Stratal overgeneration is necessary: metrically incoherent syncope in Southern Pomo

Abstract Southern Pomo (Pomoan, California) displays a process of rhythmic vowel deletion (syncope) reflecting two mutually incompatible metrical structures. This phenomenon, called metrical incoherence, can be derived by an ordered sequence of independent subgrammars, that is, strata. Metrical incoherence is under-attested crosslinguistically, and the stratal models of phonology necessary to generate it have been criticised for predicting counter-typological phenomena. Nevertheless, the Southern Pomo data cannot be generated in more restrictive frameworks. This article argues that overgeneration is a necessary property of the phonological component, and that metrical incoherence is rare because it is difficult to learn. In Southern Pomo, this difficulty appears to have caused grammatical competition and restructuring: a second pattern of syncope, occurring in only a limited context, suggests that learners have reanalysed the grammar as having consistent metrical structure across the derivation. This work thus supports the proposal that diachronic change – and therefore typology – is constrained by extragrammatical factors.


Introduction
The relationship between synchronic grammar and crosslinguistic typology is a persistent issue for phonological theory.There is considerable debate about whether grammatical or functional factors are (most) responsible for phonological typology, and how phonological theory should reflect this.One view is that many patterns are necessarily ruled out by synchronic grammar (BermúdezOtero & Hogg 2003; de Lacy 2006), and that apparent countertypological phenomena should instead be accounted for by other means, such as morphologisation (Staroverov 2020).Another approach suggests that typology is epiphenomenal, with patterns arising diachronically rather than being psychologically active (Hale & Reiss 2000; Blevins 2006).A third position posits that synchronic phonology is capable of generating countertypological patterns (de Lacy & Kingston 2013), but typology is constrained by extragrammatical factorssuch as how the distribution of evidence for a process in the lexicon affect the learnabil ity of that process (Jarosz 2016; Stanton 2016).Work within each of these theoretical traditions has also investigated how diachronic change reflects intrinsic features of the phonological component (BermúdezOtero & Trousdale 2012), extrinsic influences (Hansson 2008), or their interactions (Stanton 2016).
A general approach in phonological theory has been to limit the power of grammat ical frameworks to the level necessary to produce only attested, productive phenom ena. 1 One area of this literature has thus sought to evaluate whether certain phenomena are synchronically active, particularly for phenomena which would require more expansive architectures.Of particular concern has been the synchronic productivity of opaque phenomena -processes whose conditioning environments are not present in surface forms (Kawahara 2015).One such process is metrically conditioned syncope: the deletion of vowels in metrically weak positions (McCarthy 2008).In typical cases of syncope, vowels in weak positions (unstressed or unparsed) are targeted for deletion, as schematised in (1):2  This process poses a wellknown problem for parallel constraintbased frameworks such as classic Optimality Theory (OT; Prince & Smolensky 1993), as deletion must occur after the building of the metrical structure which conditions it.This hidden structure presents a major obstacle for parallelism, as the computation of stress assignment and deletion must be ordered (Kager 1997; McCarthy 2008).
Syncope can be called transparent if it is driven by a constraint (or rule) that is surfacesatisfied.For example, syncope driven by a constraint penalising unstressed vowels (*V; McCarthy 2008) is transparent if the output includes only stressed vowels, and all vowels that would be unstressed are deleted.(2) demonstrates this pattern in Macushi (Carib): (2) Syncope in Macushi (Hawkins 1950: 87; Kager 1997) a. [w_nàː.m_ríː]/wa.na.ma.ri/ 'mirror' b. [_wàː.n_màː.r_ríː]/u.wa.na.ma.ri.ri/ 'my mirror' Deletion in (2) has applied in such a way that there is a clear surface generalisation: vowels do not occur in weak (unstressed) positions.In other cases, vowel deletion transparently satisfies constraints on syllable weight (the StresstoWeight and Weight toStress Principles; Prince 1990) or distance from word edge (Prince 1990; see Gouskova 2003 andKager 1997 for implementations).Hidden structure is still inherent to this process, however: the conditioning environment is never visible in the output in any of these cases -the vowels in weak positions and light stressed syllables are a counterfactual.In that sense, we might think of syncope as inherently opaque.
Additionally, deletion cannot always be explained using surfaceoriented means.There are cases when rhythmic deletion does not appear to satisfy a surface gener alisation related to metrical structure, either underapplying (e.g., unstressed vowels appear in surface forms), or overapplying (occurring in locations that are not explained by surface prominence).These are instances of metrical opacity.Metrical opacity seems to provide particularly strong evidence for derivational frameworks, because metrical processes (like deletion and reduction) are conditioned by syntagmatic relations between positions within a hierarchical structure.The structural relations must exist before such deletion can occur.The output of syncope in some cases seems to prioritise faithfulness to strong positions, which must in turn be defined by an earlier derivational stage which builds hierarchical structure.A number of these cases have been analysed using restrictive derivational models like Harmonic Serialism (HS; Prince & Smolensky 1993; McCarthy 2000), which utilise a single ranking of constraints.The nature of deletion, however, may not always be amenable to such restrictions.
The present work investigates syncope in Southern Pomo (Pomoan, Northern California; ISO 639 code: peq).In this language, vowels are deleted in odd medial syllables (counting from the left word edge).Surface stress, meanwhile, is alternating and penultimate, iterating from the right word edge.The result is an inconsistent relationship between surface stress and the location of syncope, such that no clear surfaceoriented explanation is viable (Kaplan 2020; Walker 2020).Instead, this pat tern is best captured through the ordered application of multiple metrical structures, as in a multilevel derivational framework like Stratal OT (SOT; BermúdezOtero 1999;  Kiparsky 2000), the constraintbased successor to Lexical Phonology and Morphology (Kiparsky 1982; Mohanan 1982).Table 1 presents schemata and examples of odd syllable deletion in forms with four, five and six underlying syllable nuclei, with the examples crossreferenced to full glosses below. 3ather than optimising surface outputs, syncope in Southern Pomo reflects the influence of multiple opposed metrical structures.This is what is called a metrically incoherent process (Dresher & Lahiri 1991), where a metrically conditioned alterna tion does not appear to refer to surface stress.To see this schematically, (3) below schematises a process of vowel reduction affecting alternating positions -odd medial syllables, counting from the left edge.Alternating surface stress falls on odd syllables, as well.However, vowel reduction is tied to weak positions crosslinguistically.This reduction process thus seems to have misapplied.
(3) Metrical incoherence in rhythmic processes Underlying / pa.ta.ka.ba.da.ga / Vowel reduction | pa.tә.ka.bә.da.ga | Surface [ pa.tә .ka.bә .da.gá ] This pattern is difficult to explain in terms of surface structure, as the reduction process takes place in prominent syllables.This suggests that reduction has occurred at a derivational stage prior to the assignment of surface stress.It is mechanically simple to produce this kind of pattern in stratal frameworks.This is accomplished by ordering discrete, independent subgrammars such that the syllables which undergo vowel reduction are unstressed at the time that process takes place.At a following stage, stress is reparsed and those reduced syllables surface as prominent.In Southern Pomo, the earlier stage involves deletion rather than reduction, and stress is reparsed over the remaining string.
The ability of stratal frameworks to generate these patterns comes at a cost, however, as many predicted phenomena involving metrical incoherence run counter to observed typology.Indeed, the pattern above in (3), where all and only schwas are stressed, is unattested, but Wolf (2012: 6) demonstrates that this pattern is predicted by SOT.Another example offered by Wolf (2012: 3-4) produces a language in which all and only unstressed syllables start with aspirated stops.In attested languages, though, the distribution of both schwa (or reduced vowels more generally) and aspiration are asymmetrically related to prominence.Schwa can be found either only in weak syllables or in all syllables, but no language requires it in strong positions.The opposite observation holds for aspiration, which is found either in all positions, or only strong ones, but not exclusively in weak ones.Distributional asymmetries like these have been used by phonologists to motivate universal markedness constraints in CON, but, as Wolf (2012: 2) notes, 'even given an asymmetric constraint set, Stratal OT predicts a symmetric typology of stress/segmentalphonology correlations'.Strata are necessary in order to generate the attested rhythmic phenomena, but multilevel frameworks like SOT readily produce unattested, countertypological ones as well.
Additionally, despite their generability within stratal frameworks, metrically inco herent patterns are typologically quite rare (Gordon 2002(Gordon , 2016)).It is therefore important to determine how to account for attested patterns of this type.This conflict may be resolved by turning to extragrammatical explanations for the rarity of metrical incoherence: while these patterns are generable through strata, their learnability is bounded by the availability of evidence.If the language input is largely consistent with an alternative grammar which is metrically coherent, competition between grammars (i.e., between different constraint rankings) may be resolved by a learning bias for transparency (Kiparsky 1968; Prickett 2019).As a result, restructuring, in producing a new ranking of grammatical constraints, may eliminate metrical incoherence.In the case of Southern Pomo syncope, the data suggest this kind of change may have been in progress due to limited evidence for metrical incoherence, with an alternative surface optimising grammar emerging.
In §2, I detail the specifics of syncope in Southern Pomo and its interaction with metrical structure.§3 demonstrates how this pattern can be successfully captured using a stratal architecture.Following this, §4 considers alternative analyses from both derivational and parallelist perspectives, ultimately rejecting these alternatives.In §5, I discuss the stratal analysis in relation to cases of syncope in Southern Pomo which target vowels in the fourth syllable, relating this to the learnability of opaque patterns, diachronic change and restructuring.

The language
Southern Pomo is one of seven Pomoan languages of Northern California, which were traditionally spoken in the area around Clear Lake and the Russian River Valley.Kashaya (previously Southwestern Pomo) is genealogically the closest of these to Southern Pomo.Southern Pomo was the largest of the Pomo languages at the time of European contact, spoken by somewhere from 2,000 to 7,000 people in the area between modernday Santa Rosa and Sebastopol, north to Cloverdale (Walker 2020: 8).Colonial genocide of indigenous communities and the suppression of their languages left fewer than 100 speakers of Southern Pomo by 1900.

Generalisations
In Southern Pomo, vowels in odd syllables are deleted when these syllables are not at a word edge (i.e., noninitial and nonfinal; Walker 2020: 93-100).Thus, four and fivesyllable words undergo deletion in the third syllable, and sixsyllable words show deletion in the fifth syllable as well.This can be described as targeting odd syllables, with the initial syllable protected by phonotactic constraints (as in (9) below).This pattern can be seen below in (4).In these examples, the first line shows the attested surface form, with an underscore indicating a position where an underlying vowel has been deleted due to syncope.The second line shows underlying forms, decomposed into their constituent morphemes, with bolding and underlining to indicate vowels which are deleted.The segments in parentheses are the 'laryngeal increment' (Oswalt 1976), a semipredictable feature which appears in all words and has the effect of making all initial syllables heavy. 5Southern Pomo is a morphophonologically complex language with many alternations (e.g., compensatory lengthening, seen in ( 4d) and (4g)); examples highlight these processes where necessary.Unless otherwise noted, Southern Pomo examples are from Walker (2020, cited as W) or Oswalt ([1981] 2014, cited as O). 4 For information see the Western Institute for Endangered Language Documentation Southern Pomo Project: http://wieldoc.org/wp_temp/?page_id=30 5 I analyse these segments as a feature of the wordlevel phonology, rather than as lexically specified or stemlevel, on the basis of their behaviour in compounds and interaction with postlexical enclitics.I maintain Walker's morphophonemic forms.For discussion, see Halpern (1984) and Walker (2020: 119-122); see also Buckley (1992) and Oswalt (1998) for discussion of this phenomenon in closelyrelated Kashaya. 6Walker (2020) is the published form of his (Walker 2013) dissertation.The original dissertation contains additional reproduction of Halpern's unpublished fieldnote transcriptions that were omitted from the published edition.These are cited here as W13.This pattern of deletion targets alternating syllables in both lexical stems and affixes, across morphological and lexical classes.Because stems are typically two syllables (though not always, as above in (4a)), it is usually the first suffix with vocalic material which undergoes syncope due to their stringlinear position.These are commonly directional suffixes (glossed as DIR) such as /alokoʧ'/ 'up out of', /maduʧ/ 'up to' and /alameʧ'/ 'down off of' (further directional suffixes are discussed in ).9As a result, syncope frequently gives rise to intra paradigmatic vowel∅ alternations in verbal paradigms: (5) The contrast between the allomorphs seen in ( 5a) and (5c) hinges solely on the syllable added by the evidential suffix /a/.The presence of this suffix allows for deletion of the penultimate vowel /o/ in (5a); this same vowel surfaces in (5b) and (5c).
Syncope also affects root morphemes, as seen particularly in reduplication (both verbal and adjectival), and usually affects the first syllable of the second root in compounds, exemplified below in ( 6 Syncope is blocked where it would violate phonotactic constraints, including a surfacetrue ban on complex onsets and codas, even when there appears to be a viable alternative site of deletion.This can be seen in ( 9 Nonapplication can be attributed to a constraint like *COMPLEX (roughly 'no tautosyllabic obstruent clusters'). 12This phonotactic constraint, which blocks syncope when it would create a complex coda or onset wordmedially, also prohibits deletion of the vowel in the initial syllable even though this is in an oddnumbered position.All syllables in Southern Pomo surface with onsets (indicating that ONSET is another undominated constraint), and because of this, deletion in an initial syllable will necessarily result in an illicit onset cluster (e.g., deletion in /cv.cv/ results in *[c_cv]).
The examples here are demonstrative but by no means exhaustive.Oddsyllable syncope is quite regular across the lexicon, rather than lexically specific.However, we will see in §2.2.2 that this process is in conflict with the language's stress system.The alternatingsyllable deletion seen in these data is a quintessential mark of metrical conditioning, where deletion has occurred weak positions.The crosslinguistic evidence for vowel deletion in strong positions is only tenuous, and the existence of this phenomenon is disputed (McCarthy 2008).We can thus conclude that the derivation first builds metrical structure from left to right, only assigning prominence to syllables which are not ultimately targeted for vowel deletion.Next, the weak positions undergo syncope, driven by a constraint against vowels in these positions.Finally, metrical structure must be reparsed from right to left, generating the observed surface stress.(11) presents a condensed rulebased schematisation of this system; in §3, I demonstrate a formal analysis of this pattern using SOT.
( The mechanics of the constraintbased analysis differ slightly -for instance, the second and third of the steps in (11) (syncope and reparsing) will occur in tandem.The crucial similarity in these accounts is that the metrical structure which conditions syncope must be different from the one responsible for surface stress.Syncope is metrically opaque, whether this pattern is formalised with rules or constraints.

Syncope as a phrasal phenomenon
There is strong evidence that the second stage of metrical structurebuilding occurs at the phrase level: surface stress in both Southern Pomo and Kashaya is assigned in a domain larger than the word, described as the 'phrase' (Buckley 2019; Walker 2020: 49) or 'breathgroup' (Halpern 1984: 38).14This is evidenced by phrasal stress shift: (12) below shows the expected stress pattern for words in isolation vs. when phrased together, with stress shifting rightward (relative to the expected penultimate prominence seen at the word level) to avoid a lapse within a phrase.Here, the location of the shifted stress is underlined: Additionally, all morphosyntactic enclitics (discussed at greater length below) are also included in this larger stress domain, as seen in ( 13).These morphemes combine with stems at the phrase level and are prosodically deficient: not licit as independent prosodic words and thus not assigned prominence at the word level (Anderson 2005(Anderson , 2011)) The above morphosyntactic evidence indicates that enclitics are introduced into the derivation at the postlexical (phrasal) level and are not assigned stress in the word level stratum.
We thus see converging evidence that surface metrical structure must be assigned no earlier than the derivational level at which syncope occurs, that is, it is phrasal.However, the site of deletion is determined by the wordlevel stress of the input to the phrase level, not the syllable's context in a larger phraselevel unit.The derivation in (15) below shows that the predicted syncope pattern is seen even when surface stress has shifted: (15) Syncope is insensitive to position in a phrase Even when the deleted syllable would be prominent due to phrasal stress shift, the location of syncope does not change.The pair of examples below in ( 16) demonstrates that the same pattern of syncope maintains in /maʔdakaden/ → [maʔ.dak_.den]when the position of phrasal stress varies.These examples feature an allomorph of the same root seen in ( 15), /dakʰad/ 'spouse', with a different locus of deletion -conditioned by wordlevel metrical position.This root is realised as [tʰ_kʰan] in (15) (with regressive voicing assimilation and wordfinal neutralisation of /d/ to [n]) following a two syllable prefix, but surfaces as [dakʰ_d] in ( 16), following a monosyllabic prefix.The difference in syllable parity leads to a difference in which syllable is weak in the output of the word stratum.Phrasal clitics also participate in some -but not all -nonmetrical sandhi processes that occur at morpheme boundaries, providing some evidence that they are not present at all points in the phonological derivation.First, there is a process of wordfinal ejectivisation in Southern Pomo which does not occur under suffixation.The dental, alveolar and velar plosives become ejectives wordfinally, and these stops are always ejectivised before clitics. 19Compare the plain-ejective alternation, which occurs wordfinally in ( 17) and before enclitics in (18), 20  The ejectivisation before enclitics seen in (18b) suggests that clitics are outside the domain of computation for some wordlevel processes, providing more evidence for the existence of multiple morphologically conditioned levels of phonological operations, and specifically a wordlevel stratum where suffixes are counted, but enclitics are not.
Clitics in Southern Pomo do, however, participate in other sandhi processes at the phrase level, such as compensatory lengthening: heterorganic obstruent clusters at morpheme boundaries undergo deletion of the first segment, with compensatory lengthening of the preceding vowel.(Homorganic obstruent clusters typically undergo fusion.)This process is particularly instructive because compensatory lengthening has been said to require a derivational approach: predictable moraic structure must be defined prior to deletion for mora preservation to occur (Kiparsky 2011; Samko 2011).An example of this process applying wordinternally is shown in (20a).As (20b) demonstrates, this process also occurs in clusters formed between stems and enclitics: These forms suggest that a constraint preserving moraic structure (MAXµ) is active at the phrase level, where deletion is promoted by phonotactic pressures.
A number of forms in Southern Pomo demonstrate an additional process of hiatus resolution that bolsters a derivational account of syncope.In contexts where two vowels are adjacent across a morpheme boundary, the second vowel does not surface, and is also not counted in the computation of syncope.This is shown below in ( 21 The fact that this vowel is not counted by syncope indicates that it is not present at the phrase level, where syncope takes place, and instead it must be deleted prior, at the level of the word, such that the vowel is not counted in the initial computation of metrical structure.21There are no vowelinitial clitics which would allow this process to be seen occurring between stems and enclitics (such that we could ascertain whether this process is active in the phrasal stratum as well).In a rulebased derivation (elaborating on that in (11) above), this would simply precede the initial metrical parse, as seen below in ( 22). ( 22 These phenomena provide convergent evidence for a stratal derivation, and this analysis finds further justification from diachrony.While Southern Pomo has alternat ing stress assigned from right to left, stress in all other Pomo languages is leftaligned (McLendon 1973), and Buckley (2019) analyses ProtoPomo metrical structure as trochaic with initialsyllable extrametricality and peninitial primary stress.If these trochees yielded alternating secondary stresses in even syllables assigned rightto left, this is mechanically identical to the output stress of the wordlevel stratum in this account.Buckley proposes that the rightaligned prominence in contemporary Southern Pomo reflects contactinduced change influenced by the Bodega variety of Coast Miwok (Miwokan; California), a language with overwhelmingly penultimate stress (89%; Buckley 2019; Callaghan 1970.It could thus be that bilingual learners of Southern Pomo imposed this pattern of metrical organisation from Miwok at the phrase level.This would mean that, in addition to being mechanically necessary to produce this type of metrical opacity, these strata may also directly represent historical change in the layered structure of the synchronic grammar ('amphichronic explanation'; BermúdezOtero 2015; Kiparsky 2015; Gordon 2016).Sound change has been shown to develop along a trajectory, with processes phonologised at the phrase level and generalised to smaller domains over time via domain narrowing (i.e., the phonological lifecycle, BermúdezOtero & Trousdale 2012; see also domain generalisation, Myers & Padgett 2014).The development of Southern Pomo appears to present exactly this pattern, with ProtoPomoan stress still synchronically active at the word level and the laterintroduced penultimate stress pattern at the phrase level.
To summarise: rhythmic deletion in Southern Pomo targets alternating vowels in odd syllables, when these syllables are nonfinal and deletion is allowed by phonotactics (i.e., when it does not result in a complex onset or coda).The location of deletion is not consistent relative to surface stress: primary stress is penultimate and alternates leftward, such that odd syllables may be stressed or unstressed on the surface.Additionally, the domain of stress is the phrase rather than the word, which means that stress does not fall consistently even within the same word, based on its position within a multiword phrase or when it is encliticised (as clitics are also included in this phrasal domain, and bear stress).Regardless, the location of syncope remains the same.Taken together, this evidence suggests that the metrical structure conditioning syncope is distinct from and derivationally prior to that seen in surface forms.
§2.4 presents a Stratal Optimality Theory analysis of this pattern.Prior to this analysis, however, we introduce a significant wrinkle: a small portion of the lexicon which exhibits syncope in the fourth syllable, rather than the third.While these may seem at first to represent a hole in the account described above, it may be that these forms are instead a key to understanding the diachronic stability of highly opaque systems like metrical incoherence.

Fourthsyllable syncope
The derivation in ( 22) identifies the location of syncope in the vast majority of Southern Pomo forms.However, a number of words instead exhibit deletion of what would be a penultimate fourth syllable, rather than deletion of the third underlying vowel (in what would be the antepenult).Several of these are shown in ( 23 The examples in (23a)-(23c) show deletion in the fourth syllable when syncope in the third syllable would produce a licit form.(23d) shows fourthsyllable syncope occurring in a string where deletion of the vowel in the third syllable would produce an illicit cluster *[t̪ 'md], so syncope is not expected.This pattern does not appear to be related to the identity of the lexical root, as paradigmatically related words have syncope in the third position, nor is it explained by the specific suffixal morphology, as these morphemes appear in other words with the thirdsyllable pattern of deletion.Because the pattern of deletion varies within paradigms, it cannot be attributed to an itemspecific process indexed to particular lexical roots, morphemes or morphosyn tactic phases.Syncope also does not seem to 'compromise' by deleting in a position that produces a phonotactically licit form, as evidenced by the forms in (9) where syncope is blocked.Additionally, these forms show an implicational asymmetry: for any paradigm where fourthsyllable syncope occurs, there are also forms with third syllable syncope, but the reverse does not hold.Further, as discussed below in §2.4, the regularity of thirdsyllable deletion in likely lowfrequency contexts suggests that this process is productive.Deletion in the fourth syllable thus appears to be an exception to the descriptive generalisation that syncope targets vowels in the third syllable, which holds in all but a handful of words across the language.
However, we will see later in §4.3 that the grammar which produces these exceptional fourthsyllable forms also makes correct predictions for most (but not all) Southern Pomo words.Specifically, this grammar only generates the fourthsyllable syncope pattern in strings that would, without syncope, surface with five syllables and a light penult.§4.2 evaluates a parallel OT analysis where syncope is driven by the StresstoWeight Principle (SWP) and shows that it differs from the predictions of the stratal derivation for exactly these forms.To explain the variability between this SWP syncope grammar and the stratal syncope grammar, §5 approaches this conflict as a matter of grammar competition in the ranking of constraints (in particular, SWP) propelled of the scarcity of disambiguating evidence.First, I lay out a stratal analysis in §2.4 which accounts for the thirdsyllable syncope pattern.

Syncope as a productive process
Recent work has suggested that rhythmic syncope may be particularly prone to restructuring because of the computational complexity induced by alternating deletion (Hao & Bowers 2019; Bowers & Hao 2020).Rhythmic syncope systems in other languages such as Nishnaabemwin (Bowers 2019) and Eastern Slavic (Isačenko 1970, :95-6) are reported to have quickly collapsed.In Southern Pomo, syncope is further complicated by the imposition of a different structure (surface stress) which could, in theory, condition further deletion.As a result, there is no surfacetrue generalisation about weak positions which explains the process.In addition, Gordon (2016: 43) notes that phonological systems with competing iterative parses are not only rare but also unstable, with one metrical system typically fossilising in exceptions and morphologisation.
Given this, we must address whether rhythmic deletion was a synchronic process.Learning syncope requires robust evidence for vowelzero alternations across the paradigm -evidence that is clearly present in Southern Pomo, as discussed in §2.2.2 and §2.4There are robust vowel alternations in syncope with reduplication (shown in ( 6)-( 8)), as the base of reduplication surfaces with the vowel that is absent in the reduplicant and most of these reduplicative bases also surface elsewhere as non reduplicated forms.This strongly suggests that speakers are aware of the underlying phonological shape of the roots which undergo syncope.( 4) and ( 5) show a number of intraparadigm vowel-∅ alternations in highly productive functional morphemes, as well.It seems likely that learners are able to learn accurate underlying representations, with vowels still in place.That is, they have evidence that syncope is a synchronically active process.
However, as we will see, the question is not simply whether deletion per se was synchronically active, but which mechanism was responsible.This requires an analysis of the systems underlying both the third and fourthsyllable patterns of deletion.In §5, I return to this question, and propose that Southern Pomo at the time of observation was undergoing a change in progress, with ongoing grammar competition between these two systems.

