Something that bothers me about your paper, which as it stands is quite good, is that you only deal with an extremely small portion of Lojban. You don't touch at all on lujvo, which are what place structures are really for. In the case of purely compositional lujvo, e.g. kansi'u, the place structure lets you infer without any ambiguity or metaphor what the resulting place structure of the lujvo will be, based solely on the knowledge of the component gismu. As for memorizing the place structures of component gismu, one must simply realize that Lojban already *has* a case system. Simply put, rather than having named cases, we have numbered cases, and in many cases, there is a lot of parallelism between gismu of the same semantic category.
Consider {morji}, {djuno}, {smadi}, {sruma}, {jijnu}, {jimpe}, {senpi}, {birti}, {krici}, and {jinvi}.
morji = x1 remembers fact x2 about x3
djuno = x1 knows fact x2 about x3 by epistemology x4
smadi = x1 guesses x2 about x3
jijnu = x1 intuits x2 about x3
jimpe = x1 understands fact x2 about x3
senpi = x1 doubts that x2 is true
birti = x1 is sure that x2 is true
krici = x1 believes that x2 is true about x3
jinvi = x1 opines that x2 is true about x3 on grounds x4
In the "knowledge" family, which above isn't even complete, one immediately notices the parallelism. This family along with the "family" family (family-relation = x1 is the <type of relationship> of x2 by bond/tie x3; e.g. bersa) and the foods and animals family (food/animal = x1 is a <food/animal> of type/species/etc. x2) make up an extremely large portion of the gismu list.
Now, I concede that even within a family, there are some exceptions, such as birti and senpi, which don't have an "about" place at all, or jinvi and djuno which are the only ones with an x4. Even in the food and animals family, there are some notable oddballs such as lanme ( = x1 is a sheep of species x2 of flock x3). However there being a difference in the number of places *usually* isn't that much of a big deal. If someone accidentally fills in an x3 of birti, any decent listener will understand what is meant, due to the overarching systematic nature of the family to which it belongs. On the other hand, when lujvo are made (according to the rules, that is) oddball places may unfortunately need to be included, which can lead to unexpected place structures in the resulting lujvo.
e.g. jboxlajivdunsi'u = x1 are equal in that they opine that lojban is bad for x2 by standard x3 on grounds of belief x4.
If one forgets that xlali has a standard place in the x3, then if they try to fill in the belief place of jinvi, things can get messy.
Still, I don't know how your proposed case system would cooperate with lujvo (or would it simply drop lujvo from the language?) So I can't say that the status quo is any better or worse.
Finally, semantic families don't always work. {cpedu}, {minde}, and {picki} are in the same family, I'd say, but have place structures divergent in rather irritating ways.
cpedu = x1 requests x2 of x3 in manner x4
minde = x1 orders x2 to x3
picki = x1 begs x2 to x3
However, if one takes into account the emphasis difference, (when ordering something to happen, the person whom is ordered is more important, and therefore moves closer to the front,) then it becomes easier to remember. Also, cpedu2 being the action requested is more useful for translating "He asked for a glass of water." Indeed, in that case, the person of whom that action (giving a glass of water) is far less important (it could be anyone, so in truth, we're just dropping the place entirely by context).
That being said, I strongly push towards efforts to regularize the gismu list in order to make the numeric case system more self-consistent. One of Lojban's major issues is a lack of self-consistency, as pointed out by gleki in his thread about "four different vocabularies."
I started that thread with a completely different complaint, namely ....4 different but parallel sets of words/morphemes for memorising.
Not at all. It *is* possible to create a language where every gismu will have not more than two places. May be Lojban being spoken by more people will try to change into such thing. Whether such dialect of Lojban will be more easy to learn or not is another question.
mi zmadu do lo ni barda (I exceed you in property of being big) => mi poi zilkarbi do cu barda (I compared to you is big).
This is how the third place might be removed by those who hate it.
Even not touching the previous example nobody can remove metaphors from our speech. If Daniel wants to try those in Lojban why stop 'em?