Pith. sign in

REVIEW 4 major objections 5 minor 49 references

The paper claims that a single intent-specification contract expresses all 255 data-generation intents mined from 66 published measurement papers, where no existing tool covers more than 13 percent.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-01 00:02 UTC pith:3NSJH27D

load-bearing objection Solid abstraction and a useful intent corpus, but the 100% coverage figure is a fitted score and the artifacts are unreleased, so the headline numbers are promissory. the 4 major comments →

arxiv 2607.26352 v1 pith:3NSJH27D submitted 2026-07-28 cs.NI cs.AIcs.MA

Pramana: A Composable, Domain-Specific Backend for Empirical Networking Research

Jaber Daneshamooz , Eugene Vuong , Alagappan Ramanathan , Manni Moghimi , Haarika Manda , Satyam Kumar , Snithik Thode , Satyandra Guthula
show 6 more authors
Sylee Beltiukov Dongsu Han Tarun Mangla Sangeetha Abdu Jyothi Walter Willinger Arpit Gupta
This is my paper
classification cs.NI cs.AIcs.MA
keywords intent specificationthin waist architectureempirical networking researchdata-generation backendbottleneck emulationexperiment reproducibilitymeasurement intent corpusAI-assisted research
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper argues that the long lag between a networking hypothesis and the data that tests it can be removed by a single contract, the intent specification, which separates what data is wanted, where to run, and how to produce it. Pramana claims this grammar expresses all 255 in-scope data-generation intents mined from 66 published measurement papers, while no existing tool covers more than 13 percent of them. Its proof-of-concept implementation already covers 34 percent, more than twice the best specialized tool. The payoff, if true, is that a researcher states an experiment in natural language and the backend regenerates condition-labeled measurement data, including for published results.

Core claim

Pramana's central claim is that one domain-specific contract, the intent specification, can sit as a thin waist between research intents and execution substrates. The contract disaggregates an experiment into three independent axes: intent (what data), substrate (where), and mechanism (how). Because every grammar field maps to a declared capability of an existing subsystem, realizability becomes a checkable, poly-time satisfiability test rather than open-ended code generation; the only AI step is compiling natural language into the fixed grammar, and everything below that is deterministic replay. The paper demonstrates the claim by mining 255 endogenous data-generation intents from 66 papers

What carries the argument

The intent specification grammar: a composition contract in which a Foreground of application workflows runs over a Regime of imposed bottleneck conditions, with iterations and a content-hash experiment identity. Its load-bearing property is that every field is drawn from declared capability files, so the compiler can gate each model-produced field, compute coverage exactly, and keep the execution pipeline deterministic below a single compile step. The grammar also lets each sample carry the conditions it actually realized, so fidelity is reported rather than assumed.

Load-bearing premise

The paper assumes the existing subsystems can be composed on one substrate with a single shared, labeled bottleneck, but this composition is never demonstrated end-to-end for the motivating scenario.

What would settle it

Run the motivating scenario through the implementation: a real-time video call and a bulk download contending on one shaped bottleneck, with every sample carrying the conditions it actually realized. If the two flows cannot be made to share the same shaped link, or the realized-condition labels diverge from the requested envelope, the claimed 100-percent design coverage has no operational meaning.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Published measurement results could be regenerated from a single natural-language intent, turning artifact evaluation into a check that the evidence itself reproduces.
  • A researcher's active step in running a new experiment can shrink from weeks of setup to a short compile-and-confirm step, with the measurement running unattended.
  • The gap between design coverage (100 percent) and implemented coverage (34 percent) is presented as 35 shared feature additions, so community contributions amortize across many studies.
  • AI-assisted ideation, which generates hypotheses faster than humans can test them, gets a deterministic data-generation backend to validate against.
  • The capability protocol could become a machine-readable admission contract for shared testbeds that receive machine-generated experiment loads.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If the grammar truly covers a decade of measurement intents, it could serve as a standardized taxonomy for benchmarking any future network data-generation tool, not just Pramana's own implementation.
  • The crux is whether the existing bottleneck shaper, application driver, and kernel substrate can all meet on one shared, labeled bottleneck; the reported exemplar is a static bottleneck with a single transfer, so the motivating real-time-call-plus-download scenario remains the natural next test.
  • A testable extension would apply the same intent/substrate/mechanism split to other empirical systems fields, since nothing in the contract is specific to networking bottleneck links.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper proposes Pramana, a composable, domain-specific backend for empirical networking research, centered on an 'intent specification' — a small DSL that disaggregates an experiment into intent (what data to generate), substrate (where), and mechanism (how). The authors mine a corpus of 255 data-generation intents from 66 measurement papers, claim that the intent-specification grammar expresses all 255 intents (100% design coverage), that no existing specialized tool covers more than 13%, and that a proof-of-concept implementation already realizes 34% (87/255). The only end-to-end demonstration is a CCAnalyzer regeneration. The paper also presents a seven-requirement framework, a capability-based architecture composing NetReplica/NetGent/Mininet, and a roadmap for closing the abstraction–implementation gap.

Significance. If the central coverage claims are valid, Pramana's intent specification would be a substantial community asset: a single contract spanning a decade of networking measurement intents, with a measurable gap between design and implementation. The corpus of 255 hand-mined intents, even standalone, could support future DSL and tool evaluation. The design principle — separating intent from mechanism and using a fixed grammar with a single AI compile step — is well motivated and represents a credible architectural direction. The paper is also honest about its limitations, including the deferred release of artifacts. However, the evidence supporting the headline numbers is currently self-assessed, circular in part, and not independently checkable, so the significance is conditional on the authors' follow-through with releases and additional validation.

major comments (4)
  1. [§4, Fig. 4] The 100% coverage claim is circular. The grammar is 'the union of those axes' observed in the corpus, and the same corpus is used to compute the 100% figure via an unreleased 'coverage classifier' (§4, Experimental setup and §4, Fig. 4). A grammar constructed from the corpus will trivially express the corpus's own intents; this is a fitted value, not evidence of a general contract. The paper even states 'The grammar is the union of those axes and expresses all 255 in-scope experiments from 66 papers.' To make the claim load-bearing, the authors need held-out validation (e.g., a second corpus of papers not used in grammar design), inter-annotator agreement for intent extraction, and a fully specified deterministic coverage rule that an independent party can apply. The deferred release ('We release the procedure, the coverage classifier, and the corpus on acceptance') does not address the
  2. [§4, Fig. 4, baseline comparisons] The comparative numbers for existing tools (best tool ≤13%, Mininet 4%) depend on each tool's 'fixed sub-grammar,' but these sub-grammars are not specified in the paper. Coverage is computed by mapping each tool to a sub-grammar, and if that mapping is too restrictive, the baseline understates the state of the art. A reader cannot audit whether, for example, NetReplica's synthetic flows are correctly represented. The authors should release the exact sub-grammar for each tool (or the machine-readable mapping) alongside the coverage classifier, and ideally have an independent person apply the deterministic rule to a sample of intents to confirm the ≤13% numbers.
  3. [§4, implementation claim] The 34% (87/255) 'realized today' claim is a declared-capability count, not a list of executed experiments. Only the CCAnalyzer exemplar is actually run end-to-end; the paper says 'running all 255 end to end is impractical.' The word 'realizes' in Fig. 4 and the implementation discussion suggests the 87 intents have been executed, but no per-intent status or even a sample of executed experiments is provided. Without a demonstration (or at least a detailed mapping from each of the 87 intents to concrete workflows that have been run), the 34% figure is an estimate of what the implementation could do, not what it does. The authors should either execute a representative sample of the 87 intents and report results, or re-label the claim as 'implementable today.'
  4. [§2–§3, motivating composition not demonstrated] The paper's motivating scenario — a real Google Meet call competing with a bulk BBR download on one shared, labeled bottleneck (Prudentia example, §2) — is never actualized. The architecture (§3) composes NetReplica, NetGent, and Mininet, but the only end-to-end exemplar is CCAnalyzer, whose intent is a static shaped bottleneck with a wget transfer, not a real application workflow contending with dynamic cross-traffic on a shared bottleneck. This leaves unvalidated the load-bearing assumption that these subsystems can be composed on one substrate with a single shared, labeled bottleneck, which is exactly the capability that no existing tool possesses (§2). A concrete demonstration of the Prudentia-style composition — even a small-scale version — is needed before the 100% design coverage can be regarded as more than a static expressiveness claim.
minor comments (5)
  1. [§4 and abstract] The abstract says 'we demonstrate Pramana's utility' for the corpus and coverage results, but the demonstration is currently analytical, not empirical. Suggest using 'report' or 'present' until artifacts are released.
  2. [Ref. [2]] The separately reported 46-student user study is cited as 'Anonymous. 2026. Under submission' — a non-archival reference. This is acceptable as supporting evidence, but it should be clearly marked as unpublished and not used as a core quantitative pillar.
  3. [§3, formal meaning] The footnote claiming workflows compose as a commutative monoid and that realizability is a poly-time satisfiability check is asserted without proof. A brief appendix with the formal definitions and the reduction (or a citation to a full version) would let readers verify the 'computed number' claim.
  4. [Figure 4] The figure lacks error bars, confidence intervals, or any indication of annotation variability. Even with a deterministic rule, the extraction of intents is a human judgment; reporting inter-annotator consistency would strengthen the figure.
  5. [§4, '2.7 s compile' timing] The 2.7 s intent-to-spec compile time is reported without hardware/software context. A footnote on the compile setup would help reproducibility.

Circularity Check

1 steps flagged

Design coverage is computed on the same corpus used to build the grammar, so the 100% figure is a fitted value.

specific steps
  1. self definitional [§4 Evaluation, coverage of Figure 4]
    "The grammar is the union of those axes and expresses all 255 in-scope experiments from 66 papers, one intent each: total coverage is a property of a corpus we did not author, not a score we assigned ourselves."

    The grammar's axes are induced from the same 255 intents used as the test set; the paper states the grammar is 'the union of those axes.' An intent is covered when every grammar axis it needs lies within the grammar, and since each axis was added because some corpus intent required it, every corpus intent is covered by construction. The 100% design coverage is therefore a restatement of how the grammar was constructed, not an independent empirical result. The defense that the corpus was not authored by the authors does not address the lack of a held-out split: the corpus is the training set for the grammar, and no held-out papers, inter-annotator reliability, or independent implementation of the coverage rule are reported. The headline number is thus a fitted value, though the 34% implemen

full rationale

The central design-coverage claim reduces by construction: the grammar is defined as the union of axes observed in the 255-intent corpus, and the paper then reports 100% coverage on that same corpus. This is a fitted value rather than a prediction. However, the paper's other headline quantities are more independent: the 34% (87/255) implementation coverage is a declared-capability count over the same corpus but is not definitionally forced, and the CCAnalyzer replication is an executed end-to-end exemplar against an external published result. The baseline comparison (best tool ≤13%) depends on an author-defined mapping from each tool to a 'fixed sub-grammar' that is not fully specified, which is a possible understatement but not a circular reduction. No other load-bearing step is circular: the use of NetReplica/NetGent capabilities is cited from prior work by overlapping authors, but the architecture's composition is presented as a design goal and only partially demonstrated, which is an unverified assumption rather than a derivation from the assumption itself. Overall, the paper contains one clear circular step in its headline coverage statistic, but retains independent content in the implementation and replication evidence, so the score is 6 rather than higher.

Axiom & Free-Parameter Ledger

0 free parameters · 5 axioms · 1 invented entities

The central claim rests on no fitted numeric constants. It rests on domain assumptions about the representativeness of the 66-paper corpus, the decomposability of all intents into the grammar, the composability of the cited subsystems on a shared bottleneck, and the correctness of the formal realizability claim. One invented abstraction (the intent specification) carries most of the load and currently lacks independent evidence.

axioms (5)
  • domain assumption The four Lakatosian activities (re-create, extend, fix, establish) cover all empirical networking research activities.
    Invoked in §2 to justify that every experiment reduces to an 'intent'; if some research activity does not fit this decomposition, the universality of the abstraction is weakened.
  • domain assumption Every research intent can be decomposed into the grammar's axes (Foreground, Regime, substrate, telemetry) without loss.
    The 100% coverage claim depends on this decomposition; since the grammar was designed against the corpus, this is partly by construction (§3, Figure 2).
  • domain assumption The hand-mined 66-paper corpus is representative of the field's data-generation intents.
    Corpus is sampled by CCS concept/keyword from SIGCOMM, IMC, and NSDI; no inter-annotator agreement or representativeness validation is reported (§4, Experimental setup).
  • ad hoc to paper External services (NetReplica/NetForge, NetGent, Mininet) can be composed on one substrate with a shared bottleneck.
    Load-bearing for the architecture, but never demonstrated end-to-end beyond a simple static-bottleneck CCAnalyzer run (§3 architecture; §4 exemplar).
  • standard math Realizability over capability files is a poly-time satisfiability check that preserves the fidelity of the intent.
    Claimed in §3 (footnote 1) using KAT-style compositionality and conservative extension; the formal result is cited rather than proved in this paper.
invented entities (1)
  • Pramana intent specification no independent evidence
    purpose: A single contract/domain-specific language expressing what data to generate, where, and how, intended to unify networking experiment generation across substrates.
    The paper's central new abstraction; no released implementation or corpus provides an external falsifiable handle yet. The 34% proof-of-concept is self-reported.

pith-pipeline@v1.3.0-alltime-deepseek · 11516 in / 13025 out tokens · 130428 ms · 2026-08-01T00:02:56.352621+00:00 · methodology

0 comments
read the original abstract

Networking research advances by turning hypotheses into empirical evidence, so accelerating it means reducing the lag between ideation (synthesizing a hypothesis) and generating the data that tests it. Consider a concrete case: does a bulk BBR download fairly share its bottleneck with competing real-time Google Meet traffic? Validating this requires configuring a realistic bottleneck link, concurrently generating BBR's bulk transfer and Meet's real-time traffic, and collecting relevant service-quality metrics. Today this overhead is high, often forcing researchers to start from scratch for every new idea. This ideation-to-data-generation gap will only worsen in the agentic AI era, where AI-assisted ideation accelerates exponentially, yet its outputs cannot be validated without a data-generation backend. This paper explores how to bridge this gap. We envision a composable, domain-specific backend, Pramana, shaped as a thin waist, with diverse research intents at the top and disparate execution substrates at the bottom. Pramana realizes this waist through a single contract, the intent specification, which disaggregates an experiment into three independent axes: the intent (what data to generate), the substrate (where to generate it), and the mechanism (how to produce it), so one specification runs on any substrate. We demonstrate Pramana's utility by building a first-of-its-kind corpus of 255 data-generation intents mined from 66 published papers, and show the intent specification satisfies all of them, where no existing tool satisfies more than 13%. Our current proof-of-concept implementation already satisfies 34% of these intents, more than twice the best existing tool, and we lay out a roadmap for closing this abstraction-implementation gap through a broader community effort to build the envisioned data-generation backend and accelerate empirical networking research.

Figures

Figures reproduced from arXiv: 2607.26352 by Alagappan Ramanathan, Arpit Gupta, Dongsu Han, Eugene Vuong, Haarika Manda, Jaber Daneshamooz, Manni Moghimi, Sangeetha Abdu Jyothi, Satyam Kumar, Satyandra Guthula, Snithik Thode, Sylee Beltiukov, Tarun Mangla, Walter Willinger.

Figure 1
Figure 1. Figure 1: The thin waist, in three bands. Research in [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: The intent specification, as a grammar. Every [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: One model-driven compile step, then a de [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

49 extracted references · 9 linked inside Pith

  1. [1]

    Amazon Web Services. [n. d.]. Amazon Elastic Compute Cloud (EC2). https://aws.amazon.com/ec2/. Accessed 2026

    2026

  2. [2]

    Anonymous. 2026. Enabling Experiential Learning in Networking Education. Under submission; citation withheld for anonymity

    2026

  3. [3]

    Ilya Baldin, Anita Nikolich, James Griffioen, Inder Singh Monga, Kuang-Ching Wang, Tom Lehman, and Paul Ruth. 2019. FABRIC: A National-Scale Programmable Experimental Network Infrastructure. IEEE Internet Computing23, 6 (2019), 38–47

    2019

  4. [4]

    Ryan Beckett, Aarti Gupta, Ratul Mahajan, and David Walker. 2017. A General Approach to Network Configuration Verification. InACM SIGCOMM

    2017

  5. [5]

    Roman Beltiukov, Wenbo Guo, Arpit Gupta, and Walter Willinger

  6. [6]

    CAIDA. [n. d.]. Archipelago (Ark) Measurement Infrastructure. https: //www.caida.org/projects/ark/. Accessed 2026

    2026

  7. [7]

    Eric Hayden Campbell, Hossein Hojjat, and Nate Foster. 2024. Com- puting Precise Control Interface Specifications.Proceedings of the ACM on Programming Languages (OOPSLA)(2024)

    2024

  8. [8]

