The ICRA dictionary

1080 features of Gemma Scope 2's dictionary for gemma-3-27b-it (residual stream after layer 31, width 262,144), each read two ways. The Neuronpedia label is Neuronpedia's explanation, written by auto-interpretability over a broad general corpus for the base dictionary and carried to this one by index. The ICRA reading is a declared standpoint — the way the paper's human author reads a philosophical, spiritual or literary text — written down as a frame and run by a model over the feature's own firings: for every feature, thirty windows of the author's own corpus and the works he holds formative, read through this very model, each 192 tokens long with the feature's crest at token 128. Every page prints those windows whole, the feature's kind and register, its worm in the hundred-entry diary the paper reads, and its strongest crests there. Of the 1080, 222 are threads in that diary and 170 anchor a worm that returns; 952 readings are semantic by the paper's heuristic and 128 lexicographic. Each page also prints what Neuronpedia holds for the index — both of its explanations, the explainer model, the activation density on Neuronpedia's corpus, the number of activation examples and the maximum activation; 19 of the 1080 carry a Neuronpedia label written with no activations at all, from the promoted output tokens alone, and are marked.