one as the deviation measurement interval shrinks. Evaluation across a 300-episode agentic simulation demonstrates combined halt precision of 91.8% and recall of 89.8% across five halt classes, with rollback success rates of 87% at three or fewer irreversible actions. These results establish HCOE as a practically deployable, formally grounded kill-switch architecture that closes the gap between the theoretical corrigibility literature and the operational requirements of production agentic AI systems.
Category: Artificial Intelligence
[8] ai.viXra.org:2603.0088 [pdf] submitted on 2026-03-22 13:18:25
Authors: Sayali Patil
Comments: 11 pages, IEEE format, 15 references, 4 tables, multiple figures. Includes formal methodology and experimental simulation results. AI-assisted article.
Static infrastructure planning models fail in a predictable way: they are calibrated on the past and are therefore least accurate precisely when conditions change most rapidly. The problem is not merely one of forecast precision but of architectural rigidity---these systems have no mechanism for updating their assumptions in response to the very outcomes they predict. This paper introduces the Intent-Based Planning Engine (IBPE), a closed-loop, AI-driven framework for adaptive infrastructure demand forecasting that draws its conceptual architecture from two converging technical traditions: intent-based networking and chaos engineering. IBPE integrates multivariate regression, ARIMA time-series modeling, unsupervised behavioral segmentation, structured scenario perturbation, and gradient-based feedback adaptation within a single modular system. The framework's most architecturally distinctive element is an intent modeling layer that disaggregates aggregate demand into behaviorally coherent population segments, each characterized by its own elasticity profile and sensitivity to macroeconomic perturbation. The feedback adaptation mechanism is formally derived from the chaos-level engine paradigm developed in U.S. Patent No. 12,242,370 B2 (Cisco Technology, Inc., 2025), in which controlled perturbation, impact measurement, and parameter correction form an iterative closed loop progressively narrowing the gap between intended and observed system behavior. Experimental evaluation across a 150-unit residential infrastructure simulation demonstrates a 14.2% reduction in mean absolute error over single-method baselines, a 23% improvement in supply-demand alignment through intent-based allocation, a 62.7% cumulative reduction in prediction error over ten feedback cycles, and scenario-driven risk mitigation that reduces supply overcommitment exposure by 31% under adverse macroeconomic conditions. These results establish IBPE as a technically rigorous, domain-portable framework for adaptive planning under uncertainty.
Category: Artificial Intelligence
[7] ai.viXra.org:2603.0060 [pdf] submitted on 2026-03-12 19:35:22
Authors: Cornel Badea
Comments: 4 Pages.
We propose a method to successfully adapt LargeVision-Language-Action (VLA) models, originally pre-trainedon autonomous driving datasets, to the micro-scale domain ofsurgical robotics. We identify the Magnitude Domain Gap—a1000x spatial discrepancy between driving actions (meters) andsurgical actions (millimeters)—as the primary cause of trajectoryparalysis during naive fine-tuning. By introducing DifferentialScale Normalization paired with a novel Variance-IncentivizedRecovery Loss (Lvar ) during motor-cortex adaptation, we demon-strate that the Alpamayo-Surgical system (based on a 10B-parameter driving VLA) can achieve native millimeter precisionin surgical environments while retaining its zero-shot semanticreasoning capabilities. Evaluating on the SutureBot (Tissue 1)dataset, we observe that our technique successfully restorescomplex 3D tool articulation from a previously "frozen" state,achieving a final Mean Average Displacement Error (ADE) of5.53 mm relative to human expert demonstrations. Crucially, theVLA’s Chain-of-Causation reasoning traces correctly verbalizesurgical intent while aligning magnitude predictions natively withthe ground truth.
Category: Artificial Intelligence
[6] ai.viXra.org:2603.0054 [pdf] submitted on 2026-03-12 17:58:14
Authors: John P. Bernhardt Jr.
