# EigenBrains

- **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:** [Ismail ELADRAOUI](https://cerebralvalley.ai/u/Ismail_ea), [El Yazid Tebbaa](https://cerebralvalley.ai/u/yzdtr41), [Saad Qacif](https://cerebralvalley.ai/u/SaadtheApex_1)
- **GitHub:** https://github.com/coderyaz856/NOESIS-Autonomous-Information-Driven-Scientific-Discovery
- **Demo video:** https://youtu.be/x3rVdGG7zP4
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4373

Noesis is an autonomous, information-driven multi-agent scientific discovery laboratory and Research IDE. Rather than acting as a black-box chatbot that hallucinates optimistic answers, Noesis operates as a closed-loop scientific instrument: it decides where to spend intelligence, where to spend experimental compute, and when empirical evidence demands changing course.

HOW NOESIS WORKS:
1. Epistemic Honesty & Pre-Registration: When a researcher enters a question, Noesis immediately classifies it as either an "Executable Lab Study" (bound to a cryptographic SHA-256 pre-registration fingerprint) or an "Open Research Protocol—Not Executed" when outside the lab's executable domain.
2. Portfolio-Driven Multi-Agent Orchestration: Powered by the Meta Model API and Omnigent orchestration, Noesis coordinates six specialized roles—Lead Scientist (PI), Literature Scout, Experiment Designer, Experiment Runner, Critic, and Verifier. Agent allocation is governed by explicit portfolio mathematics (evaluating failure overlap, rescue probability, latency, and cost) rather than arbitrary routing.
3. Information-Theoretic Experiment Selection: Competing experimental protocols are scored on an Expected Information Gain (EIG, in bits) vs. estimated compute frontier. Every candidate experiment must declare an explicit statistical falsifier before execution.
4. Human-in-the-Loop Safety & Budget Governance: Consequential compute expenditures and confirmatory stage transitions pause for operator approval (Approve/Decline) while tracking real-time trial budget spend.
5. Scientific Rigor & Willingness to Change Its Mind: Noesis evaluates results using out-of-sample AUROC, Restricted Mean Survival Time (RMST), and CVaR, updating Bayesian posteriors P(h|D). When an intervention shows strong development signal (+0.139 AUROC) but fails on held-out landscapes (-0.006 AUROC), Noesis surfaces the negative result, flags methodological threats (autocorrelation, distribution shift), and pivots—saving over 90% of compute budget compared to exhaustive grid search.
6. One Evidence Contract, Three Ways to Do Science: Every run writes an immutable ledger audited across Python (lifecycle orchestration), Rust (high-speed entropy kernel parity), and Julia (independent statistical verification), exportable as a LaTeX-formatted PDF Evidence Report.

## More from Meta Global AI Developer Hackathon

- [Atom Industries](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4370)
- [Omni](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4371)
- [Merit](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4372)
- [erdos](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4374)
- [barqlo](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4375)
- [Fireside](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4376)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4373. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt