The left map keeps the same broad topics and camera while you explore. Its surfaces use opaque matte shading; pale colors de-emphasize unselected areas. Topic spacing translates each shell without changing its shape. The offsets reduce an approximate overlap measure, but some intersections can remain. Spacing adjusts distances for navigation only; Setting spacing to 0% restores the exact UMAP geometry. The 2D spacing slider translates each subtopic’s contours, papers and label together for navigation. Setting 2D spacing to 0% restores its UMAP coordinates. The 2D map initially frames the main density regions; zoom out to reveal distant outliers. The right panel follows a simplified Ward dendrogram built from unwhitened PCA-128 scores; the original embedding UMAP coordinates are retained. Neither map changes the grouping.

The 2D density contours use subtle directional shading for depth; this is a visual effect rather than an additional data dimension. Shells enclose approximately 75% of a topic’s Gaussian-kernel density estimated on a 3D grid. They may be disconnected or overlap. They are visual summaries, not membership boundaries or confidence intervals. Small residual collections remain in the topic list without a scientific-topic shell. The 2D panel shows precomputed density contours with additional display smoothing. Individual paper points load when the selected branch has at most 10,000 members. Paper lists load in batches; timeline counts always include all branch members. Contours use a grid approximation to the previous Gaussian density calculation. Topic labels and keywords are searchable; paper titles are not indexed. Colors distinguish sibling topics and stay consistent across the maps, lists and timeline. Topic lists are ordered by member count.

Named topics use candidate topic labels generated from keywords and representative papers; deeper topics retain keyword descriptors. Labels remain provisional. The existing adaptive size, Ward-gap and variance-gain navigation cut is unchanged. Some terminal groups exceed the 25-paper stopping threshold. This release covers 200,000 English-eligible papers and contains metadata without abstract bodies.