EDIT1: Also B5 just names 5 axes instead of MB naming (4*2) groups which always bugged me. Beyond that, because B5 does not discretize, it puts the measurement error center stage. Emphasizing measurement error is a major positive for something already so vague.
EDIT2: Also, while situation, context, how well you slept, or whatever absolutely matter, I believe this vagueness underwrites much of the "MB is pseudoscience" hostility. To some, it's so vague as to be meaningless jibber-jabber. To others, it's specific enough to be just on this side of the "science" line. Even if you think it is science, science can clearly be misapplied by those who don't understand it, and this "science" probably is..a lot (as bemoaned by the article).
As the author @dyno-might points out in fledgling steps to harden the ideas, there could be an intermediate 'x' value for each of the 4 axes. He does not say how far out from the mean one has to be "to not be in 'x'". Far enough out for each of the 4 and, boom, 99% of people are XXXX. At that point, the tests become only outlier detectors (again modulated by other, unmeasured variables) discriminating only 1% of people. Note 0.317 is also about outside 1 standard deviation for a normal curve and assuming independence, for simplicity of the example, 0.317*4 = .01. At 1.25 SDs that goes to 0.002. (Not random values - @dyno-might's distributions looked a little platykurtic.) So, it might easily be the test is only very informative (i.e. maybe predictive) for a tiny fraction of people - so not exactly unscientific, but also not exactly very useful if "over-concluding" is likely for 99% to 99.9% of the "modal" test takers. Textual profiles (or book chapters!) can take that over-concluding from just binning up a few notches.