# FrontierLab

- **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:** [Pack Kwan Low](https://cerebralvalley.ai/u/Zoe871)
- **GitHub:** https://github.com/Jaycee871/FrontierLab-Meta-Hackathon-2026
- **Demo video:** https://youtu.be/agYFr2HgZ5Y
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/436

FrontierLab is an evidence-grounded, human-gated scientific research pipeline that uses Meta models to turn literature into reviewable hypotheses and plans without hiding the evidence, uncertainty or researcher decisions in between.

A scientific language model can cite a real paper and still overstate what that paper supports. FrontierLab treats literature-to-plan reasoning as an auditable process instead of one generation step. Meta Muse models propose source-linked claims, receive adversarial critique and produce falsifiable hypotheses across five stages: Research, Evidence, Hypothesis, Evaluation, Planning. A model claim that does not cite a retrieved source ID with a contiguous 24-500 character verbatim quotation from that source is rejected before synthesis. Semantic review, method selection and planning authorization are explicit researcher gates bound to content hashes. A provenance-tagged Evidence Graph traces artifacts back to source passages, while a separate Trajectory Audit Graph records model calls, human interventions and rejected claims without treating them as scientific evidence.

On the recorded, researcher-reviewed biomedical case (pH-dependent antibody engineering, two published abstracts), the app recomputes from files in docs/runs: 100% source trace coverage (4/4 artifacts reach a known paper), 50% strict EXTRACTED coverage, 9 planning obligations carried into Stage 5, and 0 model calls to replay. The 50% is preserved, not patched: it reflects a real model/researcher provenance disagreement. Anyone can open the public Space and replay the case with no API key and no model call; tampered files or a selection that would authorize planning fail closed. A scheduled Maintenance Canary re-runs the full 407-test regression suite and compares the environment against a human-accepted baseline, with automatic repair, commit, merge and deploy deliberately disabled.

What is not claimed: no accuracy score, no independently adjudicated gold standard, no wet-lab validation. The claim is narrower and checkable: model-assisted scientific steps become inspectable, challengeable and reproducible.

Live demo: https://huggingface.co/spaces/Jaycee766/FrontierLab-Meta-Hackathon-2026
Builds on Paper2Agent (Miao et al., Nature 2026) as cited prior work; the information-management three-pillar framework (maintainability, provenance, research governance) and the Research Operating Interface are the author's own extension.

## More from Meta Global AI Developer Hackathon

- [TenderCheck](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/433)
- [NullPointerException](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/434)
- [Mendrika](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/435)
- [Kryterea](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/437)
- [workwell](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/438)
- [HeoCoin Family](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/439)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/436. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt