Apostrophe dictionary feature (feature 556) — demonstrates the limits of → Monosemanticity

The paper is explicit that its headline case-study feature, despite activating almost exclusively on apostrophe tokens, 'does not activate on all apostrophes.' Two other dictionary features (shown in Figures 14 and 15) fire on apostrophes too, but in different grammatical contexts: one for contractions like '[I/We/They]'ll' and another for '[don/won/wouldn]'t.' So what looks from the outside like a single 'apostrophe' concept is, in the model's actual feature dictionary, split across at least three separate monosemantic features, each narrower than 'apostrophe' and each keyed to a different surrounding grammatical pattern. Monosemanticity, as the paper operationalizes and demonstrates it, is therefore not a claim that one feature exhaustively covers an entire human-nameable category; it is a claim that each feature covers one coherent narrow slice, and the human-level category can itself decompose into several such slices.