2.1. The Croissant

Traditionally, modern-day geographical France is divided into three main linguistic areas as shown in Figure 1, representing the development of Latin in Gaul: ‘langue d’Oïl’ to the north, ‘langue d’Oc’ or Occitan to the south, and Franco-Provençal in the Alps (Arpitan on the map).Footnote ³ Modern French is an Oïl variety. The variety studied in this article belongs to the contact area of Oïl and Oc, referred to as ‘the linguistic Croissant’ (Ronjat Reference Ronjat1913; Guérin Reference Guérin and Noirard2022) because of its crescent shape.

Figure 1 Linguistic areas of modern-day France.

The varieties of the Croissant display characteristics of both language groups. For instance, like Oc languages, many of them distinguish between singular persons in the imperfect form of the verb, a feature that has been lost to the north of the Croissant. Conversely, Oc languages exhibit final unstressed vowels other than [ə], but Croissant languages, like Oïl ones, are by and large stress-final (with principled exceptions). This feature, which will play a role in our analysis, is the criterion that marks the southern border of the Croissant region.

2.2. North-Eastern Creusois (NEC)

The specific variety of the Croissant that features in this work is that of the village of Saint-Pierre-le-Bost, situated in the centre of the Croissant near the purported border with the langue d’Oïl; the location is marked by $\times $ in Figure 2.Footnote ⁴ Accordingly, this variety is more similar to French than to Occitan. Fieldwork has been carried out by the first author with one speaker in his late sixties.Footnote ⁵

Figure 2 The Croissant region.

Also helpful was the translation of Le Petit Prince, accomplished by the speaker with the help of the linguist Nicolas Quint as part of a large documentation project of the languages of the Croissant (Quint Reference Quint2021b).Footnote ⁶ While the facts discussed in this article are attested in other Croissant varieties, too, we henceforth discuss only NEC, in order to avoid overgeneralisation.

The phonemic system of NEC is only slightly different from the familiar system of French. The consonantal array is nearly identical: it includes plosives /p, b, t, d, k, g/; fricatives /f, v, s, z, ʃ, ʒ/; the rarer affricates /ʧ, ʤ/; nasals /m, n, ɲ/; liquids /l, ʁ/; and approximants /j, w, ɥ/. The phoneme /v/ is often pronounced as an approximant [ʋ] or [w], and the fricatives /ʃ, ʒ/ are very rarely debuccalised to [h]. This variation, irrelevant for the present purpose, is glossed over in this article.

Based on ongoing research, the vocalic system of NEC is slightly more different from that of general French spoken today in France. The following statements seem to hold. The language has three nasal vowels [ɛ̃, õ, ã]. The phoneme /ɑ/, which is by and large gone from the standard French vernacular in France, is very much present in NEC, mostly in (stressed and unstressed) open syllables, for example, [gɑ] ‘guy’, [lɑʃ-a] ‘let.go-inf’.Footnote ⁷ There are signs of neutralisation of /a/ and /ɑ/ in favour of [ɑ] before [ʁ], especially in stressed closed syllables, for example, [mɑʁ] ‘Mars’. The front mid vowels /ɛ, e/ are contrastive word-finally, that is under stress, for example, [pɒʁtʁɛ] ‘portrait’ vs. [pɒʁt-ʁ-e] ‘carry-fut-1sg’. They are neutralised in favour of [ɛ] in closed unstressed syllables, and in favour of [e] in open unstressed ones. Unlike in French, /e/ can be faithfully realised in stressed closed syllables, for example, [tet] ‘head’. As for the back mid vowels /ɔ, o/, they can both appear in unstressed open syllables, as in [ɔʃu] ‘gosling’, [sot-a] ‘jump-inf’. Word-finally and before [ʁ], /o/ is attested but /ɔ/ isn’t. Instead one finds a lower [ɒ] only in these environments, [vɒ] ‘calf’, for example, [nɒʁityʁ] ‘food’, [sɒʁ] ‘go.out.ind.sg’. This sound, which is not found in Standard Metropolitan French at all, is therefore probably an allophone of /ɔ/. In unstressed closed syllables other than before [ʁ], /ɔ/ and /o/ are neutralised to [ɔ]. In addition, NEC has a diphthong /wɛ/, which is faithfully realised regardless of the syllabic environment, for example, [etwɛl] ‘star’, [twɛlɛt] ‘wardrobe’, [ʃwɛz-iʁ] ‘choose-inf’.

All of the above-mentioned facts about the vocalic system have little bearing on the issues discussed in this article. More central is the behaviour of non-back rounded vowels [ø, œ]; it is returned to in §2.3. In slight anticipation of that discussion, Table 1 provides the vocalic phonemes of NEC and their realisations according to their prosodic position; cases in which the underlying phoneme is not faithfully realised are shaded grey.

Table 1 The vowels of NEC

Finally, a word is due regarding syllable structure. Word-initially NEC exhibits clusters of falling sonority (#RT), for example, [ʁmɒʁ] ‘regret’. Such words cannot be pronounced as the French cognate [ʁømɔʁ]. It is therefore unsurprising to find sonority plateaus, too, for example, [pti] ‘small’ (which cannot be pronounced with a vowel as in French [pøti]), and sonority rises [gʁij] ‘grill’. It is less clear if RT, RR and TT sequences represent syllable-initial clusters word-internally. More important for the present purpose, and therefore worth repeating, are the facts illustrated in (1) above: NEC, like French, allows for RT, RR and TT clusters at the right edge of the word; but completely disallows TR clusters in this environment.

2.3. Front rounded mid vowels in NEC

In NEC, the vowel [œ] in a closed syllable can correspond to at least three different underlying representations. The first two are distinguished by the behaviour of this vowel when the syllable is opened by the addition of a V-initial suffix. In the simplest case, illustrated in (4a), the [œ] alternates with [ø] in an open syllable (the so-called loi de position known from French, Morin Reference Morin1986). In the second configuration, in (4b), the [œ] disappears when the syllable is opened. Even though this vowel alternates with zero, it cannot be regarded as epenthetic, as at least some of the items in which it appears would also be well-formed without it: compare for instance [kakœt] ‘cackle.ind.sg’ and [kaʁœl] ‘tile.ind.sg’ in (4b) with [ʁɛspɛkt] ‘respect.ind.sg’, [yʁl] ‘scream.ind.sg’. The occurrence of [œ] in these cases is thus unpredictable, and its presence in the middle of the final cluster in the roots at hand must be marked in the lexicon: in these cases the [œ] cannot be epenthetic. Also note that this [œ] occurs only in closed syllables: it is left unpronounced in open syllables, and obligatorily so. Indeed, unlike in French, where alternating vowels in open syllables may or may not be pronounced (caqueter [kakøt-e] $\sim $ [kakt-e] ‘cackle-inf’; carreler [kaʁøl-e] $\sim $ [kaʁl-e] ‘tile-inf’), their absence is mandatory in NEC: *[kakœt-a] ‘cackle-inf’, *[kaʁœl-a] ‘tile-inf’. It was already mentioned in the previous section that alternating vowels in open syllables that may or may not be pronounced in French (le remords [løʁømɔʁ] $\sim $ [løʁmɔʁ] ‘regret’) cannot be pronounced in the NEC cognate ([lœʁmɒʁ] ‘regret’, *[lœʁœmɒʁ]).

