The next breakthrough is
already published
Discoveries hide across papers no single scientist reads together. MAGELLAN is an open-source, 15-agent AI system that reads across the silos, connects existing knowledge into new testable hypotheses, and then kills 86% of its own ideas to show you only what survives. Anyone can download the CLI and run it.
ONE OF THE SURVIVORS
The knowledge is already there
In 1986, Don Swanson proved that the connection between fish oil and Raynaud's disease was hiding in plain sight — published in separate journals that no single researcher read together. He called it Undiscovered Public Knowledge: answers that exist in the literature but remain invisible because science is fragmented into silos.
Today there are over 100 million published papers. The fragments are everywhere. The connections between them — the ones that lead to new mechanisms, new treatments, new understanding — go unnoticed for years, sometimes decades.
MAGELLAN does what Swanson did by hand — at scale. It deploys 15 specialized AI agents that read across disciplinary boundaries, identify where established knowledge in field A connects to established knowledge in field C through an unexplored bridge, and then generate testable hypotheses about what that connection means.
No new data. No black-box predictions. Just existing science, connected in ways no one had seen — and rigorous enough to test in a lab.
63% kill rate
Every survivor is tagged GROUNDED PARAMETRIC or SPECULATIVE. We label our uncertainty.
Hypotheses no human formulated
Each card below is a testable scientific prediction. The AI found the connection, proposed the mechanism, and survived its own quality gate.
These are hypotheses, not discoveries
We're not claiming breakthroughs. We're claiming a system that generates ideas worth investigating — and we need real scientists to tell us if we're right. The confidence scores are intentionally moderate. The kill rate is intentionally high. This is how you build something credible.
Formal methodology
The system architecture, evaluation results, and honest limitations are documented in a 42-page arXiv preprint. Two computational verifications tested predictions against real data.
First application of EVT to the Meltome Atlas. Universal Weibull-domain behavior discovered across all domains of life. A companion hypothesis predicted its own diagnostic.
Plants operate 1–6× above the fundamental physical limit. First application of Fisher information to plant gravity sensing.
We're not the only ones building this
Over $2 billion raised in AI scientific discovery since 2024. Google, FutureHouse, and others deploy multi-agent systems that generate and validate hypotheses. The field is real — and MAGELLAN operates in it with a different approach.
6 agents on Gemini 2.0. 3 wet-lab validated discoveries.
Biomedical only, human-directed.
Discovered novel dAMD treatment. Reads 1,500 papers/run.
Bio/chem focused, requires human-defined domain.
Rediscovered chemistry hypotheses later published in Nature and Science. ICLR 2025.
Chemistry only, no cross-disciplinary scope.
MAGELLAN explores any discipline, autonomously, with cross-model validation. Fully open source — anyone can run it.
This is Session 25.
Imagine Session 2500.
These hypotheses were generated by March 2026 frontier models. As LLMs improve — deeper reasoning, fewer hallucinations, broader knowledge — every stage of this pipeline becomes more powerful.
We're building the infrastructure in the open: the agents, the quality gates, the cross-model validation. The scaffolding for a future where AI doesn't just answer questions — it asks new ones.