Separability of dependents from VP in English: Beyond the argument/adjunct distinction

Abstract

This paper considers the traditional idea about English that syntactic operations targeting Verb Phrase (VP), including do so-anaphora, do what-pseudoclefting and VP-fronting, can separate adjuncts but not arguments from the VP. I argue that, in each case, the argument/adjunct distinction (A/AD) makes incorrect predictions and that the behavior of verbal dependents is more accurately explained without reference to the A/AD. With do so-anaphora and do what-pseudoclefting, I show that the behavior of a variety of Prepositional Phrase (PP) dependents is better explained by the lexical properties of the verb do: a PP’s ability to occur with do so-anaphora/do what-pseudoclefting depends on the PP’s independent compatibility with the lexical verb do. On VP-fronting, I show that apparent stranding of arguments and adjuncts poses major problems for A/AD-based analyses and suggest apparent stranding is better analyzed as extraposition. These results weaken an important motivation for the idea that adjuncts attach to a higher projection in the VP than arguments do.

1. Separability Diagnostics

It is widely believed that, for any head H, the syntactic dependents of H form two distinct classes, the arguments of H and the adjuncts to H. This Argument/Adjunct Distinction (A/AD) is associated with a wide range of empirical contrasts, called diagnostics for argumenthood: patterns of omissibility, iterability, islandhood, etc. are all thought to follow from the basic distinction between arguments and adjuncts (see Schütze 1995, Hedberg & DeArmond 2002, Hornstein & Nunes 2008, Needham & Toivonen 2011, Forker 2014, Toivonen 2016, Moura & Miliorini 2018, Bergs 2020, Bode 2020, Milway 2022, Zyman 2022, among others).

More precisely, argumenthood diagnostics serve as the evidence that motivates specific theoretical implementations of the A/AD in syntax, such that the empirical properties of adjuncts follow from formal aspects of the implementation. This paper focuses on one purported formal aspect of the syntactic A/AD in particular, namely, the idea that the A/AD conditions the relative configuration of a head’s dependents. The core idea, originating with Lakoff & Ross (1966),¹ is that the arguments of a head H occur within a smaller projection of H than the adjuncts to H do. With respect to the verbal domain specifically, I refer to this purported property of adjunction as differential attachment-heigh of verbal dependents (DAVD).

Example 1

Lakoff & Ross motivated DAVD with their now-famous ‘do so-test’. They noticed that some dependents of V, such as tomorrow in (2a), can occur alongside do so, while others, such as onto the wagon in (2b), cannot.

Example 2

Assuming that do so must substitute for a complete VP, Lakoff & Ross deduced that temporal adverbials like tomorrow must be syntactically ‘outside’ the VP, while directional/goal PPs like onto the wagon must be syntactically ‘inside’ the VP. Today, the X-bar theoretic implementation of this idea remains the rough-and-ready cross-framework mainstream for argument/adjunct configuration in the VP.²

Example 3

If we identify the ϒP of (1) with VP, then patterns of stranding under do so-substitution support DAVD and make do so-substitution an effective diagnostic for argumenthood: if a verbal dependent XP can be stranded by do so-substitution, then it must be an adjunct; otherwise, XP must be an argument.

After Lakoff & Ross (1966), several other phenomena, including do what-pseudoclefting, VP-fronting and VP-ellipsis, came to be accepted as evidence for DAVD, and hence, also serve as argumenthood diagnostics, under analyses similar to that of do so-substitution. The central premise of all these diagnostics is that they involve an operation targeting the argument-containing ϒP of (1). Crucially, given DAVD, a ϒP-targeting operation can separate adjuncts but not arguments from V.

Example 4

Example 5

Any phenomenon that involves a ϒP-targeting operation can function as an argumenthood diagnostic by the logic in (5). I call such diagnostics separability diagnostics.

Example 6

The goal of this paper is to critically reevaluate separability diagnostics as evidence for DAVD and, by extension, for the A/AD more broadly. To this end, I consider four purported separability diagnostics in English, all commonly taken to support DAVD by the logic just described:

Example 7

A complete analysis of each phenomenon doing justice to the full range of individual complexities falls far outside the scope of this paper. Instead, my approach is to focus on the DAVD-relevant properties of the phenomena, identifying problems for these specific aspects of their analysis. Sections 2 through 4 below address each of do so-anaphora, do what-pseudoclefting and VP-fronting in turn. For each one, I first introduce the DAVD-relevant aspects of its analysis as a separability diagnostic, then present data challenging that analysis and finally identify an alternative A/AD-free analysis that plausibly better captures the facts. I conclude that the data, carefully analyzed, are incompatible with the assumptions in (6) and thus that these purported separability diagnostics fail to motivate DAVD as a syntactic property of the A/AD. Section 5 turns more briefly to VP-ellipsis. I identify a constellation of analytical assumptions that must align for VP-ellipsis to succeed as a separability diagnostic, which I suggest is unlikely.

It should be noted that the literature attests a variety of views regarding the precise attachment-height of adjuncts (see, e.g. Maienborn 2001, Ernst 2001, 2020). For concreteness and consistency in discussing and analyzing the data relevant for this paper, I adopt Harley’s (2007, 2014) implementation of the A/AD in bare phrase structure. Harley’s (1995) approach draws on the root (√) analysis of Distributed Morphology (Halle & Marantz 1993, Marantz 1997). The relevant assumption is that verbs are underlyingly roots √ in the presence of an abstract verbalizer v and that the arguments in a verb phrase are sisters to (a projection of) √, while adjuncts are sisters to (a projection of) v. The argument-containing ϒP of (1), then, is vP: arguments of the verb will be strictly internal to vP, while adjuncts will be external to at least some projection of vP.³ Thus, in (8), which assumes that in the kitchen is an adjunct – and a pizza an argument – of the verb eat, a vP-targeting operation could separate in the kitchen from the verb, but could not separate a pizza from it.

Example 8

Finally for this introduction, although the point of this paper is to engender skepticism toward DAVD as a property of the A/AD (and toward the A/AD more generally), I will assume for the sake of argument that there is a valid A/AD that separability diagnostics could in principle identify.⁴ While there is no consensus on where the precise boundary between arguments and adjuncts falls, there is broad agreement on certain properties of the most clear-cut cases. For example, the most prototypical arguments must be licensed by specific verbs, while the most prototypical adjuncts are essentially self-licensing (this licensing criterion is Huddleston & Pullum’s (2002: 219) ‘most important property’ of arguments). Meanwhile, any dependent that is non-omissible and/or L-selected⁵ (in the case of PPs) is a prototypical argument. DPs are also virtually always considered arguments as a rule.⁶ These relatively uncontroversial criteria will be sufficient for the purposes of this paper.