Stratal Optimality Theory
Stratal OT is a derivational variant of classic Optimality Theory that recasts Lexical Phonology and Morphology in a constraintbased system.Phonological operations are driven by the interaction of constraints on input-output faithfulness and surface oriented markedness, and candidate forms are evaluated on their satisfaction of all of these constraints in parallel.SOT differs from classic OT in that, while there is fully parallel evaluation, there are also multiple ordered stages, corresponding to morphological domains -typically the stem, word and postlexical phrase.The constraint ranking of each subgrammar is fully independent of the others, such that marked structure may be preserved from an earlier stage by the promotion of a faithfulness constraint.
Stratal analyses of metrical opacity have been proposed to describe similar asym metries between stress and metrically conditioned phenomena, including prepausal stress shift in Tiberian Hebrew (Churchyard 1999), vowel lengthening and stress interactions in Tübatulabal (UtoAztecan, California; Benz 2018; similarly suggested in Heath 1981) and stress freezing in Washo (isolate, California; Benz 2018).The use of strata in these accounts is predicated on the observation that the relevant process cannot be predicted from surface structure and is thus metrically opaque.
Because syncope is dependent on the existence of metrical structure, the first stratum (the word level) must be responsible for assigning the prominences which then condition rhythmic vowel deletion.(This analysis will not account for stem level phonological patterns.)The second stratum (the phrase) is then able to reference existing structure, and because of this it can accomplish both syncope and stress realignment in parallel.In the remainder of §2.4,I demonstrate a SOT analysis of Southern Pomo syncope and several other phonological processes.This is certainly not the only set of constraint interactions which successfully derives syncope in SOTrather, this is merely a mechanical illustration using widelyadopted constraints.What is essential is that this analysis is able to generate the thirdsyllable deletion pattern in all cases, in contrast to the counteranalyses seen in §4.

Stratum I: the word level
As described above, the first task of the derivation must be to generate a metrical structure which defines the correct weak positions for syncope.Though this account is amenable to a footbased treatment, I use an unbracketed grid representation (Prince 1983) instead: it is prominence per se that appears to matter, rather than the locations of foot boundaries or hierarchical relationship between feet.This account also collapses across primary and secondary stress, as this distinction does not affect any phono logical processes, and assumes stress alignment constraints produce the correct edge alignment of main stress.The surface regularity of rhythmic stress is accomplished by *CLASH ('no consecutive stressed syllables') and *LAPSE ('no consecutive unstressed syllables'), which are undominated throughout (Gordon 2002).These are ranked over the faithfulness constraint IDENT(STRESS) -that is, 'don't change the stressedness of a syllable,' henceforth ID(STR) -in motivating the building of metrical structure.The undominated ranking of *CLASH and *LAPSE is assumed throughout for presentational simplicity.For this reason, candidates with stress clashes or lapses are not shown.
To define the correct weak positions that syncope acts on in the second stratum, the essential ranking is that between two TROUGH constraints: (24) Definition: TROUGH{LEFT/RIGHT}: Assign one violation just in case there is a prominent syllable at the left/right edge of the stress assignment domain.
In the word stratum, TROUGHL dominates TROUGHR, which has the effect of assigning stress from left to right, with an initial weak syllable.The first stratum thus outputs lefttoright alternating stress, and the syllables which will be targeted for syncope in the phrase stratum are unstressed.In addition, assuming a ranking of MAXV ('don't delete vowels') over *V ̆guarantees that deletion does not occur at the word level.The third syllable (ultimately, the site of deletion at the phrase level) is underlined.The tableaux in ( 25) and ( 26) demonstrate that MAXV must also outrank TROUGH R -ruling out candidate (25c) -and ID(STR), ruling out candidates (25d) and (26d).A definitive ranking of TROUGHL and *V ̆cannot be established.This gives us the ranking in Figure 1.
The tableaux in ( 27) and ( 28 The crucial aspect of these outputs is that the syllables targeted for syncope -odd syllables counting from the left word edge -are unstressed.In the phrasal stratum, these syllables are deleted by reranking *V ̆, which penalises vowels in weak syllables, above the faithfulness constraint MAXV which had preserved these positions at the word level.

