NULLGARDEN helps scientific machine-learning teams investigate shortcuts such as batch identity, duplicate subjects, and preprocessing leakage. The workbench explores negative controls and allows review of the assumptions and transformations.
The project uses NVIDIA Nemotron-3-Super through Nebius Token Factory to propose control hypotheses. A direct live test of its prompt, inference, and validation functions on a batch-confounded example returned four structurally valid hypotheses. The UI path and scientific usefulness have not been independently validated.
The repository contains the implementation and setup instructions. Results are research prototypes, not clinical guidance.
Log in or sign up for Devpost to join the conversation.