One issue I believe is the fact that EAP as it stands (and as it has been specced as at 4 cameras?) looks the world through a rather narrow box, which is fine when the lane is uniform, predictable and not too wide. In this case the lane and its markings are all over the place.
Second thing is that what we see on the IC obviously is not what EAP "sees", but a graphical approximation/averaging of some parameters coming from AP2, so as the lane "lives", the approximation "lives" too. This is made worse by the fact that EAP only sees a small portion of the road at a time anyway...
So if a lane were to go zig-zag or the lane lines made an unexpected detour, the IC still would display it as a straight line or a gently curving road or lose lane markings for a while, or alternate between a couple of states, but not show a zig-zag on the screen. Partially because EAP would only see a portion of it at a time, but mostly because the IC clearly is not getting such parameters that would allow it to draw a complex lane shape.
So no matter how accurate or inaccurate EAP may be at this time, the IC seems to be always averaging it out. Sometimes this helps give a better-than-real impression, and at other times it makes for a worse impression.
Now, the third thing is the driving algorithm, it probably also averages things out these days and does not follow every momentary stray the NN might suggest.