2. Do so-anaphora

This section demonstrates that the A/AD does not play an explanatory role in accounting for the separability of PPs from the verb in do so-anaphora. Przepiórkowski (1999; see also 2016) has previously argued for this same conclusion by compiling prior arguments that do so-anaphora resolution is a discourse-pragmatic phenomenon rather than a properly syntactic one. For example, Kehler & Ward (1999) find that do so can refer to material split across antecedents and that do so-anaphora allows certain active-passive voice mismatches. Culicover & Jackendoff (2005), Houser (2010) and Bruening (2019) also conclude that do so-anaphora resolution is discourse-pragmatic in nature on the basis of these facts and others.

Nevertheless, the do so-test remains widely accepted as a diagnostic for argumenthood. I, therefore, take a different approach here. Instead of seeking to show that do so-anaphora must be discourse-pragmatic in nature, I start with the assumption that do so-anaphora is a separability diagnostic and explore the consequences of this for the A/AD. I specifically identify three cases where the behavior of do so-anaphora is incommensurate with an independently motivated understanding of the A/AD. Only after establishing these problems for the DAVD-based analysis of do so-anaphora do I turn to the discourse-pragmatic understanding, where I show that an extension of Bruening’s (2019) analysis, incorporating the core idea of Miller 1990 with respect to do so, straightforwardly handles these problems, as well as the standard cases, without recourse to the A/AD.

2.1. Do so-substitution as a separability diagnostic

It is essential first to establish in exactly what way do so-anaphora is supposed to involve a vP-targeting operation, by virtue of which fact it may serve as a separability diagnostic for argumenthood. I will discuss three different views that the reader may hold on this, namely, (i) do so simply is a vP, (ii) do so seeks a semantically complete predicate as its antecedent and (iii) do so copies the meaning of a syntactically defined vP antecedent. Only the third view is valid for present purposes. I return to the second view in Section 2.3.

Consider the view that do so-anaphora functions as a separability diagnostic simply because it is a proform with the category vP. Under this view (assuming DAVD), the only dependents that will be able to occur with do so are adjuncts. This is essentially what is assumed by Sobin (2008) and Mikkelsen et al. (2012), among others. Unfortunately, this line of reasoning is unacceptable. The reasoning is not specific to the expression do so; any expression that we want to consider a vP can play the same role as do so in this argument. Suppose we transfer the reasoning to the verb talk: talk is a vP, and therefore any PP that can occur with talk must be an adjunct. We would then conclude that to X in ‘talk to X’ is an adjunct. This is fallacious: the assumption that talk is a vP (which it probably is) does not allow us to conclude that any PP occurring with talk must occur outside that vP. The reasoning with do so suffers the same problem. We can assume that do so is a vP, but that does not allow us to conclude that any PP occurring with do so must be an adjunct.⁷

Next, consider the idea that do so functions as an argumenthood diagnostic because it seeks an antecedent which is semantically defined to include a predicate’s arguments but not necessarily its adjuncts. This paper is about DAVD, which is a syntactic-configurational property. If do so picks out an antecedent defined along semantic lines as just described, then it not directly relevant to DAVD.⁸

In order for do so to function as a separability diagnostic, it needs to take on the meaning of a vP-constituent antecedent. This can be accomplished in two different ways: a complete vP could be built up and then replaced by do so (the classic transformational analysis) or do so could be treated as a base-generated proform whose interpretation is identified with a vP-constituent antecedent. The former understanding (with a substitution transformation) can be compared to the PF-deletion approach to ellipsis and the latter to the LF-copying approach to ellipsis.⁹ In the domain of ellipsis, these different perspectives are notoriously difficult to disentangle (see Merchant 2018). For the purposes of this paper, I will treat them as interchangeable but will keep to the copying approach in exposition.¹⁰ What is essential, then, is simply that the interpretation of do so corresponds to that of a discourse-accessible vP constituent.

2.2. Behavior of different PPs under do so-anaphora

Do so is compatible with certain kinds of PPs, and not others. This subsection considers the behavior of several different kinds of PPs under do so-substitution, each of which poses problems of varying degrees for the DAVD-based approach just discussed. Specifically, the behavior of argumental to-phrases, benefactives, comitatives and locatives highlight misalignments between the observed data and the predictions from the A/AD.

2.2.1 Argumental to-phrases

One problem for the DAVD-based approach to do so comes from what I will call ‘argumental to-phrases’.¹¹ (9) from Huddleston & Pullum (2002: 1533) illustrates the phenomenon. As they point out, the interpretation of do that in (9) is something like ‘question me for over an hour before letting me go’, with the to-phrase to me corresponding to the patient participant Jill in the antecedent clause.¹²

Example 9

Many examples of this sort with do so occur in corpora. Below are three from Corpus of Contemporary American English (COCA). Further examples can be found in COCA and other corpora, as well as in the literature (see, e.g. Miller 1990, Przepiórkowski 1999, Culicover & Jackendoff 2005, Mikkelsen et al. 2012).

Example 10

In the examples above, the to-phrases in the do so-clauses correspond to argumental dependents in the antecedents. In (10a) and (10b), the to-phrases correspond to direct objects, which are universally considered arguments (not adjuncts). In (10c) the to-phrase corresponds to a goal PP, also classed as an argument by standard criteria.¹³ Such examples undermine do so-anaphora’s ability to function as a separability diagnostic. As discussed above, this ability is predicated on the analysis whereby do so seeks a complete vP as its antecedent. Yet, DAVD predicts there to be no complete vPs in (10) that match the meanings we find for do so. In (10a), for example, do so’s interpretation is, roughly, ‘control through physical harm’. According to DAVD, there is no such vP in (10a); the minimal v P would be control others . If this vP is chosen as the antecedent for do so, an unacceptable interpretation for the do so clause would result, namely ‘control others to our children’.

Example 11

Though Przepiórkowski (1999) argues against do so as evidence for the A/AD, he remarks (290–291) that data like in (10) may be insufficient to ‘pronounce the demise of the do so test’, because one could imagine an analysis whereby the problematic material covertly evacuates the antecedent vP¹⁴ (see Mikkelsen et al. 2012 for a concrete solution along similar lines). However, I claim that this would not, in fact, save the do so-test. While Przepiórkowski’s hypothetical solution could preserve both (i) DAVD and (ii) the idea that do so seeks a vP antecedent, it only does so by introducing a uniform treatment of arguments and adjuncts, and as such, it neutralizes do so’s ability to function as evidence for DAVD or the A/AD more broadly. Crucially, if we allow overlapping argumental material, like the patient others in (10a), to covertly evacuate the vP so that do so can target the remnant, then we allow this same analysis for adjuncts. That is, just as do so could target the remnant vP in (12a), it could also target the remnant vP in (12b).