    Mark Chen, Jerry Tworek, Heewoo Jun, Qiming Yuan, Henrique Ponde De Oliveira Pinto, Jared Kaplan, Harri Edwards, Yuri Burda, Nicholas Joseph, Greg Brockman, et al. 2021. Evaluating large language models trained on code.arXiv preprint arXiv:2107.03374(2021)

    Pith/arXiv arXiv 2021

  9. [9]

    Audrey Cheng, Shu Liu, Melissa Pan, Zhifei Li, Bowen Wang, Alex Krentsel, Tian Xia, Mert Cemri, Jongseok Park, Shuo Yang, Jeff Chen, Lakshya Agrawal, Aditya Desai, Jiarong Xing, Koushik Sen, Matei Zaharia, and Ion Stoica. 2025. Barbarians at the Gate: How AI is Upending Systems Research.arXiv preprint arXiv:2510.06189(2025)

    arXiv 2025

  10. [10]

    Jaber Daneshamooz, Satyandra Guthula, Jessica Nguyen, William Chen, Sanjay Chandrasekaran, Ankit Gupta, Arpit Gupta, and Walter Willinger. 2025. NetForge: A Programmable Substrate for Bottleneck- Centric Network Data Generation. arXiv:2507.13476 [cs.NI] https: //arxiv.org/abs/2507.13476

    arXiv 2025

  11. [11]

    Jaber Daneshamooz, Eugene Vuong, Laasya Koduru, Sanjay Chan- drasekaran, and Arpit Gupta. 2025. NetGent: Agent-Based Automa- tion of Network Application Workflows. arXiv:2509.00625 [cs.NI] https://arxiv.org/abs/2509.00625

    arXiv 2025

  12. [12]

    Nate Foster, Dexter Kozen, Konstantinos Mamouras, Mark Reitblatt, and Alexandra Silva. 2016. Probabilistic NetKAT. InEuropean Sympo- sium on Programming (ESOP)

    2016

  13. [13]

    Saibo Geng, Martin Josifoski, Maxime Peyrard, and Robert West. 2023. Grammar-Constrained Decoding for Structured NLP Tasks without Finetuning. InEMNLP

    2023

  14. [14]

    Google. [n. d.]. Google Cloud Platform. https://cloud.google.com/. Accessed 2026

    2026

  15. [15]

    Michael Greenberg, Ryan Beckett, and Eric Hayden Campbell. 2022. Kleene Algebra Modulo Theories: A Framework for Concrete KATs. Proceedings of the ACM SIGPLAN Conference on Programming Language Design and Implementation (PLDI)(2022)

    2022

  16. [16]

    Diwaker Gupta, Kashi Venkatesh Vishwanath, and Amin Vahdat. 2008. DieCast: Testing Distributed Systems with an Accurate Scale Model. InNSDI

    2008

  17. [17]

    Satyandra Guthula, Jaber Daneshamooz, Charles Fleming, Kesheng Wu, Walter Willinger, and Arpit Gupta. 2025. Netburst: Event- centric forecasting of bursty, intermittent time series.arXiv preprint arXiv:2510.22397(2025). 7 Jaber Daneshamooz, Eugene Vuong, Alagappan Ramanathan, Manni Moghimi, Haarika Manda, Satyam Kumar, Snithik Thode, Satyandra Guthula, Sy...

    Pith/arXiv arXiv 2025

  18. [18]

    Pouya Hamadanian, Pantea Karimi, Arash Nasr-Esfahany, Kimia Noorbakhsh, Joseph Chandler, Ali ParandehGheibi, Mohammad Al- izadeh, and Hari Balakrishnan. 2025. Glia: A Human-Inspired AI for Automated Systems Design and Optimization.arXiv preprint arXiv:2510.27176(2025)

    Pith/arXiv arXiv 2025

  19. [19]

    Stephen Hemminger. 2005. Network Emulation with NetEm. InLinux Conf Au

    2005

  20. [20]

    Li, Emmanuel Candès, and Jure Leskovec

    Kexin Huang, Ying Jin, Ryan Li, Michael Y. Li, Emmanuel Candès, and Jure Leskovec. 2025. Automated Hypothesis Validation with Agentic Sequential Falsifications.arXiv preprint arXiv:2502.09858(2025)

    Pith/arXiv arXiv 2025

  21. [21]

