Bruening (2019) proposes an analysis of English do so anaphora that accounts for its compatibility with A-movement contexts including unaccusatives and passives. His key insight is that do so consists of the intransitive verb do and an adjunct so (cf. Bouton 1969; Ross 1972; Houser 2010), denoting λe.f(e), where function f is replaced by finding a linguistic antecedent (Hankamer & Sag 1976).

For transitive VP antecedents as in (10), Bruening’s analysis works as follows:

In the antecedent clause (10a), transitive Voice (Voice_tr) introduces an external argument with an Initiator role (agents/causers) at its specifier position. In the target clause (10b), VP_d (do so) initially denotes λe.f(e), then function f is replaced by the semantic value of the antecedent VP, yielding λe.decorate(e, the eggs). The remaining semantic composition proceeds as in the antecedent structure.

Another crucial component of Bruening’s analysis is the A-Movement Rule, which handles passives and unaccusatives:

Bruening identifies unaccusative Voice and passive Voice as A-movement heads under this rule. This rule serves as a mechanism for lambda abstraction in A-movement contexts, allowing elements that have undergone A-movement to be properly interpreted in their derived positions.

To illustrate this rule’s application, consider an unaccusative antecedent:

Unaccusative Voice (Voice_un) differs from transitive Voice in that it introduces no Initiator (12b). As an A-movement head, it triggers the A-Movement Rule, lambda-abstracting over the VP-internal trace (12c). Syntactically, Voice_un induces A-movement of the underlying object NP to the specifier position of VoiceP, where semantically it saturates the lambda-abstracted variable (12d).

Based on this process, the corresponding do so sentence derives as follows:

The copied denotation here includes a trace. Voice_un triggers the A-Movement Rule, creating lambda abstraction over the trace. VoiceP(1) becomes a predicate of individuals applying to the NP in Spec,VoiceP (13d). This is how a base-generated subject of do so is interpreted as a logical object when the antecedent supplies a trace-containing function.

In summary, Bruening’s analysis provides a unified account for do so anaphora in both transitive and unaccusative constructions through three components: (i) semantic copying, (ii) the A-Movement Rule, and (iii) distinct Voice heads.

Building upon Bruening’s (2019) framework, we now analyze Japanese VP anaphora with focus on transitivity mismatches. As established in section 2.2, Japanese exhibits both causative-to-inchoative and inchoative-to-causative mismatches, suggesting that soo anaphorically refers to a result state encoded in the antecedent, while su/nar are morphological realizations determined by the transitivity of the target clause.

To account for these observations, I propose decomposing the traditional VP into VP_Res (encoding a result state) and vP (encoding causal relation). The little v head bears either an inchoative or causative semantic value:

This analysis posits a causative relation between an event variable e and a state variable s in both lexical causatives and inchoatives, consistent with Kratzer (2005) and Schäfer (2025). The key distinction is that causative v includes an activity predicate, which is absent in inchoative v, capturing the semantic difference between these event types.

Under these proposals, the causative and inchoative variants with the root sim- ‘close’ are analyzed as follows:

These structures reveal key distinctions between the variants: transitivity-dependent morphology realizes the v head (Akimoto 2018), with v_caus realized as -e and v_inch as -ar. Semantically, v_caus introduces both a causing event and an activity predicate, while v_inch introduces only a causing event. The structures also differ in their Voice heads: causatives employ Voice_tr, which introduces an Initiator, while inchoatives use Voice_un, which does not. Following Bruening (2019), the inchoative structure contains a trace within VP_Res, associated with the internal argument that undergoes A-movement to Spec,VoiceP. Voice_un, as an A-movement head, triggers the A-Movement Rule, lambda-abstracting over this trace and thereby explaining how the internal argument becomes the subject of the inchoative while maintaining its thematic relation to the result state.7

Building on this structural analysis, I now examine how soo interacts with these VP structures to account for transitivity mismatches. I propose that the anaphoric element soo has the following denotation, parallel to English do so in Bruening (2019):

Function f is replaced by the semantic content of the antecedent VP_Res. This allows soo to preserve that result state while the surrounding verbal structure (v and Voice) determines the transitivity of the clause. Consider first a causative-to-inchoative mismatch case in (7c), reproduced in (18).

The derivation proceeds as follows: VP_soo initially has a vacuous denotation (18b). Its semantic content is derived by replacing function f with the denotation of the antecedent VP_Res (18c), resulting in (18d). Crucially, this operation copies only the result state, excluding the causative semantics introduced by v_caus in the antecedent. This mechanism preserves the core meaning (‘the closing of the door’) while allowing an event-structure mismatch: the causing event in the target is instead introduced by v_inch (18f), which lacks the activity predicate and external argument of the causative variant. Voice_un applies without introducing an external argument. Since the copied VP_Res from the causative antecedent contains no trace, the A-Movement Rule does not apply here, despite the presence of Voice_un. The inchoative v is realized morphologically as nar, contrasting with the regular inchoative morphology -ar in verbs like sim-ar.

A consequence of the causative-to-inchoative analysis in (18) is that the target clause does not admit an overt subject. Since the copied VP_Res supplies ‘the door’ as the logical subject of the target clause, this argument is already fixed, and an independently realized overt subject is not available, as (19) shows.

In (19a), the overt pronoun sore ‘it’ appears as the subject, and in (19b), the full NP ano doa ‘that door’ does; both are unacceptable. With ‘this door’ already fixed as the logical subject, there is no room for an additional overt subject, whether it re-expresses that argument (sore) or introduces a distinct one (ano doa).8

However, an apparent exception exists. Consider the following example:

Here, an overt subject (kono doa-wa ‘this door-top’) appears grammatically in the target clause (20b). The key difference is that in (20a), the antecedent object appears with the topic marker -wa, functioning as a contrastive topic rather than a simple accusative object. Topic elements undergo movement from their base position to a higher position (Saito 1985), and the VP_Res in the antecedent contains a trace of the topic-marked phrase: [VP_Res t₁ close], where t₁ is the trace of the topicalized object. When this VP_Res is copied onto soo, the unaccusative Voice head of the target clause, being an A-movement head, triggers the A-Movement Rule, thereby licensing an independently realized topic-marked subject. The structural representation and semantic derivation of (20b) proceed as follows:

This accounts for the apparent exception: a trace rather than a full NP in the copied VP_Res licenses an independently realized subject in the target clause.9

Next, I analyze an inchoative-to-causative mismatch case, reproduced from (8c) and illustrated with its structures and semantic compositions in (22):

The antecedent shows a typical unaccusative structure where the internal argument has undergone A-movement to Spec,VoiceP. VP_Res in the antecedent contains a trace t₁ (22b), which is copied onto VP_soo (22c). The target then projects transitive layers, v_caus and Voice_tr, introducing causative semantics (22e) and (22g). In this context, causative v is realized as su rather than the typical causative morpheme -e found in regular causative verbs, paralleling how nar appears in place of -ar in the causative-to-inchoative case.

This configuration differs from the causative–inchoative mismatch in an important respect. Since Voice_tr is not an A-movement head,10 the A-Movement Rule is not available here. Voice_tr instead introduces an external argument with an Initiator role in its specifier. The question, then, is how the trace t₁ in the copied VP_Res is interpreted.11 Since there is no local binder for this trace within the target VoiceP, the interpretation would be problematic if the trace t₁ were treated as an ordinary A-trace/anaphor.

As Bruening (2019: 35) shows, VP-internal traces have two possible interpretations: (i) as variables abstracted over by the A-Movement Rule, or (ii) as copies under the copy theory of movement (Chomsky 1993), interpreted as identical to the moved NP.12

In the present inchoative-to-causative case, the copy interpretation obtains. The trace t₁ in the copied VP_Res is interpreted as identical to ‘the door’ from the antecedent (22a). Accordingly, the target VoiceP receives the denotation in (23).

This interpretation specifies that the causing event denoted by soo su has the same theme argument as the inchoative event in the antecedent—namely, ‘the door.’13

Evidence for decomposing the traditional vP into VoiceP and vP emerges from the distribution of su ‘do’ and nar ‘become’ in Japanese. With non-finite stative predicates, these verbs show systematic alternation based on transitivity:

Example (26) illustrates the alternation with a canonical adjective (akaku ‘red’), while (27) shows the same alternation with a nominal adjective (kirei-ni ‘clean’). In this alternation, transitive variants (a) involve su appearing with two arguments—an external argument (nominative) and an internal argument (accusative). Intransitive variants (b) employ nar, where the internal argument becomes the subject with nominative case.

This correlation indicates that the su/nar alternation is linked to argument structure, particularly the presence/absence of an external argument. The stative predicate remains constant across variants, suggesting that the alternation is governed by syntactic configuration rather than the semantic content of the result state.

Based on such data, Sakai et al. (2004) identify su and nar in stative predicate constructions as light verbs rather than lexical verbs, demonstrating that the stative predicate—not su/nar—plays the central role in theta-role assignment to the theme argument.14 This becomes evident when we compare stative predicate constructions with su/nar to those formed with a lexical verb instead:

The crucial distinction lies in the optionality of the predicate: with lexical verbs like som-e/som-ar ‘dye (vt/vi)’, the stative predicate can be omitted, so the internal argument is theta-marked by the verb. In contrast, with su/nar, it is obligatory (30)–(31), showing that here the theme is theta-marked by the predicate, not by su or nar, confirming their light-verb status.

Based on these observations, Sakai et al. (2004) propose the following syntactic structure for stative predicate constructions, where su and nar instantiate Chomsky’s (2001) transitive v* and intransitive v, respectively:15

In this binary approach, su realizes transitive v*, which introduces an external argument and assigns accusative case, while nar realizes intransitive v, which does neither.

Sakai et al.’s (2004) analysis parallels analyses of Japanese morphological transitivity alternation proposed by Nishiyama (1998) and Hasegawa (2001), as seen in pairs like:

Here, the causative variant uses the morpheme -s while the inchoative uses -re, both attached to the root kowa- ‘break’. Nishiyama (1998) and Hasegawa (2001) propose the following structures for these alternations:

This parallel treatment suggests that just as the causative morpheme -s and inchoative morpheme -re realize v* and v respectively in morphological alternations, su and nar can be analyzed as their light verb counterparts in periphrastic constructions, providing a unified account of both alternation types in Japanese.

While the binary distinction between v* and v successfully accounts for light verb distribution in stative predicate constructions, it proves inadequate for capturing their behavior with verbal nouns (VNs). Since Grimshaw & Mester (1988), su has been widely recognized as a light verb when combined with VNs such as Sino-Japanese verbs (e.g., Miyagawa 1989; Saito & Hoshi 2000; Fukui & Sakai 2003):

The present paper focuses on the non-accusative VN-su construction of the type exemplified in (35b), as opposed to the accusative VN-o suru pattern exemplified in (35a), which is set aside here. What matters for this construction is that some VNs can be used both transitively and intransitively. As noted by Kageyama (1996: 202), this class includes items such as kakudai ‘expansion’, syukusyoo ‘reduction’, and zitugen ‘realization’. Even in these cases, however, the light verb remains su in both the transitive and intransitive uses; it does not alternate with nar.

This pattern reveals a fundamental asymmetry in the distribution of light verbs: while su participates in transitivity alternation with nar when combined with stative predicates, it fails to do so when combined with bare VNs, regardless of whether the VN itself allows both transitive and intransitive uses. This asymmetry is summarized in Table 1.

Notably, a small number of VNs also function as -ni-marked stative predicates, though speaker judgments vary. One such item is tyuusi ‘cancellation’, illustrated in (37)–(38). Since the intransitive bare-VN use in (37b) and the transitive -ni-marked use in (38a) appear to be the forms most subject to this variation, naturally attested web examples are provided for them.

Thus, tyuusi occurs with su in its bare VN use, but shows the expected su/nar alternation when used as a -ni-marked stative predicate.

Returning to the main asymmetry summarized in Table 1, it challenges Sakai et al.’s binary analysis, which treats su and nar as realizations of v* and v, respectively. Under their analysis, su in VN contexts should alternate with nar, contrary to observations. While Sakai et al. (2004: 372) suggest that “[Noun+suru] as a whole has similar functions to [V+v*] complex,” this fails to explain why [Noun+naru] cannot function similarly to [V+v] in unaccusative variants.

The key to resolving this puzzle lies in recognizing that in VN-based alternations, transitivity is an inherent property of the VNs themselves, rather than a function of the light verb (e.g., Miyagawa 1989; Kageyama 1991). While this distinction is not immediately apparent with VNs like kakudai/syukusyoo ‘expansion/reduction’ in (36), which can appear in both contexts without morphological change, most VNs show clear transitivity-dependent distribution:

Here, bakuha and bakuhatu, while sharing the core meaning ‘explosion’, differ in their inherent transitivity: bakuha is inherently transitive, while bakuhatu is inherently unaccusative.

