As mentioned in section 1, the approach of Cinque 1994, 2010, and 2023a orders conceptual classes such as color, size, shape, value, origin, and age syntactically, as a sequence of functional heads taking APs as their specifier, as in (4). Evidence against this analysis is presented by Svenonius 2008 and Truswell 2009 for English. It is true that certain classes of adjectives are ordered with respect to one another—but other classes are not. In English, the adjective pair in (13) is rigidly ordered, but those in (14) and (15) are not (Truswell 2009: 527; see also Svenonius 2008: 39).3

In other words, cartographic hierarchies like that in (4) end up undergenerating. This point is well understood from previous literature and need not be further elaborated. Let me just note that in Italian, as expected, multiple prenominal adjectives in ANA configurations, corresponding to the classes in (15), are also freely ordered:

Larson 2021 advocates retaining semantic classes as a basis for adjectival syntactic order, arguing that what is wrong with (4) is simply the cartographic structuring of the classes. Larson proposes that the features identified by cartographers are borne by a single head, for instance N, where they “undergo agreement in order from lowest ranked to highest ranked” (256–257), determining the surface order of the APs that they enter into Agree with. Since the order of features is not preencoded but rather determined by the subjectivity relation (Scontras et al. 2017, 2019), it follows that for any given pair of adjectives, the relation may be well defined or not—and if it is not, the order of the relevant adjectives is arbitrary. In this way, the empirical problem of undergeneration is addressed. At the same time, there are theoretical reasons why this proposal is problematic—which are shared with cartography. In Larson’s account, properties such as [color] or [shape] or [size] are treated as formal features since they enter into Agree exactly like formal features such as ϕ features or Q features. However note that unlike what happens with ϕ features and Q features, there is no systematic Minimality effect based on [color], no EPP effect whereby [color] is obligatorily spelled out at some phase, and so on. Therefore encoding semantic classes into formal features makes the wrong syntactic predictions.4

I therefore turn to alternatives avoiding conceptual classes. Truswell offers the conclusion that where an ordering of two adjectives is strongly preferred, the higher adjective is subsective while the lower one is intersective; (13) is an example. Similarly, Svenonius 2008, on the basis of a structure of NP that includes a lower nominal class head (n) and a higher count head (Sort), concludes that “modification of nP is essentially intersective”; vice versa, “SortP modification occurs in a different way from nP modification: it is crucially subsective” (38).

However, subsectivity versus intersectivity, despite its popularity among semanticists and syntacticians, does not seem the correct notion to employ. Take the example in (17). This is naturally interpreted with the first adjective (‘beautiful’) denoting an intersective property (i.e., beautiful as well as a linguist) and the second adjective a subsective one (i.e., ‘good as a linguist’).5 Therefore the order is intersective > subsective, contrary to the order predicted by Svenonius.

Another notion mentioned by Svenonius is scalarity. Scalarity is ultimately responsible for the fact that ‘a small elephant’ is certainly larger in size than ‘a big butterfly.’ This is because ‘small’ and ‘big’ individuate degrees (low, high) on a size scale but, in order to fix the scale, one needs to have a standard of comparison, by default the average size of elephants or butterflies (see especially Kennedy 1999, 2007).

In a strictly modular grammar, we do not expect syntax to encode semantic properties, any more than phonological ones. Intersectivity and subsectivity, which are defined by Kamp & Partee 1995 in terms of the inferences they license (see footnote 5), involve a semantic look-ahead of some sort when invoked as a principle of syntactic order. Scalarity is different, since it is overtly instantiated in the morphosyntax of comparative and superlative constructions and in structures with degree modifiers (very, etc.) or with measure-phrase modifiers (e.g., one meter tall, etc.). The analysis of adjectival order developed in the next section relies on scalarity and on phase structure.