    Jacobs, Ricardo J

    Arthur S. Jacobs, Ricardo J. Pfitscher, Rafael H. Ribeiro, Ronaldo A. Ferreira, Lisandro Z. Granville, Walter Willinger, and Sanjay G. Rao

  22. [22]

    Imre Lakatos. 1970. Falsification and the Methodology of Scientific Research Programmes. InCriticism and the Growth of Knowledge, Imre Lakatos and Alan Musgrave (Eds.). Cambridge University Press, 91–196

    1970

  23. [23]

    Bob Lantz, Brandon Heller, and Nick McKeown. 2010. A Network in a Laptop: Rapid Prototyping for Software-Defined Networks. In Proceedings of the ACM SIGCOMM Workshop on Hot Topics in Networks (HotNets)

    2010

  24. [24]

    Jiahui Li, Han Qi, Ruyi Yao, Jialin Wei, Ruoshi Sun, Zixuan Chen, Sen Liu, and Yang Xu. 2025. CClinguist: An Expert-Free Framework for Future-Compatible Congestion Control Algorithm Identification. In Proceedings of the ACM SIGCOMM Conference

    2025

  25. [25]

    Chris Lu, Cong Lu, Robert Tjarko Lange, Jakob Foerster, Jeff Clune, and David Ha. 2024. The AI Scientist: Towards Fully Automated Open- Ended Scientific Discovery.arXiv preprint arXiv:2408.06292(2024)

    Pith/arXiv arXiv 2024

  26. [26]

    George C. Necula. 2000. Translation Validation for an Optimizing Compiler. InPLDI

    2000

  27. [27]

    Ravi Netravali, Anirudh Sivaraman, Somak Das, Ameesh Goyal, Keith Winstein, James Mickens, and Hari Balakrishnan. 2015. Mahimahi: Accurate Record-and-Replay for HTTP. InUSENIX Annual Technical Conference (ATC)

    2015

  28. [28]

    Alexander Novikov, Ngân V˜u, Marvin Eisenberger, Emilien Dupont, Po-Sen Huang, Adam Zsolt Wagner, Sergey Shirobokov, Borislav Ko- zlovskii, Francisco J. R. Ruiz, Abbas Mehrabian, M. Pawan Kumar, Abigail See, Swarat Chaudhuri, George Holland, Alex Davies, Sebas- tian Nowozin, Pushmeet Kohli, and Matej Balog. 2025. AlphaEvolve: A Coding Agent for Scientific...

    Pith/arXiv arXiv 2025

  29. [30]

    Amir Pnueli, Michael Siegel, and Eli Singerman. 1998. Translation validation. InInternational Conference on Tools and Algorithms for the Construction and Analysis of Systems. Springer, 151–166

    1998

  30. [31]

    RIPE NCC. 2015. RIPE Atlas: A Global Internet Measurement Network. The Internet Protocol Journal18, 3 (2015), 2–26

    2015

  31. [32]

    Luigi Rizzo. 1997. Dummynet: A Simple Approach to the Evaluation of Network Protocols.ACM SIGCOMM Computer Communication Review 27, 1 (1997), 31–41

    1997

  32. [33]

    Damien Saucez, Luigi Iannone, and Olivier Bonaventure. 2019. Evalu- ating the artifacts of SIGCOMM papers.ACM SIGCOMM Computer Communication Review(2019). https://ccronline.sigcomm.org/wp- content/uploads/2019/05/acmdl19-308.pdf

    2019

  33. [34]

    Torsten Scholak, Nathan Schucher, and Dzmitry Bahdanau. 2021. PI- CARD: Parsing Incrementally for Constrained Auto-Regressive De- coding from Language Models. InEMNLP

    2021

  34. [35]

    Jiabin Tang, Lianghao Xia, Zhonghang Li, and Chao Huang. 2025. AI-Researcher: Autonomous Scientific Innovation.arXiv preprint arXiv:2505.18705(2025)

    Pith/arXiv arXiv 2025

  35. [36]

    University of California, Santa Barbara. [n. d.]. The PINOT Project: Programmable Infrastructure for AI/ML for Networking. https://pinot. cs.ucsb.edu. Accessed 2026

    2026

  36. [37]

    Amin Vahdat, Ken Yocum, Kevin Walsh, Priya Mahadevan, Dejan Kostić, Jeff Chase, and David Becker. 2002. Scalability and Accuracy in a Large-Scale Network Emulator. InOSDI

