AI-enabled systems classified as high-risk under the EU AI Act (AIA) are required to comply with a comprehensive set of system-level requirements that go beyond a mere AI-model-centric perspective. AI-enabled Cyber-Physical Systems (AI-CPS) are often concurrent and distributed, making it more challenging to achieve reliable, reproducible, and testable behavior as required by the AIA. We show how Lingua Franca (LF), a polyglot coordination language, facilitates compliance with selected AIA requirements by operationalizing them as architectural primitives, making them first-class concerns at the architectural level. Using an AI-enabled traffic light control system as an example, we demonstrate how LF contributes to the required technical documentation by automatically generating a visualization of the system architecture that shows how the components interface with each other. LF also enables testable and verifiable levels of system robustness via its deterministic nature and explicit timing semantics. It enables (1) deterministic design under clear assumptions, (2) detectable incorrect behavior and appropriate handling thereof, (3) precise timing assurances/guarantees, and (4) a testable system architecture. Capabilities (1)–(3) also enable the required anomaly monitoring as part of the human oversight requirement. Finally, integrating the visualization tool Rerun with LF enables real-time tracing of the system behavior and post-hoc log analysis, which aligns with the AIA record-keeping/traceability requirement. LF also ensures in-order timestamps for causally related events, addressing a significant challenge for logging and tracing in distributed systems. In conclusion, this study provides practical evidence, through an AI-CPS demonstrator, that LF’s coordination semantics facilitate regulatory compliance with selected AIA requirements regarding technical documentation, robustness, testability, anomaly monitoring, and traceability.
Regulation and industry practice have produced a sizeable inventory of safeguards for AI systems: humans placed in decision loops, oversight obligations, documentation duties, explanations, disclaimers. Whether the items in this inventory do what they are supposed to do is a separate question, and one that is easy to leave unasked, because a safeguard’s presence is far cheaper to verify than its effect. The three contributions to the Responsible and Trusted AI track at AISoLA 2026 examine that gap at three different levels of an AI system: the causal structure of the decision process into which a human is placed, the coordination layer of a distributed AI-enabled system, and the model and the hardware it runs on. Each contribution identifies a safeguard that can be satisfied nominally while failing in substance, and each responds by relocating the demand into structure that is fixed before deployment rather than delegating it to a person at the point of use. This introduction sets out the shared problem, situates the four papers within it, and marks the tensions between them that we expect the on-site discussion to take up.
This is an introduction to the proceedings for the track ‘AI Assisted Programming’ (AIAP), organized at the fourth instance of the AISoLA conference during the period October 27–31, 2026. AISoLA as a whole aims to study opportunities and risks of late advances of AI. The motivation behind the AIAP track in particular is the emerging use of large language models for the construction and analysis of software artifacts. An overview of the track presentations is provided.
Model-Based Systems Engineering (MBSE) promises structured, traceable development of complex systems, but high modeling effort remains a key adoption barrier. The textual notation of SysML v2 turns model construction into a text-to-text generation problem, opening new avenues for large language model (LLM) automation. This paper presents AMADEUS (Agentic Model-Aided Development Engineering Utilizing SysML), a conceptual framework for integrating agentic AI into systems engineering workflows in automated driving. The AMADEUS framework comprises three contributions: (1) a work package concept that operationalizes a domain-specific Systems Engineering methodology into discrete units executable by agents and maps each section to orchestration patterns such as actor-critic validation and role-based routing, (2) two complementary architecture patterns, typed operations exposed by a dedicated model server and direct SysML v2 text editing in a Git-based workflow, with a comparative analysis, and (3) an empirical benchmark evaluating three late-2025 frontier AI agents (OpenAI Codex, Google Gemini 3 Pro, Anthropic Claude Sonnet 4.5) against a human expert reference on seven SysML v2 modeling tasks of varying complexity. The evaluation provides initial evidence that AI agents achieve comparable quality (average score 3.69 vs. 3.86 on a 0–4 scale) while reducing modeling time by a factor of approximately 12 and per-task execution cost by a factor of 3,500. Transformation tasks achieve average S_total≥ 3.6 out of 4, while deduction-heavy tasks reveal remaining gaps, suggesting that the engineer’s role shifts from model creator to model reviewer.
Automatically deriving executable and efficient constraint programming models from natural language descriptions has long been a central goal in optimisation and constraint programming. Large Language Models (LLMs) appear to bring this vision closer, showing promising results in generating models for classical benchmark problems. However, part of this apparent success may stem from data contamination rather than genuine reasoning capabilities, as many standard CP problems are likely represented in the training data of modern LLMs. In this initial investigation, we study the robustness of LLM-based constraint modelling by rephrasing and perturbing a set of 10 constraint problems while preserving their underlying structure. The modifications alter contextual elements and introduce misleading or distracting information to explore whether models rely on semantic understanding or superficial textual cues. We compare the models generated by three representative LLMs across the original and modified problem descriptions. Our qualitative analysis shows that, although LLMs frequently produce syntactically valid and semantically plausible models, their performance degrades significantly under contextual and linguistic variation, suggesting strong sensitivity to wording and contextual cues.
