A Published Dictionary Reads a Fine-Tuned Diarist

A sparse-autoencoder dictionary published for a base language model comes with one-line names for its features, written by an automated explainer on web text. We ask whether such a dictionary can read a diary written by a fine-tuned copy of that model, and what its names are worth on that text. The diary is sixty entries by an adapter trained on a 2012 forum corpus; the reader is Gemma Scope 2 at layer 31. Two findings. First, coverage: read through the adapted model, the dictionary explains the diary worse than matched web text on every measure (fraction of variance unexplained 1.58×, linearly predictable error 2.14×, next-token loss increase 1.50×); read through the base model, the same text is covered as well as web text (1.03×, 0.69×). The miss belongs to the adapter's shift of the model's representations, not to the text, and the excess is of the learnable kind. Second, names: at token grain, with two control notebooks written under the same protocol by the bare model and by a second adapter, the published names are correct on his text and too coarse for it. The feature named Jungian archetypes and symbols fires on "the feminine archetype of the locally adamic games"; creation and genesis stories on "Lilith and Eve being Adam's two wives, initially formed as one"; God on "the shattering being of God to be reconstituted". No feature for Lilith, the Shekhinah, the sephirot or Adam Kadmon exists at any width, and four features with unrelated names fire on his coinage adamic and on the blank line before his closing. A dictionary's names are a reading of one corpus by one explainer; a diarist whose themes lie in the tail of that corpus is read correctly and generically. We record what the session settles, what it does not, and the design it fixes for a dictionary specialised to the diarist's own library. Full HTML text: https://icra.tanazur.org/papers/dictionary-reads-the-tailor/ — ICRA pre-print series: https://icra.tanazur.org/