For the sake of concreteness, I adopt Kennedy 1999 and 2007’s analysis of scalar adjectives, under which they are modified by a DegP (Corver 1990), whose content in the positive (noncomparative) form is conventionally designated as pos. In (18), pos introduces a relation between the degree x at which the referent of the DP (‘a butterfly’) has the AP property and the degree y at which the standard-of-comparison group (‘butterflies’ as a default) has the property.6

Adjectives high in the adjectival hierarchies (evaluative, size, age) are gradable. Adjectives lower in the hierarchies (relational, shape, color) are nongradable, as shown by their incompatibility with comparative syntax:7

We now need to ask why gradable adjectives appear higher than nongradable adjectives (and therefore to their left), both in English AAN and in Italian ANA. In the kind of syntactic framework adopted here, it is natural to assume that the organizing principle may simply be phase structure. Gradable adjectives are merged at the edge of nP and nongradable ones in the complement of n, as schematized in (20a) for English and in (20b) for Italian.

Before attempting to answer the further question of why gradability and phase structure would be matched as in (20), as opposed to any other logically possible way, I present some more evidence in favor of gradability as an order factor for adnominal adjectival modification. I use Italian examples, where, by hypothesis, the gradable class is easily isolated by its linear position to the left of N.

Relational adjectives, including nationality adjectives, are normally postnominal as in (21a, d). However, as noted by Svenonius 2008 they can be coerced into a meaning whereby ‘Italian,’ for example, does not mean ‘from/of Italy’ but rather ‘Italian-like’—in which case they are gradable.8 It is then expected that adding superlative morphology to a nationality adjective forces this second meaning and makes the adjective compatible with prenominal position, as is the case in (21c, f).

Another case in point is participial adjectives. It is often pointed out that these are generally postnominal in Romance (e.g., Abeillé & Godard 1999 for French, Demonte 1999: 189 and Bosque 1999: 299–300 for Spanish). However, some prenominal participial adjectives are possible, for instance in (22a). Bosque 1999 observes that the very same participial adjectives may be unacceptable in prenominal position in other types of examples, like (22b). Traditionally, the lexically identical modifiers in (22) are described as having an adjectival versus passive interpretation. The interpretive difference may be better understood as akin to that between causative and anticausative predicates (smarrire ‘to lose’ vs. smarrir-si ‘to be lost,’ as a psychological predicate).

What is relevant for present purposes is that the adjective in (22b) is nongradable; to be more precise, it belongs to a semantic class of (participial) adjectives such as open and closed that Kennedy 2007 describes as “absolute gradable”: in essence they cannot enter comparatives, as in (23a), though they do admit some measure modification. By contrast, the adjective in (22a) can enter comparisons, as shown in (23b).

Taking the possibility of comparatives and superlatives to be the operational definition of gradability relevant for syntax, the correlation between gradability and edge-of-nP position (pre-N position in Italian) is confirmed.

Let us then return to the question of why gradable adjectives are at the edge of nP. We may observe that Deg introduces a second term of comparison in the form of a variable (the comparison standard) that is contextually set. Nothing much then needs to be said about the positioning of scalar adjectives at the phase edge, except that all operator–variable structures are at the phase edge. Specifically with DegP at the nP edge, the default standard of comparison for the Deg relation can be set by nP. Alternatively, recall that the edge of nP is in the Workspace at the next phase up, say v*, making the larger sentential context available for the setting of the comparison standard (see Heim & Kratzer 1998: 71 for some examples of nondefault comparison standards).9

In a nutshell, the question of the relative order of adjectives must be substantially reformulated. Multimembered adjectival scales, as reconstructed in syntax by cartographic hierarchies, have a problem of undergeneration. This problem could in principle be avoided by adopting more elementary scales based on macrocategories (subsectivity, subjectivity)—but the semantic nature of the latter seems to sit uneasily with a modular view of syntax. A syntactically viable categorization of adjectives is that opposing gradable and nongradable adjectives, which is also sufficient to capture the distinction between adjectives that can appear prenominally in Italian and those that must appear postnominally. I propose that the two macrocategories are matched to the phase structure of the nP without the postulation of any dedicated functional head besides the phase head n. These results are summarized as follows.

There is in fact another class of adjectives that can be argued to require the same edge-of-nP positioning as gradable adjectives but for independent reasons. This is a small class of adjectives that can only appear prenominally in Italian. An example is mero ‘mere’; another is semplice ‘simple,’ which is a synonym of mero in prenominal position only. These facts are illustrated in (25). Prenominal mero and semplice are focalizers, like their adverbial counterpart solo ‘only’ in (26).

