# Meta Quant Auditor

- **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:** [Aung San](https://cerebralvalley.ai/u/AungSan)
- **GitHub:** https://github.com/aungsan-portfolio/meta-quant-auditor
- **Demo video:** https://www.youtube.com/watch?v=MrQomF8UIhY
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/385

Meta Quant Auditor is an AI agent that audits paper-trading order lifecycles.

The problem: I run day-trading and swing-trading bots on paper accounts and review their execution logs. A risk gate's PASS or BLOCK does not show which evidence supports it, and a later recovery can hide an earlier unsafe sequence, such as a reservation released while a cancel was still pending and a fill that arrived afterwards. AI models can also say "everything is fine" when it is not.

How Meta's model is used: Muse Spark Open (meta-models/muse-spark-open-1026, through Baseten) acts as the auditor agent. In each step it chooses one of seven allowed actions as strict JSON: run the lifecycle audit, query the order, position or protection, compare claims, ask for human input, or finish the report. Local code validates every action, computes the risk gates, keeps critical history visible after a later PASS, and checks each structured claim against the evidence as EVIDENCE_MATCH, CONTRADICTION or UNVERIFIED. When evidence is missing, the agent stops with NEEDS_INPUT instead of inventing it. The model's narrative stays labeled unverified.

Safety and engineering: the user previews the evidence and gives explicit consent before any request. Runs are capped at eight steps, actions cannot repeat, there are no automatic retries, and each request has a 30-second deadline. 232 unit tests and 14 UI tests run in Linux and Windows CI. Both demo scenarios passed 3/3 live runs.

Disclosure: my paper-trading bots and an earlier audit scaffold existed before the hackathon. During the event I built this standalone auditor, its fictional scenarios, and the Muse Spark Open agent integration. Fictional data only; no broker connection or trades.

## More from Meta Global AI Developer Hackathon

- [Nuksee](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/382)
- [FlashWatch](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/383)
- [Drawing-to-Parametric-CAD](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/384)
- [UI Evolution Engine](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/386)
- [Empu](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/387)
- [2201344667392187](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/388)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/385. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt