And funny thing, what helped me find the issue was that I scratched the whole design and started making a belt-based one where I make all the intermediate products at the start.
So I also finished that one later, and it cuts exactly at 480 science/s with no overflow at all. But it works great and was much much simpler to make. After some testing, 100 copies of it take around 4.0 ms per update in my pc. A bit better than the other one, but not that much considering it only has a third of the inserters.
Anyway, I've submitted both designs and now I just need to wait for the results! Wish me luck