ResearchJanuary 14, 2026

NCG open-source release and toolkit availability

The `ncg-torch` toolkit and GitHub v0.1.0 release make NCG reproducible for continual learning workflows and benchmark validation.

NCG grows model capacity only when novelty detectors identify genuinely new data, with meta-parameters controlling growth behavior.

Results show strong forgetting reductions on Split-MNIST and Split-CIFAR-10, with transparent reporting of optimization limitations.

The release aligns code, documentation, and research artifacts to support reproducible evaluation.