GS is an extinct Old Štokavian dialect of Bosnian/Croatian/Montenegrin/Serbian (BCMS) (see Ivić 1958 for the background on BCMS dialects). The dialect was historically spoken by the Gallipoli Serbs, a Serbian community in Bayramiç, a town near the Gallipoli Peninsula in Turkey. The Serbian population of this area is believed to have been forcibly relocated from Central Serbia in the 16th century (Filipović 1946; Ivić 1957). During the Greco-Turkish War of 1922, GS speakers migrated to Pehčevo in Eastern Macedonia, where the dialect was extensively documented in the early 1950s. The dialect became extinct soon after. The Serbian population in Pehčevo has been recently reported to have switched to Standard BCMS (Pavlović 2018).

The GS data used in this paper come from a comprehensive descriptive grammar of the dialect by Pavle Ivić (Ivić 1957). Ivić collected the data during two field trips to Pehčevo in the early 1950s, primarily from older individuals who had lived in Bayramiç before the migration. Younger generations, born in Pehčevo after the relocation, were reportedly not fully fluent in GS during Ivić’s field trips due to exposure to Standard BCMS and Macedonian. The data were obtained from two sources: narration (legends and folk stories) and spontaneous dialogues between GS speakers, recorded by Ivić. No data were collected through targeted elicitation.

Ivić transcribed and annotated the recorded data with pitch accent markings. Some of these transcribed texts, with accentual information, were published as an appendix to Ivić’s grammar (Ivić 1957: 439–465). This prosodically annotated corpus is the primary source of data for this study.

The consonant inventory of GS is presented in Table 1.

Although the consonant inventory of GS appears to only minimally diverge from that of Standard BCMS (on which see Morén 2006), the GS consonant system has undergone substantial changes. The voiceless velar fricative [x] and the palatal lateral approximant [ʎ], both part of the Standard BCMS consonant inventory, were lost historically in GS. These segments were later reintroduced through contact with Greek and are restricted to Greek loanwords in Ivić (1957)’s data. Under the influence of neighboring Greek dialects, GS neutralized the contrast between dental/alveolar and postalveolar fricatives and affricates, which led to the elimination of the postalveolar series ([ʃ], [ʒ], [ʧ], [ʤ]). Following the community’s migration to Macedonia and subsequent exposure to Standard BCMS and Macedonian, some speakers began to partially restore this contrast. This partial restoration led to variation in the realization of coronal fricatives and affricates in the dialect (Ivić 1957: 131–134).

GS features seven vowel segments and the syllabic sonorant [r̩]. The dialect’s vowel inventory is presented in Figure 1.

The central vowel [ə] occurs exclusively in Turkish loanwords (Ivić 1957: 76–77). GS differs from Standard BCMS in that it contrasts two mid front vowels: [e] and [ɛ]. Short mid vowels /ɛ/, /e/, and /ɔ/ optionally raise to their high counterparts [i], [u] in unstressed syllables (3).1

All vowels, along with [r̩], can be short and long. Vowel length is contrastive in stressed and post-tonic syllables, as well as in syllables immediately preceding stress. In pre-tonic syllables not adjacent to stress, only short vowels are allowed.

GS is a pitch accent variety. Following Inkelas & Zec (1988); Zec (1999); Zec & Zsiga (2010), GS pitch accents represent combinations of two prosodic features: stress and tone. The dialect distinguishes between two types of accented vowels: falling-accented vowels, which can be short ([ȁ]) and long ([ȃ]), and rising-accented vowels ([á]), which are invariably long. BCMS falling accents feature a pitch peak (i.e. High tone) at the onset of the stressed vowel, while rising-accented vowels are characterized by a pitch plateau, followed by a High tone on the post-tonic syllable (Zsiga & Zec 2013; Zec & Zsiga 2022).

Given that GS is an extinct dialect, instrumental data for its pitch accents are unavailable. As the best approximation, data were collected from another Old Štokavian variety—the Kosovo-Resava dialect (BCMS kosovsko-resavski), one of GS’s closest relatives. The Kosovo-Resava dialect shares the same word-prosodic inventory as GS, with short falling, long falling, and rising accents, and exhibits identical distributional restrictions: The long falling accent is unrestricted, the short falling accent is limited to non-final syllables, whereas the rising accent is restricted to the penult (Ivić 1958; 1985).

Figure 2 displays representative pitch tracks for the three pitch accents in Old Štokavian.2 The data show clear F0 peaks (i.e. High tones) on the stressed syllable in the falling accents, and on the post-tonic syllable in the rising accent. Thus, the acoustic realization of the falling-rising contrast in Old Štokavian parallels the Neoštokavian dialects of BCMS, which have been studied more extensively (Lehiste & Ivić 1986; Zsiga & Zec 2013).

Ivić (1957) uses traditional BCMS pitch accent notation to mark stress and tone. Throughout this paper, I adopt a more reader-friendly IPA-based notation in which stress and tone are rendered as separate entities. I mark High tone and leave all non-High-toned moras tonally unmarked. Example (4) shows how Ivić (1957)’s pitch accent notation translates into IPA.

In addition to pitch accent (i.e. stress and tone) information, Ivić’s prosodic annotations indicate which accent-bearing words clitics adjoin to. Ivić consistently used the undertie symbol (⌣) to link clitic forms to their hosts (5). This information is particularly important because clitics play a central role in the present study.

GS is a restricted tone system (in the sense of Hyman 2006) characterized by tonal privativity, culminativity, and obligatoriness (6).

In line with (6), configurations (7a)–(7b) are well-formed in GS. By contrast, configurations that lack High tone (7c) or display a High spanning over multiple moras (7d) constitute impermissible surface representations in the dialect.

High tone is restricted to the first mora of a syllable (8a)–(8b). Forms with a High on the second mora of a heavy syllable are prohibited (8c), as in most other Štokavian dialects of BCMS (Ivić 1958; Inkelas & Zec 1988).

GS exhibits unbounded tone-driven stress. High tone consistently attracts stress, irrespective of syllable weight (9). This is particularly evident in (9c)–(9d), where stress falls on light High-toned syllables, leaving toneless heavies unstressed.

The only exception from (9) stems from nonfinality: Stress cannot be assigned to final light syllables in GS. If a final light is High-toned, stress falls on the toneless penult (10a)–(10b). In contrast to final lights, High-toned final heavies regularly receive stress (9f)–(9h). Both CV (10) and CVC (11) High-toned ultimas are unstressable, which indicates that final consonants do not contribute weight in the dialect.

Whereas final High-toned lights are unstressable in (10)–(11), they regularly receive stress in enclisis (12). This results in a productive alternation between penultimate and final stress.

When stressed, underlyingly light toneless penults undergo vowel lengthening (13).4 The underlying quantity of lengthened vowels is preserved in unstressed position. Unstressed heavies do not shorten in GS (9c)–(9d), (9h)–(9i), which indicates that the alternation in (13)–(14) results from lengthening in stressed position rather than shortening in unstressed position.

Additional evidence for the productivity of nonfinality and penultimate lengthening comes from prosodic adaptation of Turkish loanwords. Word-final prominence in the donor language has been reinterpreted as a lexical High in GS. As in the native lexicon, final High-toned lights in Turkish loanwords are unstressable, repelling stress to the penultimate syllable (15). If the penult is underlyingly light, as (16) shows for the items in (15), it undergoes vowel lengthening.

In (17)–(18), stress and High tone alternate between the word-initial syllable and the proclitic. Forms that exhibit this alternation are underlyingly toneless and receive initial prominence by default (Inkelas & Zec 1988; Zec & Zsiga 2010).

By contrast, forms with an initial lexical High do not display stress alternation in proclisis (19)–(20).

Both proclitics and enclitics are part of the stress domain in GS, given that Stress Assignment is sensitive to their presence (12)–(18). This indicates that in GS, the stress domain comprises the morphological word and the clitics hosted by it, i.e. the Clitic Group.

In sum, GS assigns stress to the High-toned syllable unless the High-toned syllable is light and final. In the absence of a lexical High, stress defaults to the initial syllable of the stress domain. The GS stress rule is summarized in (21).

Long vowels shorten word-finally in GS (Ivić 1957: 33–35). The process is defined using SPE-style rule formalism in (22).