This lexically encoded transitivity becomes even more evident in deverbal N–V compounds, where a noun combines with the infinitive form (renyookei) of a verb. Some of these compounds take the light verb su, paralleling Sino-Japanese verbs, and exhibit morphologically transparent transitivity alternations.16^,17

The compound consisting of the noun ne ‘price’ and the verb root ag- ‘rise’ shows systematic morphological alternation: causative morpheme -e in transitive contexts and inchoative morpheme -ar in unaccusative contexts.18 This pattern precisely mirrors the simplex verb pair ag-e ‘raise’ and ag-ar ‘rise’:

These parallel morphological alternation patterns indicate that transitivity in VN-su predicates is encoded within the VNs themselves, not in the light verb su.

For stative predicates, in contrast, transitivity must be encoded in the light verb itself, as evidenced by the morphological su/nar alternation. This necessity stems from adjectival predicates lacking internal mechanisms for transitivity alternations:

Both nominal adjective kirei-da and canonical adjective atataka-i are inherently one-place predicates that can only take a subject, as in (42a). Their inability to license an external argument and accusative object is shown by the ungrammaticality of (42b).19

To summarize, Japanese light verbs exhibit two distinct patterns depending on the predicate type they combine with:

This generalization reveals a fundamental architectural property of the Japanese verbal domain: the morphological realization of light verbs systematically reflects the locus of transitivity specification in the predicate-light verb complex.

To illustrate this proposal, let us examine the structures of stative predicate constructions like those in (27a) and (27b):

In the transitive variant (44a), v_caus combines with AP to encode a causative event, while transitive Voice (Voice_tr) introduces the external argument. In the intransitive variant (44b), v_inch takes AP as its complement to denote an inchoative event, with unaccusative Voice (Voice_un) projecting no specifier position. Morphologically, these structures are realized through head movement: v merges with Voice, with v_caus surfacing as su and v_inch as nar, while Voice is consistently null (Ø). This analysis distributes the determination of argument structure across two functional heads: v specifies event type (causative/inchoative), while Voice regulates external argument presence, providing a more nuanced account than Sakai et al.’s (2004) binary v*/v distinction.

Turning to the VN-su construction exemplified in (36), I propose the following structures:

In both variants, the VN root projects its own phrasal structure (VNP), selects its internal argument, and undergoes head movement to v.20 In the transitive variant, Voice_tr introduces an external argument above vP_caus; in the intransitive variant, Voice_un merges with vP_inch without introducing an external argument. Crucially, the morphological realization pattern here is inverted compared to stative predicate constructions: with VNs, v is consistently null (Ø) while Voice invariably surfaces as su, regardless of transitivity. This contrasts with stative predicate constructions where v is overtly realized as su/nar while Voice is null.

The asymmetric distribution of the light verbs su and nar then follows from the distinct structural positions these light verbs occupy within the verbal domain. This distribution is summarized in Table 2.

This raises three interrelated questions: (i) Why does Voice realize as su with VNs but remain null with stative predicates? (ii) Why does v alternate between su/nar with stative predicates but surface as Ø with VNs? (iii) What mechanisms regulate these patterns?

To account for these patterns, I adopt the late insertion model of Distributed Morphology (Halle & Marantz 1993). In the following sections, I examine the spell-out conditions governing the morphological realization of these functional heads, demonstrating how this framework captures both invariant and alternating patterns in the data.

To account for the morphological realization of the Voice head, I propose the following Vocabulary Insertion (VI) rules:

These rules specify that Voice is realized as null (Ø) when it enters into a local relationship with v by forming a complex head, and as su elsewhere. This context-sensitive distribution follows standard locality conditions and the Elsewhere Principle (Kiparsky 1973), with the more specific rule in (46a) taking precedence over the general case in (46b). Rule (46a) accounts for Voice being realized as Ø in stative predicate constructions (44), where Voice forms a local configuration with v via head movement.

However, we must explain why Voice consistently surfaces as su in VN-su constructions (45). The VI rules in (46) suggest that in these constructions, Voice and v do not form a complex head. Following Hayashi (2015), I propose that this pattern results from a structural constraint on VNs: they are fixed at the vP level and cannot undergo raising to Voice. Consequently, Voice fails to enter into a local relationship with v, triggering the elsewhere realization as su.

This analysis suggests that the VN and the light verb su are structurally independent entities (Kageyama 1993; Miyamoto & Kishimoto 2016).21 The structural independence of VNs and the light verb su finds empirical support in ellipsis phenomena.22 Hayashi (2015) observes a significant contrast between Native-Japanese verbs (NJVs) and Sino-Japanese verbs (SJVs). Consider first:

In both target clauses, the complement of the verb—the goal PP Nihon-e ‘to Japan’—is elided, indicated by [_Compl e ]. The two clauses differ minimally in whether the NJV is present in the surface string: while (47b), with the pronounced NJV ko, is well-formed, (47c), with an elided NJV, is ungrammatical. Contrast this with SJVs:

Here, both variants—with the pronounced VN kikoku in (48b) and with the elided VN in (48c)—are grammatical, with the light verb su stranded in the latter case.

This asymmetry in anaphoric properties led Hayashi (2015) to posit differential patterns of predicate raising: NJVs undergo head movement to a higher functional position, while SJVs remain in their base position. The current tripartite VP analysis refines this insight by providing a more articulated structural account: NJVs raise successively through v to Voice, while SJVs (and other VNs) raise only as far as v, not to Voice.23

The proposed analysis explains the grammatical contrast between (47c) and (48c) under the assumption that VP ellipsis targets vP in the tripartite structure (e.g., Otani & Whitman 1991; Funakoshi 2016).24 With SJVs, the VN remains within the ellipsis domain (vP) alongside its complement (49b). Consequently, vP-level ellipsis elides the VN while stranding Voice (realized as su), correctly predicting the grammaticality of (48c). With NJVs, by contrast, V raises to Voice via v, positioning it outside the domain of vP ellipsis (49a). It thus remains phonologically realized even when vP is elided—hence the ungrammaticality of (47c).25

In sum, Voice in Japanese exhibits sensitivity to structural locality with v: when in a local configuration within a complex head, it is realized as null (Ø); when structurally isolated due to the absence of V-v to Voice movement, it surfaces as the elsewhere form su.

Before examining v’s morphological realization, two comments are necessary regarding (i) NJVs and the light verb su, and (ii) v-to-Voice movement constraints. First, NJVs systematically exclude su in canonical contexts, yet require it in object honorification:

The pattern in (50a) follows from our analysis: NJV roots undergo successive head movement through v to Voice, establishing a local relationship that triggers Voice’s null realization via rule (46a). The object honorification exception in (50b) can be explained following Ikawa (2022): the honorific head hon projects between vP and VoiceP, blocking V-v to Voice movement:

In (51), the honorific head hon intervenes between vP and VoiceP, blocking head movement of the V-v complex (containing the NJV tasuke ‘help’) to Voice. Voice is therefore not in a local relation with V-v and surfaces as su via rule (46b). More generally, when an intervening element disrupts v-Voice locality, Voice is realized as su.

Regarding the theoretical underpinnings of head movement constraints, this analysis remains agnostic about the fundamental question of why certain predicate types such as SJVs resist head movement in the grammar. A comprehensive account would require cross-linguistic comparison and investigation of a broader range of predicates—endeavors beyond the scope of the present study. Nevertheless, Hayashi’s (2015) feature-based account offers a promising theoretical direction worthy of consideration.

Hayashi proposes that verb raising is triggered by an uninterpretable feature [uNative] on the functional little v head (equivalent to Voice in our tripartite structure). Consequently, verbs bearing an interpretable feature [iNative]—specifically, verbs of Native Japanese origin—undergo obligatory head raising, while other predicates such as SJVs cannot participate in this movement operation due to their lack of the requisite feature specification. Within this framework, the light verb su is inserted precisely to value the [uNative] feature on the higher functional head.

This account encounters challenges with deverbal compounds like ne-age(-su) ‘price-rise-do’, whose head is native but still requires su. One potential refinement would posit that non-head elements in compounds function as interveners for head movement, accounting for both deverbal compounds and object honorification constructions.26 However, this refinement raises questions about what constitutes an intervener, as V-V compounds (e.g., tataki-kowas ‘hit-break’) pattern with ordinary NJVs despite their complex structure. This suggests that structural complexity alone cannot explain the observed patterns.