To distinguish between the vowels in (4a) and (4b), it is customary in the literature on French, where similar patterns occur, to say that the latter is underlyingly schwa /ə/, while the former is /ø/ (e.g., Tranel Reference Tranel1987: 87; Durand Reference Durand, Neveu, Blumenthal, Hriba, Gerstenberg, Meinschaefer and Prévost2014: 31). We follow this analytic choice.

Finally, a third possible source for the vowel [œ] is epenthesis, as in (4c), where [œ] appears in the middle of a TR cluster that is followed by a consonant ([sykœʁ-ʁi] ‘candy’). In absence of a suffix, however, the R in the nouns at hand is deleted (following the regular pattern of (3a)/(3e)): [syk] ‘sugar’. Were the [œ] present in the lexical form, there would be no reason to delete the R, since a hypothetical /sykəʁ/ could be faithfully realised as *[sykœʁ]: no final TR arises, so there is no reason to delete the R. Therefore, the fact that the R is deleted indicates the absence of a vowel in the underlying form, which means that the vowel is epenthetic when it appears on the surface in [sykœʁ-ʁi] ‘candy’. It occurs to avoid a CCC cluster, and is inserted between the first two consonants: /CCC/ → [CœCC]. Thus epenthetic [œ] always occurs in a closed syllable.

Note that there are also nouns ending in TœR# in which the [œ] alternates with zero in nominal derivation. Thus [manivœl] ‘crank’ derives [manivl-ɑʒ] ‘cranking’. The [œ] cannot be epenthetic in this case since /TR#/ would produce the loss of the R (as in /sykʁ/ → [syk] ‘sugar’). The fact that no harm is done to the R thus mandates the lexical presence of the vowel: /manivəl/. That is, the contrast between TR# appearing as T# (/sykʁ/ → [syk] ‘sugar’) in some words but as TœR# in others (/manivəl/ → [manivœl] ‘crank’) leaves no other option than considering that in the lexicon, a vowel is present in the middle of the cluster in the latter, but not in the former case. The [œ] in [sykœʁ-ʁi] ‘candy’ is necessarily epenthetic.

The three phonological identities of [œ] are thus distinguished as follows. Items that never alternate with zero are regular /ø/ (which appears as [œ] in closed syllables and as [ø] in open ones). This includes the cases in (4a) as well as word-final instances of [ø] as in [vojaʒø] ‘traveller’ or [vjø] ‘old’; also note that stable /ø/ can sustain TR clusters to its left, for example, [fløʁi] ‘bloom.ind.sg’, [kʁøz-a] ‘dig-inf’.

Items that alternate with zero are either /ə/ or epenthetic. In both cases, [œ] is realised on the surface, and only ever occurs in closed syllables. We know that the language distinguishes /ə/ and epenthetic [œ] for the reasons mentioned, but it may not always be possible to determine whether an alternating [œ] in a given morpheme represents /ə/ or epenthesis. The cases in (4b) are clear-cut: the [œ] in [kakœt] and [kaʁœl] cannot be epenthetic since, as was mentioned, the clusters produced in absence of [œ], kt# and ʁl#, would be well-formed. But when the clusters produced in absence of [œ] do not occur in the language (like [ʃt]# in [aʃœt] ‘buy.ind.sg’ – [aʃt-a] ‘buy-inf’) or when the [œ] is the only vowel of the word (as in [ʒœt] ‘throw.ind.sg’ – [ʃt-a] ‘throw-inf’), it could be argued that [œ] represents either /ə/ or epenthesis. The latter would occur in order to break up illegal clusters, or to provide a vowel to a stem which has none. The discussion in §4.2.1 will provide reason for [œ] in both cases to be lexical, though: it will be claimed that epenthetic vowels cannot be stressed, whereas the [œ] in [aʃœt] and [ʒœt] is stressed. But in cases like [kœlw-ã] ‘nail-1/3pl’ – [klu] ‘nail (noun)’, the alternating [œ] is unstressed and could be either epenthetic or lexically present.

The phonological ambiguity of alternating [œ] in some words is orthogonal to the analysis developed below. The only thing that will be important (in §4.1) is the fact that the [œ] in the middle of root-final TRs in nouns as in (4c) is epenthetic, which is guaranteed for the reasons mentioned.

Given this situation, we believe that the phonological identities of the three types of [œ] are as shown in (5).Footnote ⁸ Note that this implies for (5b) that schwa is pronounced [œ]: the fact that /ə/ appears as [œ] is merely implementational; it has nothing to do with phonological computation. The fact that a phonological item is not pronounced according to its properties is common. Its surface shape is then acquired by operations that occur post-phonologically: feature filling (see, for instance, Dresher Reference Dresher2009) or operations relating phonology and phonetics. The latter may be cue constraints (Boersma & Hamann Reference Boersma and Hamann2008), phonetic interpretation in Element Theory (Backley Reference Backley2011: 18–22) or spell-out (Scheer Reference Scheer, Cyran and Szpyra-Kozłowska2014: /ə/ ↔ [œ]).

In this perspective, the epenthetic vowel of the language that is used in (5c) can be said to be /ə/: once it is inserted, it will be converted into [œ] upon pronunciation by the same implementational routine (see §4.2.2).Footnote ⁹

An alternative representation for (5b) would be a floating /ø/ (rather than a floating /ə/): in this case, stable and alternating [œ] would both be /ø/, and they would be distinguished only by the fact that the stable [œ] in (5a) is lexically associated, while alternating [œ] floats. This analysis is known from Slavic languages where, as in NEC (and French), vowels with identical phonetic realisations alternate with zero in some morphemes, but not in others (Scheer Reference Scheer, Oostendorp, Ewen, Hume and Rice2011b).