Example 12

This eliminates the implication from strandability to adjuncthood under do so-anaphora. Therefore, while the behavior of argumental to-phrases may not undermine the idea that do so seeks a syntactically defined vP as its antecedent, it does radically undermine the idea that do so-anaphora serves as a separability diagnostic for argumenthood.

2.2.2 Benefactives

A similar problem is posed by benefactives, as in (13).

Example 13

I assume that DP-benefactives like Lee in (13) are to be considered arguments (e.g. because they are DPs and occur with specific verbs like license them). Data like (13) thus cannot be analyzed in the way necessary for do so-anaphora to serve as a separability diagnostic. Let us assume that double-object benefactives have a structure like either (14a) (based on Pylkkänen 2008) or (14b) (based on Bruening 2010, Bosse 2015).¹⁵

Example 14

For do so-anaphora to serve as a separability diagnostic, do so must seek a complete vP as its antecedent. In both structures (14a) and (14b), the minimal vP is bake Lee cookies. However, if this vP were targeted as the antecedent to do so in (13), then the do so clause would receive an unacceptable interpretation, namely, ‘bake Lee cookies for Terry’. The actual interpretation of do so in (13) is ‘bake cookies in a heartbeat’, including the temporal adjunct in a heartbeat and not including the antecedent beneficiary Lee.

As with argumental to-phrases, one could imagine an analysis that preserves both DAVD and the assumption that do so seeks a vP antecedent, by allowing the beneficiary Lee to covertly evacuate the antecedent vP. This neutralizes do so-anaphora’s ability to serve as a separability diagnostic for the same reasons as with argumental to-phrases.

One could also imagine accounting for examples like (13) with a subtly different analysis of do so: instead of being a proform seeking a vP antecedent, do so could result from an ellipsis-like derivation, wherein a complete vP structure is constructed for the do so-clause, only to be replaced by do so. The relation between do so and its antecedent would then be similar to that between elliptical clauses and their antecedents (see Harley 2007). While this might arguably be able to account for (13) (assuming do so-replacement for bake cookies in a heartbeat would be licensed by the antecedent bake Lee cookies in a heartbeat), it will not work in general because of examples like (10a)/(10b). With patient to-phrases, no appropriate vP could be constructed for do so to replace (consider *control to our children, for (10a), for example).

2.2.3 Locatives and comitatives

The behavior of benefactives and argumental to-phrases indicates that a key aspect of the DAVD-based analysis of do so-anaphora is incorrect; namely, these phenomena show that do so can have an antecedent that excludes arguments of the verb. A complementary problem is posed by locatives and comitatives, showing that there are some adjuncts that do so cannot separate from the verb. To illustrate the problem, I first distinguish locatives and comitatives that are subject-oriented from those that are object-oriented. The difference can be illustrated with (15).

Example 15

These sentences are ambiguous. One reading of (15a) is that where the 2nd-person addressee is at home, but there is another reading where the 1st-person speaker is at home.¹⁶ (15a) is similarly ambiguous; in addition to the reading where the 2nd-person addressee is with family, there is another reading where the 3rd-person referent (him) is the one with family.¹⁷

Locatives and comitatives are among the most prototypical adjuncts. It is perhaps surprising, then, to find that the object-oriented versions of these elements are incompatible with do so-anaphora.

Example 16

On their own, the antecedent clauses in (16) would be ambiguous: ‘Lee called Terry at home’ could mean that Lee was at home or that Terry was, for example. But with do so-anaphora, the object-oriented reading goes away. (16a) can only mean that Lee called Terry while Lee was at work (not while Terry was at work), and (16b) can only mean that Lee talked to Terry while Lee was with their friends (not while Terry was with their friends). Yet, on a DAVD-based analysis of do so-anaphora, assuming that adjuncts attach outside the vP, this is unexpected.

One could maintain the idea that do so seeks a vP antecedent by taking data like (16) to indicate that the subject- versus object-oriented readings of locative/comitative PPs involve different attachment heights: the subject-oriented versions attach outside the vP, and the object-oriented versions attach somewhere lower (see Zhang 2007, for example, and Maienborn’s 2001 analysis of external vs. internal modifiers). This would make the right predictions about (16), but it would neutralize do so-anaphora’s ability to function as a separability diagnostic: if some adjuncts attach within the vP, then we cannot conclude from a dependent’s non-separability from the verb under do so-anaphora that the dependent is an argument.

2.3. An A/AD-free analysis of do so-anaphora

Section 2.2 presented several problems for the idea that do so-anaphora functions as a separability diagnostic. In this section, I suggest that an A/AD-free analysis of do so-anaphora better captures the data. I specifically adapt Bruening’s (2019) discourse-pragmatic account of do so-anaphora resolution, in combination with Miller’s (1990) crucial point that the dependents of do so are subject to the selectional restrictions of the lexical verb do. I first review Bruening’s analysis, then introduce the crucial claim from Miller (1990), showing how this accommodates all the problems discussed in Section 2.2. As noted in Section 1, an analysis of the full range of subtleties exhibited by do so-anaphora falls outside the scope of this paper. Here, I focus only on the DAVD-relevant aspects of the analysis. Any adequate account of do so-anaphora must deal with two questions. First, how is the semantic value of do so determined? Second, where do the (in)compatibilities with different PPs (such as those described in Section 2.2) come from? I begin here with the first question and return to the second.

As mentioned, much prior literature demonstrates conclusively that do so-anaphora is fundamentally discourse-pragmatic in nature. I will not repeat the relevant arguments here; see Houser (2010) for a concise summary and Przepiórkowski (1999) for copious examples. Here, I adopt Bruening’s (2019) analysis treating do so as a vP with denotation λe.f _{⟨v, t⟩}(e). According to Bruening (2019), do so gets its surface semantic value simply by searching the discourse for a function f of the appropriate semantic type ⟨v, t⟩ (i.e. a predicate of events). The value of this function replaces that of f in the denotation of do so. In many cases, the value of f corresponds to the value of a vP node in an antecedent clause (see Bruening 2019: 31–34), but that need not be the case. With respect to split-antecedents, for example, Bruening (2019: 42) explicitly notes that the value of f must sometimes be ‘constructed from multiple functions in the discourse’.

Though Bruening (2019) does not address them in detail, there are many restrictions on potential values for f (which I also cannot go into here); for example, do so cannot have a stative antecedent and is limited in reference to linguistic (as opposed to situationally evoked) content. See Kehler & Ward 1999, Ward & Kehler 2005 for discussions. For the purposes of this paper, I understand Bruening’s (2019) required value f to be the maximal¹⁸ predicate of events meeting the relevant conditions (such as those identified by Kehler & Ward (1999)) that is entailed in the discourse. This analysis then straightforwardly models the problematic examples with argumental to-phrases and benefactives discussed in Section 2.2, as well as all standard examples of do so with stranded adjuncts. To illustrate, consider (17) with an argumental to-phrase.

Example 17

The antecedent for do so here is sack Carthage.

Example 18

Do so is interpreted to mean sack, excluding the patient. Under a neo-Davidsonian event representation, this value is entailed from (18) via conjunction elimination.

Example 19

Deriving the ultimate interpretation of the do so-clause involves combining the denotation (19) with the thematic content of the argumental to-phrase. Following a reviewer’s suggestion, I take these to-phrases to contribute a malefactive theta role mal, leading to the denotation λe. Sack(e) ∧ mal(e)=Corinth for the do so-clause in (17). The acceptability of an example with an argumental to-phrase thus depends on a speaker’s willingness to accept the role of the to-phrase’s correlate as compatible with the malefactive role assigned to it in the do so-clause.¹⁹

Examples with benefactives, like (13), as well as any standard example with a stranded adjunct, work straightforwardly along the same lines. What remains unexplained (so far) under this analysis is why not just any PP can be stranded by do so-substitution. For example, what makes sentences like (20) unacceptable, given that apparently appropriate values for do so could be determined straightforwardly along the lines just described?²⁰

Example 20

This brings us to the second question raised at the beginning of this section, and it is here that Miller’s (1990) insight is relevant. As he points out, the do of do so is a lexical verb (not an auxiliary), and as such, it can be expected to have selectional requirements,²¹ like all verbs have. To take a random example, a goal at-phrase is incompatible with the verb sleep not because of any DAVD-based configurational property of goal PPs, but simply because sleep is not compatible with the goal theta role. Countless other verbs have the same property, and the verb do is among them. It is selectional requirements like this that determine which kinds of PPs can occur with do so, not the A/AD.

We can test Miller’s (1990) proposal by explicitly comparing the kinds of thematic PPs that can occur with do so against those that can occur with the lexical verb do in another context (something Miller 1990, 1992 do not do). (21) illustrates with the kinds of PPs discussed above.²²

Example 21

The PPs compatible with do so align closely with those compatible with lexical do. Notice also that the locative and comitative in (21c) and (21d) necessarily have the subject-oriented reading, just as with do so.²³ These data strongly support the idea that the selectional requirements of do, not the A/AD, determine a PP’s compatibility with do so.

This section considered do so-anaphora as a potential separability diagnostic for argumenthood. While do so-anaphora has been widely cited as evidence for DAVD since Lakoff & Ross (1966), I presented several major misalignments between this diagnostic and the A/AD. Though these misalignments are problematic for the DAVD-based analysis of do so-anaphora, I showed that they are handled straightforwardly on an analysis like that of Bruening (2019), treating do so-anaphora as a fundamentally discourse-pragmatic phenomenon. The compatability of different types of dependents with do so can be handled in terms of the selectional properties of the verb do. In light of these facts, we should conclude that strandability under do so-substitution is orthogonal to the A/AD.

3. Do What-Pseudoclefting

Do so-anaphora does not provide evidence for DAVD as a property of adjunction. However, the closely related do what-pseudoclefting is also widely cited in support of this property and is sometimes claimed to provide more robust support than do so-anaphora does (see Schütze 1995, Hedberg & DeArmond 2009, Needham & Toivonen 2011, Kim et al. 2019, Zyman 2022). In this section, I show that do what-pseudoclefting also fails to provide evidence for DAVD as a property of adjunction. The argumentation is very similar to that of Section 2. I first introduce the DAVD-based analysis necessary for do what-pseudoclefting to function as a separability diagnostic. Next, I present data that is problematic for this idea. Finally, I show that an A/AD-free analysis similar to that proposed for do so-anaphora better captures the data, suggesting that strandability under do what-pseudoclefting is unrelated to the A/AD.

3.1. Do what- pseudoclefting as a separability diagnostic

Do what-pseudoclefting serves as a separability diagnostic under assumptions similar to those for do so-anaphora discussed in Section 2.1. Recall from Section 2.1 that simply assuming do what to be a pro-vP (as Kim et al. 2019 do) is insufficient to conclude that elements it can ‘strand’ are adjuncts. It is also insufficient to adopt a semantic condition on pseudocleft reconstruction, because the property at issue here (i.e. DAVD) is a syntactic one.

In order for do what-pseudoclefting to function as a separability diagnostic, it must involve a vP-targeting operation. I take the relevant operation to be reconstruction along the lines of Bošković (1997). Specificational pseudoclefts²⁴ consist of three parts, (i) a wh-clause with a gap bound by a wh-operator, (ii) a ‘counterweight’ constituent that specifies the content of the gap and (iii) a copula linking these two components.

Example 22

In an analysis like that of Bošković (1997), the syntactic derivation of (22) involves replacement of what by the counterweight at LF, for example, transforming (23a) into (23b). Reconstruction into the gap site then yields a final logical form as in (23c).

Example 23

See Van Luven (2018: 30) for application of Bošković’s (1997) analysis to do what-pseudoclefts as in (23). For present purposes, what’s critical is that the counterweight must be a complete vP, and any elements stranded at the gap site (like yesterday in (23)) must be able to attach outside this vP when it reconstructs. Compare (24a), which is predicted to be unacceptable because reconstruction to (24b) places an argument at-phrase outside vP.

Example 24

In this way, do what-pseudoclefting meets the definition of a separability diagnostic and could potentially provide evidence for DAVD.

3.2. Behavior of different PPs under do what-pseudoclefting

If do what-pseudoclefting is a separability diagnostic, then any PP that can be stranded under this operation should be an adjunct, and any PP that cannot be stranded should be an argument. This section presents several problems for that idea. These problems closely parallel those discussed with respect to do so-anaphora in Section 2 above. Therefore, for conciseness, the exposition in this section is abridged (see discussion of the parallel do so for more detail).