In sum, the realization of su correlates with the absence of v-to-Voice movement, though a full theoretical interpretation—ideally integrating feature-based and structural-intervention approaches—remains for future work.

In our tripartite analyses (44) and (45), v_caus/inch is realized as su/nar in stative predicate constructions but as Ø in VN-su constructions. However, this simple pattern fails to account for two critical phenomena: first, transitivity-alternating NJVs like kowa-s ‘break-caus’ and kowa-re ‘break-inch’, where causative/inchoative morphology surfaces in verb root-sensitive patterns; second, deverbal compounds like ne-ag-e-su ‘price-rise-caus-do’ and ne-ag-ari-su ‘price-rise-inch-do’ that exhibit transitivity alternation while still requiring su.

To account for these patterns, I propose the following VI rules:

These rules instantiate a principled hierarchy of morphological realization. When v combines with a transitivity-alternating, native verb root ([+Native]), it surfaces as the abstract morphemes -S/-R (52a)/(53a) (e.g., Oseki 2021), which undergo root-conditioned allomorphy (e.g., -S/-R becomes -s/-re with kowa ‘break’ but -e/-ar with ag ‘rise’). This allomorphic variation is mediated through morpho-phonological operations (see e.g., Inoue 1976; Akimoto 2018; cf. Miyagawa 1994; 1998; Harley 2008).27^,28 When v combines with a non-native verb root ([−Native]), it is realized as Ø (52b)/(53b). In environments where v does not establish a local relationship with any verbal root, the elsewhere forms su/nar emerge (52c)/(53c).

The proposed rules apply across four construction types: NJVs, stative predicate constructions (SPCs), SJVs, and deverbal compounds (DCs). Let us first consider the morphosyntactic structures of constructions with NJVs and SPCs, as illustrated in (54a) and (54b), respectively.

In NJV constructions (54a), a complex head comprising V-v-Voice is formed through successive head movement. Crucially, v establishes a local configuration with a [+Native] verb root (e.g., kowa- ‘break’), triggering the insertion of causative/inchoative morphemes (-S/-R) via rules (52a)/(53a). These abstract morphemes then undergo root-conditioned allomorphy, surfacing as -s and -re. SPCs (54b) present a different configuration: v forms a local relationship with Voice (realized as Ø) but not with any verbal root, as its complement is an adjectival phrase that cannot undergo head movement to v. Consequently, the elsewhere forms su/nar emerge via rules (52c)/(53c).29

The structural configurations of SJVs and DCs exhibit distinct yet systematically related patterns of morphological realization:

In SJVs (55a), the verbal root bears the [−Native] feature (e.g., kakudai ‘expansion’), reflecting their Sino-Japanese origin. This triggers the null realization of v_caus/inch via rules (52b)/(53b), resulting in no overt morphological exponent at the v node. DCs (55b), in contrast, involve a more complex architecture: a nominal element (N) merges with a verbal root (V) to form a compound base, which raises to v. Here, v establishes a local relation with a V node bearing the [+Native] feature, licensing the insertion of causative/inchoative morphemes (-e/-ar) via rules (52a)/(53a), paralleling the morphological pattern of NJVs. Despite their structural differences, both SJVs and DCs share a crucial property: their V-v complexes cannot undergo further head movement to Voice—likely due to locality constraints or intervening material. As a result, Voice remains in situ, necessitating its realization as the elsewhere morpheme su via rule (46b).

The pattern observed here—where the presence or absence of head movement determines whether a predicate surfaces as a simplex form or requires a light verb—aligns with cross-linguistic patterns of complex predicate formation. Hale & Keyser (1993; 2002) argue that many unergatives derive from an underlying N+V configuration, with languages differing in whether the nominal element incorporates into the verb. English displays incorporation, yielding morphologically simplex verbs (e.g., cry from N-to-V incorporation into do), whereas Persian exhibits the opposite pattern: the nominal remains in situ and an overt light verb is required, as in gerye kardan ‘cry do’ (Folli et al. 2005).