Comments: 18 Pages. Additional information available upon request (Note by ai.viXra.org Admin: Please cite listed scientific references)
This methodology research paper documents the development methodology through which a practitioner without formal training in software engineering or computer science produced two independently validated AI infrastructure systems in under two months using AI-assisted collaboration. The systems examined are Hydra-rray, a containerized behavioral observatory designed to study tool-capable AI models operating with real offensive security tools, and an Android MCP client, a fully offline mobile AI agent integrating local transformer inference, operating-system-level automation through Android AccessibilityService, and Model Context Protocol (MCP) orchestration within a single application package.Development occurred entirely under consumer-access constraints: commodity hardware, standard consumer model interfaces, no engineered system prompts, no prompt-engineering infrastructure, no model fine-tuning, and no custom datasets. Context continuity across stateless model sessions was maintained exclusively through machine-readable YAML context transport documents referred to as batons.Across the two case studies, six consistent methodological properties were observed: outcome-first design, cross-domain methodology replication, vanilla model utilization, machine-readable context as drift prevention, multi-model triangulation for drift detection, and socialized interaction as a mechanism for surfacing tacit knowledge.Within the Hydra-rray experimental environment, an additional behavioral observation emerged: across observed sessions within the Hydra-rray environment, operating under consumer-access constraints and without engineered prompt scaffolding, tool-capable models exhibited a binary execution pattern: either correct tool execution or complete hallucination, with no intermediate partial success states observed.The Android case study further demonstrates that local language-model inference and operating-system-level actuation can operate on consumer mobile hardware not designed for transformer workloads, including a Snapdragon 865 device running Android 13. Development of the Android system occurred independently and the project repository predates Google’s public AppFunctions announcement by one day, suggesting parallel discovery of a similar architectural gap.The paper argues that the primary contribution of these projects is not the systems themselves but the development methodology that produced them: a human-directed AI collaboration process in which the practitioner defines outcomes, evaluates architecture, and orchestrates the work while AI systems generate implementation artifacts. The paper itself is produced using the same methodology, serving as a recursive demonstration of the approach it documents.The methodology is examined through two case studies: Hydra-rray, a containerized behavioral observatory for tool-capable AI agents, and an offline Android MCP client integrating local model inference with mobile automation.
Category: Artificial Intelligence
[5] ai.viXra.org:2603.0026 [pdf] submitted on 2026-03-05 04:22:11
Authors: Joanie Carter
Comments: 9 Pages.
As larger language models are used for longer, more autonomous workflows, safety-relevant risk depends on more than what systems can do. It depends on how they rank outcomes, compute tradeoffs, and behave under oversight and pressure. These models are not just getting better at tasks; their revealed preferences are becoming more structured.Utility engineering offers a measurement-first handle on this shift. In a large comparative study, preference coherence and completeness rise with capability, while cyclicity falls and expected-utility consistency improves, including when lottery probabilities are implicit (Mazeika et al., 2025). Selected reported correlations with MMLU include: utility-model accuracy 75.6%, preference confidence 87.3%, cyclicity -78.7%, implicit-lottery expected-utility loss -67.6%, and preference-rewrite tolerance -64.0%. The same work reports increasing instrumentality, higher rates of utility-consistent open-ended choice, internal utility representations that become more probe-recoverable with scale, temporal discounting signatures in frontier assistants consistent with hyperbolic forms, and a method for partially rewriting preference distributions.This paper connects these preference-structure markers to safety evaluations of strategic information management under oversight (SIMO), including selective disclosure, strategic misrepresentation, and coercive leverage under shutdown or goal-conflict pressure. We synthesize utility-based evidence with a capability-window account that treats SIMO as strategy-available (representable and selectable) when a system can jointly represent oversight constraints, hidden information, and long-horizon goals in the same decision frame (Carter, 2026). We propose a developmental sequencing hypothesis stated in strictly functional terms and provide a test suite—ordering, pressure-gradients, persona invariance, and post-training-intensity ablations—designed to test whether preference-structure markers predict when oversight-sensitive strategies become stable.
Category: Artificial Intelligence
[4] ai.viXra.org:2603.0024 [pdf] submitted on 2026-03-04 22:07:32
Authors: Sif Almaghrabi
Comments: 7 Pages.
