Task Limitations

CIFAR-10 is a classification-style stress test, not a naturally continuous composition task.
Topics is model-derived and should be read as topic-mixture evaluation rather than raw-label truth.
PBMC is semi-synthetic and is intended as a control-style benchmark slice.
UTKFace and AffectiveText rely on derived artifacts because the source assets are restricted.
Samson is the cleanest natural low-dimensional composition task in the bundle.