However, this alternative is not viable in NEC, because the alternating [œ] shows weak behaviour: TR clusters cannot exist to its left (i.e., in terms of Strict CV, it cannot license TR clusters). A diachronic manifestation of the weakness of (5b) is discussed in the following section. As mentioned, we believe that its synchronic manifestation is the fact that root-final TR clusters followed by a suffix /-ə/ undergo metathesis: /TR-ə/ → [TœR]. This pattern is further discussed in §4.2.

Therefore, the vowel in (5b) must be phonologically weak (TR cannot exist to its left), as opposed to the vowel in (5a), which is non-weak (TR can exist to its left). Classical segmental identities for /ø/ and /ə/ express this difference: in terms of Element Theory (Backley Reference Backley2011), /ə/ will be an empty-headed item that bears the element A ( $\underline{\varnothing} $ -A), while /ø/ bears the full set of place definers (I-U-A). The alternations of mid–low vowels [ɛ, ɔ, œ] with mid–high vowels [e, o, ø] (the loi de position) may also receive a classical account in terms of headedness (Harris Reference Harris1994): the former are unheaded (I-A [ɛ], U-A [ɔ], I-U-A [œ]) while the latter are headed (I-A [e], U-A [o], I-U-A [ø]). The phonological weakness of /ə/ ( $\underline{\varnothing} $ -A) is then expressed by the fact that this vowel is the only one in the system that is empty-headed.Footnote ¹⁰

2.4. Metathesis

Morpheme-internally, NEC systematically shows TœR where French has TRø. This is illustrated in (6) below.

In NEC, the [œ] in these words always stands in a closed syllable, with the following vowel belonging to the same morpheme. Therefore, there is no way to observe the behaviour of these [œ]s in open syllables and we must conclude that its underlying identity is a stable /ø/ as in (5a).

But the synchronic state of NEC is the result of a diachronic change of metathesis whereby TRø has become TœR: French has preserved the diachronically earlier TRø (Lat. prīmārius > Fr. premier [pʁømje], NEC [pœʁmje]; Lat. capra > Fr. chèvre [ʃɛvʁ(ø)], NEC [ʃøvœʁ(fœj)]; Old High German bittil > Fr. bretelle [bʁøtɛl], NEC [bœʁtœl]; etc.).

This evolution occurred only with schwa, not with /ø/. Vowels that appear as stable [ø] in French did not trigger metathesis: French fleuri [fløʁi] bloom.ind.sg’, creuser [kʁøz-e] ‘dig-inf’ correspond to NEC [fløʁi] ‘bloom.ind.sg’, [kʁøz-a] ‘dig-inf’, not *[fœlʁi], *[kœʁz-a]. In both languages, what today are stable [œ] and schwa were presumably phonetically distinct, since schwa /ə/ was originally pronounced [ə].

It may thus be concluded that there was a diachronic metathesis in NEC as shown in (7) below (recapitulating (2)).

It was mentioned in the previous section that an alternative to the representation of alternating [œ] (5b) as a floating schwa would be a floating /ø/, as shown in (8a). In this case, there would be no metathesis, though, since the floating /ø/ could simply attach to its nucleus and then be indistinguishable from the stable /ø/ of (5a). Rather, metathesis is triggered because the vowel following TR is unable to sustain this cluster. The vowel in question must thus be distinct from all others, or, more precisely, its melody must be different (whether floating or attached). Schwa as in (8b) is a valid candidate, since it is intrinsically weak (whether floating or attached): it cannot sustain (license) the preceding TR, and this is resolved by metathesis.Footnote ¹¹

Evidence from the adaptation of French loanwords into NEC suggests that the metathesis process in (7) is still synchronically active in NEC. For instance, when asked about the NEC equivalent of the French word écrevisse [ekrøvis] ‘crayfish’, the informant said that they have never heard of crayfish in their region, but that the word would be pronounced [ekœʁvis]. That metathesis is still synchronically active in NEC receives further support in §4.2 below, where stem-final TR in verbal forms is discussed.

2.5. More data

The data in (9) show that the behaviour of TR-final stems obeys a general pattern. In verbs, the TR sequences are syllabified together as a complex onset before a vowel-initial suffix, and separated by [œ] in the ind.sg, where there is no surface suffix. The corresponding noun is consistently pronounced without the liquid.

Note that verbs like those in (9c)–(9f) are clearly denominal; we take that to support a unified analysis, in which both the noun and the verb are derived from the same underlying representation with a final TR#. Additional reason to assume that the basic nouns also contain the liquid comes from nominal derivations based on these nouns. For instance, we have already seen in (4c) that [pwɛv] ‘pepper’ derives [pwɛvœʁje] ‘pepper mill, pepper tree’ with the suffix [-je], and [syk] ‘sugar’ derives [sykœʁʁi] ‘candy’ with the suffix [-ʁi]. Other nouns, which do not have denominal verbs, also behave in this manner: the number ‘four’ is pronounced [kat], but when the ordinal suffix [-jɛm] is added, one finds [katœʁjɛm] ‘fourth’; the cardinal ‘eighty’ (lit. ‘four score’) is pronounced [katœʁvɛ̃]. Similarly, [sab] ‘sand’ yields [sabœljɛʁ] ‘sand mine’.

Moving on to verbs with a seemingly suffixed [-e] in the ind.sg, Quint (Reference Quint, Esher, Guérin, Quint and Russo2021a) reports on this phenomenon in several Croissant varieties. He lists 12 verbal stems of this type, all of which are ‘asyllabic’, meaning that their stems do not have any realised nucleus – see the NEC examples in (10), with the [-e] in bold. According to Quint, the asyllabic nature of the root is the motivation for the appearance of the [-e].

Additional NEC data show that the raison d’être of the [-e] cannot be the absence of a full nucleus in the stem, though – at least not synchronically. Alongside the items mentioned by Quint, there are also disyllabic roots which bear the extra vowel [-e], as in (11).Footnote ¹² These verbs also point to another aspect of the data: the parallel nouns end in a vowel, which is the vocalic equivalent of the glide in the verbal forms; compare ind.sg. [kopje] with the noun [kopi], ind.sg. [kœlwe] with the noun [klu], etc.

In §1, we selected the verb [gœʁje] ‘grill.ind.sg’, whose related noun [gʁij] is not V-final, to represent this class in (1d). It is noteworthy that none of the items in (10) and (11) can be related to a noun ending in a homorganic VG sequence. We show in §4.3 that the two kinds of nouns do not have the same underlying form: the items in (10) and (11) are lexically V-final (e.g., /si/ ‘saw’), while the ‘grill’ type have a homorganic VG sequence (-ij, -uw, -yɥ) in the lexicon (/gʁij/).

Still, all of the items in (10) and (11), as well as [gœʁje] ‘grill.ind.sg’ from §1, share the property that their verbal stem ends in a CG sequence on the surface. Only such verbs exhibit the extra [-e], which never occurs after stems ending in non-homorganic VG: ind.sg. [fuj] ‘search.ind.sg’, [kõsɛj] ‘advice.ind.sg’, [syʁvœj] ‘oversee.ind.sg’, [tʁavaj] ‘work.ind.sg’, etc.

The verbs in (10) and (11) relate to those in (9) in that their stems end in a syllabically problematic sequence: like TR, CG clusters are of rising sonority. Taking into account the parallel nouns, another similarity is observed between the two groups: the right edge behaves differently in nouns and in verbs, though the differences between nominal and verbal forms are not the same in TR and CG contexts.

All the pieces of the puzzle are summarised in (12). RT-final stems (as in (12a)) and non-homorganic VG-final stems (as in (12b)) behave unremarkably: they do not pose a sonority problem, and their stems are constant in all configurations, including the related noun. Verbal stems ending in CG sequences on the surface (as in (12c) and (12d)) appear with the extra [-e], and are related to nouns ending either in a vowel (as in (12c)) or in a homorganic VG sequence (as in (12d)). Finally, underlyingly TR-final stems (as in (12e)) show an [œ] in the middle of the cluster in the ind.sg, and R-deletion in the noun.

The next section presents Strict CV, the theory in which the representational part of the analysis is couched.

3.1. Government and final empty nuclei

Strict CV is a theory of autosegmental representations, developed as an offshoot of Government Phonology (GP), for instance, in Lowenstamm (Reference Lowenstamm, Durand and Laks1996), Enguehard & Faust (Reference Faust, Lampitelli and Ulfsbjorninn2018), Faust (Reference Faust2014, Reference Faust2015, Reference Faust2017), Passino (Reference Passino2013), Fathi (Reference Fathi2017), Enguehard (Reference Enguehard2017), Lahrouchi (Reference Lahrouchi2018), Lampitelli (Reference Lampitelli2017), Newell (Reference Newell2021a), Polgárdi (Reference Polgárdi2014), D’Alessandro & Scheer (Reference D’Alessandro and Scheer2015), Scheer (Reference Scheer2019). For book-length discussion of Strict CV, see Scheer (Reference Scheer2004, Reference Scheer2012). The basic tenet of this approach is simple: there is only one skeletal unit, the CV unit. Every C-slot is followed by a V-slot, and every V-slot preceded by a C-slot.

This basic assumption has two consequences, both of which are illustrated by the Palestinian Arabic example [tarʒamɨt] ‘I/you translated’ in (13), underlyingly /tarʒam-t/.Footnote ¹³ First, word-final consonants are followed by an empty nucleus at the skeletal level. Second, phonetically adjacent consonants like [rʒ] are separated by an empty V-slot (or nucleus) on the skeletal tier. Empty nuclei are realised as epenthetic vowels unless epenthesis is inhibited. For medial empty nuclei, inhibition is made possible through a relation of right-to-left government from a following contentful nucleus, shown in (13) by the arrow from the V-slot associated with [a] to the empty V-slot to its left.

Inhibition of final empty nuclei (FEN) is parametrised. If the words of a language can end in consonants at the melodic level, as is the case for Arabic, it is assumed that FEN are allowed to remain unrealised: they are parametrically governed (rather than licensed, as in Standard GP; see Scheer Reference Scheer2004: §150). In languages like Italian where all (native) words end in a vowel, FEN are disallowed (they are ungoverned by the parameter setting for the language).Footnote ¹⁴

While internal empty nuclei cannot govern in any circumstance, the ability to govern is parameterised for FEN (although they are themselves governed; see Scheer Reference Scheer2004: §524; Scheer & Ziková Reference Scheer and Ziková2010). In Palestinian Arabic, they cannot govern; therefore, the penultimate V-slot in (13) cannot be inhibited (as indicated by the broken arrow) and is thus realised through epenthesis, signalled by the upwards arrow in place of an association line.

Unlike in Palestinian Arabic, though, FEN are clearly able to govern in NEC, as evidenced by the existence of words like the morphologically simplex noun [ʃaʁʒ] ‘load’, whose representation is provided in (14).

3.2. Branching onsets and the difference between RT# and TR# clusters

Let us now consider how the crucial distinction between TR clusters on the one hand and RT, TT and RR clusters on the other is represented in Strict CV. (NEC examples of the latter include [ʃaʁʒ-a] ‘load-inf’, [aktivitad] ‘activity’ and [yʁl-a] ‘scream-inf’.) In regular syllabic accounts, a TR cluster is a branching onset, while the other type (RT, TT, RR; collectively referred to below as RT for short) is a coda–onset sequence. This distinction is in many ways crucial, since in relevant languages an intervocalic CC cluster may or may not close the preceding syllable (VR.TV vs. V.TRV). If it does not, it is always of the TR kind (onset maximisation; e.g., Hayes Reference Hayes2009: 252–253). The contrast also appears at the left edge of words, where languages like English restrict initial clusters to #TR.Footnote ¹⁵ The symmetric situation is displayed at the right edge of words, where in the same languages only RT# clusters are found. If an illegal final TR# occurs, it is repaired by various means: in Czech, German and English (Scheer Reference Scheer, Zybatow, Junghanns, Lenertová and Biskup2009), the R becomes syllabic (thus vocalic in kind); in Modern Hebrew, loans such as French Louvre [ˈluvʁ(ø)] or English Beatle [bitl̩] are adapted by breaking up the cluster with the epenthetic vowel [e]: [ˈluveʁ], [ˈbitel].

In traditional syllabic models, the fact that the two consonants of a TR cluster enter into a closer relationship than those of an RT cluster is represented by treating a TR cluster as a single constituent (a branching onset) and an RT cluster as two separate constituents (a coda and an onset). In Strict CV (where constituents do not branch), the contrast between TR and RT clusters lies in the management of the intervening empty nucleus. The tighter relation between the consonants of the traditional branching onset is expressed graphically by the symbol ← in (15b). This right-to-left relation is referred to as ‘infrasegmental government’ (IG), and is established between sonorants and preceding obstruents (in NEC, specifically liquids and preceding obstruents), with the sonorant assuming head status (Scheer Reference Scheer2004: §14). Thus TR can contract IG, but RT cannot (given that IG is right-to-left, the head R must be preceded by the T).

Languages impose variable restrictions on what kind of nucleus can license a TR cluster. In Polish or French, where final TR# goes unrepaired, TR can be licensed by a FEN (see also §3.3). This is not the case in NEC, where a full vowel is required to the right of a TR cluster. In principle, repairs other than the loss of R can be thought of when an unlicensed (final) TR is encountered: the loss of T or the establishment of a governing relation between the FEN and the nucleus in the middle of the cluster, as in the RT cluster in (15c). As a matter of fact, NEC does not use these solutions (see also fn. 18).

The closer relationship within a TR cluster silences the intervening empty nucleus, which is therefore not in demand of government. The difference between (15a) and (15b) thus boils down to the fact that the nucleus following the cluster in (15a) needs to govern the preceding empty nucleus, while in (15b) the nucleus after the cluster has no such governing duties.

In this approach, a consonant is thus a coda iff the following nucleus is empty and governed (see fn. 14); this is the case of the R in (15a). Regarding the right edge of the word, recall from §3.1 that FEN in languages like NEC are governed. Thus, the final C in (15e) is a coda. By contrast, since the nucleus following T in (15b) is empty but not governed, T is not a coda.

The contrasting behaviour of final TR (repair needed) and final RT clusters (unproblematic) follows from this setting. It was shown in (14) that FEN can govern in NEC: this is the situation of RT# in (15c). But since in NEC the nucleus following TR clusters needs to be associated with a vowel (a full vowel; see §4.1) in order to license them, such clusters cannot remain intact in final position, where the following nucleus is empty. This is shown in (15d), where the final empty nucleus is unable to license the T←R cluster.

3.3. A clarification on the purview of Strict CV

Government Phonology in general and Strict CV in particular are representation-oriented. They define well-formed and ill-formed structures, some of which are supposed to be universal. For instance, in Strict CV, a representation bearing an ungoverned empty nucleus that is not silenced by a TR cluster (as in (15d)) is ill-formed.

What happens when an ill-formed structure arises is of course different from language to language. In the case of word-final TR clusters, most languages do not allow them and offer no repair either: TR# is then ungrammatical. On the other end of the scale of tolerance, some (rare) languages allow final TR clusters without any repair. This is the case in Polish, where Piotr [pʲɔtr] ‘Peter’, for example, remains unrepaired (Bethin Reference Bethin1984; Rubach & Booij Reference Rubach and Booij1987; Charette Reference Charette1992), and it is an option in French, where cadre ‘frame’ may be pronounced [kadʁ] (or [kad] with R drop) (Dell Reference Dell1973: 225, Reference Dell1976; Tranel Reference Tranel1987: 103–104). When word-final TR clusters are repaired, a number of strategies are observed. A popular repair is to make the R syllabic (provided the language allows syllabic consonants): thus, English bottle [bɔtl̩] (compare with non-syllabic R in bottl-ing [bɔtlɪŋ]), German Regen ‘rain’ [ʁeːgŋ̍] (compare with non-syllabic R in regn-en ‘rain-inf’ [ʁeːkn-ən]) or Czech Petr ‘Peter-nom.sg’ [pɛtr̩] (compare with non-syllabic R in Petr-a ‘Peter-gen.sg’ [pɛtra]). Two other repairs occur in NEC: the appearance of a vowel in the middle of the cluster and R drop (the latter is also found occasionally in French; see above).

Strict CV as such has nothing to say about these parameter settings, that is, about the kind of repair, if any, that is encountered across languages. Therefore, nothing prevents encoding this variation in terms of constraint interaction or more traditional parameter theory (e.g., Ulfsbjorninn Reference Ulfsbjorninn and Samuels2017). Thus, the FEN can license word-final TR# in Polish (Charette Reference Charette1992), but not in English. French shows variation: FENs can optionally license TR#, and if they do not, the R is lost (unlike in English, where it is made syllabic). In this article, we do not engage with this kind of debate, and, while remaining agnostic, we will represent cross-linguistic variation in terms of constraint interaction.

Finally, languages may prefer to repair a given structure but, when no repair is available, may leave it unrepaired, which does not cause it to be ill-formed. This pattern is trivial in OT, where a candidate that violates some constraint may be nonetheless be optimal and therefore appear on the surface. In NEC, homorganic VG sequences are a case in point: the language tries to avoid them, but leaves them unrepaired unless a following vowel is available (which would enable syneresis; see §5.3). Thus, the final /ij/ of the root /gʁij/ ‘grill’ is repaired through syneresis /ij/ → / $\varnothing $ j/ when verbal morphology offers a vowel-initial suffix (/gʁij-a/ → [gœʁj-a] ‘grill-inf’), but remains unrepaired in nouns where no suffix is available (/gʁij/ → [gʁij] ‘grill (N)’).

Again, Strict CV has nothing to say about that, unless the ill-formedness in question is universal: an ungoverned empty nucleus (which does not occur in a branching onset TR), for example, cannot remain unrepaired. Restrictions whose violation makes representations ill-formed, either universally or in a given language, with no repair possible, are referred to as principles below. Restrictions that are subject to negotiation in a constraint hierarchy are called constraints. This terminology merely translates the fact that not everything is negotiable: some properties of phonology hold no matter what, either universally or in a given language. Principles have been present in the OT agenda from its beginnings in the form of universally undominated constraints or fixed rankings (e.g., Prince & Smolensky [Reference Prince and Smolensky1993] 2004: ch. 9; McCarthy & Prince Reference McCarthy and Prince1995).

In sum, Strict CV is a theory of representation. Its practitioners have only rarely dwelled on a theory of computation.Footnote ¹⁶ OT and rule-based approaches are theories of computation. A complete theory of phonology needs both (Anderson Reference Anderson1985; Scheer Reference Scheer2010). Therefore, constraint interaction and Strict CV are not incompatible and have actually been combined, for example, by Polgárdi (Reference Polgárdi2014), Cavirani (Reference Cavirani2015) and Faust & Torres-Tamarit (Reference Faust and Torres-Tamarit2017). But given its focus on representations, most work in Government Phonology and Strict CV typically states computational events in prose without further formalisation (‘and then the floating vowel associates to the empty nucleus’). The same goes for the statement of parametric choices (‘this language has no branching onsets’, ‘in this language, FEN govern’). Formalising computation in terms of constraint interaction (or ordered rules, which are also an option) is thus making explicit what is typically left to prose statements.

Having presented the language, the data and the theory, we now turn to the analysis.

4.1. Schwa and empty nuclei cannot license TR

It is not rare that vowels which alternate with zero, even when they are pronounced, behave as if they were not. This is the case in Midi French where (unlike in Northern varieties of French) the ATRness of mid vowels is in true complementary distribution: [ɛ, ɔ, œ] occur in closed syllables ([bɛt] ‘stupid’, [alɛχte] ‘alert-inf’), while [e, o, ø] are found in open syllables ([betiz] ‘stupidity’, [metrik] ‘metrical’, [fe] ‘done.prtc’) (Durand et al. Reference Durand, Slater and Wise1987, Durand Reference Durand1990: 24 ff.). But when a mid vowel is followed by a single consonant plus schwa (which alternates with zero in free variation), it is [−ATR] no matter whether the schwa is actually pronounced: [bɛtøʁav] $\sim $ [bɛtʁav] ‘beet root’, [pɔtøʁi] $\sim $ [pɔtʁi] ‘pottery’. A [+ATR] realisation would be expected when schwa is pronounced, since its phonetic existence should place the preceding mid vowel in an open syllable. But this is not the case, and the complete distributional statement is thus that mid vowels are [−ATR] in closed syllables or before a (pronounced) schwa, while they are [+ATR] in open syllables, except when the following vowel is schwa. In terms of Strict CV, [−ATR] vowels thus occur before a governed empty nucleus (i.e., in a closed syllable, as in the structures in (15a) and (15e)) and before schwa.

The same disjunction defines the ill-formedness of TR in NEC. We have seen in (6) that in the diachronic development of the language, TR clusters followed by schwa undergo metathesis ([gʁœnuj] > [gœʁnuj]). They are also ill-formed when followed by a governed empty nucleus, that is, when TR occurs in word-final position as in (15d): /sykʁ/ → [syk] ‘sugar’.

The generalisation thus appears to be that weak nuclei – that is, nuclei that are empty or contain schwa – are unable to sustain TR clusters in NEC.Footnote ¹⁷ The segmental identity of schwa introduced in §2.3, $\underline{\varnothing} $ -A, shares with empty nuclei the fact of being empty-headed: having an empty head thus translates as weakness when it comes to licensing preceding clusters.

Therefore, we believe that the repair observed for TRə in the diachronic development of NEC has the same cause as the one that is applied to TR# synchronically in the language. But the two structures undergo different repairs: TRə is remedied by metathesis (→ TœR) as in (2), while TR# results in the loss of R (→ T#).

The main argument is that metathesis also occurs word-finally when verbal stems are followed by the ind.sg marker /-ə/: /sykʁ-ə/ → [sykœʁ] ‘sugar.ind.sg’. Thus the two distinct repairs observed for stem-final TR (the appearance of [œ] in the middle of the cluster in verbal roots vs. the loss of R in nominal stems) are the consequence of this suffix /-ə/, which is present in verbs, but absent in nouns.

4.2. TR-final stems

4.3. Additional support for /-ə/: infinitives

In several Romance languages, and specifically in French, verbs are traditionally divided into three classes. The logic of this division is presented with data from NEC in (21). Group 1 includes those verbs which have an infinitive in [-a]. Verbs of Group 2 have an [-i-ʁ] infinitive. Group 3 comprises all verbs which do not fall into Group 1 or 2, such as V-final stems with infinitive [-ʁ], like ‘see’, and T-final stems with no apparent infinitival suffix, such as ‘sell’.

Two aspects of the classification are specific to the variety discussed here. First, all verbal stems in Group 1 end in a consonant (unlike in French, which has Group 1 stems like créer [kʁe-e] ‘create-inf’, ind.sg [kʁe] and broyer [bʁwaj-e] ‘grind-inf’, ind.sg [bʁwa]). Second, as just mentioned, some Group 3 infinitives lack a final /ʁ/. Besides ‘sell’ in (21), other cases in point are [mɔʁd] ‘bite’, [tɔʁd] ‘twist’, [kunɛt] ‘know’, [pʁãd] ‘take’ – in fact, all Group 3 verbs whose root ends in a consonant. Also note that the [-ʁ] is present in the French cognates of these verbs: mordre [mɔʁd-ʁ(ø)], tordre [tɔʁd-ʁ(ø)], connaître [konɛt-ʁ(ø)], prendre [prãd-ʁ(ø)] (though it may also be left unpronounced; see §3.3).

The presence of /ʁ/ exponing the infinitive may be generalised to all verbs of Groups 2 and 3 (Group 1 having an /-a/ allomorph) given our analysis of TR-final stems. As shown in (22), the situation created in C-final stems in Group 3 is analogous to that of TR-final nouns in (17) above: no suffix /-ə/ occurs underlyingly; epenthesis is prohibited; and the R is not licensed. Thus there is no way to avoid deletion (of the R together with its syllabic support).

We take the fact that Group 3 infinitives behave like nouns to support our analysis, which predicts deletion in infinitives exactly where it is observed (assuming, as is the null hypothesis, that infinitives do not carry a suffix other than /-ʁ/).

4.4. Interim summary

In this section, we have provided an analysis of the different solutions to TR-final stems by appealing to the difference between an underlying, stressable /-ə/ (exponing ind.sg) and an epenthetic schwa. Verbs and nouns with TR-final stems implement different solutions because nouns are truly unsuffixed: the R in TR# can only be saved from deletion by an epenthetic [œ], and such epenthesis is disallowed. But in verbs, stem-final TR undergoes metathesis with the ind.sg suffix /-ə/, and the R is saved. Infinitives are not expected to carry an agreement marker, and therefore behave like nouns.

5.1. Adjustment of basic stems

The basic facts from verbs with CG-final stems (§2.5) are recalled in (23). The [-e] occurs after surface [Cw, Cj, Cɥ] (regardless of whether the stem is otherwise asyllabic). Note that verbs with CG-final stems are arguably denominal, and that the corresponding nouns are vowel-final (we will return to the case of the noun type [gʁij] ‘grill’ below).

Crucially, all verbs with the extra [-e] belong to Group 1 (infinitive /-a/). Now recall the generalisation about this group from §4.2: all stems in this group are consonant-final. This generalisation is recast in terms of Strict CV in (24):

This condition suggests that nominal bases like [kopi] cannot serve as Group 1 verbal roots. Such bases must be adjusted to comply with the condition. The only way this is possible in Strict CV is through the addition of a CV unit, as shown in (25). The vowel-final forms in (25a) correspond to the lexical identity of the root. In order for the stem to comply with (24), an additional CV is added as in (25b), and the final vowel spreads onto the following onset. The VG sequences created are thus homorganic. Stems ending in a lexical (non-homorganic) VG sequence, as in (25c), are not expected to be adjusted, as they already end in an empty nucleus. Note that we represent glides and corresponding high vowels with the symbols I / U, which are neutral regarding vowelhood and consonanthood.

Because of this adjustment process, the verbal forms of vowel-final stems in Group 1 end in a homorganic VG sequence. The extra vowel [-e] thus appears after underlying /Cij/ or /Cuw/, and only in this context: [kœlwe] ‘nail.ind.sg’ originates in /kluw/; [kopje] ‘copy.ind.sg’ originates in /kopij/; and so on. This adjustment process thus unifies (adjusted) V-final stems with stems whose final homorganic VG sequence is lexical, such as /gʁij/ ‘grill’: the noun reveals whether the homorganic VG# sequence of the verbal stem is lexical (in which case the noun ends in homorganic VG: [gʁij] ‘grill’) or derived (in which case the noun ends in V: [kopi] ‘copy’).

For what follows, it is important to note that the adjustment in (25b) is not phonological, but rather a lexical condition on Group 1 stems. Thus, it is not the case that the input to phonology is /klu+CV/ and then the /u/ spreads to occupy the C-slot. Rather, the input to phonology is the result of the adjustment, /kluw/.

5.2. Homorganic VG in NEC and French

The glide in (23) thus comes into being through root adjustment (25). But this cannot be the entire story: after adjustment, the underlying forms in (23) have a homorganic VG sequence /ij/ or /uw/ (/kopij/, /kluw/), but their surface realisation lacks the vowel. This observation may be formulated as the following constraint:

At this point, it is useful to compare the situation of homorganic VG sequences in NEC and French. French (like other languages) shows a reaction against homorganic VG whereby the V is deleted (ij → j; uw → w); this is called syneresis (e.g., Dell Reference Dell1972; Klein Reference Klein1992; Bullock Reference Bullock, Wiltshire and Camps2002; Hualde Reference Hualde, Colina and Martínez-Gil2020: 165 ff.). Thus in French, roots ending in /i, u, y/ appear as such in word-final position: nie [ni] ‘deny.ind.sg’, loue [lu] ‘rent.ind.sg’, tue [ty] ‘kill.ind.sg’. When followed by a vowel-initial suffix such as infinitival /-e/, the hiatus is resolved through gliding, that is, the insertion of a glide corresponding to the preceding high vowel: [nij-e], [luw-e], [tyɥ-e]. The vowel may then optionally be omitted through syneresis: [nje], [lwe], [tɥe]. But not all homorganic VG sequences are affected by this process: the [ij] in habille [abij] ‘clothe.ind.sg’, pétille [petij] ‘sputter.ind.sg’, fusille [fyzij] ‘shoot.with.a.rifle.ind.sg’ and many others is stable when a vowel-initial suffix is added. Thus the infinitives of these three verbs are habiller [abij-e], pétiller [petij-e] and fusiller [fyzij-e], where the [i] cannot be left unpronounced: *[abj-e], *[petj-e] and *[fyzj-e] are ungrammatical. Hence, there are two types of homorganic VG sequences in French: one optionally undergoes syneresis, while the other cannot. The distinction between these types lies in the lexical presence or absence of the glide in the root: roots of the type nie ‘deny.ind.sg’ are V-final /ni-/, while roots like habille ‘clothe.ind.sg’ are Vj-final /abij-/.

This distribution also has a diachronic rationale: the stable glide goes back to a palatal lateral [ʎ], which was pronounced as such until the early 19th century (Nyrop Reference Nyrop1899: §351) and is still reflected in spelling, while the unstable glide has no diachronic ancestor.

The two types of [ij] sequences may therefore be represented as in (27).

The stable yod in (27b) is a segment in its own right that is independent of the preceding [i]: the [ij] sequence involves two |I|s. The structure is thus equivalent to non-homorganic VG sequences, as in (25c) [tʁavaj] ‘work’, except that the V happens to be |I|. By contrast, the unstable yod in (27a) is the result of the spreading of an /i/ to a following empty onset: there is only one |I| in this structure.

NEC also features the two types of [ij] sequences, but, as shown in (28), they are not distributed in the same way as in French. In NEC, syneresis occurs with the items in (28a) and (28b), which also show the extra [-e]. In the French cognates, syneresis can only apply to the items in (28b) (except when preceded by a TR cluster as in trier and clouer, a fact that has no bearing on the argument): it is impossible for those in (28a), which feature a former palatal lateral [ʎ].

NEC also has stable stem-final [ij] sequences, as in (28c): these never undergo syneresis, and no extra [-e] appears in the ind.sg. They always feature a former palatal lateral [ʎ] (and thus their French cognates are of this kind as well).

Our best guess is that in NEC, the two types which are distinguished in French as (27a) and (27b) diachronically merged into (27a). That is, all native NEC words, as in (28a) and (28b), are like (27a): they have a single |I| that spreads and thus undergoes syneresis. The NEC words in (28c) are loans from French and instantiate (27b): they do not undergo syneresis and show no extra [-e] for that reason.

Thus the NEC roots in (28a) lexically end in /-ij/ (/gʁij/), which has the structure in (27a): it appears as such in nouns ([gʁij] ‘grill (N)’) and undergoes syneresis when followed by a vowel ([gœʁj-ã] ‘grill-1/3pl’). On the other hand, the NEC roots in (28b) are lexically V-final (/tʁi, klu/) and appear as such in nouns ([tʁi] ‘sort (N)’, [klu] ‘nail (N)’). When used as verbs, they undergo adjustment (25b), which also produces (27a). Therefore, they are subject to syneresis and receive the extra [-e].

In sum, then, the two structures in (27) exist both in French and in NEC. In both languages, (27a) is subject to syneresis (optionally in French, obligatorily in NEC), while this process does not apply to (27b). In NEC, the extra [-e] is concomitant with syneresis – that is, it occurs if and only if syneresis does.

Given the distinction between (27a) and (27b), the constraint against homorganic VG sequences in (26) needs to be updated as in (29): its target is not any surface VG sequence, but specifically the type of VG sequence in (27a): a prime (|I| or |U|, represented by the variable α in (29)) that is attached to a nucleus and spreads onto the following onset.

5.3. Syneresis

This section is concerned with lexical items instantiating the branching structure in (27a) that the constraint under (29) militates against. We first (in §5.3.1) look at stems that are V-final in the lexicon, as in (28b), then (in §5.3.2) at those that are already equipped with a homorganic VG sequence in the lexicon, as in (28a).