Final Shortening is observed in both monosyllabic (23a)–(23d) and polysyllabic forms (23e)–(23f).

The original quantity of shortened final vowels is preserved in word-internal position, as well as in enclisis (24). The latter fact suggests that enclitics are part of the domain of application of Final Shortening.

GS assigns stress to the word’s only High-toned syllable (9), except when the High-toned syllable is light and final, in which case stress falls on the penult (10)–(15). In addition, long vowels regularly shorten word-finally (23). In forms with an underlyingly long final vowel that bears a High, Stress Assignment and Final Shortening crucially interact.5

Final Shortening renders the ban on stressed final lights non-surface-true: High-toned final syllables that become light by the shortening of originally long final vowels bear stress on the surface. Consider the forms in (25), which display final rather than penultimate stress. This runs afoul of the transparent pattern in (10)–(11). Final lights brought about by Final Shortening are therefore treated differently from originally light High-toned ultimas, which are barred from receiving stress.

That the final vowels in (25) are underlyingly long is indicated by their realization before enclitics (26).

Particularly instructive is the existence of minimal pairs that differ solely in the length of their final vowels at the underlying level. Since the contrast in vowel length is neutralized in final position, these items are distinguished on the surface only by stress position. Forms with an originally light final syllable (27a)–(28a) are subject to the ban on final prominence, displaying penultimate stress. By contrast, items that undergo Final Shortening (27b)–(28b) exhibit final stress, defying the prohibition against stressed final lights. If stress were assigned transparently in the latter group, the underlying contrast between the two sets of forms in (27)–(28) would be lost.

The original length of the final vowels in (27b)–(28b) is revealed in enclisis (29).

The opaque interaction between Stress Assignment and Final Shortening lends itself to a straightforward rule-based analysis (30), which makes crucial use of extrinsic process ordering. On this account, Stress Assignment applies before Final Shortening. This ordering explains why the context-changing effect of Final Shortening is invisible to stress: Final Shortening applies too late in the derivation to affect the application of Stress Assignment.

The interaction in (30) is a showcase of countershifting, i.e. misapplication opacity (Rasin 2022; Pruitt 2023; Baković & Blumenfeld 2024). Countershifting occurs when a later-ordered process 𝔹 changes the environment of an earlier-ordered process 𝔸. Crucially, the context-changing effect of 𝔹 happens too late to affect 𝔸’s application. Instead, 𝔹 renders the previously transparent environment for 𝔸 opaque. Shifting interactions and countershifting opacity differ from more familiar interaction types, such as (counter)bleeding and (counter)feeding, in that 𝔹 neither prevents 𝔸 from applying nor enables its application. Rather, 𝔹 only modifies the environment of 𝔸. Therefore, what is at stake in (counter)shifting is not whether or not 𝔸 applies, but rather how it applies.

In GS, Final Shortening (𝔹) countershifts Stress Assignment (𝔸). In forms like /ʒɛ.nɛ́ɛ/ (30), stress is initially assigned to the final High-toned heavy, in line with the dialect’s stress rule: Final heavies with a High-toned first mora regularly receive stress (9f)–(9h). The subsequent application of Final Shortening opacifies this hitherto transparent stress environment, causing stress to occur in a suboptimal environment—on a final light syllable—where it is otherwise avoided. Reassigning stress to the penult to eliminate this marked configuration is not possible, since Stress Assignment has already ceased to apply at the time Final Shortening applies.

In this section, I present an OT analysis of the data in Section 3, identifying the active constraints and establishing their ranking.

Stress-to-tone mapping falls out from Head-H (Yip 2001), the markedness constraint requiring that foot-heading moras bear a High.

Default initial stress follows from All-Feet-Left (Hyde 2012), which assigns a violation for every syllable intervening between the foot-heading syllable and the Prosodic Word’s left edge. High insertion in the initial syllable of underlyingly toneless forms satisfies Head-H at the expense of Dep-H, which requires that every High tone in the output have an input correspondent (32).

In non-default cases, High-toned syllables receive stress regardless of their distance relative to the Prosodic Word’s left edge. This shows that Head-H outranks All-Feet-Left (33).

Head-H also outranks Weight-to-stress Principle (WSP; Prince 1990), defined in (34). This is demonstrated in (35): Compliance with Head-H leads to the violation of WSP.

