CWT architecture ditches residual stream for structured workspace

CWT is a fascinating structural rethink that forces a model to use explicit memory partitions instead of a black-box continuous vector stream. By using a central shared channel mediated by decay gates, the model must explicitly decide what to remember and forget. The massive 45% reduction in core compute suggests traditional transformers waste significant capacity constantly overwriting states. The structured workspace inherently enables 3D visualization of the model's "thought process," offering a huge leap for mechanistic interpretability. This self-funded project provides a compelling, open-source blueprint for more efficient and transparent foundational models.