There’s no possible way to determine that number, the reliability of polygraphs as lie detectors in situations where there’s a penalty for failure. To do that would mean that you’d have to know, independently of the polygraph, what was true and what wasn’t, and if you already know that, then the results of the polygraph can’t be high-stakes.
And why is the human operator so essential, and how do you know how good an operator is? How do you train them? If you can train a human to interpret a polygraph, then you can train a machine to interpret it, too (or contrapositively, if you can’t train a machine to do it, then neither can you train a human to do it). Polygraph data is a heck of a lot simpler than CCTV data, and even CCTV data, we’ve successfully made machines that can interpret it.