HERD Framework Combines Harness-Based Skills with Recursive Distillation

HERD uses a hybrid continual learning approach that captures patterns in a non-parametric harness while performing drift-constrained distillation into model weights. This dual mechanism aims to mitigate catastrophic forgetting and plasticity loss by retaining high-residual examples in bounded memory.