Here is my post on reliability amplification, which requires some constant error probability to work (the threshold might end up being < 1/3, but it doesn’t really matter). In practice, I think that the only way security amplification is plausible is by using a lot of randomness and reliability amplification.