1 article
Same model, same task, 25-point approval gap. The difference isn't prompt engineering. It's what your AI agent can access.