I believe science challenges any technique, even those based on science, that are not able to predict with certainty and repeatability. Behavioral analysis is based on statistics and patterns of behavior, which as you point out are not universal across all persons, cultures, etcetera. That is why the detection of lies is described in terms of likelihood rather than certainty.

It is also true that the more samples you get for a given person, to establish baseline truth-telling as well as variances from thst baseline that suggest lying, the higher the probability of being accurate in lie detection.

You are correct that a person who tells an untruth that they believe to be true is not lying, just wrong, so there is no "lie" for behavioral analysis to detect.