3.2.1 Argumental to-phrases

As with do so-anaphora, a do what-pseudocleft can host a to-phrase that corresponds to an argument in the counterweight. Many examples occur in corpora, and examples are also easy to construct. Here is one example from COCA and one constructed example:

Example 25

The wh-clause in sentence (25b) is what they did to the children, where the to-phrase to the children that appears stranded at the gap site corresponds to an argument NP in the counterweight vP (i.e. the patient them). However, if the counterweight in (25b) reconstructs into the gap site as in (23), then an illicit vP would be formed, namely, they chased them through the park to the children. This is problematic for the idea that do what-pseudoclefting is a separability diagnostic.

3.2.2 Benefactives

Benefactives in do what-pseudoclefts pose a similar problem. Consider a sentence like (26).

Example 26

The wh-clause here contains a benefactive PP corresponding to an indirect object benefactive them in the counterweight. If the counterweight bake them cookies were to straightforwardly reconstruct into the gap site as in (23), the result would be (27).²⁵

Example 27

This paraphrase is acceptable for some speakers, if the for-phrase is interpreted as a ‘deputative’ benefactive (a benefactive for-phrase indicating that the event was carried out on someone else’s behalf, in this case, on behalf of the students, see Bosse 2015: 121–122). If reconstruction yields (27), it is predicted that the for-phrase in the wh-clause in (26) is interpreted as a deputative benefactive; thus, in addition to being the recipient of the cake, the students in this example should also be the party on whose behalf the baking was carried out. That is not necessarily the case, though. The pseudocleft in (26) does not require an interpretation in which cookies were baked on behalf of the students. Instead, (26) can be interpreted as simply saying that the cookies were baked for the students to eat.²⁶ To the extent that for the students in (26) needs not be interpreted as a deputative benefactive, it is a problem for the DAVD-based understanding of do what-pseudoclefting.

3.2.3 Locatives and comitatives

Just as with do so-anaphora, benefactives and argumental to-phrases show that the elements apparently stranded at the gap site in do what-pseudoclefting do not necessarily have to correspond to adjuncts of the counterweight vP. Also as with do so-anaphora, locatives and comitatives pose a complementary problem, that is, showing that not all adjuncts can be stranded by this operation. To illustrate, consider (28).

Example 28

The reconstructed versions of these sentences would be ambiguous between subject- versus object-oriented readings of the locative and comitative (compare (15) above). If object-oriented locative and comitative adjuncts could occur outside the vP, as DAVD would predict, we would expect the object-oriented readings to remain available under pseudoclefting. Yet, in these pseudoclefts, only the subject-oriented readings are available: (28a) can only mean that the speaker shouldn’t be called when the 2^nd-person addressee is at home, and (28b) can only mean that the 3rd-person referent him shouldn’t be talked to when the 2nd-person addressee is with his family. This is problematic if do what-pseudoclefting is supposed to be a separability diagnostic.

3.3. An A/AD-free analysis of do what-pseudoclefting

I have presented several problems (paralleling those discussed for do so in Section 2.2) for the idea that do what-pseudoclefting is a separability diagnostic. In this section, I argue that an A/AD-free understanding of do what-pseudoclefting better accounts for the data, contradicting the idea that do what-pseudoclefting provides evidence for DAVD. In brief, I adopt the analysis of Den Dikken et al. (2000), according to whom the relevant pseudoclefts have the structure of ‘concealed questions’. Do what-pseudoclefts, therefore, do not involve reconstruction of the counterweight into the wh-clause, instead being interpreted in the same way as a question-answer pair. The key insight of Miller (1990) is also relevant here in the same way as with do so-anaphora.

Den Dikken et al.’s (2000) concealed-question analysis is necessitated by the existence of what O’Neill (2015) terms ‘amalgam pseudoclefts’. In this construction, the counterweight duplicates part of the material of the wh-clause and is often a fully articulated clause itself. Several examples are given in (29).

Example 29

Den Dikken et al. (2000) propose that pseudoclefts of this sort comprise a topic-comment structure mediated by a Topic Phrase (TopP) projection. The wh-clause subject of the pseudocleft occupies Spec-TopP, and the counterweight occupies Comp-TopP.

Example 30

This structure is claimed to be present even when overlap between the wh-clause and the counterweight is not overt; in these cases, Den Dikken et al. (2000) (citing Ross 1997 as precedent) argue extensively that overlapping material in the counterweight is elided (see also O’Neill 2015, Van Luven 2018).

Example 31

Den Dikken et al. (2000) propose that, rather than involving syntactic reconstruction, such sentences are interpreted in the manner of question-answer pairs like (32), which can similarly involve overlap between the question and the answer.

Example 32

The ability of a PP to occur at the gap site in a do what-question like (32) is determined by the selectional properties of the lexical verb do. Compare (33) with (21) above.

Example 33

The same PPs that can occur in do what-questions are the ones that can appear ‘stranded’ at the gap site in do what-pseudoclefts. All the questions in (33a) make acceptable do what-pseudoclefts when paired with appropriate counterweights (i.e. answers), while none of the questions in (33b) do. For example, What I’d like to do to Lee is surprise them, *What I’d like to do at Lee is toss the ball.

Therefore, DAVD is not needed to explain which PPs can occur with do in the wh-clause of a do what-pseudocleft. Instead, the ability of a PP to occur in a do what-pseudocleft depends on that PP’s compatibility with the lexical verb do. This account naturally resolves all the problems discussed in Section 3.2. For example, the overlap that occurs with argumental to-phrases and some benefactives works by the same mechanisms as when overlap occurs in pseudoclefts or question-answer pairs generally, as in (29) or (32). As for locatives and comitatives, the reason why only the subject-oriented versions are compatible with do what-pseudoclefting is that the object-oriented versions are incompatible with lexical do, as discussed in Section 2.3.

This section considered patterns of apparent stranding under do what-pseudoclefting potentially providing evidence that DAVD is a property of the A/AD. I argued that this phenomenon provides no such evidence. Instead, I presented several misalignments between the A/AD and the set of PPs that appear separable from the verb under do what-pseudoclefting, and I argued that an analysis accurately capturing these data has no need to invoke DAVD or the A/AD more generally.

4. VP-Fronting

This section considers the phenomenon of VP-fronting as a potential motivation for DAVD as a property of the A/AD. Specifically, the idea is that a PP’s ability to be stranded under VP-fronting is predictable from its argumenthood. I demonstrate that attested examples are incompatible with the A/AD-based predictions and that contemporary approaches to VP-fronting predict that the A/AD need not be involved at all. Instead, PPs are separated from the verb under VP-fronting by extraposition, applying to arguments and adjuncts equally. VP-fronting, therefore, does not provide evidence in favor of DAVD. To show this, I first discuss how VP-fronting is thought to be a separability diagnostic in Section 4.1, then report data from prior literature which pose problems for A/AD-based approaches to VP-fronting in Section 4.2 and conclude by explaining how an A/AD-free analysis could account for the stranding facts in Section 4.3.