We know that focus operators sit at phase edges, as do wh operators and negative operators: that is, the operators that Chomsky 2013: 46 suggests unifying under an F (Force) feature. The connection between these operators and phase edges is provided by the postulate that phase heads host formal-feature probes, including [Q], [neg], and [F], which may trigger EPP effects. Overt realizations of operators (question particles/wh phrases, negative clitics/adverbs, etc.) are therefore hosted by phase edges. I suggest that this standardized picture is sufficient to force focalizer adjectives to be merged at the nP edge:

Recall from example (5) that reversing the gradable > nongradable adjective order in prenominal position in English does not necessarily yield ungrammaticality. A better description of the resulting pattern is in terms of markedness. Thus, a well-formed reading can be obtained if the leftmost adjective is focused:

In present terms, focusing is construed at the edge of the phase headed by n, allowing the nongradable adjective (normally attached to NP) to surface in leftmost position.

Let us then return to the conclusions in (24). The one in (24b), as it now stands, does not take into account the fact that Italian also has NAA orders, mentioned in section 1 but not analyzed so far, where gradable and nongradable adjectives freely mix. I will come back to unordered adjectives in section 4. I will also come back, in section 5, to a different issue that arises if one considers variation among the Romance languages (Bernstein 1993), namely whether the gradable–nongradable distinction is in fact too coarse and ends up undergenerating, like cartographic hierarchies but for opposite reasons.

In the final subsection of section 3, I focus on (24a). Assuming that a coarse (or partial) hierarchical order of adjectives is empirically correct, is there evidence that the n phase head is effectively responsible for it? In principle, the relevant order could be determined by a reduced functional hierarchy, consisting for instance of SortP and nP, as posited by Svenonius 2008. In what follows, I provide evidence for (24a) from so-called lazy plurals in Romance, based on the assumption that phases are the structural chunks undergoing Transfer and Spellout.

A relatively well-known correlation between adjective positions and phonological rules concerns French liaison, where AN is a possible context of liaison but NA is disfavored if not impossible (e.g., Bouchard 1998: 150–151). Here I focus on so-called lazy agreement (Cinque 1994: n. 28): a differential realization of plural morphology on pre-N modifiers versus post-N modifiers, attested in Romance varieties, including Italo-Romance varieties and neighboring Rhaeto-Romance ones. The distribution is not lexically determined, in that the same adjective will or will not bear the relevant morphology depending on its prenominal or postnominal position. In dialects of Brazilian Portuguese, as illustrated in (29), the plural inflection -s surfaces only on determiners and prenominal modifiers. In Ladin, as (30) illustrates, the mirror distribution prevails, in that -s surfaces only on N and on postnominal adjectives.10

Discounting morphological accounts (Distributed Morphology, Optimality Theory), two main types of explanation have been proposed in the literature: in terms of syntactic Agree and in terms of Spellout. Agree explanations include Rasom 2008 on Ladin as well as cross-linguistic accounts by Stark & Pomino 2009 and Cyrino & Espinal 2019. For instance, Cyrino & Espinal propose that in Ladin (30) the [plural] probe is in n and takes as goals only nP-internal constituents, so that nP serves as the top boundary for Agree. By contrast, in Spanish and Italian, the [plural] probe is in D, so that all terminals in DP check the relevant morphology. The lower boundary for pluralization observed in Brazilian Portuguese (29) is determined by an independent parameter concerning the overt realization of [plural]. In short, the patterns in (29) and (30) are not given a unified analysis. Rather, Ladin is accounted for as a core-syntax (Agree) phenomenon while Brazilian Portuguese is given a different explanation in terms of Spellout.

Savoia et al. 2020 and Manzini et al. 2021 propose to unify the two patterns in (29) and (30) under a single parameter at Externalization. Their idea is that [plural] is uniformly represented in the syntax of the languages under consideration but undergoes differential Spellout. What is crucial for present purposes is that according to these works, the left–right asymmetries in the distribution of -s seen in (29) and (30) reflect phase boundaries, as expected under the minimalist assumption that Transfer takes place at the phase level.

Savoia et al. 2020 presents empirical evidence against syntactic Agree analyses of lazy plurals. In Friulian varieties, plural is spelled out on all members of the DP; however, on prenominal categories it is represented by a purely vocalic morpheme -i, while on N and postnominal modifiers it is spelled out as -i- followed by -s morphology:

Savoia et al. conclude that in Friulian, syntactic agreement in plural affects all nominal nodes. At the same time, left–right asymmetries are observed concerning the spellout of -s morphology, determined at Transfer.

In short, it is desirable to treat (29–31) as different outcomes of the same parameter. Savoia et al. 2020 and Manzini et al. 2021 suggest that such a result can be achieved if the whole set of lazy-agreement phenomena is taken to result from an Externalization parameter. Specifically, they argue that the left–right asymmetry, which remains constant in (29–31) (despite the contrasting distribution of -s morphemes), reflects the fact that the relevant nominal structures are sent to Spellout in two chunks, roughly corresponding to complement versus edge of phase. I simply update Savoia et al. and Manzini et al.’s analysis, showing its relevance for the conclusions in (24).

Under the Phase-Impenetrability Condition of Chomsky 2001: 14, upon merger of the phase head Z the complement of the next phase head down, X, becomes inaccessible to syntactic operations. This is because the relevant domain is assumed to undergo transfer to the interfaces. Recently Chomsky 2021 suggests a revision of the classical Phase-Impenetrability Condition under which “v need no longer be at the edge of the vP phase, but can be within the domains of [the Phase-Impenetrability Condition] and Transfer” (36). I adopt this definition here:

What is relevant for present purposes is that upon merger of the phase head Z the immediately lower phase head X is transferred together with its complement. Thus, assuming that all ANA configurations in Romance have the same structure as in Italian (20), we predict that Transfer will chunk the DP exactly as required by the contrasts in (29–31). Upon merger of a superordinate phase head, n and its complement (including the noun in n and the modifiers of NP) undergo Transfer. By contrast, the edge of nP (including the modifiers of nP, as well as determiners, quantifiers, and higher functional categories) undergoes Transfer as part of the next phase up. As already indicated, I assume that the [plural] feature is represented on all nodes of the DP. At Transfer, a parametric choice active in some Romance languages determines whether the [plural] feature is in fact spelled out in a given phase or instead undergoes ellipsis, construed as No Transfer (Saab 2008). So, in principle [plural] could be spelled out as -s in the nP chunk but not in the chunk that includes the edge of nP or vice versa.

Concretely, consider the Brazilian Portuguese examples in (29), with the structure in (33). In (33a) the nP phase (shaded) is packed off to the interfaces on the basis of (32). A language like Brazilian Portuguese enforces a parametric choice to the effect that within nP, [plural] is elided, that is, simply not sent to Spellout. The prenominal adjective is at the edge of nP and therefore escapes the nontransfer of [plural]. In (33b), on the other hand, the adjective is postnominal. By present assumptions, since the adjective is merged with NP, it is predicted to be sensitive to the same parameter as the noun in n, that is, nontransfer of [plural], correctly.

Consider next Friulian (31), which presents the mirror distribution of -s in Brazilian Portuguese.11 While both -i and -s instantiate [plural], I propose treating them as exponents of two different word-internal categories. The vocalic morpheme -i externalizes plural as part of nominal class features (including gender); in other words, plural -i occupies the same slot as feminine singular -a. In contrast, -s is a pure exponent of plurality, arguably merged in a separate word-internal node insensitive to gender. We can call the two nodes Class and Number (Num) as in Picallo 2008. When they combine, they yield the following morphological structure.

At Transfer, Friulian mirrors Brazilian Portuguese, as shown in the structures in (35). The shaded portion of the structure, namely nP, is transferred with the -s exponent of plurality duly spelled out. Transfer of the edge of nP and of higher material takes place at the next phase up. At that point, some parameter of Externalization requires nontransfer of the higher plural node. The postnominal adjective in (35b) undergoes Transfer as part of nP and therefore displays -s morphology. The higher adjective in (35a) is at the nP edge and reflects the different conditions that characterize Transfer at the higher phase (no -s).12

One may surmise that uniform lexicalization of features across Spellout domains would be a simpler choice. On the other hand, note that there is a functional advantage to breaking uniformity at the phase boundary, namely that it makes the phase structure maximally transparent. This is in fact the standard “third-factor” explanation of phonosyntactic rules like French liaison (whether expressed in terms of phases or major constituents or something else).

