Zex / ZEROg — Space biology with the evidence attached
Spaceflight produces valuable biological data, but turning those tables into useful research questions requires careful interpretation. A compelling hypothesis can still be statistically weak, contradicted by the data, or already established elsewhere.
ZEROg makes that reasoning visible. It is an interactive research explorer built around saved outputs from Helix, an existing open-source Python research engine. Researchers can inspect a proposed lead, see the statistics attached by code, read the critic’s judgment, and examine a separate cross-study comparison.
The demo includes a searchable catalog of 243 NASA Open Science Data Repository studies, 92,959 gene rows across four bundled tables, and a saved mouse-thymus run from OSD-421. That run contains 60 proposed leads, of which 38 passed the saved critic. Users can replay the run, inspect accepted and rejected leads, and search a research graph containing 294 findings and their connected evidence.
The Loop page demonstrates why evaluation matters. It compares saved rulebooks v0, v2, and v4. Although v4 performs better on the training set, v2 performs better on the saved holdout evaluation and is selected for the demo. These evaluations use the upstream project’s LLM-jury labels; they are not presented as independent human validation.
The website also includes an Inspiration4 study directory, creating a starting point for future human spaceflight research. Those records are clearly marked as listed only: this demo does not analyse Inspiration4 data or infer crew-health outcomes from the mouse results.
The app uses Next.js 16, TypeScript, and Tailwind CSS 4, with searchable evidence views, saved-run playback, persistent light and dark themes, and downloadable source artifacts. It works without API keys or live model calls. Original scientific files are preserved, and checksums verify their integrity. A discrepancy between the saved summary’s sample counts and the underlying CSV is documented rather than silently corrected.
The central design principle is simple: a research lead is not experimental validation. Table statistics, model judgments, literature connections, and replication checks each remain identifiable.
This submission contributes the web experience and evidence presentation around the existing Helix engine. The next step is a reviewed human-data adapter and a human-labelled evaluation set for Inspiration4.
ZEROg is an independent project with no SpaceX affiliation or endorsement.
Repository: github.com/Invincible1234/zerog
Log in or sign up for Devpost to join the conversation.