4.1. VP-fronting as a separability diagnostic

VP-fronting is standardly analyzed as movement of a verbal projection assumed here to be vP. This straightforwardly makes VP-fronting a separability diagnostic capable of motivating DAVD as a property of the A/AD. Given DAVD, the targeted vP will have to contain all the arguments of the verb but could potentially exclude adjuncts to the verb. Extraction of vP, therefore, must carry along the arguments of the verb to the landing site but could leave adjuncts behind at the extraction site.

Example 34

According to this analysis of VP-fronting, any instance of Internal Merge targeting _vP represents VP-fronting, not just VP-topicalization.²⁷ Therefore, VP-fronting in though-preposing can also be considered as relevant data, as it has in prior literature (see, e.g. Phillips 2003: 77 fn. 33, Baltin 2006, 2017, Landau 2007, Janke & Neeleman 2012, Bruening 2018, Culicover & Winkler 2019, Thoms & Walkden 2019). For example, the fact that (56b) is acceptable with the same interpretation as (56a) would be taken to indicate that the durational PP throughout the entire day is an adjunct to eat.

Example 35

4.2. Behavior of different PPs under VP-fronting

This section collects judgments from a variety of sources contradicting the generalization that VP-fronting can strand adjuncts but not arguments. Judgments on VP-fronting are notoriously variable. Some speakers reject VP-fronting outright. Among those that accept it, some do not accept any stranding at all. However, in the judgments reported in the literature from speakers who do accept stranding under VP-fronting, there is little reason to believe that strandability is conditioned by the A/AD. The literature attests to many examples in which conventional argument PPs are stranded. This includes addressee to-phrases, recipients, goals and certain of-phrases.

Example 36

Example 37

Example 38

Example 39

All these kinds of PPs would be considered arguments, not adjuncts, by standard criteria – for example, none of them has the self-licensing property that is characteristic of prototypical adjuncts (see Section 1 above). If DAVD is a property of the A/AD, these PPs should be situated within the vP. The vP in (36a), for example, would have a structure along the lines of (40).

Example 40

With this structure, it is not possible to separate the PP from the verb simply by targeting the vP for Internal Merge. We could stipulate that the to-phrase is first dislocated to a position external to vP, with movement then targeting the remnant vP (compare Lechner 2003), but such a move would make it possible to separate anything that can be dislocated, including arguments.²⁸ VP-fronting would not then be expected to treat arguments differently from adjuncts and would not provide evidence for the DAVD hypothesis. I conclude that data from VP-fronting do not provide evidence for the hypothesis that adjuncts attach externally to the vP while arguments attach internally to the vP or that the A/AD otherwise plays a role in determining which PPs VP-fronting can separate from the verb.

4.3. An A/AD-free analysis of VP-fronting

Many analyses of English VP-fronting do not assume that the A/AD constrains separability. I will not attempt to adjudicate between the various options, which would require discussion of a vast literature.²⁹ Instead, in this subsection, I will consider one state-of-the-art analysis of English VP-fronting (that of Thoms & Walkden 2019), showing that within this analysis, the A/AD cannot parsimoniously be invoked to account for the possibilities of separation from the verb. With a minor extension, this analysis more parsimoniously accounts for the possibilities of separation without requiring DAVD as a property of the A/AD.

Thoms & Walkden (2019) propose that VP-fronting does not involve movement of vP at all but instead involves base generation of a vP in fronted position. The fronted vP is linked to the ‘extraction site’ via ellipsis plus operator movement. Their proposed analysis is exemplified in (41) for a sentence like eat the pies he did. In this example, strikethrough indicates ellipsis, and angled brackets (< >) indicate an unpronounced copy created by Internal Merge.

Example 41

According to Thoms & Walkden (2019), the subject of the lower vP is merged with an operator Op forming a complex DP. The operator, co-indexed with the subject of the lower vP, moves to Spec-CP, where it binds the PRO subject of the fronted vP (ensuring the subject of the fronted vP covaries with the subject of the lower vP). Thoms & Walkden (2019) argue that connectivity effects between the vP in fronted position and the lower vP arise through identity requirements on ‘high predicate ellipsis’, that is, ellipsis which applies relatively ‘high’ in the vP (they assume the relevant domain is VoiceP, a detail I omit from the notation below).

This analysis has a variety of advantages. Specifically, Thoms & Walkden (2019) show that their analysis enables an account of (i) the restricted licensing conditions on VP-fronting, (ii) the possibilities of morphological mismatch between the fronted vP and the elided vP and (iii) the restricted set of connectivity effects occurring with VP-fronting. However, Thoms & Walkden (2019) do not consider the way their proposal deals with the instances of stranding considered in this section. I argue that, according to Thoms & Walkden’s (2019) proposal for VP-fronting, such examples cannot be treated as stranding involving the A/AD but are better treated as involving extraposition from the fronted vP (ignoring the A/AD).

To reach this conclusion, first consider Thoms & Walkden’s (2019) critical diagnostic for high Verb Phrase Ellipsis (VPE): retorts and question tags. The reason VPE in these constructions is considered ‘high’ VPE is that voice mismatches are not possible.

Example 42

Example 43

This suggests that retorts and question tags involve ellipsis of Voice (compare Merchant 2013). Thoms & Walkden (2019), following Sailor (2014), take the target of deletion in these constructions to be (the maximal projection) VoiceP. They argue that VP-fronting involves high VPE because, for instance, the possibilities of morphological mismatch in VP-fronting mirror those in high VPE (see Thoms & Walkden 2019: Section 2.3).

High VPE does not strand dependents of the verb. According to Sailor (2014), high VPE is preferred to low VPE when the antecedent and ellipsis site are structurally equipotent, as in (44).

Example 44

Sailor (2014) notes that the interpretation of (44) preferentially includes the manner adverb carefully, so that the interpretation of the elliptical clause is ‘carefully review the book’ rather than just ‘review the book (not necessarily carefully)’. Sailor (2014) thus assumes that the adverbial is contained in the minimal projection targeted by high VPE.

The idea that manner adverbs are contained within the minimal target of high VPE is supported by their behavior with retorts and question tags.

Example 45

Example 46

The retort of (45) seems to necessarily include the unfairly part of the antecedent; it asserts that, in fact, John will penalize Molly unfairly. The question tag of (46) includes the carefully part of the matrix clause; the answer would be ‘no’ if Jordy had in fact reviewed the book carelessly. Thus, because retorts and question tags involve high VPE, (45)–(46) suggest that the domain of high VPE includes manner adverbials.

