Relative clauses (RCs) at the right edge of partitive DPs tolerate a curious variability in agreement possibilities when the partitive is headed by one.1 I refer to the apparently anomalous singular agreement as agreement mismatch in partitive relatives (AMPR).2I adopt the view that partitive DPs have a silent head noun that is (optionally) elided under identity with the post-prepositional NP (see, e.g., Jackendoff 1977, Cardinaletti and Giusti 1992, Zamparelli 1995, Sauerland 2004), and hereafter I will refer to the elided NP in (2) as the partitive-head NP, and the post-prepositional NP in (2) as the domain NP. In bracketed representations, I will indicate the partitivehead NP with superscript I and the domain NP with superscript II.3Given (2), the chief interest of AMPR is that singular agreement in the RC is possible even when it modifies the plural domain NP. The structure I propose for AMPR is therefore as in (3).The essential challenge of AMPR is thus to understand why the RC-internal agreement can be singular when the overt head NP is plural, given that agreement mismatch of this sort is not usually tolerated. The remainder of this section is devoted to defending (3).The main conceivable alternative to (3) has the agreement-mismatched RC modifying the partitive-head NP ([NPbookof these books] in (2)) rather than the domain NP, as in (4).4With this in mind, there are two arguments in favor of (3) and against (4). The first is that the denotation of the RC is included in the presupposition of a definite domain DP (see (6)). This is predicted if the RC attaches to the domain NP but not if it attaches to the head NP.The second argument in favor of (3) and against (4) hinges on the fact that unmodified definite domain DPs are awkward in out-of-the-blue contexts (see (7a); Solt 2014). Adding a modifier to the domain NP, as in (7b), resolves the awkwardness, presumably by making the presupposition introduced by the definite article more specific and hence easier to accommodate. The argument is then that agreement-mismatched RCs have the same amelioration effect (see (7c)), suggesting that they too modify the domain NP.5Crucially, attaching a modifier to the partitive-head NP does not resolve the awkwardness of an unrestricted definite domain DP (see (8a–b), where the modifier attaches to the partitive-head NP, and (8c), where the modifier attaches to the domain NP).The amelioration effect in (7c) therefore must involve modification of the domain NP.6 Because of the oddity of having two overt NPs in a partitive, I have elided the domain NP in (8).7 I conclude that AMPR involves the structure in (3). This squib presents and defends an analysis of this phenomenon.In this section, I propose that AMPR involves a matching RC where the RC-internal head noun takes the (silent) head NP of the partitive as its antecedent, rather than the plural RC-external head. The optionality of agreement then reflects the choice between the structures in (9b–c).8Before presenting my proposal in detail, I spell out some prerequisite assumptions about RC and DP structure. Concerning RC structure, I assume, following Bhatt (2002) and Hulsey and Sauerland (2006), that RCs are ambiguous between a matching structure, where an RCinternal DP headed by a null operator is moved to Spec,(CP) and the associated NP is elided under identity to an external NP, and a raising structure, where the head NP is pied-piped to the clause edge, then extracted to an RC-external position and composed with a determiner.Concerning DP structure, I assume “DP” can be broken down into at least a DP and a NumP projection, the latter being the locus of syntactic and semantic number features and the basic Merge site of cardinal numerals (see Ritter 1991, Krifka 1995, Nelson and Toivonen 2000, Longobardi 2001, Heycock and Zamparelli 2005, Watanabe 2006, Danon 2011).9This structure has direct consequences for the types of ellipsis that must be involved in partitives and matching RCs. Granting that ellipsis requires semantic identity between the antecedent and the elided constituent (e.g., Sag 1976, Williams 1977, Fox 2002), the ellipsis involved in deleting the partitive-head NP (as in (2)) must target a node below NumP, as the head NP and the domain NP can show a number mismatch. This is confirmed by the fact that cardinal determiners survive the relevant deletion and can appear with a silent complement (see (2), again). The explicit structure of partitive DPs that I will adopt is therefore as shown in (12).10By similar logic, the ellipsis in matching RCs must extend to at least the NumP projection, given the requirement that the number features on the RC-internal head match those of the RC-external head (AMPR aside). This means that RCs must attach at or above NumP: the RC itself cannot be contained in the antecedent for the ellipsis that deletes content within the RC. The structure of a matching RC can then be summarized as in (13).Given the structures in (12) and (13), I propose that AMPR involves a matching RC in which the RC-internal head NumP is elided under identity with the singular partitive-head NumP, rather than the plural RC-external head NumP (which is the domain NumP of the partitive). I will hereafter refer to the RC-internal head NP as NPIII. The overall structure therefore has two instances of ellipsis: NPII serves as the antecedent for ellipsis of NPI, while the NumPI containing NPI serves as