Determiner removal in Balinese nonpivot agents

Patient-voice clauses within the symmetric voice sys-tem of Balinese disallow any extraction from the external-argument position, while definite external arguments are blocked from occurring altogether. The former fact is traditionally taken as evidence for syntactic ergativity in Austronesian. The latter fact has recently been argued to provide evidence for postsyn-tactic case licensing via adjacency with the verb. In this article, we offer a simple alternative explanation for the in-situ properties of patient-voice agents in Balinese—one that does not make reference to case. We argue that patient-voice heads come with a feature that triggers removal of the external argument’s DP shell, resulting in the loss of a determiner and a category-D feature that would otherwise enable extraction.

(3) dagang trader celeng pig uli from Badung Badung ento that onya all 'all of the pig traders from Badung' (Arka 2003: 106) Levin 2014a and 2015 discuss novel empirical data on weak (existential) quantifiers such as liu 'many' that can occur to the right or left of the head noun. In addition to the examples given by [183][184] with liu following the noun, Levin provides data where it precedes the noun: (liu). many 'The dog bit many children.' (Levin 2015: 76) Note that similar constructions exist for strong (universal) quantifiers: (5) (Onya) all cerik-cerik-e child-child-def (onya) all meli av.buy jaja. cake 'All the children bought cake.' (Arka 2003: 44) Arka, however, attributes the Q-DP order to quantifier float, a phenomenon well known from languages like German (Giusti 1990) and French (Sportiche 1988). In (6), we give further examples that demonstrate the existence of quantifier float independently. The quantifier and associated DP are in bold, separated by an adverb. (6) Quantifier float in Balinese a. Ia 3 n-yemak av-take nyuh-e coconut-def ibi yesterday makejang/liu/dadua. all/many/two 'She/he took all/many/two coconuts yesterday.' b. Makejang/liu/dadua all/many/two ibi yesterday nyuh-e coconut-def jemak=a. pv.take=3 'She/he took all/many/two coconuts yesterday.' (Arka & Dalrymple 2017: 277) From a crosslinguistic perspective (Fitzpatrick 2006, Ko 2014, Balinese quantifier float is remarkable in two ways. First, universal as well as existential quantifiers are able to undergo quantifier float and pattern alike. Second, the associated noun phrase does not have to precede the quantifier (6b), thereby making a stranding analysis (Sportiche 1988, McCloskey 2000 very unlikely.

Patient-voice in-situ agents
Balinese patient-voice constructions show a definiteness effect, in that definite DPs are illicit in the external-argument position, as shown in (7a). Definite DPs contrast with indefinite noun phrases, pronouns, and proper names in this respect, as shown in (7b) and (7c). (Wechsler & Arka 1998: 401) (Wechsler & Arka 1998: 401) (Levin 2015: 77) External arguments also cannot be dislocated from postverbal position; that is, they cannot extrapose, as in (8a), nor can they undergo topicalization, as in (8b), or wh movement, as in (8c). 4 4 We will not consider wh movement any further for two reasons. First, Kim et al. 2019 shows that Balinese wh arguments are obligatorily in situ. Note that the grammatical counterparts to (8c) that Erlewine et al. 2015, Levin 2015: 73-74, and Erlewine et al. 2017 provide, such as (i), always constitute string-vacuous wh movement, thus are compatible with a wh-in-situ analysis. Hence, the unacceptability of (8c) could be attributed to a more general constraint. (Wechsler & Arka 1998: 405) (Levin 2015: 73) (Levin 2015: 74) Levin 2014a and 2015 note two further restrictions that apply to patient-voice agents: neither weakly quantifying determiners nor adverbs, as shown in (9a) and (9b) respectively, can intervene between in-situ agents and the verb.  (Levin 2015: 73) Second, our own elicitation reveals a general dispreference for wh agents in patient-voice sentences-not only ex situ but also in situ (we thank Mike Berger for his help with elicitation and discussion): (ii) Context: At a rally, you see a child being carried away. You ask yourself who arrested the child. As will become clear in section 4, the theory proposed in this article can account for the general dispreference in more than one way. For example, wh features assumed to be situated on the D head (Ouhalla 1996, among others) will become unavailable via removal of D, which will subsequently lead to a crash at the interfaces. Alternatively, under a Q-particle theory (Cable 2010), for wh words the particle would intervene, so that the removal feature on the patient-voice head would not be able to access the D head, per the locality condition in (21).  (Udayana 2013: 52) The word-order effects in (9) have been taken as crucial evidence for an approach involving case licensing under adjacency with the verb. In our account, they will follow from a restriction on adverb placement that forms an interesting parallel to the quantifier-float properties of Balinese. More importantly, our theory provides a straightforward explanation for the restriction in (10), in contrast to postsyntactic-case-licensing theories.

PATIENT-VOICE IN-SITU AGENTS VIA POSTSYNTACTIC CASE LICENSING
Recent theories of case assignment and nominal licensing have tied word-order restrictions to an adjacency constraint. This allows nominals without a dedicated case assigner to be licensed postsyntactically via adjacency to V (Erlewine et al. 2015, Levin 2015, Branan 2017, Erlewine et al. 2017, Erlewine 2018, Erlewine et al. 2019, van Urk 2019. Within Austronesian voice systems, the movement restrictions on nonpivots are argued to follow from the lack of a case assigner/nominal licenser, which forces them to remain in a position adjacent to the verb. Balinese is of special interest since the language seems to give rise to a head-to-head-adjacency requirement, resulting in a definiteness effect. Whereas Erlewine et al. 2015Erlewine et al. , 2017 and 2019 assume the lack of a case assigner throughout the derivation, Levin 2015 pursues a derivational approach in which the licenser is lost due to a Distinctness violation (Richards 2010). We will discuss each system in turn.

A parametric account (Erlewine et al., 2015, 2017, 2019)
The main empirical observation all accounts aim to capture is that the highest nominal projection of the agent in patient voice seems to require surface adjacency with the lexical verb. While Balinese generally displays free word order, agents in patient voice must be postverbal and linearly adjacent to the verb: recall the data set in section 2.2. Erlewine et al. 2019 derives these word-order restrictions by assuming that abstract case is licensed by T and v but that Balinese patient voice is defective, so that only T acts as a licenser. 5 An additional assumption ensures that T licenses the patient and not the agent: the patient-voice head is equipped with an EPP feature, attracting the patient to an outer specifier, as shown in (11), thereby making it the closest target for licensing by T. (Erlewine et al. 2019) Since the agent is not licensed by a case assigner, it can only get case licensed postsyntactically. Erlewine et al. follow Levin 2015 in classifying this operation as local dislocation-a type of adjunction that is only licit between two elements if they are linearly adjacent (Embick & Noyer 2001). In (12), encodes immediate precedence. The bans on displacement and on intervention of adverbs, illustrated in (8) and (9b), follow directly from the linear-adjacency requirement. In order to derive the classes of nominals that are licensed as in-situ agents, Erlewine et al. 2015Erlewine et al. , 2017 and 2019 note a distributional constraint, again adopted from Levin 2015, which in turn is based on a theory of extended projections: 6 (13) Highest-nominal-head constraint The highest overt head in the extended nominal projection, whatever it is, must be linearly adjacent to the verb. (Levin 2015: 47) Erlewine et al. and Levin treat weak quantifiers as adjectives, presumably because they can occur pre-or postnominally. Thus, a prenominal weak quantifier with an in-situ agent is blocked (9a) since N as the highest overt head is not linearly adjacent to V. A definite in-situ agent is not licensed (7a) since the highest nominal head, in this case D, is not adjacent to V due to NP's interference. Pronouns and proper names (7c) are analyzed as (monovalent) D heads, 6 Levin 2015: 114-117 argues that case licensing is a subcase of the requirement that all categories must be part of a complete extended projection, which is KP for the nominal domain; see also (14) in the next section. For Erlewine et al., case-valued DPs serve as a complete nominal projection. Under the assumption that adjunction creates neutral categories, nominal arguments can become part of the verbal projection via local dislocation of the highest nominal head to V. thereby obeying linear adjacency. Finally, indefinites are licensed because they constitute NPs by assumption. 7 Crucially, the analysis cannot account for the illicitness of NP drop in patient-voice agents (10b) since the absence of an overt N head is orthogonal to (13) and thus is not excluded by it.
In addition to the possibility of postsyntactic case licensing, the theory relies on three assumptions: that the patient-voice head is not a case licenser and comes with an EPP feature; that verbs can c-select both DP arguments and NP arguments; and the constraint in (13).

