Congratulations to the authors of the following accepted papers. Accepted papers are non-archival and will be discussed in themed panels led by Program Committee members.
Long Papers
-
Long Paper
A course on responsible digital developments: a teaching experience report
Daphne TUNCER (Ecole nationale des ponts et chaussees, Institut Polytechnique de Paris)
Abstract
The paper presents a course that was developed in 2021 and that has been delivered since then on responsible digital developments. The course is built around the notion of responsibility. It aims to help students practice their critical-thinking skills with respect to the implications and consequences of digital developments. It consists of a taught lecture and a tutorial activity, and is driven by real-life examples taken from different sources, including the press, research projects and art work. The course is currently implemented in a format of either three or six hours, and is delivered in three university-level programs in the UK and in France.
-
Long Paper
A First-Principles Framework for Networking Education
Jaber Daneshamooz (UCSB), Sanjay Chandrasekaran (UCSB), Arpit Gupta (UCSB)
Abstract
Generative AI can now do many of the production tasks that traditional networking courses test, producing correct-sounding TCP walkthroughs or CSMA/CA traces in seconds. What remains difficult to automate is claim evaluation: recognizing when a confident-sounding claim about an unfamiliar system is wrong. We evaluate a first-principles framework designed to teach this skill. The framework consists of three components: four invariants (State, Time, Coordination, and Interface), the Environment-Measurement-Belief (E-M-B) decomposition, and binding-constraint cascade reasoning. We deployed the framework as the foundation of an upper-division networking course and surveyed students at the end of the term. Of the 54 students who consented to research use, 53 had previously completed the protocol-survey prerequisite at the same institution, allowing direct comparison between the two courses. Students reported substantially deeper understanding of why systems are designed as they are, which was the largest effect observed in the survey. They also reported being better able to analyze systems they had never encountered before. Most notably, more than a quarter described a specific instance in which they identified a plausible-but-incorrect claim, with several independently rejecting the same WiFi PHY-rate claim by citing fixed per-frame overhead. We also identify a limitation of the framework: students found it easier to decompose an existing system than to predict an architecture from a constraint, suggesting a direction for future refinement.
-
Long Paper
A Lightweight JupyterHub Testbed on k3s for AI and Networking Education in Low-Resource settings: Experience from Senegal
Faly GAYE (UCAD), Bamba GUEYE (UCAD), Ibrahima DIANE (UCAD), Damien SAUCEZ (INRIA), Nikos Makris (University of Thessaly & CERTH)
Abstract
Limited infrastructure and equipment hinder hands-on AI and networking education in resource-constrained environments. In this paper, we present a lightweight, sustainable testbed based on 𝐽𝑢𝑝𝑦𝑡𝑒𝑟𝐻𝑢𝑏 deployed on a 𝑘3𝑠-managed edge/cloud continuum, designed to support practical training under low-bandwidth and intermittent connectivity conditions. The underlying 𝐵𝑙𝑢𝑒𝑝𝑟𝑖𝑛𝑡 leverages a heterogeneous, low-cost architecture optimized for smartphone access and minimal bandwidth consumption. Compared to traditional 𝐷𝑜𝑐𝑘𝑒𝑟 -based orchestration, our architecture significantly reduces 𝐶𝑃𝑈 usage by 93 % for 𝐾𝑒𝑦𝑐𝑙𝑜𝑎𝑘, 82 % for 𝑃𝑜𝑠𝑡𝑔𝑟𝑒𝑆𝑄𝐿, and 74 % for 𝑅𝑒𝑑𝑖𝑠 while maintaining reliable performance under constrained connectivity. Furthermore, the stack is replicable by any African university with modest hardware and connects natively to the SLICES-RI identity fabric via 𝐾𝑒𝑦𝑐𝑙𝑜𝑎𝑘 federation. These results highlight the potential of lightweight, cloud-native infrastructures to democratize access to advanced digital skills and support scalable, context-aware education in underserved regions.