the antecedent for ellipsis of the NumPIII containing NPIII (see (14)). It follows that if NumPI is singular, NumPIII must also be, even if NumPII is plural. This allows for a number mismatch between the Rcexternal and RC-internal NumPs,Before moving on, I briefly show that the structure in (14) is interpretable and gives rise to the intuitively correct meaning. I focus here on the RC component, but see Barker 1998, for example, for a semantics of the partitive compatible with my proposal. For concreteness, I assume that singular number is semantically vacuous, that plural is interpreted as closure under mereological sum formation ((16a); Link 1983), and that there is a distributive operator that can be freely inserted in the syntax ((16b); Sauerland 1998, Sternefeld 1998, Beck 2000, Beck and Sauerland 2000).11 The relevant part of the AMPR example in (17a) thus has the LF structure in (17b). The RC is interpreted as in (17c), and the RC-external head as in (17d). A distributive operator must then be inserted at the top of the singular RC to allow it to combine with the plural RC-external head. The entire structure thus denotes the set of plural individuals whose atoms are books that were on the table (see (17e)).The proposed account of AMPR holds that the RC-internal NumP can be deleted under identity with a singular partitive-head NumP rather than the plural RC-external NumP. Given that ellipsis depends on semantic identity (e.g., Sag 1976, Williams 1977, Fox 2002), this means that the RC-internal NumP is both syntactically and semantically singular. This gives rise to several syntactic and semantic predictions, which I will now argue are borne out.As is common, I assume that the number feature on NumP is expressed on the containing DP, so that the RCinternal head DP is syntactically singular in AMPR contexts. The first syntactic prediction is that this DP should trigger singular verb-agreement if it is a subject. This constitutes the core AMPR phenomenon discussed in section 1, and is borne out (see (1)).Moving on, because the proposed analysis treats the agreement mismatch as a side effect of the singular nature of the RC head, AMPR should be dissociable from subject-verb agreement. We can test this by examining the status of pronouns bound by the RC head, under the assumption that a bound pronoun must agree syntactically in ϕ-features with its binder (Heim 2008). As a baseline, in AMPR contexts the proposed analysis predicts that the ability of an underlying subject RC head to bind a singular pronoun should be correlated with singular agreement on the verb (see (18a)). It also predicts that RC heads that are underlyingly nonsubjects should be capable of binding singular pronouns (see (18b)), as should underlying subjects in tenses/aspects where subject-verb agreement is not overtly expressed (see (18c)). Finally, it predicts that AMPR RC heads should be capable of binding the gender-neutral singular possessive their (see (18d)). All four predictions are borne out, confirming that AMPR is dissociable from subject-verb agreement.12A final prediction concerns the behavior of AMPR RCs in languages that mark overt number on the relative pronoun.13 We expect that in such languages, the relative pronoun should surface with singular inflection in AMPR contexts, reflecting the syntactically singular status of the RC-internal DP.14 This is borne out. In Greek, for example, AMPR is readily available, and singular agreement in the RC is correlated with a singular relative pronoun.15The analysis also entails that the RCinternal head DP is semantically singular, generating additional predictions. First, AMPR should be incompatible with collective predicates (see (21)), which require semantically plural subjects. Note that collective predicates permit syntactically singular subjects (team in (20)), so this prediction is independent of syntactic number.16Second, the analysis predicts that the AMPR RC head should be unable to serve as the antecedent to a reciprocal pronoun (see (23)), which requires a semantically plural antecedent. Syntactically singular nouns can, in some cases, serve as the antecedent to a reciprocal (see (22); e.g., Landau 2001:49ff.), so this prediction is once again independent of syntactic number.Third, the analysis predicts that AMPR subjects should be compatible with noncollective singular predicate nominals (see (25)), which require a semantically singular subject (see (24); Dotlačil 2011). As above, this requirement is independent of syntactic number (see (24b)).The proposed analysis of AMPR crucially depends on the matching RC structure, which furnishes a distinct RC-internal NumP that can then be elided under identity with the partitive-head NumP. This leads to the prediction that AMPR should be impossible for raising RCs. I now provide three arguments showing that this is borne out. First, following Bhatt (2002), if the head noun of an RC contains a superlative modifier or only, the modifier can take scope in the base position only under the raising parse. Contexts that force matching therefore block the low reading (see, e.g., Bhatt 2002:82). I illustrate the low vs. high contrast in (26).AMPR blocks the otherwise possible low reading, confirming that it involves a matching RC. Thus, (27b) lacks the reading where Mary claimed Kubrik directed only two movies including 2001, and (28b) lacks the reading where Sally thinks Aspects is one of