This is a point I already mentioned here, but I'd like to amend it.
This is what happened some days ago:
Now this is even worse:
This comes out of a somewhat huge Total Recall TM with lots of TOCs (where term and number are separated by a tag). Excluding TR from AC is not an option here, I think.
If a kind of filter does not work here or if it is too resource-consuming, how about giving them a very low priority in the algorithm behind AC, setting them nearly automatically at the bottom of the list?
I can see the advantage of having AC with sentence marks in some case scenarios (e.g. larger chunks of text, subsegments), but to include a word + number does not makr sense (or: it is so seldom that you can easily create non-translatables or glossary entries for them).
Next stop: Filtering by words with numbers, so at least the maintenance of TMs would be easier (or does "List words with unknown spelling" catch these?)..