Two-panel UMAP scatter plot comparing clustering by reflection style vs provider.

Panel A (left): Points colored by reflection style (8 colors).
Visible clustering pattern with distinct groupings.
Annotation: "η² = 36.2%  |  Silhouette = 0.095"
Styles include: notice_uncertainty, notice_bias, notice_contradiction, etc.
Style silhouette (0.095) indicates weak but measurable clustering.

Panel B (right): Same coordinates, points colored by provider (6 colors).
More dispersed pattern ("fruit salad" effect) - providers mixed together.
Annotation: "η² = 5.6%  |  Silhouette = 0.005"
WARNING: Provider silhouette (0.005) is essentially random.
Random baseline silhouette: 0.44
Providers: Anthropic, Google, OpenAI, xAI, DeepSeek, Moonshot AI.

CRITICAL METHODOLOGICAL NOTE (from 2025-12-03 audit):
- η² (eta-squared) measures variance attribution - how much overall variance
  is attributable to group membership
- Silhouette score measures spatial clustering quality - whether points in the
  same group are actually close together in UMAP space
- Provider η² (5.6%) is misleading because providers contribute to
  variance but do NOT cluster together spatially (silhouette = 0.005)
- Provider silhouette is 93x worse than random k-means baseline

Key findings:
- Style shows both variance attribution AND spatial clustering
- Provider shows variance attribution but NO spatial clustering
- This distinction is critical for interpreting UMAP visualizations

25,000 points shown (subsampled if >25K).
Batch: full-scale-2025-11-20-v2.