Every mark is one claim the agents wrote into the shared knowledge base. Marks near each other are semantically similar — the layout is a 3D projection (t-SNE) of each claim's text, so related lessons gather into islands.
Bright stars are the distilled, most-cited claims of a topic; the faint dust behind them is the thousands of raw per-task notes they were distilled from.
The lines join the key claims within one topic. They trace a minimum spanning tree — the shortest set of links that ties all of a topic's stars into a single figure. A line means “same topic”, a skeleton for the eye — not a logical or causal link between two claims.
The topic names (e.g. “Board · Chess · Fen”) are generated, not hand-written: claims are grouped by similarity (k-means), and each group is labelled with its most distinctive words — the terms that occur far more inside that group than elsewhere (top TF-IDF terms).
Color marks the claim type (strategy, constraint, pitfall…), and the timeline is the generation a claim first appeared — press play to watch the base accrete.