WSP is likewise dominated by All-Feet-Left. The two constraints clash in (36), where All-Feet-Left is satisfied at the expense of WSP.

The moraic version of Nonfinality (Hyde 2007) explains the avoidance of final prominence. This constraint, formally defined in (37), penalizes foot-level grid marks on the rightmost mora of the Prosodic Word. Specifying final moras rather than final syllables as the locus of Nonfinality violation correctly ensures that only stressed final lights, but not stressed final heavies, are prohibited in GS.

In GS, Nonfinality dominates Head-H. This follows from cases like (10)–(15), where stress falls on the toneless penult to avoid the violation of Nonfinality (38).

Putting the pieces together, GS stress rule emerges from the constraint hierarchy in (39).

Final Shortening is triggered by the markedness constraint *Vː#, which penalizes long vowels in final position. *Vː# outranks the antagonistic faithfulness constraint Max-μ, which militates against vowel shortening (40).

The independently motivated rankings in (39)–(40) fail to capture opaque stress in Final Shortening environments. The current constraint hierarchy incorrectly selects penultimate stress as optimum for inputs with shortened final vowels. This is because the transparent contender fares better on top-ranking Nonfinality than the intended opaque winner. The failed OT account is presented in (41).

For the intended winner with final prominence to prevail against its transparent contender, Head-H has to dominate Nonfinality. This is at variance with (38). The two patterns are therefore irreconcilable, requiring conflicting rankings of Nonfinality and Head-H.

The upshot of the rule-based analysis, summarized in (42), is that Stress Assignment must apply before Final Shortening. This ensures that the context-changing effect of Final Shortening is invisible to Stress Assignment, which results in opaque final stress in forms like (10)–(15).

RBP models opacity using extrinsic process ordering, stipulated in the phonological grammar. RBP therefore offers a unified account of opaque process interactions, deriving all instances of this phenomenon from a purely phonological device—extrinsic process ordering, although the picture may be more complicated (see Baković 2007; 2011).

By contrast, Stratal OT (Kiparsky 2000; 2015; Bermúdez-Otero 1999; 2003) directly links phonological opacity to morphosyntactic structure. This multi-level derivational framework retains the parallel architecture of classical OT but assumes that phonology operates across multiple serially ordered domains—strata. Each stratum constitutes a parallel OT grammar with its own constraint ranking. Most versions of Stratal OT posit three strata in the phonological grammar: the Stem, Word, and Phrase strata.

The multi-level architecture enables Stratal OT to replicate RBP’s analysis of opacity without recourse to extrinsic ordering. The output of each non-terminal stratum is fed back to the phonological grammar as the input to the next stratum. This cyclic mechanism ensures that opaquely interacting processes are crucially ordered by virtue of their stratal affiliation.

A key property of Stratal OT (at least some versions of it) is the principle of stratum-internal transparency (Bermúdez-Otero 2003; Kiparsky 2000; 2015), which holds that input-output mappings are transparent within each stratum (43).

Each stratum of phonology, being a parallel OT grammar, selects a transparent winner as its optimal output. However, the transparent output of a non-terminal stratum can become opaque at later strata. For example, a word-level process applies before, and can consequently be opacified by, a postlexical process. Between-stratum ordering obviates the need for RBP-style extrinsic process ordering.

Some proponents of Stratal OT maintain the strict version of (43), according to which opacity only arises between processes that apply at different levels. As a corollary, between-stratum process ordering is argued to be the only possible source of phonological opacity (Kiparsky 2015). This view has been challenged by more recent work in Stratal OT, which adopts a less restrictive perspective on the sources of opacity (Bermúdez-Otero 2019).

Stratal OT sidesteps the ranking paradox encountered by the failed parallel OT account in (41), where no self-consistent constraint grammar could be found which outputs both transparent and opaque stress in GS. In Stratal OT, Stress Assignment and Final Shortening need not apply in the same stratum of GS phonology. Both processes can be transparent in their respective strata, with opacity emerging from their stratal affiliation.