A derivational approach (Levin 2015)
The original proposal of postsyntactic case licensing is put forward in Levin 2015 and provides a uniform analysis for pseudoincorporation processes found in Sakha, Tamil, and Niuean on the one hand and for in-situ agents in Balinese and Malagasy on the other. In contrast to Erlewine et al. 2015Erlewine et al. , 2017 and 2019, Levin situates the case licenser on the argument itself in the form of a K head. If noun phrases are smaller than a KP, say a DP or an NP, then the head of the highest nominal projection must get licensed by forming a complex head with the lexical verb via postsyntactic local dislocation (12). This step obviates the case filter, defined as in (14), since the nominal becomes part of the verbal projection. Non-KP nominals have to be linearly adjacent to V because this is the only configuration where local dislocation is permitted. The advantage of this account over the one presented in section 3.1 is that it dispenses with the assumption that languages like Balinese may have defective voice heads that cannot act as licensers. Instead, certain syntactic constellations can lead to the loss of a KP shell. Hence, the need for postsyntactic licensing arises in the course of the derivation. Levin proposes that reduction of in-situ agents to DP size in Balinese is triggered by a Distinctness violation of the form ⟨KP, KP⟩, created by moving the patient to an outer specifier of vP for EPP reasons. Per Richards 2010, the Distinctness constraint on syntactic structures prohibits the occurrence of identical categories too close to each other: (15) Distinctness If a linearization statement ⟨ , ⟩ is generated, the derivation crashes. (Richards 2010: 5) Although Levin does not formalize the operation, he proposes (p. 132) that a Distinctness violation is remedied by removing the agent's KP layer; compare the nodes in boldface in (16). The resulting DP status requires agents to stay merged in their in-situ positions so that they can get case licensed by local dislocation at PF. 8 7 Levin 2015, however, argues that indefinites come with an empty D head that is invisible to the application of local dislocation, thereby making N the highest overt head adjacent to V. 8 In Levin's account, the EPP feature on v is not special to patient voice; agent voice also promotes the patient to an outer specifier of vP. The in-situ position of patient-voice agents is thus not tied to the feature specification of v.
Apart from the fact that the derivational approach encounters the same issue as the parametric approach when faced with the restriction on NP drop (10b), 9 Levin's analysis faces two additional problems, both with respect to postsyntactic removal of KP layers. Firstly, contra Richards 2010: 7, Levin posits that unpronounced lower "copies are relevant to Distinctness" (p. 138). The relevant environment for Distinctness must be spec,vP and cannot be the entire phase domain, which could compute Distinctness on overt copies only. This is necessary to derive the lack of restrictions on agent-voice pivots; KP removal in this case applies to the unpronounced lower copy of the agent in order to satisfy Distinctness. The relevance of lower copies is unexpected for a constraint that was originally motivated to avoid contradictory linearization statements. In most accounts of lower-copy deletion (Bobaljik 1995, Brody 1995, hierarchical structure is a necessary ingredient to distinguish lower from higher copies. Nunes 2004: 27 explicitly argues that movement chains are reduced to the highest copy in order to avoid contradictory linearization statements. This makes it unclear how lower copies can enter linearization statements to begin with and cause Distinctness violations. Moreover, lower copies cannot be visible at the point when local dislocation applies since the intermediate copy would count as an intervener between the in-situ agent and V in (16). Consequently, Levin must assume that lower copies are visible to some PF operations but not to others. Note that these are additional assumptions that are needed apart from structure removal. This contrasts with the approach put forward in this article, where no reference is made to additional postsyntactic PF operations.
The second issue concerns the argument choice. It seems coincidental that it has to be the agent whose KP shell gets removed. Levin 2015: 140-142 addresses this issue and suggests two solutions: (i) KP-shell deletion targets the argument whose case value is not valued, or (ii) KP-shell deletion always targets the argument merged in the highest base position, that is, in the inner specifier of vP. While (i) has the potential to receive independent motivation since it has the effect of eliminating unvalued case features, (ii) is a stipulation that cannot be made to follow from anything else in the system. 9 Recall from footnote 7 that Levin postulates zero D heads for indefinite arguments, which are ignored by local dislocation. Weakly quantified arguments can be analyzed with the same empty D head. This structure, however, does not prevent NP drop from taking place since the complete absence of overt heads in the nominal projection does not violate (13).
In order to investigate solution (i), we have to take a closer look at the case system. Since Levin 2015 does not spell out the case-assignment system in Balinese, we turn to Levin 2014b, where an analysis of the Balinese asymmetric voice system is provided. This analysis crucially makes use of the assumption that the patient-voice head does not assign case to the internal argument, which thus remains an active goal and thereby undergoes case assignment by T and movement to spec,TP. In contrast, the agent-voice head assigns case to the internal argument, and although it moves to spec,vP (due to the EPP feature), it is not active anymore, thus allowing T to skip the patient and assign case to the agent, accompanied by attraction to its specifier. The two derivations are given in (17) and (18), following Levin 2014b: 297-300; the dashed arrow indicates case assignment while the solid arrows signal attraction due to case assignment. 10 10 Treating case assignment by T and movement to spec,TP as two separate operations with the former preceding the latter would not be able to derive the asymmetry of the voice system. If case assignment were to apply as an independent operation, there would be no way to exclude the possibility in (17) of case being assigned to the agent while the patient is attracted to spec,TP, ultimately resulting in unattested OVS clauses with agent-voice morphology.
It is important for the analysis of patient voice in Levin 2014b that the agent's and patient's case features are unvalued in spec,vP but that, since the patient is remerged higher, it serves as the goal for case assignment and attraction to spec,TP, effectively deriving the pivot properties of patients in patient voice. Crucially, this assumption is incompatible with solution (i). Since both external and internal argument are unvalued for case, there is no reason to pick one KP over another to circumvent Distinctness. This leaves us with option (ii)-a specific assumption about the type of argument that must undergo KP-shell deletion, which needs to be hardwired into the system. Note that this assumption is also crucial in predicting the lack of adjacency effects in agent voice, since it is here the lower agent copy whose KP shell is deleted with no consequences for overt syntax (Levin 2015, 139).
Let us take stock and summarize the assumptions needed under Levin 2015's original proposal. In addition to the option of postsyntactic case licensing and the constraint in (13), KP-shell removal must always target the argument merged in the highest base position. Moreover, the restriction on NP drop (10b) is not accounted for, forcing us to stipulate that restrictor nouns must always be spelled out in patient-voice agents; see also the discussion in section 4.3.
In the rest of the article, we will develop an idea that makes Levin's trigger for surface adjacency, namely removal of a nominal shell, the main component of the analysis. We dispense with Distinctness and encode structure removal via a feature on the patient-voice head. Thus, no effects are expected in agent-voice clauses. In-situ properties of the agent in patient voice fall out from removal of a DP layer; hence no reference to postsyntactic case licensing under linear adjacency with V is needed. In fact, our proposal derives the Balinese facts without making reference to the presence of any particular case system. The choice of targeting the agent over the patient for structure removal follows from locality considerations, without the need for an EPP feature.

PATIENT-VOICE IN-SITU AGENTS VIA STRUCTURE REMOVAL
Structure removal is formalized in Müller 2017 as an operation that removes structure within a syntactic derivation, triggered by designated [-F-] features on syntactic heads, either phrase-removal features, [-F 2 -], or head-removal features, [-F 0 -]. We derive the properties of Balinese patient-voice agents by assigning [-D 0 -] to patient-voice heads. Note that Levin 2015 also makes use of a structure-removal operation, but as a last-resort solution, in reaction to a Distinctness violation. This contrasts sharply with our implementation of structure removal as a syntactic operation to which we tie syntactic consequences such as the ban on dislocation. We will follow Müller 2009 in assuming that features on heads are ordered, thus enabling heads to merge with an argument and later remove part of that argument.
Although structure removal is a relatively recent idea, it has proven fruitful for a number of phenomena such as German complex prefields (Müller 2018) and passives (Müller 2019), 'tough' movement in English and German (Schwarzer 2016), restructuring in Russian (Dschaak 2017), and pro drop in Breton (Weisser 2019), among many others. Similar ideas have been pursued under the names Exfoliation (Pesetsky 2019, Stojković 2019) and tree pruning (Ross 1967, Embick 2010).

Removal of D
We claim that the Balinese data can be derived straightforwardly under the assumption that patient v heads bear a [−D 0 −] feature that removes the DP shell of their first-merged specifier. This naturally leads to a neutralization of definite and indefinite agents in patient voice, since that distinction is overtly encoded by the D head. The ban on definite agents in patient-voice clauses, illustrated in (7a), is thus only apparent: they are permitted in principle but exhibit a short life cycle-they are only accessible to other operations within a narrow window of time during the derivation. We illustrate our proposal in (20). In the derivation of a patient-voice sentence, a vP is built by first merging a patient v head with VP, then merging a DP in the specifier of vP. Each operation is triggered by a categorial structure-building feature, which we notate as [•X•]. The next feature on v's feature stack is the removal feature [-D 0 -]. Removing the DP shell of a head's specifier is a strictly local operation; no other D head can be found in a sufficiently local domain. This operation only applies to patient-voice constructions since only patient v heads bear a removal feature. Other v heads do not bear such a feature. We make this assumption explicit in (19).
Feature stacks for voice heads Crucially, the removal feature can only be discharged by removing the D head of the external argument. This is ensured by the Strict-Cycle Condition (Chomsky 1973): (21) Strict-Cycle Condition Within the current XP , a syntactic operation may not exclusively target some item in the domain of another XP if is in the domain of . (Müller 2017) Removing the D head of the internal argument violates (21) since syntactic operations cannot affect a proper subdomain of the vP. In other words, the D head that is removed in that operation is inside the VP, which is itself contained by the maximal projection at that point, namely the vP.
Removing the D head of the external argument, on the other hand, is licensed since the specifier is not contained by any other XP but the vP, which is the highest phrase at that point in the derivation.
Interestingly, this approach captures the intuitive similarity between a regular passive voice and Balinese patient voice. Müller 2017 and 2019 derive the German passive by placing [-D 2 -] on v, which triggers removal of a maximal DP. Hence, the difference between these two voices is reduced to different removal features on v heads. In both cases, the external argument is demoted and, thus, is less prominent syntactically. In patient voice, this is manifested as a restriction on possible argument types and a ban on movement. In regular passive, the external argument is completely banned from its base position. It should be noted that Balinese additionally has a regular passive voice with its own dedicated morphology. This is expected in our approach: it reflects the presence of another voice head bearing a [-D 2 -] feature, thereby exactly mirroring Müller's approach. 11 One of the advantages of our proposal is a simplification of the operation responsible for movement to pivot position within the Balinese voice system. Pivot properties in Austronesian voice systems have been proposed to follow-among other ideas-from base generation of the pivot in pivot position with a coindexed empty operator in the argument position (Pearson 2005) or from movement of the pivot from argument position to the phase edge of vP over potential nonpivots, making it accessible for further movement into pivot position (Rackowski 2002, Aldridge 2004, Rackowski & Richards 2005. In contrast, we derive the pivot-nonpivot asymmetry via a categorial DP-NP distinction. DP movement is often argued to be derived by a categorial feature [•D•], be it for scrambling generally or for EPP movement and object shift specifically (Chomsky 1995, Kitahara 1997). In our account, the agent has lost its DP shell; therefore only one argument DP is left. Hence, the patient DP can easily be targeted for movement to spec,TP, bypassing the agent NP. Movement into pivot position is schematically shown in (22), which constitutes the underlying structure for (1a). We assume, with Levin 2015: 104, that head movement of V via v to T ensures that the verb precedes the in-situ agent. DP movement requires a DP, but at the point where the movement-inducing feature enters the derivation the highest accessible argument, the in-situ agent, is not a DP anymore. 11 The definiteness restriction does not hold for arguments under causatives, applicatives, and intransitives (58 Udayana 2012, Udayana 2013. This is expected in our approach since different v heads can bear different feature specifications. Since functional sequences are established by feature-driven Merge, a v head involved in, for example, a ditransitive structure must select ApplP (Marantz 1993) instead of VP, thereby constituting a feature bundle that can but does not need to include a structure-removal feature. Hence, agents can be DPs.
There are two possible ways to explain the prohibited dislocated structures in (8). Either topicalization and extraposition (regarding wh movement, see footnote 2.2) are triggered in the same fashion as movement to pivot position, or these operations are triggered by a category-neutral feature that nevertheless targets the closest argument, which again is the pivot in (22).
Let us now turn to the definiteness effect (7). We claim that the presence of [-D 0 -] on the patient-voice head leads to neutralization of definite and indefinite agents. Under the assumption that definiteness is encoded as the morphosyntactic feature [±def] on D, this feature is deleted within agents of patient-voice heads. As shown in (23), deletion of the D head, including the [±def] feature, neutralizes the definiteness distinction such that the structure and featural content of a definite DP and an indefinite DP become identical. We assume that this neutralization always leads to an indefinite interpretation; see section 4.2 for more details. (23) Neutralization of definite and indefinite agents in patient-voice constructions Removal of D and subsequent neutralization explains why (i) definite noun phrases can never occur as nonpivots in patient voice (7a) and (ii) indefinite nominals are licensed (7b). In order to extend the analysis to proper names and pronouns (7c), we follow Abney 1987, Szabolcsi 1987, Longobardi 1994, and many others in assuming that all arguments constitute DPs. In line with Longobardi 1994: 650, we treat determiners co-occurring with proper names as "expletive" articles, which is meant to suggest that they do not contribute semantic content. These D heads are different from the D heads shown in (23), in that they do not change the type of the embedded NP and can be spelled out as zero. Supporting evidence for the presence of D comes from Udayana 2013, which shows that proper names and pronouns can optionally co-occur with ento/ene, as in (24a) and (24b), but not when they are patient-voice agents, as in (24c) and (24d) (Udayana 2013: 56-57) Hence, proper names and pronouns behave identically, in that they can both be accompanied by the determiner ento/ene, whose spellout is optional and whose meaning is expletive. Structure-removal derivations are given in (25) The next section will spell out the semantic compositions necessary after structure removal.

