Sometimes “Loft” will create some inconsistent results. Just had a situation in which it happened again. I use totally the same mirrored objects as an input geometry, so there should not be an issue. However, if you check the video, Rhino seems to get confused about the isocurve direction to which it matches the tangency of the lofted surface to the target surface edge:
This issue looks quite similar to the one that happens with the “Match target isocurve direction” option of the “Match surface” tool: