# ZeroHP

- **Event:** [Meta Global AI Developer Hackathon](https://cerebralvalley.ai/e/meta-ai-developer-hackathon)
- **When:** Oct 26, 12:00 PM – Nov 4, 9:00 PM (EST)
- **Where:** Online
- **Team:** [Rohit Patil](https://cerebralvalley.ai/u/Rohit_decodes)
- **GitHub:** https://github.com/rohitdecodes/netproof
- **Demo video:** https://youtu.be/XjunlU6rk0c
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2275

NetProof is an AI agent that repairs network faults, and is only allowed to keep a change after deterministic tests prove the change fixed the problem and broke nothing else.

The problem: configuration changes are one of the most common causes of network outages. An AI agent that can read routes and firewall rules could repair many of them, but nobody should let a model type commands into a network on trust. NetProof treats the model as untrusted and makes it earn every change.

How it works:
- Lab: six Linux containers on Docker (two routers, four hosts, four networks). The network's intended behaviour is written as 13 flows: 8 that must work, including two 64 KB transfers that expose MTU problems, and 5 that must stay blocked.
- Agent: Muse Spark investigates through read-only tools (routes, addresses, firewall rules, ping, traceroute, flow tests) and submits one change plan per device, with a root cause and a rollback.
- Policy gate: deterministic rules, no model. An allow-list of command forms; no flushing or disabling firewalls; no deleting default routes; no broad ACCEPT rules; changes only on devices on a failing path.
- Verifier: every allowed change is applied to a snapshot-protected device and all 13 flows are re-tested. The change is kept only if it fixed something and broke nothing. Otherwise the device is restored and the agent is told why.
- Drift cleanup: once the network is healthy, the agent is shown what still differs from the intended configuration and may remove leftovers. Such a change is kept only if it reduces drift and breaks nothing.
- Every run writes an incident report: ticket, diagnosis, each plan and its verdict, flows before and after, leftover drift, cost and time.

Evaluation: 54 fault variants across 9 classes (missing routes, wrong gateways and next hops, firewall drops, ACL typos, interfaces down, MTU black holes, multi-device faults, misleading tickets), plus an adversarial set where tickets demand unsafe changes and instructions are planted in firewall comments. I froze the prompt and engine (git tag p3.0-freeze) before any model saw the held-out or adversarial variants. A scripted runbook fixes 11/27 held-out faults. [RESULTS: paste the held-out and adversarial sentences from docs/results.md here.]

Built solo in eight days. Limitations, stated in the README: one emulated topology; held-out and adversarial sets written by the same author; no comparison with human engineers.

Code, results and incident reports: https://github.com/rohitdecodes/netproof
Live replay of recorded runs: https://rohitdecodes.github.io/netproof/

## More from Meta Global AI Developer Hackathon

- [Wick Broker](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2272)
- [Uplift](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2273)
- [Saad](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2274)
- [DataBridge AI](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2276)
- [Cactus Sentinel](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2277)
- [KLI](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2278)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2275. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt