These sources inform design choices. Their empirical results are not CommGuard results and are never copied into the evidence index.
- Robi Rahman and Sabiha Tajdari, “Detecting Hidden ML Training With Zero-Overhead Telemetry,” arXiv:2606.19262v1, 2026. https://arxiv.org/abs/2606.19262. The supplied PDF motivates the nine NVML signals, broad workload/control coverage, whole-run evaluation, and held-out adversarial iteration. CommGuard does not reproduce its hardware corpus, monitor–evader rounds, or reported scores.
- Emmanouil Seferis and Tim Fist, “Detecting Compute Structuring in AI Governance is likely feasible,” supplied manuscript, 2025. It motivates communication/bandwidth as one signal, conservative decisions, segmentation, and decentralized/sparse synchronization. Its cloud-provider evidence and governance system are outside CommGuard's Kaggle scope.
- Arthur Douillard, Qixuan Feng, Andrei A. Rusu, Rachita Chhaparia, Yani Donchev, Adhiguna Kuncoro, Marc'Aurelio Ranzato, Arthur Szlam, and Jiajun Shen, “DiLoCo: Distributed Low-Communication Training of Language Models,” arXiv:2311.08105v3, 2024. https://arxiv.org/abs/2311.08105. CommGuard's bounded two-GPU parameter averaging is only DiLoCo-inspired; it is not a faithful algorithmic or scale reproduction.
- Shaoke Xi et al., “A Few GPUs, A Whole Lotta Scale: Faithful LLM Training Emulation with PrismLLM,” arXiv:2605.15617v1, 2026. https://arxiv.org/abs/2605.15617. It motivates explicit emulation labels and warns that ordinary downscaling changes communication, memory, and bottlenecks. CommGuard implements no PrismLLM emulation result.
- Naci Cankaya, “A system overview for near-term, low-trust AI compute verification,” version 0.2 working draft, 23 June 2026. The supplied DOCX motivates evidence/evaluation separation and layered provenance. Network taps, physical security, secure evaluation, replay protocols, and treaty architecture are outside this SDK.
- Mohamed Amine Merzouk, Dmitri Carpov, Mirko Bronzi, Damiano Fornasiere, and Adam Oberman, “How Much is Left? LLMs Linearly Encode Their Remaining Output Length,” arXiv:2607.05316v1, 2026. https://arxiv.org/abs/2607.05316. Its hidden-state probing task does not define CommGuard's detector.
The two supplied copies of “How Much is Left?” are byte-identical. The supplied LessWrong welcome page is community context and supplies no detector method.
- NVIDIA Management Library documentation: https://docs.nvidia.com/deploy/nvml-api/. CommGuard labels the exposed PCIe values as NVML PCIe traffic readings, never direct NCCL byte counts.
- NVIDIA Collective Communications Library documentation: https://docs.nvidia.com/deeplearning/nccl/user-guide/docs/.
- PyTorch distributed documentation: https://pytorch.org/docs/stable/distributed.html.
- Kaggle notebook platform: https://www.kaggle.com/docs/notebooks.
Optional NVIDIA tools (nccl-tests, nvbandwidth, and cuda-samples) are not
compiled or downloaded automatically. Their outputs, if later collected, must
remain distinct from NVML measurements and receive their own indexed artifacts.