A type shift as a semantic rescue operation
Let us first address contexts involving proper names and pronouns. As already hinted at in the previous section, the absence of the expletive D head does not affect the semantic interpretation of pronouns and proper names, and neither does its presence. We model expletive D as an identity function from entities to entities, as in (26). The removal operation, in this case, targets nodes that are semantically recoverable.
(26) Determiner for proper names D = e [ ] (Following Longobardi 1994: 650) Pronouns denote indices, and proper names individuals. Hence, they both provide the right input for D. We sketch the semantic composition for the relevant removal contexts in (27). For the purpose of illustration, we adhere to a simple semantics, where V introduces the internal argument and v the external argument. As is apparent from the derivations in (27) We adopt an elaborated nominal structure for pronouns (Cardinaletti & Starke 1994, Ritter 1995, Déchaine & Wiltschko 2002. Balinese pronouns constitute full DPs as in (i). Importantly, features are encoded below the D layer by P, so that removal of the D head leaves the -feature structure intact.
(i) Balinese pronouns [ DP [ P NP ] D] (Longobardi 1994: 650) According to Arka 2003: 166 andDalrymple 2017: 267, Balinese lacks plural pronouns; features on pronouns include person and status. Gender can only be encoded on second-person pronouns. 13 A semantic framework that makes reference to event variables (Kratzer 1996, Kratzer 2000 is fully compatible with our approach. The interpretation of quantifiers has received different treatments in event semantics, ranging from obligatory QR out of the event domain (Landman 2000) to in-situ approaches (Champollion 2015).
We now turn to the more interesting definite/indefinite cases. D [+def] and D [−def] , in contrast to expletive D, are crucial in ensuring argumenthood. While the former constitutes a function from properties to individuals and is only defined for singleton properties, the latter takes a property as an argument and returns an existential quantifier. The denotations are given in (28). Both (28a) and (28b) output semantic objects directly composable with v ′ . Now, let us consider what happens when structure removal takes place. Both D [+def] and D [−def] take NPs of type ⟨e, t⟩ as arguments. If these heads are removed, we end up with semantic incompatibility between the NP agent and v ′ . We propose that the type clash can be avoided by a type-shifting operation of the a kind (Partee 1986), as formulated in (29). Together with THE and BE, (29) is argued to be a natural type-shifting operation and thus widely expected to be lexicalized across languages.
(29) A type shift ⟨e,t⟩ ⇒ ⟨e,t⟩ . ∃ [ ( ) ∧ ( )] (Partee 1986: 358) Compare (29) to (28b): the result of NP undergoing A type shift is equivalent to the result of NP being taken as an argument of D [−def] . Thus, removal of either D [+def] or D [−def] results in an indefinite interpretation. Proper names and pronouns do not have to type shift, as they are already of the right type to serve as an argument. The structure-removal derivations for DP [+def] and DP [−def] agents are given in (30) Since quantifiers can undergo QR, we expect in-situ agents to take flexible scope with respect to other operators. This prediction is borne out. Levin 2015 reports for (32) that the indefinite can take scope above or below negation. Under the assumption that negation applies at the vP level (Chung & Ladusaw 2004, Penka 2010, both readings are accounted for. 14 14 Flexible scope properties exclude an alternative repair operation for NP agents after structure removal, namely one that assumes that NPs combine with v ′ via Predicate Modification with subsequent existential closure at the vP level. This approach would predict obligatory low scope of in-situ agents in patient-voice constructions, contrary to fact. (32) Nyoman Nyoman sing neg gugut pv.bite cicing. dog 'A dog didn't bite Nyoman.' ∃¬, ¬∃ (Levin 2015: 77) One issue we have not discussed so far is the availability of different type-shifting operations. If structure removal can trigger A type shift to ensure successful semantic composition, why do we not find other type shifts, for example, THE or iota, which can also create arguments? These type shifts would result in definite interpretations of in-situ agents, contrary to fact. We argue that they are not permitted due to the Blocking Principle of Chierchia 1998: (33) Blocking Principle For any type-shifting operation and any , * ( ) if there is a determiner such that for any set in its domain, ( ) = ( ). (Chierchia 1998: 360) Chierchia proposes (33) in order to account for the fact that English bare arguments receive a kind interpretation and not a definite or indefinite interpretation, whereas Russian bare arguments allow all three interpretations. In contrast to Russian, English has overt lexical entries in the form of the and a(n). According to (33), overt lexicalized determiners win out over covert type-shifting operations. We can paraphrase the Blocking Principle along the lines of Don't do covertly what you can do overtly.