-
Long Paper
Agentic AI for Accessibility Remediation in Learning Management Systems
Ali Mazloum (University of South Carolina), Jose Gomez (Fort Lewis College), Elie Kfoury (University of South Carolina), Jorge Crichigno (University of South Carolina)
Abstract
Networking instructors increasingly rely on Learning Management Systems (LMSs) to distribute slide decks, PDFs, lab handouts, and other course materials, yet remediating those artifacts for accessibility remains labor-intensive. This paper presents an agentic AI system that improves LMS-hosted networking courses through a four-phase workflow: course navigation, Ally-based assessment, document remediation, and re-upload. The design combines parallel document processing for throughput with a three-agent validation pattern for reliability. In an initial evaluation on three networking courses, the system improved Blackboard Ally scores from 73% to 95%, from 82% to 96%, and from 66% to 89%. On a representative course, it collected 92 documents, processed 513 candidate images, used up to 20 concurrent subordinate agents, and completed the semantic remediation stage in about 28 minutes, corresponding to an estimated 20x speedup over a single-worker baseline. The validator accepted 82 of 83 documents without manual intervention, while a byte- equality shortcut bypassed the validator for about one quarter of cases. These results suggest that agentic AI can provide practical instructional support for networking education, while exposing a central tradeoff: higher throughput and higher confidence require additional agent invocations and therefore higher operational cost.
-
Long Paper
An Agentic AI Approach for Hands-on Networking Education
Akhil Gorthi Bala Sai (University of South Carolina), Ali Mazloum (University of South Carolina), Jose Gomez (Fort Lewis College), Samia Choueiri (University of South Carolina), Amith GSPN (University of South Carolina), Elie Kfoury (University of South Carolina), Jorge Crichigno (University of South Carolina)
Abstract
Hands-on networking education relies on laboratories where learners configure networked systems, observe behavior, and debug failures. Virtual labs make this experience scalable, but support remains difficult because useful guidance often requires knowing both the intended procedure and the current state of a learner's environment. A lab manual describes the topology, setup steps, commands, and expected outputs, but it cannot inspect whether a learner has missed a step, configured the wrong interface, or produced an inconsistent routing state. General-purpose Large Language Models (LLMs) can explain networking concepts, but they are not grounded in a specific lab manual or in the live state of a running lab instance. They may also provide final commands too early, reducing the opportunity for learners to reason through the debugging process.
This paper presents an agentic AI approach for hands-on networking education. The system is developed around an IPv6 configuration and routing training sequence, with a static-routing lab used as the running example, while the design targets virtual networking labs more broadly. The agent builds a lab-specific knowledge base from manuals and structured notes, uses Retrieval-Augmented Generation (RAG) to retrieve procedural and diagnostic context, and uses read-only live-state tools to inspect the learner's lab instance when an answer depends on current state. A human-audited evaluation over 100 lab-specific questions compares three conditions: LLM-only, document-only RAG, and the full agentic system. The full system answers 91 questions correctly and 9 partially, compared with 79 correct and 9 partial answers for document-only RAG and 34 correct and 17 partial answers for LLM-only. The results suggest that agentic AI can offer a promising direction for scalable support in hands-on networking education.
-
Long Paper
An AI Orchestrator Model for Troubleshooting Skills Training in Network Engineering
Kapil Mehta (Cisco Systems), Prashant Sumanprasad Bhadoria (Cisco Systems), Jaypal Baviskar (Cisco Systems)
Abstract
When a network goes down or resiliencies are at stake, the pressure to restore the network to a business-as-usual state is immense. In that moment, engineers often fall into the trap of “solution-jumping” relying on quick fixes, such as reloads or using Generative AI to provide a quick fix rather than digging for the actual root cause. While these one-shot solutions might bring the network back to its desired state, they often fail to solve the underlying problem, which leads to repeat issues and a steady decline in customer trust. Our method looks at how we can better prepare network engineers for these high-pressure moments. Instead of relying just on traditional, theory-heavy training or only expertled trainings, we’ve blended a framework that focuses on cognitive conditioning.we propose the KTO-AI methodology (Kepner-Tregoe, Topology awareness, and OSI-layer mapping) with the 4E’s model (Education, Experience, Exposure, and Environment) in a simulated environment, we’re shifting the focus from speed to accuracy. The core of our approach is a mandatory, rigorous “Assess and Acquire” phase that must happen before any “Act” phase is permitted. We’ve validated this methodology with experts in the network troubleshooting domain, and the results were clear: when engineers receive real-time customer satisfaction scores (CSAT) and Question Credit feedback for each reasoning step, they stop guessing and start solving. To scale this cognitive conditioning, we propose an AI Agent Orchestrator as a design model to provide Experience, Exposure, and Environment. The orchestrator is intended to automate repetitive training functions: presenting failure scenarios, tracking Question Credit, recording CSAT trends, enforcing Assess/Acquire phase discipline, and providing structured feedback. This enables organizations to run continuous, personalized troubleshooting skills training at scale. We believe that by providing these safe spaces for practice regularly through AI orchestration, organizations can build the technical and psychological resilience their teams need to handle critical incidents effectively, ultimately leading to a more stable network and superior customer experience.
