Legal reasoning operates on propositions, not topics.

A holding is a specific rule announced by a specific court, often in twenty words inside an opinion of ten thousand. Everything a lawyer needs to know about it is a property of that proposition: what it says, whether it is holding or dicta, which party it favors, whether this court established the rule or applied one from an earlier case, what doctrinal category it belongs to, and whether it is still good law.

Retrieval by text similarity works on topics. Chop an opinion into chunks and embed them, and the chunk containing the critical twenty words is represented mostly by the material around them. The vector captures that the passage is about hospital negligence. It does not capture what the holding says about it. Two holdings with opposite stances on landlord liability sit in the same neighborhood of the space, because they are about the same thing.

This is why topical proximity should earn no credit in a filing-grade evaluation. The useful answer to a research question is not "cases about X" but the rule that says Y, established by court Z, still good law, favoring party P, citable in forum F. A system that surfaces the right case inside a pile of topically similar results has not answered the question, because the lawyer still has to find the proposition and test it.

For knowledge engineering, the implication is that the unit of legal knowledge should be the proposition with typed attributes, not the document or the chunk. Structure built at that level generalizes: the same holding serves a motion to dismiss, a summary judgment brief and an appeal, because the proposition does not change. Only the context of its use does.