    2002

  37. [38]

    Ranysha Ware, Adithya Abraham Philip, Nicholas Hungria, Yash Kothari, Justine Sherry, and Srinivasan Seshan. 2024. CCAnalyzer: An Efficient and Nearly-Passive Congestion Control Classifier. InProceed- ings of the ACM SIGCOMM Conference. doi:10.1145/3651890.3672255

    arXiv 2024

  38. [39]

    Brian White, Jay Lepreau, Leigh Stoller, Robert Ricci, Shashi Gu- ruprasad, Mac Newbold, Mike Hibler, Chad Barb, and Abhijeet Joglekar

  39. [40]

    Willard and Rémi Louf

    Brandon T. Willard and Rémi Louf. 2023. Efficient Guided Generation for Large Language Models.arXiv:2307.09702(2023)

    Pith/arXiv arXiv 2023

  40. [41]

    Ferreira, Arpit Gupta, Roman Beltiukov, Satyandra Guthula, Lisandro Z

    Walter Willinger, Ronaldo A. Ferreira, Arpit Gupta, Roman Beltiukov, Satyandra Guthula, Lisandro Z. Granville, and Arthur S. Jacobs. 2025. When Something Looks too Good to be True, it Usually is! AI is Caus- ing a Credibility Crisis in Networking.ACM SIGCOMM Computer Com- munication Review55, 1 (2025), 10–15. doi:10.1145/3727063.3727067

    arXiv 2025

  41. [42]

    Yutaro Yamada, Robert Tjarko Lange, Cong Lu, Shengran Hu, Chris Lu, Jakob Foerster, Jeff Clune, and David Ha. 2025. The AI Scientist-v2: Workshop-Level Automated Scientific Discovery via Agentic Tree Search.arXiv preprint arXiv:2504.08066(2025)

    Pith/arXiv arXiv 2025

  42. [43]

    Yan, Hudson Ayers, Chenzhi Zhu, Sadjad Fouladi, James Hong, Keyi Zhang, Philip Levis, and Keith Winstein

    Francis Y. Yan, Hudson Ayers, Chenzhi Zhu, Sadjad Fouladi, James Hong, Keyi Zhang, Philip Levis, and Keith Winstein. 2020. Learn- ing in situ: a randomized experiment in video streaming. InUSENIX Symposium on Networked Systems Design and Implementation (NSDI)

    2020

  43. [44]

    Yan, Jestin Ma, Greg D

    Francis Y. Yan, Jestin Ma, Greg D. Hill, Deepti Raghavan, Riad S. Wahby, Philip Levis, and Keith Winstein. 2018. Pantheon: the training ground for Internet congestion-control research. InUSENIX Annual Technical Conference (ATC)

    2018

  44. [45]

    Tao Yu, Rui Zhang, Kai Yang, Michihiro Yasunaga, Dongxu Wang, Zifan Li, James Ma, Irene Li, Qingning Yao, Shanelle Roman, et al

  45. [46]

    Wang, Sadjad Fouladi, Francis Y

    Yajie Zhou, Jiajun Ruan, Eric S. Wang, Sadjad Fouladi, Francis Y. Yan, Kevin Hsieh, and Zaoxing Liu. 2026. NetArena: Dynamic Benchmarks for AI Agents in Network Automation. InProceedings of the Interna- tional Conference on Learning Representations (ICLR). 8

    2026

  46. [2002]

    An Integrated Experimental Environment for Distributed Sys- tems and Networks. InOSDI

  47. [2018]

    InProceedings of the 2018 conference on empirical methods in natural language processing

    Spider: A large-scale human-labeled dataset for complex and cross-domain semantic parsing and text-to-sql task. InProceedings of the 2018 conference on empirical methods in natural language processing. 3911–3921

    2018

  48. [2021]

    InUSENIX Annual Technical Conference (ATC)

    Hey, Lumi! Using Natural Language for Intent-Based Network Management. InUSENIX Annual Technical Conference (ATC)

  49. [2023]

    InProceed- ings of the ACM SIGSAC Conference on Computer and Communications Security (CCS)

    In Search of netUnicorn: A Data-Collection Platform to Develop Generalizable ML Models for Network Security Problems. InProceed- ings of the ACM SIGSAC Conference on Computer and Communications Security (CCS). doi:10.1145/3576915.3623075

    arXiv