Show HN: Hound – Relation-First Knowledge Graphs for Complex-System Reasoning

1 point

9 months ago

Hound is a code security auditing tool that draws inspiration from human cognitive processes to enhance reasoning about large, complex systems. By modeling the target application as relation-first knowledge graphs (e.g., monetary/value flows, authentication/authorization roles, call graphs, and invariants with compact annotations), the agent enables multi-granular attention: it can "zoom in" on specific graph slices for detailed analysis while maintaining a summarized view of the entire system through broad mappings.

Complementing this is a persistent belief system that tracks hypotheses with explicit evidence and confidence levels, refining them over time as new evidence emerges—ensuring a disciplined lifecycle for findings that mirrors iterative human reasoning and belief updating.

Evaluated on a subset of the ScaBench benchmark, Hound shows improvements in vulnerability detection (31.2% true positives versus 8.3% for a baseline LLM analyzer) and F1 score (14.2% versus 9.8%).

While tailored for security audits, Hound's analyst-defined graphs and cognitive-inspired framework provide a solid basis for general complex-system reasoning. Released on September 15, 2025, the full paper is available on [Zenodo](https://zenodo.org/records/17129271), with the implementation hosted on [GitHub](https://github.com/scabench-org/hound) for further exploration and reproduction.