This benchmark intentionally does not stage public TCGA-LUAD matrices in input/.

At solve time the agent is expected to:
- download the public TCGA-LUAD data from GDC
- perform the survival analysis locally on the VM
- write only the required outputs under output/

The final evaluation environment will provision the required R runtime separately from these Stage 1 assets.