For Stress Assignment to be transparent in the output of its stratum, Final Shortening must be inhibited at the level at which stress is assigned, which I tentatively label Stratum 1. Thus, the shortening-inducing markedness constraint *Vː# ranks below Max-μ in Stratum 1 (44). Top-ranking Max-μ knocks out both candidates with shortened final vowels in (44). Stress falls on the final heavy, consistent with (9f)–(9h).

Subsequently, in Stratum 2, *Vː# gets promoted over Max-μ, which gives rise to Final Shortening. Opaque stress could be avoided by stress retraction. However, the winner of Stratum 2 retains the stress position established in Stratum 1, in compliance with the NoFlop constraint (Alderete 1999):

NoFlop_stress dominates Nonfinality, which inhibits stress retraction. Constraint grammar (46) selects the opaque form as the winner.

Taking stock, the gist of the Stratal OT analysis of GS opaque stress is that Stress Assignment and Final Shortening must be assigned to separate stratal domains. In the version of Stratal OT that assumes universal transparency within strata, opacity is not attributed to a purely phonological mechanism, being treated as an epiphenomenon of phonology–morphosyntax interleaving. This approach makes a strong prediction about opacity: Stratum-internal opaque interactions should not exist (Kiparsky 2015). Whereas Kiparsky (2015) demonstrates that many oft-quoted cases of within-stratum opacity actually involve between-stratum processes interactions, ample evidence has emerged for the existence of within-stratum opacity (Kavitskaya & Staroverov 2010; Broś 2016; Broś & Nazarov 2023; Stanton 2023; Obiri-Yeboah & Rasin 2025). The remainder of this paper tests whether stratum-internal transparency in its strongest form withstands the empirical challenge posed by countershifting opacity in GS.

In Section 3.1, it has been established that both proclitics and enclitics affect the position of stress. Two pieces of evidence support this: (i) proclitics bear stress when the host lacks a High-toned mora (18), and (ii) enclitics trigger stress alternations in words with final High-toned light syllables (10)–(12).

The stress domain in GS corresponds to the Clitic Group. However, not all clitic forms affect the position of stress. A subset of pronominal enclitics resists shifting stress onto the host’s final syllable, defying the pattern in (12):

In (47), stress falls on the toneless penult in enclisis, contrary to (12). This discrepancy follows from the fact that GS distinguishes between inherently clitic forms and forms that become prosodically deficient in the course of derivation. This distinction sheds further light on process ordering in GS phonology.

GS distinguishes between strong and weak personal pronouns. Strong pronouns are minimally bimoraic and contain a lexical High, thereby behaving like content words: They form a Prosodic Word and receive stress. Meanwhile, weak pronouns are monomoraic and lack a High. Due to the prosodic minimality constraints in BCMS (Zec 1999), weak pronouns cannot form permissible Prosodic Words and receive stress, thereby being enclitic.

All case forms of personal pronouns have strong variants. Only the genitive, dative and accusative forms also have weak variants (Ivić 1957: 199–205) (Table 2).6

Strong pronouns with an underlyingly long final vowel surface faithfully in enclisis (48a) and otherwise undergo Final Shortening (48b).

Looping back to (47), note that the 1sg.nom personal pronoun does not have a weak variant. However, in (47), this form functions as an enclitic. Unlike inherently clitic forms (12), the enclitic in (47) is stress-neutral. This distinction suggests that GS treats “real”, i.e. inherently weak pronouns, differently from “contextual”, or derived clitics.

Derived clitics (47) include pronouns that are originally strong but become prosodically deficient by an optional deaccentuation process. Strong pronouns can optionally lose their pitch accent (stress and tone), under circumstances that are beyond the scope of this paper. Without stress and tone, the defining attributes of prosodic wordhood in GS, deaccented pronouns cannot function as standalone Prosodic Words. Consequently, they must cliticize to a neighboring accent-bearing word.

Derived clitics can be either proclitic or enclitic. Encliticization is the default strategy, whereas procliticization occurs only when a deaccented pronoun is CP-initial because no clausemate host is available to its left (49).

Example (50) shows a clear prosodic difference between inherent and derived clitics. Inherently clitic (i.e. weak) pronouns are visible to stress: They enable word-final High-toned lights to bear stress, as shown in (12). Contrastingly, derived clitics (i.e. deaccented strong pronouns) do not affect the position of stress in their hosts (50).

