Triple redundancy works because you know that under nominal conditions each computer would produce the correct result independently. If 2 out of 3 computers agree, you have high confidence that they are correct and the 3rd one isn’t.

With LLMs you have no such guarantee or expectation.