The same diagnostics can be applied to various other adjuncts to the verb, suggesting that all kinds of verbal dependents are included in the domain of high VPE. (47)–(48) illustrate with benefactives, comitatives and temporals, all typically considered adjuncts.

Example 47

Example 48

The interpretation of the retort in (47) and the question tag in (48) necessarily includes the ‘adjuncts’ in their antecedents. The retort in (47), for instance, asserts that ‘you will buy cookies for me/with me/on Tuesday’. The question in (48) would be answered ‘no’ if the buying didn’t take place ‘for me/with me/on Tuesday’. Thus, we should conclude, based on the diagnostics from Sailor (2014) and Thoms & Walkden (2019), that adjuncts of various kinds are contained in the domain of high VPE.

Because Thoms & Walkden (2019) argue that VP-fronting involves high VPE in the ‘extraction site’, this predicts that arguments and adjuncts cannot be stranded in situ by VP-fronting. That is, the arguments and adjuncts in the lower vP are necessarily contained in the domain of high VPE, as (47)–(48) show; they are, therefore, necessarily deleted. Thus, a representation like (49b) could not be derived for (49a) by high VPE.

Example 49

Such sentences are acceptable, however, as illustrated in Section 4.2 above. How can these sentences be derived on Thoms & Walkden’s (2019) analysis? The A/AD-based hypothesis that arguments are attached within the vP while adjuncts are attached outside it are of little help here. The domain of high VPE seems to cover arguments and adjuncts alike. Attachment height of arguments versus adjuncts thus can’t be used to explain what can be stranded and what can’t be.

I propose here that separation of dependents from the verb under VP-fronting can be derived via extraposition from the higher vP. That is, an instance of VP-fronting like (49a) would be analyzed as in (50). Both the fronted vP and the lower vP contain the stranded PP. High VPE deletes the PP in the lower vP, and the PP in the fronted vP is extraposed from Spec-CP.³⁰

Example 50

Importantly, there is independent evidence that extraposition can take place from Spec-CP in English. Consider (51)–(52).

Example 51

(Adapted from Reeve 2012: 77 fn. 19 ex. (i))

Example 52

(Adapted from Culicover & Rochemont 1990: 43 ex. 50)

In fact, the idea that dependents ‘stranded’ from a fronted vP are actually extraposed is suggested by Baltin (2017: 242). Baltin (2017) suggests specifically that the apparently stranded dependents in (53) are separated from the vP via extraposition.

Example 53

This analysis is supported by Culicover & Winkler’s (2019) observation that constituents which can be extraposed can also be ‘stranded’ by VP-fronting. PP dependents of the verb, arguments and adjuncts alike, can be extraposed, explaining core examples like those in (36)–(39). It also explains why clausal complements to the verb can be stranded by VP-fronting, as illustrated in (54a). Strikingly, it additionally explains why heavy NPs can be stranded, while their light counterparts cannot be (Janke & Neeleman 2012, Culicover & Winkler 2019), as in (55): heavy but not light NPs can be extraposed.³¹

Example 54

Example 55

With this extension (i.e. allowing extraposition from the fronted vP), Thoms & Walkden’s (2019) analysis of VP-fronting can accommodate possibilities of separation without the DAVD hypothesis. Since both arguments and adjuncts of different kinds can be extraposed, there is no expectation that VP-fronting should be sensitive to the A/AD.

5. VP-Ellipsis and Pseudogapping

The preceding three sections of this paper considered three different phenomena commonly cited as separability diagnostics for argumenthood: do so-anaphora, do what-pseudoclefting and VP-fronting. As noted in Section 1, another phenomenon sometimes cited (e.g. by Hornstein & Nunes 2008) as evidence for a configurational distinction between arguments and adjuncts is VP-ellipsis (VPE).³² As advocates point out, if we assume (i) DAVD and (ii) that VPE targets a vP constituent for either syntactic deletion or PF-non-realization (see Merchant 2018), then VPE makes a natural diagnostic for argumenthood. Ellipsis would be able to strand adjuncts to the verb but not arguments of it.

Example 56

Under the right analysis along these lines (see below for more discussion), VPE would meet the definition assumed in this paper for separability diagnostics and could potentially provide evidence for DAVD as a property of the A/AD.

VPE is even more vast a topic, with more competing analyses and intricate confounds than the other three diagnostics considered above. The potential relation between VPE and DAVD hangs on the interaction of a variety of distinct aspects of the analysis of ellipsis, no one of which can be addressed in detail in this paper. As such, it would not be feasible for this paper to examine the DAVD-relevant aspects of VPE along the same lines as in Sections 2 – 4 above. The status of VPE as a separability diagnostic, therefore, cannot be settled here.

Nevertheless, I conjecture that the prospects that VPE will make a successful separability diagnostic are not promising. The reason is that, for VPE to succeed in this way, a very specific constellation of analytical stars must align, some aspects of which seem particularly unlikely. First, VPE must be a syntactic operation, crucially involving the derivation of a vP which becomes suppressed in the syntax or at spellout. If an alternative non-structural approach like that of Culicover & Jackendoff (2005) were adopted, then VPE would not make a successful separability diagnostic. Second, VPE must have the effect of silencing the complete vP constituent that it targets, such that any material stranded by ellipsis must be located outside the vP. In other words, we need to adopt a ‘move-and-delete’ approach to ellipsis remnants. If we were to adopt an approach allowing non-constituent deletion, then VPE would not be relevant for DAVD. Third, our analysis must distinguish the remnants of ordinary VPE from the remnants of pseudogapping. That is, constructions like (57a), where a direct object survives ellipsis, must be derived by mechanisms distinct from those by which in five minutes is stranded in (57b). Otherwise, argument- and adjunct-remnants alike could survive ellipsis by the same move-and-delete procedure.

Example 57

Further, we must assume that argument-PP remnants of ellipsis result from pseudogapping, not from ordinary VPE, which is not obviously true. Levin (1979: 17) excluded PP remnants results from the pseudogapping construction on the basis of the observation that examples with PP remnants tend to have a much higher acceptability than the prototypical pseudogapping examples with NP remnants. Lasnik (1999) also develops an analysis whereby pseudogapping leaves NP remnants only. Argument and adjunct PPs alike would then have to be remnants of ordinary VPE, as explicitly assumed by Janke & Neeleman (2012), for example.

With the necessary analytical assumptions in place, it would have to be shown, using robust criteria for distinguishing pseudogapping from ordinary VPE (criteria which prove difficult to pin down in the first place; see Miller 2014), that argument PPs are systematically unacceptable as remnants of ordinary VPE, while adjunct PP remnants are systematically acceptable. In ordinary contexts, however, this contrast is not obvious. I do not detect a systematic difference in acceptability between examples like (58) and examples like (59). The former involves conventional argument PPs and the latter conventional adjunct PPs.

