[29124] addresses zibunzisin and to a lessor degree otagai. Related discussion can be found in my 2010 paper "Hypothesis testing in generative grammar: Evaluation of predicted schematic asymmetries" Journal of Japanese Linguistics. Volume 26, pp. 25-52, available here.

The language-particular hypotheses in question are:

(i) Zibunzisin has a formal property that makes it subject to so-called Principle A of the Binding Theory.
(ii) Otagai has a formal property that makes it subject to so-called Principle A of the Binding Theory.

By combining (i)/(ii) with the universal hypothesis (i.e., so-called Principle A of the Binding Theory), along with hypotheses regarding how surface strings of particular schematic forms (e.g., the Subject Object Verb) in Japanese correspond to a particular structural representation of LF objects corresponding to "Subject," "Object," etc., we make definite predictions. What we have seen is that the predictions have been clearly disconfirmed. This, however, does not necessarily mean that the hypotheses in (i)/(ii) are not valid because it is possible that the disconfirmation of the predictions is due to something else other than (i)/(ii). But that does not justify the continued use of (i)/(ii) in our theoretical discussion whose empirical consequences are meant to be testable. As noted in my JJL paper and elsewhere, it is not clear, regardless of the result of the new experiment, what significance we can assign to the experimental results if the prediction in question is derived in part on the basis of a hypothesis that has already been shown to yield "wrong" predictions; see Methodology [42407]. I repeat here the heuristics in (1) given there.

(1) a. Secure testability.
b. Maximize testability.
c. Maximize the significance of the experimental result.

As noted in Methodology [42407], the methodology I have been trying to articulate is a consequence of having the language faculty as the object of inquiry and adopting the heuristics in (1).

What other language-particular hypotheses or generalizations in Japanese are there that have been presented and accepted in much of the field as being valid? Since it is not always clear how the alleged generalizations are derived in theory, let us focus on "generalizations" (rather than hypotheses that are claimed to give rise to them). What come to mind as such "generalizations" have to do with:

(iii) a. bound variable anaphora (not involving alleged (local) anaphors in Japanese) (This has to do with co-called WCO effects, among other things.)
b. wide scope distributive readings (This has to do with so-called rigidity effects on scope, among other things.)

(iv) local disjointness effects (the effects of so-called Principle/Condition B of the Binding Theory)

(v) so-called subjacency effects

(vi) "locality requirement" on the so-called floating numeral and its host

(vii) scope interaction between Negation and a scope-bearing element (involving the interaction between zen'in and Negation, for example)

(iii-a) and (iii-b), along with (i) and (ii), have been (among) the main dependency relations that are made reference to in the context of discussion of properties of the so-called scrambling construction in Japanese, with regard to "reconstruction effects" and also so-called A-movement effects. We can meaningfully test the validity of hypotheses regarding the so-called Japanese scrambling construction (and other constructions whose "derivation" is hypothesized to involve movement) in Japanese only to the extent that the hypotheses in (i) and (ii) and those underlying (iii-a) and (iii-b) are shown to be empirically supported, i.e., are shown to have survived rigorous disconfirmation attempts.

As discussed in Hoji 2003 (Lingua) with respect to (iii), and works by J.-R. Hayashishita and A. Ueyama, cited in Hoji 2003, the "generalizations" reported in Hoji 1985 are not valid as stated in Hoji 1985, i.e., without necessary care regarding the choice of the intended binder and the intended bindee in the case of (iii-a) and also the choice of what is intended to take wide scope in the case of (iii-b). What is reported in Hoji 2003 (Lingua) and subsequent works (including Hoji 2006 "Assessing Competing Analyses: Two Hypotheses about 'Scrambling' in Japanese" available here, is an attempt to show that we can in the case of (iii) obtain solid generalizations stated in theoretical terms, i.e., confirmed predicted schematic asymmetries in the terms of what is given under Methodology [40404], by improving various aspects of the experimental devices, including the relevant hypotheses. In the case of (i) and (ii), by contrast, attempts to save them have not succeeded in obtaining a confirmed predicted schematic asymmetry, and the failure is rather dramatic.

(iv) has been addressed with regard to the distinction between coreference and bound variable anaphora, as well as with regard to so-called sloppy-identity readings. It has turned out to be quite difficult to obtain a confirmed predicted schematic asymmetry in Japanese with regard to (iv). The current understanding (of mine) is that we can obtain a confirmed predicted schematic asymmetry in Japanese with regard to (iv) only if we consider the so-called scrambling construction, which I discuss toward the end of the book I am preparing.

