We build capable, safe,
and interpretable AI.

Vellum is an applied AI safety lab. Our work focuses on alignment, constitutional methods, and the long-horizon behavior of frontier systems. The papers below are written for working researchers — and for the public that pays for the field.