Regular cliticization, which applies to inherent clitics, precedes Stress Assignment. This explains why inherent clitics are part of the stress domain. Secondary cliticization, which targets deaccented strong pronouns, applies after Stress Assignment. This leaves derived clitics outside the stress domain. This ordering accounts for the stress neutrality of derived clitics.

Secondary cliticization is relevant to the analysis of opaque stress because it provides evidence for the ordering of Stress Assignment relative to postlexical processes. Although the exact conditions for the deaccentuation of strong pronouns remain unclear, the process is evidently linked to information structure and sentence-level prosody, which suggests that deaccentuation is postlexical. Deaccentuation feeds secondary cliticization, the output of which is inaccessible to Stress Assignment because stress has already been assigned in an earlier phonological domain.

The stress domain encompasses the morphological word and any inherent clitics it hosts, i.e. the Clitic Group. Secondary clitics are incorporated into a larger structure, thereby being external to the stress domain. The different behavior of inherent and derived clitics therefore stems from their adjunction height. Stress-shifting, inherent clitics are word-level clitics, adjoining to the Prosodic Word domain (51a). Stress-neutral, derived clitics are phrase-level clitics which attach to the Prosodic Phrase (φ) node (51b). Phrase-level clitics thus constitute free clitics (Selkirk 1996; Anderson 2005).

Inherent clitics can be analyzed in two ways. Under the internal clitic analysis (52a), inherent clitics attach directly to the ω-node projected by their host and are therefore internal to the lowest ω-projection. Alternatively, under the affixal clitic analysis, clitic-host combinations form a recursive ω-structure (52b).

The data presented thus far are indeterminate between (52a) and (52b). I present evidence in favor of (52b) in Section 5.3.1.

In sum, inherently clitic forms in GS differ from forms that become prosodically deficient in the course of derivation. Inherent clitics undergo word-level cliticization, which precedes Stress Assignment, and are therefore visible to stress. Derived clitics undergo phrase-level cliticization, which takes place after Stress Assignment, and are therefore invisible to stress. Crucially, this distinction provides evidence for the ordering of Stress Assignment relative to postlexical processes: It shows that Stress Assignment is pre-postlexical, given its “blindness” to context-changing effects in postlexical phonology.

Final Shortening is sensitive to regular, word-level clitics, as evidenced by (24). That is, word-level enclitics are part of the domain of Final Shortening.7

The available evidence suggests that Final Shortening is word-bound. Word-final long vowels shorten phrase-internally in (53). The data in (54) illustrate faithful realization in enclisis in similar phrasal environments.

Conspicuously, Final Shortening applies to the same domain as Stress Assignment: the Clitic Group, i.e. the topmost projection of the ω-domain (55) (the Final Shortening domain is delimited by dashed rectangles).

Consistent with this finding is the fact that Final Shortening interacts opaquely with a postlexical consonant deletion process. In GS, stops and fricatives delete before a homorganic obstruent, modulo fricative-stop clusters such as [st] (Ivić 1957: 143). Homorganic Cluster Simplification (56) applies at multiple phonological levels: word-internally, i.e. at morpheme boundaries (57), at clitic-host boundaries (58), and across word boundaries (59).8

Crucially, postlexical Homorganic Cluster Simplification counterfeeds Final Shortening. Homorganic Cluster Simplification introduces new instances of final long vowels, which, unlike original final long vowels (23), resist Final Shortening (60).

The underapplication of Final Shortening in (60) cannot be attributed to prosodic minimality. Although the relevant items in (60) happen to be monosyllabic, recall that monosyllables regularly undergo Final Shortening and become monomoraic (23a)–(23d).

Homorganic Cluster Simplification is the only source of final long vowels in surface forms in GS. All other final long vowels shorten without exception. The fact that Final Shortening cannot apply to final long vowels introduced by a postlexical consonant deletion process suggests that the latest instances of Final Shortening must occur before postlexical phonology.9

Stress Assignment and Final Shortening apply before postlexical phonology, because both processes are insensitive to changes in their environments at the postlexical level. At the same time, the two processes apply before regular word-level processes like Final Devoicing, given their sensitivity to clitics.

