GEMEO Architecture (Reference)
This repository holds the GEMEO reference architecture: the specification, the reference implementation of the three pillars, and the adaptation guide. Access is gated, and downloads are granted on request.
Model weights (CC-BY-NC-4.0): https://huggingface.co/Raras-AI/gemeo-sus Benchmark (CC-BY-NC-4.0): https://huggingface.co/datasets/Raras-AI/rarebench-br-trajectory Paper (concept DOI, always resolves to the latest version): https://doi.org/10.5281/zenodo.20092130
What GEMEO is
The first world-model architecture for rare disease: a three-pillar pipeline (Propose, Simulate, Verify) that grounds onset candidates in a biomedical knowledge graph, simulates patient trajectories with a recurrence-aware Causal Diffusion Forcing transformer, and adjudicates every prediction with an agentic verifier that carries a traceable evidence path per decision.
What is in this repository
gemeo_architecture_spec_v2.md: the architecture specification.reference_impl/: reference implementation (Causal Diffusion Forcing transformer and components).pillar_a_kg_proposer.py,pillar_c_swarm_verifier.py: Pillar A and Pillar C reference code.ADAPTING_TO_A_NEW_DATABASE.md: how to adapt GEMEO to a new clinical database.gemeo_bench.py: benchmark harness.
License
- Code and reference implementation: AGPL-3.0 (open-core). Commercial licenses available: dimas@raras.ai.
- Trained weights (gemeo-sus) and benchmark: CC-BY-NC-4.0.
How to use
Download the files from the Files tab above (access is granted on request), then
follow ADAPTING_TO_A_NEW_DATABASE.md.
Research preview, not a medical device.