In any event, there is a good match between the structure for adjectival modification motivated here and the evidence provided by partial spellout of plural morphology at the phonology interface. Specifically, the noun in n and lower adjectives belong to a different Spellout domain than higher adjectives (at the edge of nP). This elementary structure is necessary. I return in section 5 to the question of whether it is sufficient.

In section 3, I have considered adjectives that are ordered with respect to one another—leaving out the various instances where adjectival modification yields grammatical results under any linear order. Two subcases of this are considered here. The first one is the English pattern illustrated in (14) and (15), reproduced in (36) for ease of reference. In present terms, two gradable adjectives, as in (36b), or two nongradable ones, as in (36a), are free to occur in any order.

The second case of free order that I consider is the free intermixing of gradable and nongradable adjectives in Italian, which constitutes the main topic of this section (i.e., section 4).

Let me however begin by addressing the data in (36). Chomsky 2019 and 2021 introduce the descriptive problem of unbounded unstructured sequences of modifiers, including adjectives as well as PP modifiers. These sequences seem to be unbounded except for the obvious limitations imposed by memory; they are also unstructured in the sense that no discernible syntactic order prevails among them. Hence, their linear order is free. Chomsky 2024 proposes that unbounded unstructured sequences are generated by an operation FormSet along the following lines (cf. Chomsky 2021: 31–32).

According to Chomsky 2024 “binary [FormSet] is distinct from Merge in lacking its special theta-related properties.”13 Furthermore, “within sentence grammar, sets formed by [FormSet] are introduced into derivations in the position of single XP’s.” In this context, Chomsky’s statement that “order is imposed in Externalization” implies that linearization can work with syntactic structures that are not preordered by asymmetric c-command.

I propose adopting the theoretical constructs just reviewed and applying them to the cases in (36), without much modification. This means that the underlying syntax is the same for both surface orders observed in (36a) and (36b), namely an unordered two-member set merged as a single syntactic object. The only difference between (36a) and (36b) is that, in line with the discussion in the previous sections, the nongradable-adjective set is adjoined to NP, as in (38a), while the gradable-adjective set is adjoined to nP, as in (38b).

The linear order of the two adjectives is determined by Externalization and therefore is expected to be insensitive to semantic classes, though it may be sensitive to phonological properties (e.g., shorter vs. longer; Bošković 2009: 191) or to usage-based considerations (e.g., frequency of occurrence with N; Dyer et al. 2023).14 Once linear order is factored in, structures like (38) are equivalent to non-binary-branching (n-branching) trees, for instance (39) for (38a). This kind of structure is explicitly proposed and motivated by Neeleman et al. 2023 for coordinations, including (coordinated) adjectival modifiers.15

Similarly, in the Italian ANA order, two gradable adjectives in prenominal position are unordered with respect to one another, as in (16), repeated in (40a) for ease of reference. In present terms, they form a two-member set merged with nP as a single object, as in (40b), and their precedence order is not fixed by syntax but only by Externalization.

We then get to the NAA configuration in Italian. The basic facts about this configuration are that all adjectives (except functional ones) can be postnominal and that in postnominal position gradable and nongradable adjectives are freely intermixed, as exemplified in (7), repeated in (41a) for ease of reference. Suppose we apply the FormSet treatment to (41a). We obtain a structure like (41b) where a set containing the gradable adjective and the nongradable adjective is adjoined to NP as a single syntactic object. At Externalization, the set surfaces postnominally. Both precedence orders between the two adjectives are grammatical; criteria influencing the choice are to be found only in extragrammatical or extrasyntactic factors.

But if a structure like (41a) is possible in Italian, why wouldn’t it be possible to have the same structure in English, leading to free order to the left of N?