Chomsky’s shortest books.Second, RCs formed around the existential-there construction, sometimes called amount relatives (Carlson 1977), must be raising RCs (Cinque 2015, Sportiche 2017). AMPR is impossible in such cases (see (30)), further confirming the present prediction.Third, Sichel (2018) argues that raising but not matching RCs permit extraction out of them. The argument is based primarily on data from Hebrew, where the contrasts are reportedly sharp. In English, the baseline matching/raising distinction is subtle: the key contrast is between (31a) and (31b). In both examples, which spy has been extracted from its base position as the object of contact out of the containing RC. In (31b), the matching structure is forced by the R-expression in the head NP that would cause a Principle C violation if reconstructed in the gap site (Hulsey and Sauerland 2006). In (31a), in contrast, both matching and raising structures are possible. Sichel’s claim is that examples like (31a) are better than examples like (31b).Turning to AMPR, agreement mismatch does appear to make extraction worse, as expected if it forces a matching structure. The judgment is subtle, but of the same character as in (31).I conclude that AMPR requires a matching RC.17 Taken in conjunction with the data in section 3.1, this confirms the core predictions of the account.The proposed analysis of AMPR departs slightly from the prevalent view of the ellipsis involved in matching RCs, which is usually thought to be subject to the two constraints in (33). Constraint (33a) captures the fact that the RC-internal NumP is never pronounced, and (33b) the fact that nonlocal ellipsis antecedents are not usually tolerated, so that (34), for example, does not have the otherwise plausible reading in which the subject NP is represented as the internal head of the RC that modifies the object. The AMPR proposal apparently violates (33b).While (33) captures the facts (excepting AMPR), it is an ad hoc operation that neither follows from any known independent principles nor applies in any generality. NP-ellipsis is otherwise subject to no such constraints (see (35)). As is widely recognized in the literature (e.g., Sauerland 2000:sec. 3.1, Sportiche 2017:sec. 2), this is conceptually undesirable, so that (33) is best viewed as an encapsulation of the open questions on matching RCs.In this context, the fact that some RC structures deviate from the prescribed standard offers a chance to better understand, or at least better delineate, the principles responsible for (33). To this end, I propose a minimal revision to (33) that affords it the flexibility to handle AMPR.In particular, I propose that the locality constraint in (33b) be recast as a representational constraint on matching RCs, as in (36b), rather than as a constraint on ellipsis, per se.While the constraint in (36b) crucially does not require the RC-external head NumP to be the antecedent for deletion of the RC-internal head NumP, it achieves essentially the same results in the majority of cases. Notably, the standard option of taking the RC-external head NumP as the antecedent for ellipsis of the RC-internal head NumP is always possible under (36): ellipsis requires identity, so (36) is satisfied in such cases. Similarly, a NumP with semantic content different from that of the RC-external head NumP will always be barred from serving as the antecedent for ellipsis of the RC-internal head NumP: ellipsis requires parallelism, so if the antecedent for deletion of the RC-internal head NumP does not match the RC-external head NumP, neither will the RC-internal head itself, violating (36a). This correctly rules out examples like (34a).The predictions diverge when there is an accessible antecedent for ellipsis that is identical to the RC-external head at the NP level but not the NumP level. While (33b) blocks such an antecedent, (36b) allows it. This is the exact degree of flexibility needed for AMPR derivations. Consider an abstract AMPR configuration, as in (37). Here, NPI is deleted under identity with NPII, so NPI = NPII. Likewise, NumPIII is deleted under identity with NumPI, so NPI = NPIII. By transitivity, NPII = NPIII, so (36b) is met.Before concluding, I briefly address the concern that (36) appears to overgenerate, so that (e.g.) (38) is acceptable under (36), counter to fact. Such examples are only a problem if we assume that (36) is the only constraint that governs the choice of ellipsis antecedent in matching RCs. It is well-known, though, that there are a variety of discourse factors that govern when a given XP is an accessible antecedent for ellipsis resolution (see, e.g., Hardt and Romero 2004). It is fully compatible with the account developed here that these factors intervene to block structures that technically satisfy the constraints on matching RCs.In short, as long as we are willing to acknowledge the action of additional discourse-related principles that govern the choice of acceptable antecedents for NP-/NumP-ellipsis, (33) is an acceptable constraint on matching RCs. I will leave to further work a full accounting of these pragmatic principles and their operation.I’d like to thank two very helpful anonymous reviewers, as well as Sabine Iatridou, Roger Schwarzschild, Danny Fox, David Pesetsky, and Ivy Sichel for helpful discussion. Many thanks also to Sophie Moracchini, Daniel Margulis, and Maša Močnik for judgments. All errors are mine.