-
Long Paper
Bringing Research Infrastructure into the Networking Classroom at a Public Teaching College
Ilknur Aydin (Farmingdale State College), Fraida Fund (New York University)
Abstract
Public research testbeds make sophisticated networking and cloud computing capabilities available to educators and students at non-R1 and teaching-focused institutions without requiring those institutions to own or maintain the infrastructure. However, most reports on educational use of networking testbeds come from research-intensive institutions. This experience report instead examines adoption in a required undergraduate computer networking course at Anonymous Teaching College, a public teaching-focused institution where many students commute, work while enrolled, and have limited prior exposure to Linux command-line tools or research infrastructure. Since Fall 2022, one instructor has used public testbed-based labs across 16 course sections serving 373 students. We describe this experience and its outcomes, with a focus on the conditions that made adoption in this setting feasible: mentoring, reusable lab materials, instructor-created scaffolding, detailed student-facing instructions, and platform choices that reduce unproductive friction.
-
Long Paper
Crawl, Walk, Run: A Scaffolded Framework for AI Integration in Computing Education
Ramakrishnan Durairajan (University of Oregon and Link Oregon)
Abstract
AI is rapidly transforming computing practice, creating new challenges for educators across computing, systems, and networking disciplines. While unrestricted AI use can undermine conceptual learning, prohibiting AI risks disconnecting coursework from modern professional practice.
This experience report presents a scaffolded framework for generative AI integration in computing education that balances foundational skill development with AI-assisted learning. With this framework, students progress through three stages: foundational skill development without AI assistance, structured AI-assisted pair programming, and AI-enabled project development where AI serves as a productivity multiplier. We describe the design rationale, implementation, and lessons learned from a pilot deployment of this framework. Our experience suggests that integrating AI into project-based coursework can help students develop critical awareness of AI’s limitations and the professional habits needed to evaluate AI-generated code. However, students frequently reported tension between the speed AI affords and the depth of understanding that coursework is designed to build.
-
Long Paper
Exploring Digital Twin Classroom: Extended Reality for Immersive Learning
Chaowei Wang (Beijing University of Posts and Telecommunications), Yunze Zhang (Beijing University of Posts and Telecommunications), Wenyuan Wang (Beijing University of Posts and Telecommunications), Lexi Xu (Research Institute, China United Network Communications Corporation), Fan Jiang (Xi'an University of Posts and Telecommunications)
Abstract
The development of new technologies has driven the transformation of teaching models. Extended Reality (XR) technologies, which bridge the gap between the virtual and the real world, have further evolved to incorporate digital twin and semantic technology within the immersive learning framework, creating innovative approaches for higher education. This paper formulates a Semantic-enhanced Immersive Digital Twin (SIDT) technology framework, introducing new, knowledge-driven concepts for teaching and curriculum design. The constructed system features high-fidelity interaction, cognitively-aware personalized learning, and remote distributed collaboration, encompassing the holistic design and optimization of educational objectives, teaching methods, learning environments, and instructional activities. In the teaching practice of the course \textit{Fundamentals of Wireless Communication Technology}, this model was applied to classroom instruction with the integration of innovative assessment methods to evaluate student learning outcomes. A comparative analysis between traditional teaching and digital twin–based immersive learning environments was conducted. The study provides exploratory practice and valuable insights into the application of digital twin technology in higher education.
-
Long Paper
Flipped Team-Based Learning and Labs for a Course on Computer Networking
Thomas Zinner (NTNU), Katrien De Moor (NTNU), Stanislav Lange (NTNU), David Palma (Takinobori), Trond Vatten (NTNU)
Abstract
Introductory computer networking courses must help students move beyond memorizing protocols toward reasoning about packet-level behavior, configuration, and debugging. This experience report presents a flipped Team-Based Learning design that integrates weekly preparation, readiness assurance tests, team activities, and reproducible hands-on labs in a 13-week undergraduate networking course. We describe the course workflow, assessment model, and support structure, including an oral exam for individual accountability. Drawing on student feedback and teaching observations, we identify perceived strengths, operational trade-offs, and lessons for scaling active, lab-centered networking education in mid-sized cohorts.