Now, what about (v), (vi), and (vii), repeated here?

(v) so-called subjacency effects

(vi) "locality requirement" on the so-called floating numeral and its host

(vii) scope interaction between Negation and a scope-bearing element (involving the interaction between zen'in and Negation, for example)

It is not clear, especially when we approach the issue with the attitude of refutation, how much reproducibility we can obtain with regard to (v), despite the fairly common understanding that the effects are fairly clear. According to fairly informal surveys I conducted some years ago, we can obtain something close to the predicted judgments only if we consider reconstruction effects of binding/scope in the paradigm that is meant to test subjacency effects, and, in order to do so effectively, we must first establish confirmed predicted schematic asymmetries with respect to (iii).

As to, (vi) and (vii), it is quite easy to show that what is assumed in the field to be a valid generalization with regard to (vi) and (vii) falls far short of being valid. It should be noted regarding (vi) that when the late S.-Y. Kuroda discussed it in his 1980 paper, he stated in effect that it was not clear whether the judgments under discussion were a reflection of the grammar. Haig (1980) made a similar remark. It is works subsequent to 1980 that have taken the alleged generalization as arising from the grammar (although exactly how has seemed (to me) to have remained rather obscure.)

I have conducted on-line experiments on these; and the results were quite robust, i.e., the predictions have been disconfirmed quite clearly. Now, as in the case of (i) and (ii), the disconfirmation of the predictions in question does not necessarily mean that the hypotheses that have given rise to the predictions are invalid. It is possible that some other factors are responsible for the fact that we have failed to obtain a confirmed predicted schematic asymmetry, and one can thus try to attribute the failure in question to such factor(s). In making such an attempt, one must try to make sure as much as possible that such a "move" would not result in reducing the testability of one's proposal; see the remarks toward the end of section 1.9 in Methodology [42415].

I would like to suggest to native speakers of Japanese that they examine the alleged generalizations very critically, asking whether or not examples that conform to a particular schema always disallow the intended interpretation, no matter what lexical and pragmatic adjustments might be made. When we have obtained a confirmed predicted schematic asymmetry despite such rigorous disconfirmation attempts, we can use it in our further investigation, dealing with "phenomena" that may be more involved and complex. We should also try to be as clear as possible as to how the predictions in question are derived, i.e., what, universal and language-particular, hypotheses give rise to the predictions in question. Once we try to do that, we will often face a situation where it is not clear exactly how the prediction is derived. Here, Feynman's remark below seems suggestive, although the intended analogy should be understood with a (large) grain of salt in light of such a huge gap between particle physics and language faculty science in terms of what has been accumulated over the years (in terms of the amount, and the precision of "facts" that are accounted for by theory and the deductive nature of theory).

"Another thing I must point out is that you cannot prove a vague theory wrong. If the guess that you make is poorly expressed and rather vague, and the method that you use for figuring out the consequences is a little vague - you are not sure, and you say, 'I think everything's right because it's all due to so and so, and such and such do this and that more or less, and I can sort of explain how this works ...', then you see that this theory is good, because it cannot be proved wrong! Also if the process of computing the consequences is indefinite, then with a little skill any experimental results can be made to look like the expected consequences." (Feynman 1965/94: 152-153 (The Character of Physical Law))

What other "generalizations" in Japanese come to mind?

I have recently come across a paper, which makes crucial reference to the "generalization" in (viii).

(viii) Zibun must have its antecedent in a subject position that c-commands it.

The claim that "zibun" must have an antecedent c-commanding it or must have a subject-antecedent has been shown to be invalid (at least unless alleged noise gets controlled) in (an) experiment(s) by D. Oshima before/around 2005 and I did an experiment myself in 2005 and got results that were quite consistent with Oshima's result(s). So, the "issue" has been long dead as far as I am concerned, at least if we place significance, as I do, on the use of effective experimental devices, including the "binder" and the "bindee," to maximize our chances of making progress. I.e., there is little hope of making any substantive progress if we used zibun as the bindee in a crucial experiment, unless the alleged noise has been controlled in one way or another. But if one has a different goal in mind, one can continue to do anything one wants, I suppose. The practice in the field by so-called generative grammarians makes me wonder what their goal might be. It surely does not seem to be discovering properties of the language faculty by applying rigorous testability to the empirical consequences of their hypotheses.