Dead features — hollows out the count of → Dictionary feature

A dictionary feature is defined structurally as one row of M -- it exists as soon as the matrix is initialized, regardless of whether it ever actually fires. A dead feature is exactly this: a row that occupies dictionary capacity and counts toward R but never activates on the training corpus, so it contributes nothing to any reconstruction and is never scored for interpretability. This gap between nominal and effective dictionary size can be severe: Appendix C.3 reports that an MLP dictionary with twice as many features as neurons (R=2) can end up with fewer than half of them alive enough to run autointerpretability on, meaning the dictionary is effectively less complete than the raw neuron basis it was meant to expand on. The paper's proposed fix -- periodically reinitializing dead features rather than leaving them permanently inert -- is framed as future work precisely because dead features are wasted rows, not evidence that a smaller true dictionary exists.