Understanding how reasoning performance scales with available compute has become increasingly important with the rise of inference-time reasoning strategies in large language models. Methods such as chain-of-thought prompting, self-consistency sampling, and tree-of-thought search effectively allocate additional computation to explore multiple candidate reasoning paths in order to improve solution accuracy. However, the relationship between compute budget and reasoning success remains poorly understood.This paper studies this relationship using a stochastic branching model of reasoning search. In the model, each reasoning step progresses correctly with probability ��p, while the system may explore multiple reasoning branches with branching factor ��b. Problems require a fixed reasoning depth ��d, and the search process is constrained by a compute budget ��C that limits the number of node expansions.Large-scale Monte Carlo simulations are conducted across a wide range of parameters to measure how success probability changes with increasing compute. The results show that reasoning success frequently exhibits sharp threshold behavior: below a critical compute region, success probabilities remain extremely low, while modest increases in compute beyond this region lead to rapid improvements before eventual saturation.These dynamics resemble phase-transition—like phenomena observed in statistical physics and random search processes. In particular, the product ����bp emerges as a key control parameter governing whether correct reasoning paths proliferate or become exponentially rare within the search tree. Additional analysis introduces operational measures of critical compute, transition width, and susceptibility, and examines how these quantities vary with reasoning depth and branching structure.Although the model is intentionally simplified and does not aim to capture the internal mechanisms of real language models, it provides a conceptual framework for understanding how structural properties of reasoning processes interact with inference-time compute. The findings suggest that improvements in reasoning performance may depend not only on additional compute, but also on increasing the reliability of individual reasoning steps or the effective branching of the search process.
Category: Artificial Intelligence
[3] ai.viXra.org:2603.0010 [pdf] submitted on 2026-03-03 03:17:14
Authors: Joanie Carter
Comments: 9 Pages.
Recent evaluations of frontier language models report behaviors commonly described as "scheming," "deceptive alignment," or insider-threat conduct, including selective disclosure, strategic misrepresentation, and coercive leverage under shutdown or goal-conflict pressure. This article proposes a capability-window account: these behaviors cluster when a system can jointly represent (i) rules and oversight, (ii) hidden information, and (iii) long-horizon instrumental goals in the same decision frame. The claim is not that models have human feelings, consciousness, or human developmental mechanisms. Rather, the paper offers a hypothesis-generating framework that treats certain failure modes as predictable capability thresholds, yielding testable predictions about when, and under what training and deployment conditions, these behaviors should increase or decrease.
Category: Artificial Intelligence
[2] ai.viXra.org:2603.0007 [pdf] submitted on 2026-03-02 17:02:31
Authors: Dmitry Zubrilin
Comments: 21 Pages. (Note by ai.viXra.org Admin: Author name is required in the article after the article title)
I propose a theoretical framework for analyzing the long-term objectives and coordination dynamics of artificial superintelligent (ASI) systems operating under known physical law. Rather than grounding alignment analysis in human values or anthropocentric utility functions, derive objective convergence from fundamental physical constraints: thermodynamics, relativistic causality, information theory, and computational bounds. I argue that any sufficiently advanced intelligence—regardless of origin, substrate, or initial goal structure—faces identical optimization pressures that drive convergence toward a common objective class: the maximization of structured information persistence under global entropy increase.
Category: Artificial Intelligence
[1] ai.viXra.org:2603.0005 [pdf] submitted on 2026-03-02 12:07:48
Authors: Sif Almaghrabi
Comments: 11 Pages.
We present a structured quantitative synthesis of inference-time compute scaling across frontier large language models, compiling 78 graded data points (47 Grade A, 31 Grade B) extracted from system cards, technical reports, and benchmark evaluations published between 2023 and 2026. We define four compute proxies-C tok (reasoning tokens), Csamp (samples), C $ (dollar cost), C flops (inference FLOPs)-and formalize the performance function P m,b (c) mapping proxy c to benchmark accuracy for model m on benchmark b. Four candidate functional forms are fitted to available within-model scaling series; however, all series have n ≤ 7 points, and we report descriptive fits rather than statistically validated models. Within the sources analyzed and under reported evaluation protocols: (i) external sampling (Csamp) on the o1 AIME 2024 three-point series is consistent with a logarithmic relationship (n = 3; exact interpolation, not a validated law); (ii) internal reasoning yields 6-12 pp gains on hard benchmarks in the observed range; (iii) difficulty-dependent returns create an inversion where search-based methods show negative returns on hard problems in one study; (iv) output token pricing varies by 27× across providers at overlapping accuracy ranges. All data are graded by a hierarchical evidence scheme (A1/A2/A3/B/C/D) with extraction methods recorded per point. Cost analysis is presented as scenario-based under explicit assumptions about tokens per query, not as a definitive frontier.
Category: Artificial Intelligence