Japanese fits naturally within this typology. NJVs undergo full head movement to Voice (54a), producing morphologically integrated forms without an overt light verb, while VNs such as SJVs move only to v (55a), necessitating the overt realization of Voice as su. Thus, the degree of head movement directly determines morphological realization, paralleling the incorporation vs. light verb contrast observed cross-linguistically.30

The preceding analysis has focused on configurations where v and Voice align in transitivity: v_caus combines with Voice_tr, and v_inch with Voice_un. However, the tripartite architecture logically permits mismatches: v_caus could combine with Voice_un, and v_inch with Voice_tr. This subsection demonstrates that at least the former type of mismatch is well-motivated in the literature, providing further evidence for treating v and Voice as independent functional projections.31

The first type of v_caus + Voice_un mismatch involves causative constructions where the external argument is interpreted as an experiencer or affectee rather than as an agent or causer. Consider (56):

In (56a), the subject Taro is not construed as an agent or causer but rather as an experiencer or affectee—someone affected by the burning event. This non-agentive interpretation is similar to that of the standard inchoative in (56b). Asami (2024) proposes that experiencer subjects in causatives like (56a) are licensed by the functional head Affect (Bosse et al. 2012), which projects between v_caus and Voice_un in the present framework:32^,33

Under this analysis, v_caus introduces an event that brings about the eventuality denoted by its complement (i.e., the event of Taro’s house burning down), but Voice_un does not introduce an Initiator. Instead, Affect licenses the external argument as an experiencer/affectee. The structure in (57) can be semantically paraphrased as: there is a set of events that causes Taro’s house to burn, and this situation affects Taro.34 This configuration exemplifies v_caus + Voice_un, demonstrating that v and Voice need not align in transitivity.

It should be noted that in (57), the structure for (56a), Affect projects between v_caus and Voice_un, yet it does not block V-v from reaching Voice. The question, then, is why Affect, intervening between v and Voice, does not block this movement, given that hon—in the same structural position in object honorification (51)—does, forcing the realization of su on Voice.35 I suggest two possibilities. One is phonological: head movement has often been argued to be sensitive to overt intervening heads (see e.g., Kandybowicz 2015 on Asante Twi). hon is an overt head, being realized as the prefix o-, whereas Affect lacks any phonological exponent.36 The other is categorial. Affect is responsible for introducing an argument (i.e., experiencer) just like Voice. hon, by contrast, introduces no argument and contributes nothing to event structure. If head movement depends on a local relation among verbal heads, hon may disrupt that relation whereas Affect may not. Notably, -tari, which will be argued in section 5 to be an intervener, patterns with hon in being both overt and non-argument-introducing. On either view, Affect is correctly predicted not to block this movement, whereas hon and -tari, being both overt and non-argument-introducing, are predicted to do so; which property is operative I leave for future research.

The second type of v_caus + Voice_un mismatch is what Schäfer (2025) terms transitive anticausatives, which are observed cross-linguistically:

(58a) takes a transitive structure but is semantically equivalent to the unaccusative (58b). The nominative subject in (58a) does not denote an agent or causer; rather, both sentences describe a spontaneous change event. Japanese exhibits a similar pattern with causative-inchoative alternating verbs:

In (59a), the verb takes the causative form kaw-e, yet the nominative subject kumo ‘clouds’ is not a causer. This is evidenced by the fact that the actual causer can be expressed as an adjunct PP kaze-de ‘by wind’ in both examples—a pattern unexpected if the subject were introduced by Voice_tr as a causer. In line with Schäfer (2025), I analyze transitive anticausatives as involving v_caus combined with semantically contentless Voice_un:

In (60), v_caus introduces the change-of-state event affecting the shape of the clouds, while the subject kumo ‘clouds’ is not assigned an Initiator role by Voice. Rather it originates in the possessor position of the theme katati ‘shape’ and raises to Spec,VoiceP via possessor raising.

These two types of v_caus + Voice_un mismatches demonstrate that v and Voice can vary independently. The converse pattern—v_inch + Voice_tr—is also logically possible within the tripartite architecture but lacks clear empirical instantiation in Japanese.37

Nevertheless, the well-established existence of v_caus + Voice_un patterns already demonstrates that v and Voice can vary independently in transitivity. This flexibility—whereby event structure (causative vs. inchoative) and Initiator licensing (present vs. absent) are controlled by distinct functional heads—would be unexpected under bipartite approaches that conflate v and Voice into a single head. The tripartite architecture naturally accommodates such mismatches, providing a more empirically adequate framework.