Stratum II: the phrase level
The second stratum inherits the metrical structure assigned by the first stratum as its input.With this, it must both propel deletion in the weak positions defined by that structure, as well as reassign prominences from the right edge leftward, beginning with an unstressed syllable.The second of these effects is accomplished by inverting the ranking between the trough constraints, such that TROUGHR is reranked above TROUGHL.
To generate syncope, *V ̆is reranked above MAXV.However, the targets of deletion are motivated not by surface stress, but by the metrical structure from the first stratum.This entails a constraint demanding faithfulness to stressed syllables in the input, MAXV́('don't delete vowels which are stressed in the input'), which has the effect of protecting those vowels from deletion regardless of whether they remain stressed in the output of the second stratum. 2229) Definition: MAXVF or any stressed vowel α in the input, assign one violation if that vowel has no correspondent β in the output.This constraint is undominated at the phrase level and may be considered undom inated at the word level as well; there, it is vacuously satisfied, irrespective of the input, because deletion is prevented by the high ranking of MAXV.As a more specific constraint, MAXV́is less stringent than MAXV -it is violated by only a proper subset of the deletions which would violate MAXV.Thus, vowels that are unstressed in the input may be deleted, while deletion of their stressed counterparts is barred.The crucial phraselevel ranking to motivate syncope in the appropriate syllables is therefore MAXV́≫ *V ̆≫ MAXV.This ranking ensures that only syllables that are unstressed at the word level are deleted at the phrase level.The tableaux in ( 30) and ( 31) demonstrate this result for even and odd syllable parities, again with undominated *CLASH and *LAPSE, and the assumption that phonotactic constraints bar the deletion of vowels in the initial syllable or consecutive syllables: (30) Phrasal stratum syncope and stress realignment, even parity In these tableaux, we can see that higherranked TROUGHR successfully reverses the direction in which metrical structure is built, *V ̆propels deletion in weak syllables, and MAXV́prevents deletion from occurring in syllables which were stressed in the input.(30) demonstrates several rankings: candidate (30b) is eliminated by the ranking of TROUGHR over both MAXV and ID(STR), and (30e) similarly shows that TROUGHR outranks *V ̆.Further, we see that TROUGHL is still active at the phrase level stratum and must be ranked above MAXV to eliminate candidates without deletion like (30c).Interestingly, it is this alignment constraint, rather than *V ̆, which motivates syncope for evenparity forms.This can be thought of as a peculiarity of this OT implementation, rather than an important feature of the system; as we have seen, this can instead be rendered in terms of rulebased processes.(31) demonstrates two further critical rankings: *V ̆, which motivates deletion by dominating MAXV, must also dominate TROUGHL and ID(STR) in order to rule out candidate (31b), which violates neither TROUGH constraint.The full ranking is shown in the Hasse diagram in Figure 2.
These constraint rankings demonstrate that TROUGHR and *V ̆have been re ranked above TROUGHL and MAXV, and further articulate the ranking between these constraints.This propels both deletion and the reconfiguration of the metrical structure in a way which must satisfy *CLASH by shifting stress away from derived clashes (e.g., in forms like *[ʔah.tìm._kó.ʧin]).Because TROUGHR now outranks TROUGHL, stress appears to be 'shifted' leftward onto the initial syllable (in oddparity inputs); in even parity inputs, this results in one less stressed syllable.We thus see deletion (in the odd medial syllables) and alternating stress shifted to the penult, iterating leftward.The tableaux below show that this results in the desired outputs for our two example words from §3.2: (32) Phrasal stratum for /pʰuht̪ oput̪ ow/ We have now seen that the Southern Pomo data are generable using a twostage derivation in SOT.The first of these strata, corresponding to the word level, defines weak positions from the left edge, and the second stratum (phrase level) deletes vowels in these positions and reassigns prominence from the right edge.The positions where syncope occurs are determined by wordlevel metrical structure assigned in the first stratum, but deletion itself is applied at the phrase level.This seems at odds, however, with the observation that the location of syncope is insensitive to a word's position within a larger prosodic constituent.Syncope occurs in the same syllable regardless.This outcome is in fact consistent with the strata above.Vowel deletion is determined by an outputoriented markedness constraint against unstressed vowels (*V ̆), but the targets and extent of deletion are governed by higherranked faithfulness constraints.MAXV́ensures that deletion occurs in positions determined by wordlevel metrical structure, even though deletion takes place in the phraselevel phonology.Recall the form in (16b) (repeated below), which shows that syncope is not affected by the placement of phrasal stress: Compare the prediction for the morpheme string /maʔdakaden/ in isolation (in (34)) to the form predicted when this is followed by three syllables, as in (16b), shifting stress to the ultima (in (35)).These tableaux demonstrate that the locus of deletion does not change, and the outputs are both [maʔ.dak_.den]; the account predicts the observed invariance in the locus of deletion.
(34) Phrase level, showing syncope for /maʔdakaden/ in isolation We have also seen that wordfinal syllables do not undergo syncope even when they do not receive wordlevel stress in Stratum I, for example, the wordfinal syllable in [ts'íh.ta]'birds' in (12) above, reproduced here in (36).This constraint is undominated, with the result that vowels in final syllables in the word domain are not eligible for deletion in the phrasal domain. 23 The discussion of enclitics above indicated that these are introduced into the derivation at the phrase level and are not assigned stress in the wordlevel stratum.This means that, in the latter stratum, MAXV́(which causes deletion to be rhythmic elsewhere) will be vacuously satisfied no matter the pattern of deletion in the clitic field.Syncope in these morphemes is therefore free to occur anywhere -any vowel in the clitic field is liable to be deleted.Combined with the activity of *CLASH, *LAPSE and the syncopedriving constraint *V ̆, this predicts that the clitic field will undergo as much deletion as is allowed by phonotactics.This deletion is predicted to occur regardless of the position of encliticised material relative to either word or phrase level stress.That is, clitics should not show rhythmic vowel deletion, nor the vowel-∅ alternations that wordinternal suffixal morphemes do.Clitics should instead surface with the minimal amount of phonological material allowed by phonotactics.This prediction is illustrated in (38): 24 (38) Maximal deletion is predicted in the clitic field This prediction cannot be tested, because the clitics of Southern Pomo are pre dominantly monosyllabic and appear to be irreducible: any further vowel deletion 23 Syllables which are final in some domain are not always immune to deletion -for example, we see root final deletion in (4a) [ʃiː.báːt̪ʰ._mʰuj].This is not contradictory to the model proposed here: bracket erasure is a prediction of this model -'phonology applying within higherorder stems in a word does not make reference to deeply embedded morphological structure', (Inkelas & Zoll 2007: 145; see also Bermúdez Otero 2012: 82).Bracketing at the root level in noun-noun compounds is not visible at the phrasal level (two stages later) where syncope occurs. 24Here, I assume any candidate […=c_c_] with two deletions in the enclitic will be blocked by phonotactic constraints, as this results in a complex coda.results in phonotactically illicit sequences.There are thus no synchronic vowel-∅ alternations in the clitics themselves, and the tableau above represents an earlier stage in the language's development.However, some evidence emerges from comparison to the patientive pronominal clitics [=ʔt̪ o] '1.SG.PAT' and [=mt̪ o] '2.SG.PAT', both of which have unreduced forms (and cognates in other Pomo languages) that appear as independent words, [ʔaːt̪ o] and [miːt̪ o]; I analyse the underlying realisations of these morphemes as /ʔat̪ o/ and /mit̪ o/, respectively (i.e., I attribute vowel length and thus syllable weight in the initial syllables of these two forms to the laryngeal increment).When encliticised, these pronouns are always reduced, irrespective of the position of surface stress.Similarly, Walker proposes that 'regular syncope' is responsible for the synchronic forms of many clitics which were historically multimorphemic (380), though this did not apparently rely on the rhythmic properties of the host word like syncope elsewhere in the language.
This suggests that the clitic field was once subject to deletion of unstressed vowels, but that this was a nonrhythmic process, as predicted by this stratal analysis.Deletion has seemingly occurred wherever possible and is nonalternating.It therefore appears that the lexicon has undergone diachronic restructuring, what has been called lexical or input optimisation (Prince & Smolensky 1993).The learner, as a default, assumes that the surface form of a word violates as few faithfulness constraints as possible -that is, that the surface form and underlying representation are maximally similar.Without alternations providing evidence for the presence of underlying vowels, learners acquire the reduced forms of these morphemes, the only observed surface forms, as underlying representations (BermúdezOtero & Hogg 2003).The synchronic forms of these clitics represent historically frozen patterns of deletion.
I have now shown how SOT accounts for the thirdsyllable pattern of syncope in Southern Pomo.The theoretical consequences of this analysis are significant: with the stratal framework needed to account for metrically incoherent syncope in Southern Pomo, we allow for countertypological overgeneration.We should thus consider whether this syncope pattern can also be achieved in parallel, without a derivation, or in a more restrictive framework with a static constraint ranking, like Harmonic Serialism (HS), OT with Candidate Chains (OTCC; McCarthy 2006), or Transderivational Faithfulness (Benua 1997).§4 considers such analyses, which are ultimately unsuccessful at generating this pattern.However, we will also see that the parallelist analysis utilising the STRESSTOWEIGHT PRINCIPLE (Prince 1990) generates fourthsyllable deletion.

Nonstratal derivation
In HS and OTCC architectures, opacity arises from satisfying constraints sequentially, from highest to lowestranked, rather than in parallel.This system does not allow those constraints to be violated in later stages of the derivation; there is only one constraint ranking in the derivation, and constraints are not reranked (unlike SOT).Metrical structure in an HS or OTCC account is determined by highly ranked constraints, and no candidate can harmonically improve if it is not fully compatible with these constraints (McCarthy 2008; Pruitt 2010; Elfner 2016; Calamaro 2017).HS is thus unable to produce metrical reversals due to inherent features of its architecture.In syncope processes, HS enforces an inherent ordering of metrical parsing before the deletion that it conditions.This captures rhythmic vowel deletion like that seen in Awajún (Aguaruna) and Tonkawa (McCarthy 2008), but by design is unable to produce the Southern Pomo pattern: even in the absence of syncope, surface stress in Southern Pomo must be realigned to the right edge of a postlexical phrase, which necessitates metrical reparsing.These factors militate for an account in Stratal OT, with free re ranking of constraints between strata.
Another set of analyses, utilising base-output correspondence (i.e., Transderiva tional Faithfulness; TF), have been put forward for the Tripoli and Palestinian dialects of Levantine Arabic (Kager 1999).These analyses derive opacity in morphologically complex forms by enforcing faithfulness to their lesscomplex stems.TF was devised to handle opacity arising from morphologically conditioned opaque inheritance, such that transparent phonological alternations from base forms appear in more morpho logically complex forms as over or underapplication of a phonological process.In Southern Pomo, there is no form in the paradigm with transparent application of deletion that can serve as such a base, and as such there is not a clear way to discuss this pattern of opaque deletion in terms of faithfulness to such a base.
Derivational frameworks which depend on a single ranking are not able to realign stress without making a candidate less harmonic.However, these frameworks may be able to generate the attested pattern using disjoint metrical tiers (also called disjoint footing; Rappaport 1984; Halle & Vergnaud 1987; Parker 1998; Aion 2003; González 2007; Vaysman 2008).Such analyses utilise multiple tiers of distinct metrical structure that coexist within a word, each conditioning the application of one of the two rhythmic phenomena in question.
(39) Disjoint footing for stress and syncope Stress: R→L σ (σ σ)(σ σ) Deletion: L→R {σ σ}{σ σ} σ Here, unlike in analyses involving incoherent surface prominence (e.g., Gordon 2016), one metrical structure corresponds to prominence, and the other to a rhythmic segmental process (e.g., vowel deletion).However, because surface stress in Southern Pomo is alternating and postlexical, it must be calculated based on the syllable string after deletion.Therefore, disjoint footing processes must still be derivationally ordered, syncope followed by stress assigment.Though similar analyses have been proposed for stress and syncope in Tiberian Hebrew (Rappaport 1984; cf. Dresher 2009), the interaction of vowel lengthening and stress in Tübatulabal (Heath 1981; Aion 2003; cf. Benz 2018), and alternating epenthesis and stress in Huariapano (González 2007; Parker 1998; cf. Bennett 2013b), the necessity of sequential deriva tion makes this analysis qualitatively different from other uses of disjoint footing.This approach thus appears to merely recapitulate the predictions of SOT, including the same set of countertypological rhythmic processes criticised by Wolf (2012).In addition, Disjoint footing is unable to account for nonrhythmic derivational phenomena such as compensatory lengthening.For further argumentation against this approach, I direct the reader to Bennett (2013a), Churchyard (1999) and the references therein.

