Yeah. Also, we think that in certain scenarios, it's possible for "cryptographic patterns" to arise in model weights via normal training, especially if models are trained against things such as honesty probes.