-
Long Paper
Forty Years of Teaching Computer Networks: A Retrospective
Shivendra Panwar (New York University), Fraida Fund (New York University)
Abstract
We present a forty-year retrospective on teaching computer networks, focusing on a graduate course in an Electrical and Computer Engineering department that evolved from a lecture-based course into a lab-intensive course centered on hands-on experimentation. Over this period, the Internet became the substrate for everyday computing, cloud services, mobile systems, security, and AI infrastructure. We describe how the course content has evolved, and why many core networking ideas have remained stable despite changes in the field. We then examine how course infrastructure changed and how the delivery of the course has improved as a result. Finally, we discuss how the student population, student motivation, and student engagement may have shifted as networking has become a core systems area rather than a specialized frontier. We close with reflections on the role of networking education in the generative AI era.
-
Long Paper
From Module Consumers to Module Producers: A Mentored Summer Cohort for AI-Generation, Data-Driven Cybersecurity Education
kc Claffy (CAIDA/UC San Diego), Ricky K. P. Mok (CAIDA/UC San Diego), Alexander Marder (Johns Hopkins University), Bradley Huffaker (CAIDA/UC San Diego), Rocky K. C. Chang (Calvin University)
Abstract
For students, hands-on exposure to large-scale Internet mea-surement data is rare because the data are too big, too sen-sitive, and/or too operationally awkward to fit on a laptop. We describe the design of a ten-week mentored summer hy-brid program that treats course module production as both a research-mentorship experience and a curriculum pipeline. Participants are paired with measurement-community men-tors, and work on national cyberinfrastructure (the National Research Platform and SDSC Expanse) with real datasets (BGP, active topology, network telescope traffic). Our am-bitious goal is that each participant contributes to design-ing, implementing, and testing open-source teaching mod-ules that use real Internet measurement data, and gains first-hand experience of the AI-era data-science workflow we want future students to learn. The ten-week program ends with an in-person hackathon where modules are integrated, tested, and peer-reviewed before release as open educational resources. We position the design within a central tension of CS education—integrating AI so that it deepens rather than shortcuts learning—and report here as a work-in-progress. We describe the program’s motivation, the intended mod-ule catalog, the assessment infrastructure that makes the modules adoptable at scale, and invite discussion on how to maximize utility of the results for educational use.
-
Long Paper
From Protocols-to-Prompts (P2P): Mapping the Agentification of Open Networking Education
Venkat Sai Suman Lamba Karanam (Bowling Green State University), Zahmeeth Sakkaff (West Virginia Institute of Technology), Shubham Sundriyal (Bowling Green State University)
Abstract
AI coding assistants and agentic software workflows since late-2022 have transformed how networking is taught, practiced, and learned. Formal curricula, however, have yet to address this transition. In this paper, we present a large-scale measurement study using 1,318 GitHub repositories to gauge how open networking education is evolving in the generative-AI era. The mined 1,318 GitHub repositories span the 2009–2026 period and are scored each along three axes, each measuring a pedagogical signal: (1) Conceptual Depth (CD) for protocol-level vocabulary density; (2) Operational Abstraction (OA) for orchestration and API-layer presence; and (3) Agent/Tool Dependence (ATD) for explicit AI framework instrumentation. We compare pre- and post-generative-AI artifacts using a November 2022 cutoff — at the dawn of ChatGPT release.
Using extensive analysis we report the following findings. Within 208 repositories spanning the cutoff, we show a significant decline in mechanism-centric signals (ΔCD = −0.085, p < 0.001) — with a concurrent and consistent rise in orchestration-layer indicators (ΔOA = +0.063, p = 0.008). Interestingly, ATD scores remain uniformly negligible across the entire corpus (mean ≈ 0.002) — a key finding that indicates that agentification is a shift in the tooling substrate rather than of explicit AI instrumentation.
Additionally, our complementary analyses show that Docker adoption increased from 9.6% to 14.3% post-cutoff, GNS3 displaces Mininet as the dominant emulation tool (1.2% → 4.3%). From a problem-structure/activity perspective, we report that Parsons-style problems double in prevalence (1.2% → 2.5%). Networking and computing testbeds, specifically the NSF PAWR testbeds of FABRIC, COSMOS, POWDER, NRP remain largely absent from open educational repositories in both pre- and post-cutoff epochs. We manually validate our repository-level case studies across five spanning repositories to confirm that these aggregate findings (over 1,318 repos) match at the per-repository level. Finally, we present a taxonomy of pedagogical regimes and concrete recommendations for AI-aware networking curricula with the goal of preserving mechanistic understanding while embracing AI contemporary tooling.