Parallelist approaches
A number of successful analyses of syncope -such as those for Southeastern Tepehuán (Kager 1997; Gouskova 2003), Tonkawa, Hopi (Gouskova 2003; cf. Blumenfeld 2006), Bedouin Hijazi Arabic and Central Alaskan Yupik (Gordon 2001) -rely only on surfaceoriented constraint interactions and are generable in parallel OT.Some of these analyses utilise FOOT or STRESSTOEDGE ALIGNMENT constraints, which penalise intervening syllables between a word edge and foot edge or stressed syllable.Others make use of the STRESSTOWEIGHT PRINCIPLE (SWP, which penalizes stress on light syllables; Prince 1990), or other quantitysensitive constraints on stress.All of these constraints maximise the wellformedness of surface metrical structure and can promote deletion to either minimise the number of syllables between a prominence and word edge (ALIGNMENT) or ensure that stress only falls on syllables containing enough phonological material (SWP).
As I demonstrate below, parallel OT and surfaceoriented constraints -those that militate for stress to fall on certain types of syllables, and those that militate for reducing phonological material -cannot generate the observed thirdsyllable deletion pattern, precisely because thirdsyllable deletion is not surfaceoptimising.Instead, this syncope pattern requires reference to metrical positions defined at an earlier derivational stage.While syncope in the stratal account is conditioned by a surface oriented markedness constraint (*V ̆), it is faithfulness to metrically strong positions in the input that determine where deletion occurs.
Given that surface metrical structure in Southern Pomo is aligned to the right edge of the word, we might hypothesise that syncope can be explained by the ranking of ALIGNRIGHT(σ,ω) (McCarthy & Prince 1993) over MAXV.This ranking would have the effect of promoting deletion when this results in a prominence closer to the right word edge.As a result, this constraint, and its sibling ALIGNLEFT, suffer from the same shortcoming: deletion in any syllable can improve harmony by reducing word length, and worse yet, this promotes deletion of as many syllables as possible (to the extent allowed by other constraints).Perhaps SWP will fare better?
Syncope appears to always result in fewer stressed light syllables, due to the phonotactic restrictions of the language and to the fact that initial syllables in the language are always heavy due to laryngeal incrementation (discussed above in §2.2.1).Deletion of the vowel in a given syllable results in the resyllabification of that syllable's onset as a coda consonant in the preceding syllable.That is, /cv 2 .cv 3 .cv4 / becomes [cv 2 c._cv 4 ].Therefore, deletion in the third syllable creates a heavy peninitial syllable, and deletion in the fifth syllable creates a heavy fourth syllable: SWP, which assigns violations to stressed light syllables, is thus a plausible constraint to motivate syncope, and maximally wellformed words will avoid light stressed syllables.Impressionistically, this appears to reflect the data.Syncope is blocked in closed syllables by phonotactic constraints, and so words which have undergone syncope invariably have one less light syllable.The resulting heavy syllable is frequently stressed.As the tableaux in (41) show, the syncopated candidates are in fact more harmonic than their faithful competitors: (41) a. Forms that have undergone syncope are more harmonic for SWP than candidates without deletion.As previously mentioned, this is typically because deletion shifts stress leftward: in words with an odd parity of underlying syllables, this results in the heavy initial syllable receiving a secondary stress; in evenparity words, this results in one less stressed syllable overall, and stress falls on a nowheavy peninitial syllable.The resulting wellformedness is not necessarily local, as can be seen in ( 42), where the derived heavy syllable is not stressed: Here, main stress remains on a light syllable, and the heavy syllable formed by vowel deletion is itself unstressed.The word, however, is globally more harmonic with respect to SWP, because stress is shifted leftward onto the heavy initial syllable.Syncope does not result in a heavy stressed syllable locally, but there is one less light stressed syllable.This is an example of how parallelist approaches to syncope have generally resolved apparent opacity: a process may seem unmotivated or opaque locally, but results in a more harmonic form under global evaluation (Gouskova 2003; Blumenfeld 2006).
Syncope always increases global harmony with respect to SWP, and SWP accu rately predicts the majority of Southern Pomo surface forms.But we cannot conclude that syncope is transparent: like the alignment constraints, SWP is not able to correctly predict the location of syncope in every case.The optimal candidate for SWP is always one where all stressed syllables are heavy.As shown in (43), deletion in the fourth syllable, rather than the third, would result in a maximally harmonic candidate for SWP: ( As this tableau shows, the attested form (43a) is indeed harmonically bounded by a competitor (43c) in which V 4 has deleted, resulting in a candidate without light stressed syllables.When V 4 is light, deletion of V 4 shifts stress leftward to σ 3 , which is then heavy, having resyllabified the onset of σ 4 as a coda.The optimal location for deletion in these cases is V 4 , the vowel in the even penult, but syncope instead targets vowels in the odd (third and fifth) syllables of the word.The only forms where we observe this are foursyllable words like those in (44), derived from underlying fivesyllable strings: (44) Example forms suboptimal for SWP a. i.
[ The predictions of an SWPdriven system are usually in line with the observed Southern Pomo data, but SWP cannot be responsible for driving the thirdsyllable pattern of rhythmic deletion because a subset of forms is not surfaceoptimising.
These forms -surface strings bearing stress on odd, light penults -provide crucial evidence for opacity, but are only a small portion of the language.How heavily do learners weigh this evidence when most of the input is equivocal?Tellingly, the subpattern of fourthsyllable syncope discussed in §2.3 appears to delete in exactly these surfaceoptimising positions, and these forms resist analysis by stratal derivation.I propose that this reflects restructuring of the synchronic grammar, due to the limited learnability of the opaque stratal pattern.

Parallelism and fourthsyllable syncope
Examples of these forms are given again in (45), along with the expected forms (based on the SOT account) that do not appear.
( Compare the forms in (45a), where syncope results in forms without light stressed syllables, to those in (45b), where syncope is not surfaceoptimising for SWP.The pattern of syncope in thirdsyllable forms is not surfaceoptimising because they have light stressed penultimate syllables on the surface.In contrast, the fourthsyllable forms have syncope of the underlying fourth vowel, resulting in a heavy stress bearing penult, and thus transparent surface optimisation relative to SWP (as no stresses fall on light syllables).These forms cannot be morphemespecific, as shown in (23).They instead suggest a reanalysis of the grammar, where syncope and metrical parsing occur in parallel, such that syncope is metrically coherent (in the sense that the language makes use of a consistent metrical structure across derivational stages; Dresher & Lahiri 1991; see also Bennett 2013b).The constraints driving syncope here are outputoriented, unlike the thirdsyllable deletion pattern seen in similar forms.
This analysis still requires strata because of phrasal stress shift, and two separate computations of stress are thus necessary to produce the observed structure. 25How ever, under this analysis, syncope does not involve interstratal dependencies, and instead takes place at the word level, alongside word level stress.It is only in larger prosodic constituents that metrical opacity is introduced by shifts in stress.An example derivation is presented below; the tableaux in ( 46 The word stratum outputs the isolation forms; in most cases these look identical to those derived by the derivation outlined in §3.The notable exception is the winning candidate (a) in (46b), which has no light stressed syllables, whereas the opaque derivation results in candidate (c), with a light penult that is not surfaceoptimising.(46c), showing sixsyllable forms, demonstrates how SWP optimises globally, rather than locally.Deletion in the winning candidate seems to overapply, in that the second syllable is heavy because of deletion, but stress does not actually fall on it.Instead, deletion causes stress to fall on the obligatorily heavy initial syllable, reducing the overall number of light stressed syllables.
The second stratum in this alternative derivation maintains the same ranking; the key difference is the size of the input.Further deletion in the stem is prevented by phonotactics: deleting any medial vowels would result in illicit onset or coda clusters.The tableau in (47) demonstrates this with a (derived) foursyllable word followed by a threesyllable word.
(47) Phrasal Stratum IIb, alternative SWPdriven account (five underlying sylla bles) This second stratum outputs the shifted stress in phrases without necessitating a reranking, with further syncope nevertheless prevented by phonotactic constraints.That is, the language requires a single uniform parsing algorithm across strata, making it metrically coherent.26

