Two passages about war. Traditional text embeddings place them close together — same topic, similar vocabulary. But one is a battle report and the other is a soldier's letter home. The function is different. The transition structure — what comes before and after, how the prose moves — distinguishes them in a way that semantic similarity cannot.
Dury trains a 29.4 million-parameter contrastive model on 374 million co-occurrence pairs from Project Gutenberg. The resulting association space clusters passages not by what they are about but by how they function within a text. “Direct confrontation” at one granularity level. “Sailor dialect” and “courtroom cross-examination” at finer levels. The clusters span thousands of books — the same functional pattern appearing across genres, centuries, authors.
The comparison to standard embeddings is sharp. Embedding similarity groups by subject: all war passages together, all romance passages together. Association-space similarity groups by register and literary mode: all confrontation passages together regardless of whether the confrontation is military, legal, or domestic. The same text gets two different addresses depending on whether you ask what it discusses or what it does.
Unseen novels map to existing clusters without retraining. The structural patterns are transferable — they describe how prose moves, not what any particular text says. But the mapping is selective. Some passages receive coherent cluster assignments; others do not. The association space is not a universal classifier. It recognizes structure where structure exists and stays silent otherwise.
The structural insight: topic is a property of content. Transition structure is a property of form. Standard NLP conflates them because both produce word co-occurrence patterns. Separating them requires training on temporal co-occurrence specifically — what follows what — rather than spatial co-occurrence — what appears near what.