Leakage-Proof Evaluation#
NeuroTwin v1 treats split construction as part of the scientific claim.
Rules:
Build
SplitManifestfrom recording-level metadata before preprocessing, augmentation, or windowing.Use held-out subject, site, and dataset splits as primary tests.
Run record-ID reuse checks for every split.
For subject-held-out claims, no subject can appear across train/val/test.
For site-held-out claims, no site can appear across train/val/test.
For dataset-held-out claims, no dataset can appear across train/val/test.
Clinical prediction is secondary and cannot rescue a weak Neural Translation result.
Main failure modes to catch:
Subject identity leakage.
Scanner/site leakage.
Repeated-session leakage.
Stimulus or clip leakage.
Metadata labels that encode the target.
Training windows generated before split assignment.