Reproducibility Commitment: All code, data, manual validation sample, and artifacts will be released upon publication.
-
Long Paper
Ludwig: Autonomous Multilingual Synthesis of Correct, Engaging Explainer Videos
Alexander Krentsel (UC Berkeley)
Abstract
Building the right mental model is central to teaching and understanding systems. Well-animated explainer videos help build these mental models effectively, but are expensive and time-consuming to produce, requiring pedagogical design, scripting, animation, narration, and review work. We argue that recent capability shifts enable multi-modal models to effectively produce such content by autonomously authoring animation code from a style guide, auditing rendered frames for defects, and applying multilingual neural text-to-speech.
We prototype an agent-driven system, Ludwig, structured around a highly capable producer agent that plans, narrates, and animates, and an independent vision-model reviewer that closes the loop by auditing the rendered output. Run for two months as a public channel, Ludwig has produced 150 videos across five languages with over 45,000 views at $2 to $10 and about 25 minutes of wall-clock time each, orders of magnitude below a human-authored equivalent. We discuss the implications of such a system for systems education. All videos are publicly available on the live channel at www.youtube.com/@ludwigexplains.
-
Long Paper
Networking Unplugged: Hands-On, Inquiry-Based Learning for Undergraduate Networking Courses
Aaron Gember-Jacobson (Colgate University)
Abstract
This paper presents a collection of hands-on, inquiry-based learning activities for introducing undergraduate students to important concepts in networking without using computers. Our _networking unplugged_ activities–inspired by the Computer Science (CS) Unplugged"approach of using kinesthetic-based puzzles and games to introduce CS concepts–frame important concepts in networking as physical communication tasks between humans. By brainstorming and experimenting with different ways of completing the tasks, students "discover" and compare important networking algorithms/protocols. Over the past decade, we have used these activities multiple times at different levels of the curriculum at a primarily undergraduate institution. Based on our informal observations, the activities have a positive impact on student engagement and understanding.
-
Long Paper
Open Source Networking Textbooks: A Clean Slate
Bruce Davie (Systems Approach), Larry Peterson (Systems Approach)
Abstract
After three decades of producing networking textbooks, we observe that the Internet has changed enough to warrant a fresh approach to educating the next generation of students and practitioners. We propose a clean-slate refactoring of our networking textbook that aims to give students the necessary perspective to form a big-picture view of networking and the skills to deal with the complexity of modern networks. We treat networking as a set of system design problems and draw on existing artifacts such as Ethernet, HTTP and IP to illustrate how these challenges have been solved in the past, while aiming to equip readers with the perspective and systems thinking skills to shape the networks of the future.
-
Long Paper
Teaching Computer Networks by Building Networks: An Experiential Learning Approach Using Docker
Ram Rustagi (UMBC)
Abstract
Most Computer Networking curricula at the undergraduate level aim to familiarize students with key networking tech-nologies and standards traditionally taught through a com-bination of lectures, analytical exercises, simulations, and packet trace analysis. These approaches help students deelop a basic understanding of inherently complex yet fundamen-tal concepts, such as, complexities of web communication, reliable delivery and packet forwarding in IP networks etc. Many students struggle to connect theoretical models with actual protocol behavior observed in real systems and often unable to translate their limited understanding into diagno-sis and action when troubleshooting networks. This paper presents a container-based experiential learning approach based on a four-stage pedagogical cycle: Prediction, Imple-mentation, Observation, and Explanation (PIOE). The frame-work uses Docker-based networking laboratory, in which routers, switches and end hosts are instantiated as docker containers. It enables students to experimentally validate protocol behavior across multiple layers of the networking stack. With the increasing adoption of generative AI in edu-cation, the framework emphasizes experiential learning by requiring students to relate conceptual understanding to the observed behavior of actual network implementations.
We illustrate the framework using representative assign-ments covering Network Delay, TCP Flow Control, TCP Fairness, Longest Prefix Match based forwarding, IPv6 de-ployment and Layer-2 switching. Preliminary classroom ob-servations and feedback collected from our students confirm that these real time experiential activities improve student engagement and facilitate deeper understanding of network-ing concepts.
