Principles of Privacy for Alignment Research — LessWrong