Applying the Blocking Principle to the determiner system of Balinese, it becomes apparent why in-situ agents can never be definite (unless they are pronouns or proper names). Any covert type shift that could create definite readings is blocked due the presence of the overt D [+def] head. In other words, the presence of ento in the lexicon of Balinese blocks covert type shifting to a definite interpretation. Overt definiteness, on the other hand, is blocked in selected syntactic configurations due to the structure-removal feature on the patient-voice head. 15

Adjacency effects and the restriction on NP drop
Finally, let us address the additional adjacency effects pointed out in (9) and the NP-drop facts in (10). Recall that Levin 2014a and 2015 use adverbial distribution, illustrated in (9b), as one piece of evidence in favor of the surface-adjacency requirement between the agent and the verb in patient voice. Levin argues that this intervention is related to the voice head given that there is no such ban observable for agent voice. The agent-voice counterpart to (9b) is given in (34), where an adverb is able to occur in immediate postverbal position (boldface). 15 A reviewer asks if the availability of QR in (32) commits us to covert LF movement, against the view of a single-output syntax (Bobaljik 1995, Bobaljik 2002, since type shift enables QR. The answer to this question depends on where type shifts driven by the Blocking Principle take place. Since this principle relies on spellout information, available type-shifting operations are not necessarily relevant only to LF. Further research is needed to settle this question. (34) (Sanget-sanget) very-very ia 3 (sanget-sanget) very-very nepak av.hit (sanget-sanget) very-very gamelan-e gamelan-def (sanget-sanget). very-very '(S)he was hitting the gamelan really hard.' (Wechsler & Arka 1998: 394) Since we dispensed with the surface-adjacency requirement, our theory overgenerates at this point. We account for the contrast between (34) and (9b) with an assumption about the adjunction site of adverbs. In patient-voice clauses, where the postverbal argument is the in-situ agent, an adverb could only appear in immediate postverbal position by adjoining to vP. We claim that vP does not constitute a possible adjunction site. 16 In agent-voice clauses, however, T attracts the agent, and the patient stays in situ. This derivation leaves enough space for the adverb to occur between the verb and the patient, for example, as an adjunct to VP. The two underlying structures are given in (35), with pivot movement indicated.
While this assumption might seem stipulative, it creates an interesting parallel to the second type of adjacency effect, illustrated in (9a), where prenominal quantifiers are blocked.
Recall that Levin assumes weak quantifiers to be adjectives, which lets them adjoin to the left or to the right of the NP, as in (4), the former is blocked in patient voice since it prevents the NP from undergoing local dislocation with V. There are, however, two reasons to doubt the modifier analysis of prenominal quantifiers. One is the empirical observation that modifiers such as adjectives and PPs uniformly appear to the right of the noun phrase, as was shown in (2a)  Another complication that arises with a nominal-modifier analysis involves the fact that existential quantifiers can participate in quantifier float, as was shown in (6a). Stranding the quantifier, 16 See also Sternefeld 1995 for a similar approach to Toba Batak. In order to explain the ordering properties of adverbs in Toba Batak, Sternefeld assumes that they cannot adjoin to vP. Vikner 1995 offers a similar explanation for Germanic word-order facts in terms of adjunction constraints on adverbs. as in (37), would require movement of the lower segment of an NP to which the quantifier is adjoined, presumably violating Minimality.
In light of these problems, we adopt an adverbial analysis of quantifier float (Bobaljik 1995, Doetjes 1997 in Balinese, such that the quantifier does not directly quantify over the nominal but instead forms a constituent with pro, coindexed with the associated nominal. This constituent then adjoins like and patterns in its distributional properties with other adverbs of the language. In particular, we assume that prenominal quantifiers can only result from adverbial quantification, adjoining either to TP or to VP. The structures for the prenominal quantifiers in (4) and (5) are sketched in (38), again with movement to pivot position indicated.
This analysis allows other constituents to intervene between the floating quantifier and the associated DP, a possibility seen in (6). 17 Most importantly, prenominal quantifiers are disallowed for agents in patient-voice clauses, for the very same reason that adverbs generally are prohibited in this position; compare (39) to (35a).