-
Long Paper
Teaching Wireless Networking in the Age of Softwareization and AI
Serge Fdida (Sorbonne Université), James Kurose (University of Massachusetts), Akihiro NAKAO (The University of Tokyo), Larry Peterson (Systems Approach)
Abstract
IIn this paper, we describe the learning goals, course topics, and teaching experiences associated with a course on the principles and practice of modern wireless networks at the advanced undergraduate or master’s level, including the open educational resources we have developed for such a course. We emphasize 5G and Wi-Fi networks, but also include Bluetooth, IoT, and LEO satellite networking technologies. Unlike a wireless networking course that focuses on radio communication at the edge, our courses and materials focus on a softwareized (and programmable), full-stack mobile system that includes wireless access networks, as well as core networks, network control, management, and orchestration. While the radio access network (including the physical radio layer) is a critical part of any wireless network course, we also emphasize the levels of abstraction at which students can understand the radio channel and learn how it is programmed and managed, notably through concepts such as Radio Resource Blocks. We also address teaching challenges and opportunities in the age of AI that extend beyond wireless networking.
-
Long Paper
Think First, Code Later: Enhancing Student Reasoning and Learning in Practical Engineering Education through Guided Generative AI
Musab Ahmed (Hamburg University of Technology), Koojana Kuladinithi (Hamburg University of Technology), Aliyu Makama (Hamburg University of Technology), Frank Laue (Hamburg University of Technology), Jonas Bozenhard (Hamburg University of Technology), Maximilian Kiener (Hamburg University of Technology), Andreas Timm-Giel (Hamburg University of Technology, Germany)
Abstract
The use of Artificial Intelligence (AI) in education has triggered discussions ranging from redesigning courses around AI to restricting its use entirely. However, before fundamentally restructuring curricula, it is essential to establish how students can be trained to use AI effectively within existing courses to strengthen reasoning and critical thinking skills. This is particularly important in practical engineering education, where simulation plays a central role in enabling students to connect theoretical models with observable system behavior through experimentation and analysis. In such courses, generative AI offers new opportunities for personalized, dialogue-based support on demand, but its value depends on how it is pedagogically integrated. Our primary objective is to shift students from answer-seeking code generation toward a reasoning-first learning approach. To this end, this paper evaluates a Think-First workflow that restructures simulation exercises to emphasize problem formulation, result prediction, and pseudocode development prior to implementation. We examine students’ AI interaction patterns, simulation reasoning, and perceptions of the workflow in a master’s course on Simulation of Communication Networks.
-
Long Paper
Transport It Your Way: Rethinking Networking Projects to Emphasize Design
Alexander Krentsel (UC Berkeley), Tess Despres (UC Berkeley), Sylvia Ratnasamy (UC Berkeley)
Abstract
AI is quickly shifting what it means to be a computer scientist in networking. As coding agents become more capable, the day-to-day practices of building systems are quickly evolving. These shifts force us to reexamine how we teach networking, raising many open questions including: should syntax still be emphasized, how should we scaffold student coding assignments, and how should we assess what students have learned from artifacts? With these questions in mind, we focus on revamping an upper division networking and systems class project in this proposal. The previous project, which centers implementation, can now be completed end-to-end by a coding agent. To mitigate the impact of this reality, we propose a redesigned project that shifts effort towards system design and tradeoff analysis, while granting students the autonomy to use modern coding tools for their implementations. We also discuss how this approach could generalize to other projects in upper division systems classes.
-
Long Paper
Trust, but Verify: AI Interventions in Computer Networking Education
Tanya Shreedhar (TU Delft), Christoph Lofi (TU Delft)
Abstract
Generative AI assistants are now part of how undergraduates write and debug code, and computing education has moved from asking whether to allow them to asking how to teach with them. On networking tasks, those whose answers turn on the behavior of a running network the student can measure, these assistants are often fluent and wrong, and a beginner has no reliable way to tell. We call this failure mode vibe networking, accepting an AI artifact because it looks plausible, without checking whether it behaves as claimed. The networking course already holds the antidote, because its core method is measurement, the discipline’s way of settling a claim against the system itself. We propose that the course give a generative model three roles across the term, shifting as the student’s competence grows. At the start it is an adversary the student learns to catch by redoing its computation by hand, in the middle an assistant whose proposed experiments the student runs and measures on the live network, and at the end an examinee whose diagnosis the student confirms or refutes from a capture they take themselves. Measurement is the constant across the three roles: an observation the course produces in its normal run decides whether each AI answer holds, and the grade rewards the student’s check, never the model itself. We work out the three roles and design one classroom intervention for each, sized for a 9-week BSc course of around 500 students at TU Delft.