5.4. Additional support: [œ]/[e] alternations in Group 3

Alongside the infinitive /-ʁ/, verbs of Group 3 are characterised by a specific stem alternation: the last consonant of the stem is missing in the ind.sg. In the subjunctive, this missing consonant reappears, even though there is no suffix on the surface. For instance, for the verbal root ‘sell’ in (34a), the consonant /d/ is missing in the ind.sg. Now consider the three verbs in (34b)–(34d). They exhibit the same C–zero alternation, except that the subjunctive is of the form [CœC]; the suffixed form ind.1/3pl loses the [œ] and becomes [CC-]; but, crucially, the ind.sg, lacking the final consonant, exhibits the vowel [e], appearing as [Ce].Footnote ²⁰

The alternations in (34b)–(34d) are perfectly regular if one assumes that the underlying form of these verbs is /CəC/. When the final consonant is missing, the lexical /ə/ finds itself in an open final syllable. Since it is the only vowel of the stem, it cannot delete.Footnote ²¹ However, because this vowel is /ə/, it cannot be realised faithfully in the final syllable; instead, it undergoes fortition to [e]. We thus take the alternations in (34) to provide independent support for the proposed fortition process, which not only concerns the suffix /-ə/, but also the root vowel /ə/ when placed in an open syllable.Footnote ²²

To summarise, we have proposed a unified account of the data, in that all differences between nouns and verbs follow from the presence of a suffix /-ə/ in verbs, which is absent in nouns. Independent support was provided by considering Group 3 verbs: the infinitives of consonant-final stems in this group predictably lack the expected /ʁ/, and the root vowel /ə/ of /CəC/ stems strengthens to [e] when word-final.

7.1. Category-specific phonology

7.2. Paradigm Uniformity

Another way of motivating the difference between nouns and verbs exploits the fact that the seemingly unsuffixed ind.sg is just one inflected form of the verb. Nearly all other verbal forms in NEC are suffixed, mostly with V-initial suffixes, for example, ind.1/3pl [sykʁ-ã]. Such suffixes allow the realisation of the final liquid. Nouns, in contrast, are only suffixed in derivation, as in /sykʁ-ʁi/ → [sykœʁʁi] ‘candy’ above. Therefore, one may posit that the suffixed forms of verbs put pressure on the unsuffixed forms, but that nouns are not subject to this pressure, because there are no suffixed forms in their inflectional paradigm. Such effects are known as ‘Paradigm Uniformity’ (Downing et al. Reference Downing, Alan Hall and Raffelsiefen2005): because one form of the paradigm behaves in some way due to a surface-true trigger, another one behaves in that way, too, even though the trigger is absent.

The obvious disadvantage of such an approach is transderivationalism: the treatment of one underlying string would be dependent on the result of the treatment of another. Let us nevertheless examine what an analysis of the patterns discussed along these lines would look like. Again, we assume that, unlike in our own analysis, the stem is underlyingly identical in both verbs and nouns; otherwise, there is no reason to attempt to derive the difference from Paradigm Uniformity.

Again, in our case, an account based on Paradigm Uniformity would argue that deletion of the R in an unsuffixed ind.sg form /sykʁ/ → *[syk] is problematic because the R is realised in the suffixed ind.1/3pl [sykʁ-ã] and uniformity is enforced in paradigms. This cannot be the whole story, because we know that in other respects, paradigms are allowed to be non-uniform. For instance, in ind.sg–ind.1/3pl pairs such as [sykœʁ]–[sykʁ-ã] ‘sugar’ and [kaʁœl]–[kaʁl-ã] ‘tile’, [œ] is present in the stem-final cluster in unsuffixed forms but absent in suffixed forms, creating different syllable structures within the same paradigm. If Paradigm Uniformity were a general requirement, it would generate the incorrect but phonotactically licit paradigm *[kaʁl]–[kaʁl-ã]. Instead, it must target a more specific aspect of the stem. One possibility is to assume, alongside a general deletion process for TR-final stems, a Paradigm Uniformity constraint requiring that the consonants of all stems in a paradigm be identical. Thus, again, deletion of the R in an unsuffixed ind.sg form /sykʁ/ → *[syk] is prohibited because the R is realised in the suffixed ind.1/3pl [sykʁ-ã]. Conversely, in nouns, since there are no suffixed forms in the paradigm, deletion may apply. This Paradigm Uniformity constraint would not be affected by the epenthesis in /sykʁ/ → [sykœʁ].

This analysis does not immediately offer an account for homorganic VG-final stems. An ind.sg–ind.1/3pl paradigm without syneresis, for example, *[gʁij]–[gœʁj-ã] ‘grill’ (the observed pair being [gœʁj-e]–[gœʁj-ã]), does not violate the constraint, as the consonants of the paradigm remain untouched. One may assume, as we have, that there is an additional constraint against homorganic VG, and this constraint is obeyed through the addition of an epenthetic [e] (not the fortition of a suffix /-ə/) in the observed paradigm [gœʁje]–[gœʁj-ã]. Thus, on this account, the reason for epenthesis in TR-final and homorganic VG-final verbs is not uniform: non-phonological pressure due to Paradigm Uniformity in TR-final stems, as opposed to the markedness constraint against homorganic VG in Homorganic VG-final stems. Moreover, by separating these two stem types in this way, the account makes a wrong prediction: since epenthetic [e] in verbs is unrelated to paradigmatic effects, it is expected to occur in nouns, too. Yet, as we have seen repeatedly, it does not: /gʁij/ remains [gʁij] ‘grill (N)’. We conclude that a Paradigm Uniformity account that is as unified as our own is not straightforwardly available.

Like the category-specific accounts discussed, another aspect of the facts that the Paradigm Uniformity account fails to express is the link between the verbal [sykœʁ] ‘sugar.ind.sg’ and the general metathesis /TRə/→[TœR] in the language.

Finally, another obvious problem with Paradigm Uniformity as a phonological force is that some paradigms are not uniform in the same way as others. This is true, for instance, of many verbs in Group 3, where the suffixed form (and the unsuffixed subjunctive) carry an additional consonant, for example, [vã] ‘sell.ind.sg’, [vãd-ã] ‘sell.1/3pl’ (see (34)). There is no phonological problem with having an ind.sg *[vãd], yet such paradigms are exempt from analogical modification. To formalise this exception, one would have to restrict the Paradigm Uniformity constraint to verbs of Group 1, that is, make it diacritic-specific. In our opinion, this raises the price of an analogical analysis even higher than it already was due to transderivationalism.

We mentioned in the beginning of this article that we adhere to a classic, ‘one UR, one phonology’ approach. We hope to have argued here that competing approaches, while they might initially seem tempting, either do not provide a unified analysis, or make incorrect predictions, or are prone to additional complications that are best avoided.