Learnability, grammar competition and diachronic change
In the preceding section, we saw that the most promising grammatical competitor to a metrically incoherent stratal account is one in which deletion is instead driven by avoidance of stress falling on light syllables (SWP).This can nearly account for the Southern Pomo pattern of thirdsyllable vowel deletion, but SWP is unable to produce light stressed syllables in attested forms like (8a) ʔàhkʰaptáka 'sea lion'.However, there are some forms where deletion of a vowel in the fourth syllable prevents a light stressed penult from arising, precisely as predicted by SWP -for example, (23d) dàkk'at̪ 'máʧʧ'in 'bring several here!'.
The exact historical trajectory of these patterns cannot be confirmed; there is not diachronic data to prove that Southern Pomo went through contactinduced change, rather than an endogenous stress restructuring due to syncope itself.In the latter case, the possibility is that, beginning with a leftaligned peninitial stress system (as in Proto Pomo), deletion might have rendered surface stress ambiguous (for threesyllable forms) between peninitial and a rightaligned, penultimate system.Learners might then have settled upon a penultimate interpretation.This seems to be unlikely, as the language features a large number of words that, due to enclisis, are quite long and have multiple stresses.There are also no shortage of disyllables (not remarked upon here because they never display vowelzero alternations), where stress would remain final.The majority of input would thus be unambiguous.For this reason, the proposal that contact with Bodega Miwok spurred the development of rightaligned phrasal stress is more plausible.The later development of a wordlevel SWP grammar therefore represents restructuring, flattening the metrically incoherent system.
If fourthsyllable syncope is due to restructuring of the grammar, what promoted this reanalysis?Prior work has suggested that there is a diachronic bias towards more transparent systems (e.g., Kiparsky 1968; Prickett 2019).This bias is hypothesised to arise from the relative ease of learning how processes (or constraints) interact.The difficulty associated with learning opaque interactions is at least partially due to the hidden structure problem: learners have access to underlying representations (mor phemes or concatenations of morphemes from the lexicon) and surface realisations, but do not see direct evidence of the intermediate representations involved in the map ping between them (Jarosz 2016).The learner must accumulate sufficient evidence to infer how the map works and rule out alternative hypotheses.Multilevel grammars in particular are only learnable given robust evidence for the stratal affiliation of a process (Nazarov & Pater 2017).
From the learner's perspective, both the surfaceoptimising and stratal analyses of syncope are possible analyses of the same data in almost all cases.The predictions of these analyses appear to diverge exclusively in forms with underlyingly light fourth syllables.The circumscribed set of forms in which the predictions of these models do not overlap is suggestive: because there is minimal evidence from the input to the learner that could clearly support one account over the other, learners are likely to converge on whichever generalisation is easier to extract.Thus, opaque patterns may with the gradualness condition.If we assume that HS is able to produce syncope in the appropriate position, it must still concede that there are strata because of phrasal stress shift.(Combined Stratal HS approaches have been proposed; see Calamaro 2017.)prompt restructuring by learners, who acquire a different constraint ranking -or induce ad hoc constraints -to transparently account for diachronic opacity in the input (e.g., Hayes 1999).Fourthsyllable syncope provides evidence that Southern Pomo was in the midst of such a restructuring.
We must note that the synchronic status of metrical opacity in Southern Pomo is clear, regardless of whether the language remained metrically incoherent.As demonstrated in §2.2.3, the loci of deletion and surface stress are independent.This means that any analysis of metrically conditioned vowel deletion -whether derivationally or outputoriented -must be able to generate surface structures where deletion is opaque.The locus of difference between the *V ̆and SWP grammars is at the word level, specifically in the relative ranking of SWP and MAXV.The wordlevel SWP grammar presented in §4.3 requires a phraselevel grammar to realign surface stress, and the phraselevel grammar presented in §3.3 accomplishes exactly this: despite nominally encouraging deletion (via promotion of *V ̆over MAXV), cluster restrictions prevent any further deletions from occurring.A diachronic shift towards fourthsyllable deletion can therefore be thought of as domain narrowing of syncope from the phrase to word level, in line with the theorised trajectory of the phonological lifecycle (BermúdezOtero & Trousdale 2012).
What, then, do learners hypothesise about the conditioning of syncope?We must ask whether and how they come to learn that this process is opaque.What assumptions do learners begin with?In this case, there are two possibilities: either the learner expects opacity, and finds it, or they expect transparency and are met with opacity.Let us suppose that the learner begins with the assumption that any given process is transparent until they have positive evidence to the contrary.This is a conservative assumption, and the opposite approach would encounter a familiar problem.Much like errordriven constraint demotion (Tesar & Smolensky 1993, 2000), starting with the most permissive option -an assumption that a derivation, and specifically stratal architecture, is necessary -will prevent the learner from finding positive evidence for a more restrictive system.In this way, the 'stratalfirst' learner is akin, by analogy, to a starting ranking of Faithfulness over Markedness.To successfully converge on a grammar, they must instead start from a more conservative system -one which is parallel until necessary.
The learner might then approach this syncope pattern in parallel, as above in §4.2, and will successfully derive syncope transparently in the great majority of cases by using SWP -except in words like [ʔàh.t̪ im._kó.ʧin], with stress on light penults.These forms exhibit deletion that is nonoptimising, and thus present the learner with counterevidence for the SWP hypothesis.If too few of these cases are encountered, however, the learner is then able to settle on this transparent derivation of syncope rather than posit a derivational metrical structure to condition it.It is important to note that the 'transparency' of this reanalysis is, in fact, still rendered opaque in the surface phonology when phrasal stress shift occurs -that is, in an estimated half of all tokens for any given form, whenever an odd number of syllables follow it within the phrasal domain.There is no evidence of an asymmetry in the syllable parity of words following syncopated forms, so we may assume that learners of Southern Pomo received roughly equal amounts of input with and without phrasal stress shift.That is, syncope could not be conditioned by surface stress in roughly half of learners' input.Allomorphs of the suffixes are conditioned by wordlevel metrical structure; however, this conditioning may be rendered opaque by phrasal stress.For any given allomorph, phrasal stress shift will ensure that there are two distinct surface realisations: (48) Phrasal stress interacts with syncope a. Preceding even parity in phrase i. [hà:.ʧat._lók_ʧ'a# σσ] ii.[hàʧ.ʧ_al.kó.ʧ'in # σσ] iii.[hà:ʧat̪ ._ló.koj # σσ] b.Preceding odd parity in phrase i. [ha:.ʧàt._lok_ʧ'à# σσσ] ii.[haʧ.ʧ_àl.ko.ʧ'ìn # σσσ] iii.[ha:ʧàt̪ ._lo.kòj # σσσ] Given this, the learner is inevitably faced with a ranking paradox if they attempt to proceed in parallel; two instantiations of the same lexical item would require different grammars to generate this pattern, in the forms where it is possible at all.The interaction of phrasal stress shift with deletion thus demonstrates that the metrical structure which conditions syncope cannot be built at the phrasal level, whether or not syncope itself is a phraselevel process.This is supported by the data: speakers did not demonstrate transparent alternations in the location of syncope based on phrasal surface stress, regardless of whether the word produced had deletion in the third or fourth syllable (Walker, p.c.).The issue is then whether syncope is computed subsequent to the building of this wordlevel structure, or in parallel with it.
In addition to providing converging evidence for the stratal account, the behaviour of the clitic field and the boundary between wordlevel stems and enclitics (described in §2.2.2) is evidence for the learner about opaque structure.Domainfinal phonolog ical alternations demonstrate the existence of these domains.For example, the stem final ejectivisation process shows alternations when there is following wordinternal material, but not when the following material is enclitic -showing phonological operations applying to the word level, exclusive of enclitics.In contrast to this, the participation of enclitics in phraselevel stress assignment and compensatory lengthening demonstrates the existence of a postlexical stratum.Thus, the learner receives a significant amount of evidence that multiple levels of phonology are active in the synchronic grammar.
However, with respect to the learnability of the interactions between these strata that result in opaque syncope, the data are largely equivocal or even biasing towards a 'simplified' grammar in which syncope is driven by surface optimisation and is not influenced by positional faithfulness.The evidence in the input comes only from four syllable words of the form [σσσ L σ] that also have related forms in which an additional vowel appears.In one of the few transcribed Southern Pomo narratives, Rock Man and Grey Squirrel (Walker 2020: 345-361), comprising 71 sentences, there are only eight [σσσ L σ] forms, none of which have intraparadigmatic vowel∅ alternations.Such words likely did not comprise a large portion of the input.When these words do occur, deletion is dictated by wordlevel faithfulness constraints.This patterning could lead the learner to suppose that syncope is a wordlevel process, for lack of sufficient positive evidence to the contrary.The clitic field meanwhile features exclusively heavy syllables, and thus heavy stressed syllables when those bear stress, satisfying SWP while providing no evidence in support of the stratal grammar.The paucity of evidence is not the only challenge for the learner: even when provided with surface alternations, learners may fail to extract the necessary grammatical generalisation (Morley 2018; Bowers 2019).This is thus an instantiation of the credit problem (Stanton 2016): most of the Southern Pomo surface forms are equally consistent with deletion driven by *V ̆or by SWP, such that both markedness constraints are promoted.SWP is just as good a fit for the majority of the data.Stanton (2016) demonstrates that this problem prevents artificial learners from acquiring typologically predicted grammars despite being provided with the necessary constraints in CON.Limited learnability, rather than nongenerability, is sufficient to curtail the observed typology of stress.Gordon (2002) suggests that many unattested metrical patterns should be considered accidental gaps rather representing the bounds of the grammar itself, and these gaps have been attributed to learnability (Heinz 2009) and its interplay with patterns of phonologisa tion (Hyman 1977).The evidence for stratal opacity in Southern Pomo was limited, with most of the vowel-∅ alternations in the input being interpretable as a transparent deletion process driven by surfaceoptimising constraints (specifically SWP).It is likely that this led to reinterpretation by learners acquiring the language, resulting in innovative forms (Hayes 1999).
We must therefore ask whether the stratal vowel deletion pattern was a syn chronically active process at the time of observation.The alternative is that this process was fully superseded by monostratal, surfaceoriented deletion, and that the vestiges of the older pattern are simply memorised exceptions (see, e.g., discussion of Mojeño Trinitario in Rose 2019: 21).This interpretation is, to a degree, a mirror of the discussion about learnability; if speakers encounter very few forms which can disambiguate between these two grammars, the forms contrary to their hypothesis might simply be stored as exceptions.Furthermore, some of the forms which might disambiguate are noun-noun compounds with idiosyncratic interpretations, for exam ple, (8a) ʔàhkaptáka 'sea lion' (lit.'water bear').This idiosyncrasy suggests the forms must be memorised rather than generated on the fly.On the other hand, there is not clear evidence to suggest that all thirdsyllable forms are listed irregulars.Many forms displaying the thirdsyllable syncope pattern are lowlikelihood collocations with highly specific, compositional meanings.(5c) hà:ʧat̪ lókoj '[they] fly out [of something]' and (10b) ʔàht̪ imkóʧin 'put foot back!' are unlikely to have been high frequency forms, which militates against the argument that these were memorised.We see this in verbal reduplication, as well, where thirdsyllable syncope forms can be transparently compositional -for example, (6d) t'uʔbut'bulaw 'to run down[wards] (bent way over)'.These metrically opaque forms appear to have equally compositional interpretations when compared to the fourthsyllable forms.This seems to be strong evidence that the thirdsyllable syncope process, driven by stratal interaction, was still synchronically active.While an appeal to memorised irregulars is possible, it is not clearly motivated by these data.
The next question, then, is how both third and fourthsyllable deletion can be syn chronically real processes in Southern Pomo. §2.3 demonstrates that the distribution of these forms cannot be tied to particular morphemes, roots or lexical categories, which seems to rule out analyses relying on indexed constraints or cophonologies.This leaves grammar competition (Kroch 1989; Anttila & Cho 1998; Fruehwald et al. 2009) as the most plausible explanation.Specifically, there appears to have been a degree of stochasticity in the relative ranking (or weighting) of SWP (propelling surface optimising deletion) and MAXV (preventing said deletion) in the wordlevel grammar.
Figure 3 shows Hasse diagrams for these partial rankings.
The two partial rankings are only appreciably different in their effect on underlying fivesyllable strings.We might think of this as instantiating the credit problem at the level of partial rankings, rather than individual constraints.A grammatical computation would be necessary only when encountering a novel string of this typeit is reasonable to imagine that, given the small number of such forms, these could then be stored, explaining the lack of variation in the site of deletion for any single string.
Altogether, the picture emerges that this metrically opaque grammar was unstable.The phonological component is capable of representing complex stratal derivations and does not impose restrictions on how strata may differ with respect to metrical structure.That is, the mechanisms that constrain the typology and frequency of these processes are not grammarinternal: typologically rare patterns like metrical opacityand perhaps other unusual patterns predicted by the ranking independence of strataare within the generative capacities of phonology (de Lacy 2006; de Lacy & Kingston 2013).Learners, however, are limited by the quality and consistency of the input data they receive, and the ability of these data to (dis)confirm hypotheses about the grammar.Successful acquisition of stratal syncope requires consistent evidence which differentiates this pattern from one in which syncope is a surfaceoriented process.It is apparently the case that learners received too little input of this sort, and instead settled on the grammar which maximises transparency and metrical coherence.
Another point to consider is the ontogenesis of these patterns.If indeed the Southern Pomo stress shift originated in a context of MiwokPomo bilingualism as Buckley (2019) suggests, this is an 'unnatural' change (in the terminology of evolutionary phonology) that emerged from language contact rather than phonetic principles (Blevins 2004; Hansson 2007).One might theorise that it is only such unnatural processes that can lead to metrical incoherence in sound change, and that languages otherwise hold to the principle of metrical coherence in derivational layering.Under this interpretation, language contact may give rise to unnatural changes that push the limits of the phonological system (Hansson 2007(Hansson , 2008)).This would further constrain the likelihood of such systems being widely documented in the phonological typology.
Southern Pomo demonstrates that metrical incoherence must be representable and transmissible in the synchronic grammar, though we are left with open questions about what diachronic pathways allow for patterns of this nature to arise.
The fact that the opaque pattern arose at all, as well as Southern Pomo's pattern of phrasal stress shift, are evidence that the phonological component must be able to generate strata and learn stratal patterns.Further, it demonstrates that these strata have a degree of independence with respect to their constraint rankings (or component rules) which goes beyond the promotion of faithfulness constraints (cf. KoontzGarboden 2001).With respect to the existence of metrical incoherence across derivational stages, this case presents strong evidence that such patterns are both real and, at least to some degree, transmissible.However, metrically opaque processes are likely to be unstable if there are possible transparent analyses with a high degree of empirical coverage -even if these are imperfect.Syncope in Southern Pomo thus demonstrates that the typology of these patterns is limited by grammarexternal factors, in particular their learnability.

