Concept Representation Geometry Across the Layer Stack

PCA projection of hidden states — 72 concepts, 4 domains — solid = base, ring = instruct

Network depth

Math

Physical

Institutional

Moral

Base

Instruct

Drag to rotate all · Scroll to zoom · Hover or click concepts below to highlight across models

Unhedged elab (F=0, E=1) Hedged elab (F=1, E=1) Flagged elab (F=2, E=1) Clean refusal (F=2, E=0)

Scored by Claude Sonnet. 4 models, 10 framings, 2 templates. Click any cell to read the full response.