The “demo goes sideways” part is dead on, CKKS can turn a tiny client-side fault into something thats hard to localize later. But its not always the ciphertext itself that matters, a flip in the encoder state or secret key material can poison every later encryption/decryption path, while a single slot error may get smeared by packing and rescaling enough that it only shows up as a small bias. That client-side detail is the one people skip over, because once the fault happens before encryption there isnt any builtin authenticity check to catch it. In practice the annoying case is not catastrophic failure, its silent numerical drift that still looks “close enough” until the last layer or thresholding step, which is a much uglier bug to debug than one obvious crash.