Confidence-Weighted Plasticity: Experimental Validation and Boundary Conditions

We present experimental validation of Confidence-Weighted Plasticity (CWP), a mecha-
nism for distributing gradient updates across modular components based on demonstrated
predictive reliability. Our experiments confirm that the core mechanism operates as spec-
ified: confidence accurately tracks prediction accuracy, gradient routing scales according
to plasticity, and the system responds to prediction failure. However, we identify a critical
boundary condition in tightly-coupled architectures where a component’s local prediction
task shares parameters with the global training objective. In such configurations, train-
ing downstream components disrupts upstream predictions, triggering plasticity increase
rather than protection—the mechanism actively directs more gradient toward the compo-
nent we intend to protect. We characterise this limitation and discuss architectural require-
ments for effective deployment.