The key insight on unordered adjectives goes back to Sproat & Shih 1991. There it is observed that in Chinese (where adjectival modification is consistently prenominal) bare APs are ordered by semantic classes but adjectives embedded under the particle de are not. Sproat & Shih take the latter to be relative clauses. Following their lead, Cinque 2010 accounts for the fact that adjectival order is not observed in Italian NAA by assuming that postnominal adjectives can be reduced relative clauses, which are unordered. The rationale behind this kind of analysis emerges perhaps most clearly from Larson 2021. Larson observes that in English adjectival-modification structures, for example (42a, b), a head H to which the AP is attached “will have strictly local access to A and its syntactic and semantic features.” However this access fails “where AP is embedded within a larger CP,” that is, the relative clause in (42c, d). Relative clauses are therefore freely stacked, despite embedding the very same modifiers rigidly ordered in (42a, b).

In other words, adjectives that are embedded under a complex enough structure (typically, a relative clause) may be opaque to the element they modify, hence impossible to order with respect to it and to one another.

Let us then focus on Cinque 2010’s implementation of the reduced-relative-clause analysis for Italian. One key property of Cinque’s account is that an NP projection always moves to the left of reduced-relative-clause adjectives (generated as specifiers of some functional category F), so that reduced relative clauses are always postnominal, as in the following derivational schema (RC = relative clause).

The second key property of Cinque’s implementation depends on a further cartographic postulate, Uniformity (in the sense of Culicover & Jackendoff 2005), to the effect that a given semantics necessarily corresponds to a fixed syntax. Therefore, different syntactic structures are expected to have different semantic reflexes. In the case at hand, reduced relative clauses are mapped to a predicative semantics (intersective, restrictive, stage-level interpretations of the adjectival modifiers) as opposed to the nonpredicative semantics of bare APs (subsective, nonrestrictive, individual-level). The prediction is thus that prenominal adjectives in Italian can only have a nonpredicative semantics, since the structures that are responsible for predicative semantics, namely reduced relative clauses, are obligatorily postnominal. Manzini 2024 argues in some detail that the prediction is incorrect. In (44a) I reproduce just one example where the prenominal adjective is clearly restrictive; (44b), already mentioned in (22), shows a prenominal stage-level adjective; and finally (44c) is ambiguous between intersective and subsective readings of the prenominal evaluative adjective, essentially just like its English counterpart.16

Therefore if we take prenominal APs to be direct-modification APs, we must conclude that direct-modification APs can have predicative interpretations.

Scontras et al. 2024 also argues that Cinque 2010 has a problem with respect to English. Recall Cinque’s explanation for the free order of Italian postnominal adjectives: it is due to derivation (43b), under which only postnominal adjectives can be reduced relative clauses, which are freely ordered. At the same time, structure (43a) is universal, holding for English as well as for Italian; it is simply the case that in English no movement takes place. Now, since reduced relative clauses can produce free postnominal orders in Italian, we expect them to do so in English prenominal position. Take for instance a big blue box/a blue big box in (5). On the basis of (43a), we should be able to construct both prenominal adjectives as reduced relative clauses. At this point, we would predict free order—incorrectly. Alternatively, one must abandon the idea that English has prenominal reduced relative clauses. If so, we must conclude all over again that direct-modification APs can have predicative interpretations.

Conversely, we may wonder whether an implication may hold to the effect that reduced relative clauses, when present in a language, must have predicative meanings. Here the counterargument is trivial: languages that overtly have functional embeddings of some sort for adjectives, such as Chinese de, typically embed core subsective and functional adjectives in the relevant structures (e.g., former; see Paul 2012: 193 for Chinese examples involving de). This means that syntactic properties of order (rigid vs. free) or structure (AP vs. reduced relative clause) are dissociated from semantic properties (predicative vs. nonpredicative) and are best discussed separately. This is what I will do in section 4.3.