Hence, the two adjacency effects pointed out in (9) result from the same restriction, namely a ban on vP adjunction. 18 Lastly we turn to the restriction on NP drop in patient voice, illustrated in (10b). The present theory can account for this restriction by analyzing the NP drop in (10) as NP ellipsis. With Merchant 2001, we assume that ellipsis is triggered by an E feature on a functional head, in the present case D, which results in elision of its complement. If ellipsis is a postsyntactic process, that is, if the E feature on D instructs PF not to pronounce the restrictor NP (Merchant 2001), then structure 17 In (6a), an additional movement operation takes place that lets the associate precede the floating quantifier. This order, in fact, constitutes a common pattern of quantifier float crosslinguistically. Fitzpatrick 2006: 53 argues that cases where a floating quantifier instead precedes the associate provide compelling evidence against a stranding analysis. Hence, (6b) as well as (4) and (5) support the account of adverbial quantification. 18 A reviewer remarks that postsyntactic case licensing is superior to the present approach since it extends to pseudo noun incorporation, this being another phenomenon where structurally reduced nominals can only appear surface adjacent to verbs. There is, however, reason to doubt that surface adjacency is a requirement both in the pseudo-noun-incorporation languages that Levin 2015 considers and in languages that are not the focus of the literature on postsyntactic case licensing. Focus adverbs have been shown to interrupt surface adjacency between caseless nouns and verbs in Tamil (Lehmann 1993: 112) and Turkish (Öztürk 2009: 337). Moreover, pseudoincorporated nominals can undergo intermediate scrambling in Turkish (Öztürk 2009: 339) as well as in Hindi (Dayal 2011: 137); see also Driemel 2020a and 2020b for an overview. The present approach does not make any predictions with respect to pseudo noun incorporation. removal in syntax proper is predicted to bleed NP ellipsis, thus accounting for (10b). A bleeding interaction can also be modeled if NP ellipsis takes place in syntax, for example as proposed by Aelbrecht 2011, which argues that E features have to be licensed by c-commanding functional heads; this results in the ellipsis site being opaque. If the licenser enters the derivation after [-D 0 -] becomes active, NP ellipsis is again bled by structure removal.
Note that NP ellipsis cannot be modeled in the postsyntactic-case-licensing frameworks discussed in section 3 since the E feature present on the D or K head would mark the entire NP for elision, including any weak quantifier, contrary to what we see in (10a). Placing the E feature on the adjoined weak quantifier instead predicts NP ellipsis to be licit in not only (10a) but also (10b), contrary to fact. Crucially, the current approach is able to generate weak quantifiers outside of the nominal domain in adverbial position, so that they are unaffected in (10a). NP ellipsis is blocked in (10b) due to removal of D and thereby of the E feature.

CONCLUSION
We have presented a structure-removal approach for in-situ agents in Balinese patient-voice constructions. Removal of the DP shell through a [-D 0 -] feature on patient-voice heads neutralizes the distinction between definites and indefinites but leaves pronouns and proper names intact. Expletive D heads on pronouns and proper names are semantically vacuous; their removal does not influence the semantic derivation. Noun phrases whose DP shell has been removed undergo type shift to an indefinite meaning. Type shift to a definite meaning is blocked since Balinese lexicalizes an overt definite marker. Our approach can be extended to other voices in Austronesian, such as locative voice in Tagalog: one can posit [•P 2 •] and [-P 0 -] features that trigger merger of a locative PP into spec,vP and subsequent removal of its P head.
The main advantage of our proposal lies in its independence from any assumptions about the presence of any particular case system. Overcoming the problems with Levin 2015's last-resort KP-removal operation, our implementation of structure removal is able to capture the definiteness restriction on patient-voice agents as well as their nonpivot properties with one single assumption. Whereas the argument choice is largely stipulated in Levin 2015 or achieved by postulating an EPP feature as in Erlewine et al. 2015Erlewine et al. , 2017, and 2019, we make it follow from the Strict-Cycle Condition. Additionally, the current approach allows a straightforward explanation of why NP ellipsis is banned with patient-voice agents.
We do not claim that Balinese does not have case. The properties of Balinese patient-voice agents, however, do not necessitate an approach that makes reference to case; in particular, it is not necessary to invoke postsyntactic case licensing via local dislocation with V. Finally, both the ban on immediately postverbal adverbs and the ban on prenominal modifiers of agents in patient-voice clauses can be traced back to a general ban on vP adjunction in Balinese.

DATA-AVAILABILITY STATEMENT
Most of the data presented in this article are taken from published sources. Additional elicitation on our part was conducted in Indonesian. The data generated are available from the corresponding author upon reasonable request.