My company uses Claude through our own private data centers behind our own proxy that logs all requests and responses in and out. However, Anthropic heavily steers these models during RL to respond a certain way to certain prompting, so that's basically the "secret sauce" you're thinking of.