In view of the recent work in Cinque 2023a, one may also wonder whether cartographic models have a different avenue of explanation open to them. Cinque discusses the well-known typological generalization according to which more nominal-modifier orders are attested postnominally than prenominally. In Cinque’s analysis this follows from the fact that XYN sequences correspond to the base order while N-initial sequences are generated via movement. Depending on the different types of movement that apply, different postnominal orders can be derived. One may then wonder whether one may derive free adjectival order in the Italian NAA structure simply from the fact that different types of movements can apply to derive NAA from AAN. Cinque himself very clearly associates each language with one single sequence—for instance, Italian has mirror order and only mirror order. Indeed, if one admitted that movement could freely scramble postnominal adjectives within one and the same language, then one would expect the same of {Dem, Num, A} sequences—which is not the case. Therefore, as things now stand, this avenue of explanation is closed.

So far, I have not questioned the assumption that what appear to be simple adjectives (e.g., in Italian) are reduced relative clauses. However, Belk 2017, specifically addressing the possible reduced-relative-clause status of adjectives, states that “adnominal adjectives, reduced relatives and full relatives are derived from three separate sources” (34). For Den Dikken 2006 and Den Dikken & Singhapreecha 2004 functional elements embedding APs in many languages (including de in Chinese) are linkers, that is, copula-like elements that define elementary predications. Furthermore, treatments of linkers as agreement heads have also been entertained in the literature (Lekakou & Szendrői 2012, Philip 2012, Franco et al. 2015). I therefore make no assumptions about the reduced-relative-clause status of some adjectives but take up the whole question from scratch in the next section.

In minimalist theory, each lexical head is associated with a phase head, as discussed in section 2 in relation to the n-N structures in (9). It is also standardly assumed that there is variation as to whether the phase head is defective. Following standard practice, then, I assume that there is a phase head a associated with the lexical head A. I further assume that a can in principle be defective, as in (45a) as opposed to (45b).

Let us assume that the a* head is phasal in that it supports an external argument. The argument introduced by the phase head is notated as φ in (46b) to indicate a minimal pronoun interpreted as identical to a local antecedent and therefore not undergoing Spellout (i.e., descriptively a PRO). Specifically, the minimal pronoun is bound by D, assuming that D is the element that saturates the R-role of N (Higginbotham 1985). Nonphasal adjectives have the simpler structure in (46a).

I propose that the derivation outlined in section 3 holds for nonphasal adjectives. Recall that under this derivation, in order for the standard of comparison required by the degree modifier to be interpretable on gradable adjectives, it is necessary for AP to be positioned at the edge of nP. In English, then, gradable adjectives are forced to the left of nongradable adjectives, to the extent that they are nonphasal.

On the other hand, in Italian gradable adjectives can appear in the same adjectival sets as nongradable ones, ordering freely with respect to them. To be more precise this free order is observed in postnominal position: that is, the sets are generated in the complement of n, as already proposed in (41) and repeated in (47) for ease of reference.

The crucial case to be considered is that of a gradable adjective, for example grosso ‘big’ in (47), since nongradable adjectives occur in the complement of n in any event. I propose that the gradable adjective is legitimate in (47) to the extent that it has the richer phase structure in (46b). The minimal pronoun φ provides an explicit representation, internal to the structure of nP, of the argument of N. Therefore, the standard of comparison necessary for the correct interpretation of Deg can be fixed by reference to φ. This means in turn that there is no need for the gradable adjective to be at the edge of nP.

In short, I propose that adjectives that display free-order properties are phasal. Italian is a language with phasal adjectives occurring unordered inside the nP. Recall on the other hand that I have accounted for ANA Italian patterns along the same lines as English AAN orders. The obvious conclusion to draw about Italian, then, is that it has both phasal and nonphasal adjectives. The latter, being merged obligatorily at the phase edge, yield ANA. This is stated in (48a).17 In English, APs are nonphasal and obligatorily ordered by gradability. This is stated in (48b).

We of course expect that there should be languages that only have phasal adjectives. These would be languages where adjectives are obligatorily embedded under linker structures and are not ordered by conceptual classes. A case in point is Semitic languages (Fassi Fehri 1999 on Arabic, Shlonsky 2004 on Hebrew), whose adjectives are inflected for definiteness (i.e., they are embedded under linkers) and are unordered by semantic classes (Trainin & Shetreet 2021 on Hebrew, Dyer et al. 2023).