One possible way of capturing this intermediate position of Stress Assignment and Final Shortening is by positing an independent Clitic domain situated between the Word and Phrase strata (in the spirit of Rubach 2011; 2016; 2019). However, the available evidence suggests that both Stress Assignment and Final Shortening apply in the Clitic stratum. Thus, within-stratum opacity persists even under this more flexible four-level architecture.

The representational alternative à la Bermúdez-Otero & Luís (2009) is likewise unable to eliminate within-stratum opacity, but for a different reason: It shifts stratum-internal opacity to the postlexical stratum. Ultimately, no extension of the standard Stratal OT framework salvages the universality of stratum-internal transparency.

What does the existence of within-stratum opacity in GS phonology show about the typology of opaque interactions and approaches to opacity within OT?

Most directly, it challenges the strong version of the stratum-internal transparency principle adopted in some versions of Stratal OT. Approaches to within-stratum opacity in Stratal OT fall into two camps. The strict stratum-internal transparency approach rejects the existence of within-stratum opacity altogether (Kiparsky 2015): All instances of phonological opacity result from between-stratum process ordering and are therefore morphosyntactically confined. No cases of opacity arise from a purely phonological device, such as constraint interaction within a single stratum or enriched representations. A more relaxed approach acknowledges the existence of within-stratum opacity but maintains that, as long as such interactions do not involve ordering paradoxes, they do not undermine Stratal OT’s cyclic, multi-level architecture (Bermúdez-Otero 2019). This approach allows within-stratum opacity to be captured in OT through additional, non-default mechanisms.

This paper supports the position that in its strong form, the stratum-internal transparency principle is overly restrictive and empirically untenable. Within-stratum opacity exists and is likely more common than traditionally recognized. While the existence of stratum-internal opaque interactions does not fundamentally challenge the main postulate of Stratal OT, namely cyclic organization of phonology, it does show that Stratal OT’s built-in machinery is insufficient for capturing the full spectrum of opaque interactions.

More broadly, these findings suggest that phonological opacity arises from multiple sources and is not universally a by-product of phonology-morphosyntax interleaving. Accordingly, it is unlikely that a single mechanism can capture the full range of opaque interactions. Many OT-based approaches pursue a unified analysis of opacity (Kager 1999; McCarthy 1999; 2007; Goldrick 2000; Jarosz 2014, a.o.). However, opacity appears to be too diverse a phenomenon to be amenable to a unified account (Itô & Mester 2003; Baković 2007; 2011; Bermúdez-Otero 2019; Pruitt 2023; Nazarov 2025).

HS is another derivational variant of OT. Rather than partitioning the phonological grammar into multiple serially ordered domains, as done in Stratal OT, HS maintains a single self-consistent constraint hierarchy throughout the derivation. HS produces stepwise derivations because its Gen is limited to emitting candidates that undergo at most one structural change (McCarthy 2000) (70). What counts as a permissible structural change in a single HS step is nonetheless a topic of ongoing debate. Structure-changing operations in HS are therefore introduced incrementally, one at a time.

At each step, the winner is the most harmonically improving form out of the restricted candidate pool. The order of operations is determined by the ranking of process-inducing markedness constraints: The change that satisfies the highest-ranked constraint takes precedence over other changes. Once determined, the winner loops back to Gen and Eval as the input to next step. The derivation terminates when no further harmonic improvement is possible, that is, when the faithful candidate wins.

HS successfully models countershifting opacity (Rasin 2022; Pruitt 2023), performing particularly well on deriving stress in suboptimal environments (Elfner 2016; Ryan 2020), and shows limited success in modeling certain counterbleeding interactions (Torres-Tamarit 2016). A key feature shared by these phenomena is that they involve structure-building processes such as (re)syllabification and metrical structure assignment. This success stems from the fact that prosodic structure is built incrementally in HS (Pruitt 2010; Elfner 2016). The stepwise character of HS derivations allows for prosodic structure assigned at earlier steps to be rendered opaque by structural changes at later steps. In Section 6.2, I show that this success carries over to opaque stress in GS.

