Definition
Plain language
A crosscoder with labeled shelves — some features reserved for one model, some for the other, some shared — enforced so a feature assigned to one model physically cannot affect the other.
As stated in the literature
A crosscoder variant (DFC) that partitions the dictionary into model-exclusive and shared bins, mechanically zeroing cross-connections during training; reaches the same behavioral ceiling as a plain crosscoder with far fewer pointable features, improving legibility rather than capability.
Also called: DFC
Why it matters: By cleanly separating which features belong to which model, it makes the comparison far easier to read and trust without sacrificing how well it captures behavior.
For example, a concept assigned to the fine-tuned model's private shelf is physically wired so it cannot influence the base model's reconstruction at all.
Heard on the show
“Which is where the architecture under test comes in — the Dedicated Feature Crosscoder, the DFC.”Episode 175 — One Crosscoder Feature Flips a Stalling Chatbot Into a Working Agent