PDGA
Predictive Dendritic Gathering Adapter
Adaptive fine-tuning for small language models. Standard methods overwrite what a model already knows, so each new domain costs a large part of the accuracy on the last one. PDGA predicts the effect of an update before it is applied and routes it so prior knowledge is protected while new learning still lands.
Every routing decision is logged in machine-readable form, which makes an adapted model auditable rather than opaque, and the method runs with under 3% latency overhead on consumer hardware.