However, HS faces serious challenges with other types of opacity, much like parallel OT (McCarthy 2007; Wolf 2011; Staroverov & Kavitskaya 2017; Broś 2020; Pruitt 2023, a.o.). To accommodate problematic types of opacity, HS adopts supplementary machinery, including serial markedness constraints (Jarosz 2014) and contextual faithfulness (Hauser & Hughto 2020).

HS provides a straightforward account of opaque stress in GS using only constraints independently motivated in the dialect. The sole addition to the constraint set established in Section 4.1 is Lx≈Pr (Prince & Smolensky 1993/2004) (71), which triggers Stress Assignment.13

As in the parallel OT account (38), transparent penultimate stress results from Nonfinality’s dominance over Head-H (72).

The HS account of opaque final stress is shown in (73). Stress Assignment and Final Shortening each constitute a minimal operation. Since (70) mandates that structural changes be introduced one at a time, these two operations cannot take place in a single HS step. The precedence of operations is determined by constraint ranking. Stress Assignment applies first because Lx≈Pr outranks *Vː#.

On the first pass through Gen and Eval (73a), Stress Assignment is prioritized over Final Shortening. Stress is assigned transparently to the High-toned final syllable, which is still heavy at this point. The shortening candidate, which would pave the way for transparent stress on later passes, fails to satisfy top-ranking Lx≈Pr.

Final Shortening is enforced on the second iteration in response to *Vː# (73b). This is possible because *Vː# outranks Nonfinality.

Final Shortening results in a stressed final light, which GS otherwise avoids (10)–(15). Nevertheless, stress cannot be relocated from this suboptimal environment. Following McCarthy (2011); Lamont (2025), I consider stress shift a sequence of two minimal changes: removal and subsequent reassignment of prominence. Stress removal as the necessary first step in stress shift is not harmonically improving, because it violates top-ranking Lx≈Pr. Since no available minimal change is harmonically improving, the derivation converges on the opaque candidate at Step 3 (73c).14

To summarize, opaque final stress is captured in HS by the following constraint rankings:

To accommodate problematic cases of opacity in HS, Hauser & Hughto (2020) propose a special class of faithfulness constraints known as contextual faithfulness constraints. These are reminiscent of the more widely used positional faithfulness constraints (Beckman 1998). Both constraint types penalize unfaithful mappings only in specified environments. A key innovation in Hauser & Hughto (2020)’s approach is that the relevant environments for contextual faithfulness constraints can be defined over the form’s underlying representation, rather than the immediate local input.

Standard input-output (IO) correspondence constraints (McCarthy & Prince 1995) evaluate candidates against their immediate local inputs in HS. By contrast, Hauser & Hughto (2020)’s contextual faithfulness constraints, commonly referred to as UO faithfulness, assess violations relative to the underlying representation. This property makes UO faithfulness particularly promising for handling counterfeeding opacity.

The underapplication of Final Shortening in the postlexical phonology of GS can be analyzed as a contextual faithfulness effect. To block Final Shortening from applying to final vowels created by consonant deletion, I introduce the UO constraint in (75):

In the relevant environments, word-final consonants are deleted by Homorganic Cluster Simplification.15 As a result, the context that (75) references is not visible in the local input to a given HS step. What matters is that the long vowel that (75) protects is underlyingly followed by a final consonant. The analysis is outlined in (76). The first two steps, which involve Stress Assignment, are omitted.

Both *HomCl and *Vː# outrank the antagonistic IO-faithfulness constraints. This ensures that Homorganic Cluster Simplification and Final Shortening apply on their own. At Step 3 (76a), Homorganic Cluster Simplification satisfies *HomCl. At Step 4 (76b), Final Shortening applies, removing one of the violations of *Vː#. However, the remaining violation of *Vː# cannot be eliminated by Final Shortening at Step 5 because the long final vowel originally preceded a final consonant. Thus, the derivation reaches convergence at Step 5 (76c).

The HS analysis proposed here derives the two opaque interactions in GS phonology using two different mechanisms. Opaque stress falls out from the baseline HS architecture. The ranking of the relevant markedness constraints ensures that Stress Assignment precedes context-changing Final Shortening. The underapplication of Final Shortening results from contextual faithfulness (Hauser & Hughto 2020). The constraint hierarchy in (77) derives all opaque interactions in GS.