ROUND 4: The 350M token floor is devastating for the primary use case (personal knowledge management with scrivener-mcp). A novelist has maybe 500K-2M tokens total. Can we reduce the floor? Ideas: (1) Pre-seed with a publicly-available word co-occurrence matrix (not an external model, just static frequencies). (2) Use document-level co-occurrence (sparser but faster convergence). (3) Accept lower quality at small corpus and improve incrementally. Derive: what quality (Jaccard between cat and feline) can we achieve at 1M tokens? At 100K tokens?
