AI coding agents repeatedly commit the same high-cost failures. Pre-action enforcement intercepts tool calls before execution and decides allow, warn, or deny. This paper presents a practical evaluation framework for infrastructure firewalls on coding agents, instantiated by ThumbGate, with seven proof dimensions and re-runnable evidence paths.

Files

main.pdf — compiled preprint
main.tex — LaTeX source (arXiv-ready)
arxiv-src.tar.gz — source bundle for arXiv upload

Linked artifacts

Space: https://huggingface.co/spaces/IgorGanapolsky/ThumbGate
Dataset (sample feedback schema): https://huggingface.co/datasets/IgorGanapolsky/thumbgate-agent-feedback-sample
GitHub: https://github.com/IgorGanapolsky/ThumbGate
HTML white paper: https://thumbgate.ai/whitepaper

arXiv

Primary category intended: cs.SE (cross-list cs.AI).
After arXiv assigns an ID, this card will be updated and the paper will be indexed at https://huggingface.co/papers/<arxiv-id> and claimed on the author profile.

Citation

bibtex

@misc{ganapolsky2026preaction,
  title={How We Know Pre-Action Gates Work: An Evaluation Framework for Infrastructure Firewalls on AI Coding Agents},
  author={Ganapolsky, Igor},
  year={2026},
  howpublished={Hugging Face Hub preprint},
  url={https://huggingface.co/IgorGanapolsky/thumbgate-pre-action-gates-paper}
}