Conclusion
In this article, I have shown that Southern Pomo demonstrates a pattern of rhythmic vowel deletion which is not predictable from the surface metrical structure.Syncope and stress assignment are evaluated over different rhythmic structures, with syncope aligned with the left edge, and stress with the right edge.This poses a significant problem for both parallelist and nonstratal derivational frameworks.I analyse this pattern in a SOT framework involving two strata: the word level defines the weak positions targeted for deletion, while the phrasal stratum enacts both deletion and reassignment of prominence.Independent evidence for the proposed strata is found in phrasal stress shift and several nonrhythmic phrasal processes, as well as the diachronic development of this pattern (Buckley 2019).While stratal architectures have been criticised for countertypological overgeneration, this work suggests that such overgeneration is necessary to analyse metrical opacity, in particular given the contact context in which this pattern arose.Southern Pomo also provides evidence that these countertypological systems may be unstable due to extragrammatical factors.The language features a pattern of fourthsyllable syncope in a small set of forms, which appears to show grammatical restructuring as the result of limited learnability and ambiguous input.Based on the data presented, I conclude that grammar competition between this and the older system resulted in inconsistencies where they conflicted.The full picture suggests that the phonological component must be capable of generating complex stratal interactions that result in opacity, even in the syntagmatic relations of hierarchical metrical structure.It is the limits of learnability, rather than of the grammar itself, which constrain the attested patterns and result in systematic asymmetries in language typology.

( 36 )
Final vowels are not deleted [ts'íh.ta]ω [min.náːn.t̪i] ω → [ts'ih.tà# min.náːn.t̪i] φ 'trapping birds' (W13: 487)This suggests the activity of an independent constraint which protects positions at word edges from deletion.To this end, I propose an Anchor constraint, ANCHOR RIGHTEDGEWORD, which assigns violations for deletion of vowels occurring word finally in the input:(37) ANCHORREDGEWD (adapted fromMcCarthy & Prince 1995) Assign one violation for each wordfinal vowel in the input that has no correspondent in the output.

Table 2 .
Syncope is irregular relative to stress.

)
Rulebased derivation with hiatus resolution preceding syncope ) demonstrate this ranking using attested Southern Pomo forms.
Syncope is insensitive to stress shift Syncope increases harmony for SWP (evenparity) σ H .cv 2 .cv 3 .σSWP MAXV 45) Surface optimising and nonoptimising syncope in /σσσσ L σ/ strings ) show wordlevel Stratum Ib, and the one in (47) shows a sample derivation of phrasal stress (Stratum IIb).*CLASH, *LAPSE and phonotactic constraints are again assumed to be undominated and are not shown.