Example 58

Example 59

Some of the necessary assumptions just described are quite plausible (see, e.g. arguments for a structural approach to ellipsis in Van Craenenbroeck & Merchant 2013, Merchant 2018). However, others are less likely. For example, there are many challenges for the move-and-delete approach to ellipsis remnants (see, e.g. Ott & Struckmeier 2018, Broekhuis & Bayer 2020, Ott & Therrien 2020, Griffiths et al. 2023). Additionally, the necessary distinction between argument and adjunct PPs with respect to pseudogapping is not obvious, as reflected in Levin’s (1979) explicit choice to exclude PP-remnants from the pseudogapping construction and as reflected in the similar status of (58)–(59).

For the purposes of this paper, I conjecture that VPE will not make a successful separability diagnostic and will not revitalize DAVD as a property of the A/AD. Concretely, this conclusion is compatible with the analysis of PP-stranding VPE put forth in Janke & Neeleman (2012) – they argue that PPs stranded by VPE attach to the VP in an ‘ascending’ fashion (i.e. right-adjoined to VP), with VPE targeting the full VP constituent to which the PP adjoins.

Example 60

Crucially, their proposal allows for both argument and adjunct PPs to attach in this way, so that arguments and adjuncts alike can be stranded by the same VP-deletion operation. In contrast, DP arguments must attach in a ‘descending’ fashion (i.e. as complement to V or as specifiers of higher VP shells).

Example 61

Thus, DP-stranding VPE could not be accomplished simply by deleting a single VP constituent without prior evacuation of the DP.

6. General Discussion

Separability diagnostics for argumenthood are predicated on DAVD, the idea that arguments systematically attach within a particular syntactic projection of the verb and that adjuncts systematically attach outside that projection. This paper examined the evidence in favor of this hypothesis, taking the relevant projection, for concreteness, to be vP: arguments attach within vP, adjuncts outside vP. Sections 2–4 of this paper demonstrated that do so-anaphora, do what-pseudoclefting and VP-fronting fail to provide consistent evidence in support of this hypothesis. The evidence shows that none of these phenomena is uniformly sensitive to a purported syntactic difference between argument versus adjunct PPs. Section 5 considered VP-ellipsis as further potential motivation for DAVD but suggested that the combination of analytical assumptions this requires is unlikely.

These results support the null hypothesis regarding DAVD, that is, the hypothesis that there is no systematic difference in attachment height between arguments and adjuncts. The two competing hypotheses can be formulated as in (62).

Example 62

The null hypothesis H₀ is strictly more parsimonious than the alternative H₁, which involves an added stipulation which is absent from H₀. The idea that arguments and adjuncts are systematically distinguished in terms of attachment height, therefore, bears a burden of proof – it must be motivated through demonstration of empirical necessity. Traditionally, do so-anaphora, do what-pseudoclefting, VP-fronting and VP-ellipsis have been invoked as the necessary sort of motivation.

Some authors have proposed that the A/AD be reconceptualized as a ternary distinction (e.g. Hedberg & DeArmond 2009, Needham & Toivonen 2011) or as a gradient distinction (e.g. Forker 2014, Rissman et al. 2015, Kim et al. 2019). One might suspect that DAVD could be salvaged if stated as a property of one of these more sophisticated approaches. However, I maintain that the arguments of this paper apply even if we adopt a ternary or gradient A/AD. The purview of such approaches is the set of dependents for which informal intuitions of argumenthood are not always clear, like instrumental with-phrases and perhaps goal PPs. However, I showed above that the most prototypical arguments (e.g. patient direct objects, recipient to-phrases) behave counter to DAVD-based expectations. This is just as problematic for a ternary/gradient A/AD as for the conventional A/AD.

Therefore, if the conclusions of this paper are correct, and the phenomena examined above are not, in fact, sensitive to the A/AD, then the question arises to what extent H₁ is justified. Given that these phenomena fail to provide evidence against the null hypothesis, what reason is there to accept the idea that the A/AD is instantiated by a systematic distinction in attachment height? The full answer to this question lies beyond the scope of this paper and would involve considering in detail a range of other properties associated with the syntactic A/AD, including omission, iteration, islandhood and more. However, to the extent that separability diagnostics comprise a primary source of evidence for this particular aspect of the A/AD, the case for rejecting H₀ is substantially weakened.

At the very least, then, the results of this paper substantially weaken the case in favor of H₁. This conclusion supports theories of clause structure which do not posit DAVD as a property of the A/AD, consistent with H₀. In fact, despite the perceived importance of the A/AD in current theory, DAVD is absent from many current approaches to clause structure. For instance, Larson (1988, 1990) argued for a VP structure such that arguments and adjuncts alike are treated as specifiers to V⁰ (save for the rightmost dependent which is the ‘innermost complement’ to V⁰). Later developments (e.g. Pesetsky 1995, Schweikert 2005, inter alia) have a similar property. Culicover & Jackendoff’s (2005) treatment of VP also draws no configurational distinction between arguments and adjuncts; all are complements to the verb in a flat structure. The argumentation against such A/AD-free analyses has been sourced in part from the phenomena discussed in this paper (see Sobin’s 2008 arguments against Culicover & Jackendoff’s flat structure and Jackendoff’s 1990 arguments against Larson’s VP shells). This paper supports the A/AD-free analyses (Larson 1988, Culicover & Jackendoff 2005, etc.) by eliminating these potential counterarguments.³³

This paper considered a particular source of evidence for the syntactic A/AD, separability diagnostics. This is only one of a number of different purported motivations for the syntactic A/AD. Other phenomena, including islandhood, iteration, extraction from weak islands, Condition C reconstruction and selection, have all been taken to motivate further properties of syntactic argumenthood. The results of this paper alone, therefore, should not be taken to cast doubt on the syntactic A/AD as a whole, since all these other properties are widely accepted as independently supporting the distinction. Nevertheless, this work contributes to a trend alongside other detailed investigations showing these additional properties also fail to provide evidence for the A/AD (e.g. Payne et al. 2013 on one-substitution, Miliorini 2019 on weak islands, Bruening & Al Khalaf 2019 on Condition C reconstruction, McInnerney 2023 on the adjunct island effect; and see Przepiórkowski 1999, 2016 for critique of the A/AD involving a variety of diagnostics). If this trend continues, it will become increasingly plausible that the A/AD could play a weaker role in syntax than previously assumed.