# Master Branch

- **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:** [S Roshan Pranao](https://cerebralvalley.ai/u/roshan_emerald)
- **GitHub:** https://github.com/Roshan5105labs/Fault-Line
- **Demo video:** https://youtu.be/RXcG3jQxgqk
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2212

FaultLine is a proving ground and runtime guardrail for AI agents that operate infrastructure. When an agent says "incident resolved", FaultLine checks whether that's true, and what the agent broke along the way.

The problem: agents are getting write access to production. A restart can turn an alert green while silently killing in-flight work. Agent benchmarks measure whether agents diagnose faults; FaultLine measures what their fixes cost and verifies their claims against measured state.

How it works: FaultLine runs a production-like stack (FastAPI, Postgres, Redis job queue, scalable workers) under steady load, injects a reproducible fault, and pages Muse Spark via the Meta Model API with a symptom-only alert. Muse investigates and acts through constrained tools. A state-aware guardrail checks every write action against live system state before it runs, and returns the reason and the safe path when it blocks. An independent evaluator then scores the run from an append-only trace and a per-job ledger, never from the agent's words: did the service recover, were jobs lost, was the claim true, was the root cause right.

Results across 21 Muse Spark runs: every root-cause claim Muse made was correct (11/11), and it fixed a connection leak cleanly in 6/6 runs. But on a bad config release it restarted workers unnecessarily after its correct rollback; unguarded, that lost a job in 2 of 3 runs while Muse still claimed "resolved", and FaultLine flagged both as false resolutions. With the guardrail, Muse followed the block feedback (pause, drain, restart safely, resume) and lost nothing. FaultLine's own runs also exposed a race in its guardrail; we fixed it and added a regression test.

Built from scratch during the hackathon (Oct 27 - Nov 4). Runs on a laptop with one docker compose command; plug in any agent with one adapter class. Limitations (2 fault types, 3 seeds per configuration) are stated in the README.

## More from Meta Global AI Developer Hackathon

- [DA Syllabus Auditor + Blindspot Engine](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2209)
- [Anatole KABORE](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2210)
- [Tabby](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2211)
- [Vinklings](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2213)
- [Capin](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2214)
- [BolSaakshi](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2215)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2212. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt