Outward-sensitive phonologically-conditioned suppletive allomorphy vs. first-last tone harmony in Cilungu

We present a case study of grammatical tone allomorphy in Cilungu (Bantu). Tense/Aspect/Mood designations (TAMs) are realized via co-exponence of prefixes, suffixes, and floating tones. In a minority of TAMs, there is allomorphy with the floating tones. For example, in the Recent Past one allomorph involves floating tone targeting the final mora of the stem () versus one targeting the stem’s second mora (2). For all such allomorphic TAMs, the alternation is conditioned by the tone of subject agreement markers (SMs) at the left edge of the word. If the SM is high-toned the variant occurs, but if it is toneless then 2 occurs. We present two competing accounts of these data. Under a morphological account, we posit contextual realizational rules with multiple suppletive exponents conditioned by SM tone. In contrast, under a phonological account a ‘first-last tone harmony’ applies here, morphologically restricted to the context of SMs with a small set of TAMs. Such a harmony rule captures a generalization of these alternations: if the SM is high at the left edge then there is a grammatical high at the right edge, but if the left edge is toneless then grammatical tone does not fall on the right edge. We present several arguments in favor of the morphological analysis (suppletion) over a phonological one (harmony). Specifically, the patterns are not subject to phonological locality, other TAMs involving similar tone patterns are not subject to this harmony, and the proposed first-last tone harmony would be a highly phonologically-unnatural rule with little cross-linguistic support, and exceeding the computational properties of all well-known and established phonological operations. We conclude by discussing a major theoretical implication of the morphological account: this constitutes outward-sensitive phonologically-conditioned suppletive allomorphy, standardly argued to be unattested and/or impossible. Ultimately, we hold that under either account Cilungu presents a novel and important contribution to linguistic theory.


Introduction
The topic of allomorphy is a major focal point for linguistic inquiry across morphological frameworks. Most issues remain without consensus, such as what constitutes allomorphy vs. morphologically conditioned phonology (Kiparsky, 1996;Paster, 2014), and what are the restrictions on directionality and locality between the allomorphic trigger and target (Bobaljik, 2000). These discussions are largely based on segmental allomorphy, and absent is incorporating allomorphy found in suprasegmental exponence such as grammatical tone. At least half the world's languages are tonal (Yip, 2002) and virtually all African tonal languages exhibit something which can be classified as grammatical or morphological tone (Hyman et al., 2021). Tone is also known to be different from other phonological phenomena both quantitatively and qualitatively (Hyman, 2011), as well as computationally (Jardine, 2016). These factors conspire to make tone a ripe area for developing empirically rich yet constrained morphological theories, and ideal grounds for testing them.
This paper seeks to incorporate grammatical tone into these theoretical debates, presenting a case study of grammatical tone (GT) allomorphy in the Bantu language Cilungu (Bickmore 2007(Bickmore , 2014. In this language, tense/aspect/mood (TAM) designations are realized via a combination of segmental and tonal exponents. For example, the Remote Perfect is realized as a circumfixal complex /a-. . . -a/ plus a floating tone which associates from the second to the final mora of the stem; we render this grammatical tone H 2-F . We focus on a class of TAM designations with grammatical tone alternations which cannot be reduced to the general phonology. For example, the Recent Past is segmentally realized as /á-cí-. . . -il-e/ but also must appear with one of two floating tone allomorphs, H F or H 2 . The first allomorph associates to the final mora of the stem and the other to the second mora of the stem. The alternation is conditioned by the tone of the subject agreement markers: if the subject marker (SM) is high-toned then the H F variant occurs, but if it is toneless then H 2 occurs. An example is in (1), where the grammatical tone and the position it associates to are boxed, and where the trigger of the alternation (the subject marker) is bolded and underlined.
(1) Recent Past /á-cí-. . . -il-e/ grammatical tone alternation a. / tú-á-cí-| STEM sópolol-il-e H F / → tú-á-cí-sópolol-il-é SM-TAM-TAM-untie-TAM-FV TAM 'we recently untied' b. / u-á-cí-| STEM sópolol-il-e H 2 / → u-á-cí-só pó lol-il-e SM-TAM-TAM-untie-TAM-FV TAM 'he/she recently untied' Only a few TAM designations exhibit such GT alternations in the language; the majority of other TAM designations either have consistent grammatical tone or no grammatical tone. We present two possible accounts of these data, one morphological and the other phonological. The morphological account analyzes these as contextual realizational rules with suppletive allomorphs, e.g. different exponents for the tense feature [RECENT] are inserted depending on the tone value of the subject marker. In this ac-count, phonology is maximally regular and natural. In contrast, under the phonological account, morphology is maximally regular and there is no suppletive allomorphy. Instead, it is phonology which is irregular, having a phonological operation which only applies in the context of subject markers and a small set of TAM features. Such a phonological operation must capture a striking pattern of (1). If the subject marker is high-toned at the left edge, then there is a grammatical high tone at the right edge; equally, if the left edge subject marker is toneless, then the right edge must also be toneless, requiring that grammatical tone not fall on the right edge. We refer to this pattern as 'first-last tone harmony'.
The goal of this paper is to argue for the morphological account involving suppletive allomorphy. We develop a series of arguments that undermine the phonological account, centered around two aspects. First, this tonal alternation is highly restricted, triggered only by a single morphological class (subject markers) and only targeting a subclass of TAM designations which involve exponents of the aspect value [PERFECT] and the tense value [RECENT]. The second concerns the phonological operation 'first-last tone harmony' which would derive the patterns without suppletion. We show that such an operation is cross-linguistically rare if not non-existent, and phonologically unnatural from the view of laboratory and computational phonology (e.g. with respect to the Chomsky Hierarchy and the 'Subregular Hypothesis'). The result is that a morphological account with suppletion emerges as a superior analysis compared to a purely phonological account without it.
Under this morphological account, Cilungu GT suppletive allomorphy constitutes phonologically-conditioned suppletive allomorphy (PCSA) because it references the tonal value (a phonological feature) of the subject marker (the trigger). An important ramification is that because the target (the TAM tone) is inward compared to the trigger (the subject marker tone), this pattern constitutes PCSA which is outwardsensitive, a rare type often claimed to be unattested/impossible (Bobaljik, 2000;Paster, 2006Paster, , 2009Paster, , 2015Embick 2010Embick , 2015. If we are correct that this is outward-sensitive PCSA, this requires us to reevaluate theories whereby exponence proceeds strictly inside-out, and instead allow for exponence to be simultaneous within a given domain. At the same time, if the alternative phonological account involving first-last tone harmony were adopted (against our analysis), this would force phonological theory to reevaluate the kinds of phonological operations which are possible, permitting a wider set under morphologically-conditioned phonological patterns. Regardless, this case highlights the fact that 'restrictiveness' is not a uniform measurement across a grammar, and arguments relying on restrictiveness must take into account concomitant changes to other modules. This paper is structured as follows. Section 2 lays out the basic facts on the Cilungu language to help understand the morphological patterns, and Sect. 3 describes the details of the grammatical tone alternations. Section 4 presents two accounts of these facts, one morphological and the other phonological. Section 5 debates both accounts and concludes that the morphological account is superior. Section 6 discusses the implications. After our conclusion and references, we include two appendices summarizing some data complications.

Preliminaries on Cilungu
Cilungu [IPA: tSílúúNgù; ISO 639-3: mgr] is a Bantu language spoken in Zambia and Tanzania (Bantu zone M14). The data in this paper come from Bickmore (2007Bickmore ( , 2014, and each data point is accompanied by its relevant source. This section provides necessary background to help understand the allomorphic patterns, such as basic facts about the verbal template, the tone system, tense/aspect/mood (TAM) designations, and tone specification of subject markers.

Verbal template
As stated, our empirical focus is the set of grammatical tone patterns which are part of tense/aspect/mood (TAM) inflection. To explain these patterns, we introduce the Cilungu verbal template, familiar across Bantu languages: The more or less obligatory parts of the template (in bold) are the initial position for subject markers (agreement with the subject), tense/aspect/mood in one of several TAM positions, the root, and the Bantu final vowel whose shape co-varies with specific inflectional contexts. Note two important boundaries within the verbal template. The macro-stem boundary (MS) is between TAM and OM, and plays an important role in demarcating tone spreading domains (discussed immediately below). The other is the stem boundary (STEM) between the OM and the root, which will be important in understanding grammatical tone assignment.

Basics of tone system
Cilungu is classified as a privative tone system (Hyman, 2001), contrasting /H/ vs. Ø (toneless). Low tone is realized by default on toneless tone-bearing units (TBUs), which in Cilungu is the mora. No /L/ toneme exists. The position of /H/ tone is not predictable, and thus must be pre-associated to specific TBUs in the underlying representation. Surface forms exhibit a number of phonological processes. Most relevant to this paper is high tone spreading. There are three main types of high tone spreading which are general across the language (i.e. not morphologically conditioned): PWord spreading, MStem spreading, and binary spreading. Each is sensitive to whether the domain of spreading is final within the intonational phrase (ÌP). These are shown in Table 1.
PWord spreading is illustrated in (3). 1 If a high tone originates before a macrostem boundary | MS within a phonological word ω, it spreads unbounded up to the Binary spreading e. Bounded spreading: a Note that MS = macro-stem, τ = tone bearing unit (TBU) which in Cilungu is the mora, ω = phonological word, and Ì = intonational phrase penultimate syllable (3a). However, it shows bounded spreading up to the final syllable before the macro-stem boundary if the word is non-final in the ÌP (3b).
(5) Binary spreading: / tú-ku-| STEM sí-a Choola / → (tú-kú-! shá) ω (Chóólà) ω ] Ì we-TAM-leave-FV Chola 'we are leaving Chola' [B07:172] Bounded spreading occurs even if it were to incur an OCP violation, which is repaired by a downstep. Downstep is indicated throughout by a raised exclamation point. As stated, these three spreading rules are phonologically general and orthogonal to the grammatical tone allomorphy which is the focus of this paper. We identify classes: gú-i-lí-yá-cí-ví-i-zí-lú-ká-tú-gú-kú-pá-kú-múthem because they alter the tone of the surface form, sometimes drastically. For each example in this paper, they are 'factored out' in order to reveal underlying tone and morphologically assigned tones (such as the relevant ones from tense and aspect).
Despite these complications, what remains is a unique phonological factor that unifies these SMs: either tonelessness, or the fact that they sponsor a floating tone but do not have pre-associated tone themselves. In either case, they remain a grammatically unnatural class.

The target of allomorphy: The TAM system and grammatical tone
In Cilungu, individual categories of tense/aspect/mood are expressed by a unique combination of TAM morphs. These consist of prefixes, suffixes, the final vowel and/or tone, as shown in (7) (recall the positions from (2) above). We refer to unique combinations of morphs as 'TAM designations'.
(7) Component morphs of TAM designations: a. 0, 1, or 2 prefixes in pre-radical TAM position b. 0 or 1 suffix in post-radical TAM position c. Shape of the final vowel (-a/-e) d. A grammatical tone which associates within the stem For example, consider the TAM designation 'Yesterday Past Progressive' /á-. . . -anga H F /, in (8). This is composed of a prefix morph /á-/ (with inherent high tone), a suffix /-ang/, the final vowel /-a/, and a grammatical high tone which associates to the final vowel of the stem, H F . 3 SM-TAM-OM-root-DER-TAM-FV TAM 'we were burying for him/her' [B07:258] This TAM grammatical tone is referred to as 'Melodic Tone' or 'Melodic Highs' in the Bantu literature (e.g. Odden & Bickmore, 2014, and papers therein), which we analyze as floating tones not pre-associated to a TBU in the input. Following Yip (2002) and others, we circle floating tones to differentiate them from pre-associated high tone. 4 There is a four-way contrast for TAM grammatical tone in Cilungu, shown in Table 3. Individual TAM designations may have either no grammatical tone (denoted with Ø), a grammatical high tone on the final TBU of the stem ( H F ), a high tone on the second TBU of the stem ( H 2 ), or a high tone which stretches from the second to the final TBUs of the stem ( H 2-F ).
Grammatical tone patterns are consistent in both toneless verb roots and hightoned roots. In (9), forms in the Remote Perfect TAM designation show a H 2-F pattern with both toneless roots (9a) and high-toned roots (9b).

(9)
Remote Perfect TAM designation with consistent H 2-F GT pattern a. / yá-a-sukilil-a H 2-F / → \ yá-a-sukílíl-á \ → [ yáásùkílílá ] 'they have already accompanied' [B14:48] 'we have already untied' [B07:290] An important observation is that while the general tonal spreading rules are sensitive to the macro-stem boundary (Table 1 above), the domain to which grammatical tone is assigned is the stem rather than the macro-stem. The Far Past TAM designation /a-. . . -il-e H 2-F / is shown in (10), where its grammatical tone targets the second TBU of the stem (10a), rather than the second of the macro-stem (cf. 10b).

Local summary
We summarize the Cilungu TAM designations in Table 4, splitting them into five groups based on their tonal behavior. The first group has no grammatical tone (Ø), the second group has a final grammatical tone ( H F ), and the third group has grammatical tone from the second to the final TBU of the stem ( H 2-F ). The patterns of these groups are invariant. In contrast, the fourth group (in bold) shows three patterns of grammatical tone allomorphy. It is this group which is the focus of our study. Group five shows a different kind of grammatical tone allomorphy with Subjunctive and Imperative moods, allomorphy which is both grammatically-and phonologically-conditioned. Crucially, allomorphy in this last group is not conditioned by the tonal designation of the SM and is thus outside of the scope of this paper. Further, certain TAMs change the tone of subject markers, e.g. in the Past Inceptive the subject marker is always high (regardless of its underlying value; cf. Table 2). We indicate this with a superscript H before the TAM prefix. Further, in the Potential with /ngá-/ the subject marker is always toneless, which we indicate with a superscript Ø before the prefix. These tonal changes on the subject marker do not interact with the allomorphy of focus, and we will not treat them further.
It is not possible to associate an individual grammatical tone pattern with a consistent semantic meaning by itself in the absence of segmental co-exponents. Bickmore in fact is explicit on this point, stating "it does not seem possible to assign the [grammatical tone] any consistent meaning that it contributes the form" (Bickmore, 2007:314). For example, H F appears in the Potential but also in temporally more recent situations (e.g. Yesterday Past) and in certain Subjunctive/Imperative contexts; H 2-F appears is temporally more remote situations but also the Perfect and some Subjunctive/Imperative contexts as well. These do not form natural classes such that we can say the GT by itself indicates any semantic meaning. All relevant generalizations appear to involve covariation between specific segmental morphs and specific GT patterns, e.g. the prefix /a-/ co-varies with H 2-F , the prefix /á-/ with H F , and morphs with a long /áa-/ or /aa-/ with Ø.

The allomorphy patterns
With this background established, we now present the basics of the allomorphic patterns involving grammatical tone. In Table 4, the only TAM designations showing allomorphy triggered by the subject marker are those that appear with a prefix /á-/ or with the suffix /-il/. All other comparable TAM designations -hereafter simply called 'TAMs' -are unaffected.

Allomorphy pairing H F /Ø
We begin by examining TAM allomorphy pairing H F with Ø, such as Yesterday Past (YP) /á-. . . -il-e/. Examples are in (11) with the toneless root /fuk/ 'harvest'. In the context of a high-toned subject marker (SM) /ú-/ 'you (sg.)', this TAM is coexponed with a H F grammatical tone (11a), while in the context of a toneless subject marker /u-/ 'he/she' there is no TAM grammatical tone, Ø (11b). In this example and thereafter, the trigger of GT allomorphy is in bold and underlined (the SM), and the relevant TAM morph which is the target of allomorphy is in a dotted box. These examples illustrate 'first-last tone harmony', introduced above. This refers to the fact that if the verb form begins with a high tone then it ends with a (grammatical) high tone / H F /, as well. However, if it begins with a toneless SM, it ends with no grammatical tone (realized in the surface form with default low tone). At this juncture, we only employ the term 'first-last tone harmony' to capture this descriptive generalization; in §4-5, we debate whether it should be interpreted as a bona fide phonological harmony operation in the language. These allomorphy facts are replicated across the full range of Yesterday Past contexts. The same allomorphy patterns are shown with the H-toned root /fúk/ 'turn up hem' in (12), which forms a minimal pair with (11). As above, the form with SM /ú-/ surfaces with high on the initial and final TBU (12a), while the form with SM /u-/ surfaces with low on both (12b).
Furthermore, all and only toneless SMs trigger the special allomorphy. Table 5 lists SMs of classes 3-15 with the toneless verb /ful/ 'wash' in the Yesterday Past. Like toneless class 1 /a-/∼/u-/, toneless SMs /i-/ (class 4) and /i-/ (class 9) trigger the Ø form of the Yesterday Past (rows shaded gray). All other classes are high-toned and trigger the H F form.
Parallel patterns are seen with other H F /Ø TAMs from Table 4, namely the Yesterday Past Progressive, the Recent Perfect and Recent Perfect 2, and the Recent Past Progressive. All of these also contain the TAM morph /á-/, denoting more recent situations. A representative example set is in (13)  Taken all together, H F /Ø allomorphy is consistent across (i) different TAMs (ii) different toneless subject markers, (iii) different root types (H-toned vs. toneless), as well as (iv) several other morphological contexts (not shown) such as in relative clauses, with intervening object markers, with the passive extension /-u/, inter alia. See Bickmore (2007Bickmore ( , 2014 for full details on these other contexts. 5 There is one further complication relevant to our discussion concerning the H F /Ø allomorphy: regular PWord spreading is unexpectedly blocked with the Ø allomorph. Table 1 above schematized PWord spreading whereby a high tone originating outside of the macro-stem spreads to the penult TBU if the word it is contained within is final in the ÌP, i.e. (τ τ | MS τ τ τ) ω ] Ì → (ττ | MSττ τ) ω ] Ì . With TAMs which do not sponsor any grammatical tone (group 1 from Table 4), high tone from outside of the macro-stem undergoes PWord spreading. However, in contexts of the H F /Ø pairing, a high tone before the macro-stem undergoes only bounded binary spreading (i.e. spreading one TBU to the right). This is shown in (14), with the allomorph Ø of the Recent Perfect TAM designation. (14) Lack of MWord spreading with Ø allomorph of Recent Perfect /á-.
These data warrant at least two possibilities. One is that allomorphy here is not between H F and Ø, but rather between H F and another special floating tone which never docks, e.g. H 0 where the superscript 0 indicates that it cannot associate to a TBU. This abstract tone H 0 would still block MWord spreading. A second is that Ø should be replaced with a grammatical L tone, which would also block MWord spreading. This would, however, complicate the fact that Cilungu is otherwise a straightforward H vs. Ø Bantu language. We do not provide an account of these data in this paper. What is important is that regardless of the interpretation, there is complementary allomorphy between H F versus some other GT allomorph. 6

Allomorphy pairings H F / H 2 and H 2-F / H 2
Let us now examine the other two relevant allomorphic pairings from Table 4, occurring in the Recent Past (segmentally: /á-cí-. . . -il-e/) and the Perfect (/-il-e/). In these cases, the trigger of allomorphy is the same -the tone value of the subject marker -but the allomorphic patterns are different. The Recent Past shows a H F / H 2 allomorphy pairing, while the Perfect shows a H 2-F / H 2 pairing. Examples with the Recent Past are in (15), with a high-toned root (15a-b) and a toneless root (15cd).
5 In Appendix 1, however, we discuss one morphological context which does systematically disrupts normal GT allomorphy, namely negative clauses. 6 Cilungu is not the only Bantu language displaying bounded spreading in certain contexts where unbounded spreading is expected. For example, consider closely related M Zone languages Aushi [auh] and Copperbelt Bemba [bem]. In these languages there is no evidence of grammatical tone allomorphy triggered by toneless SMs. In both, the TAMs corresponding to those here are exponed segmentally as /á-cí-/ but without grammatical tone. And in both of those languages, the H on /cí-/ exceptionally undergoes bounded rather than the expected unbounded spreading.
\ a-yá-lásíl -e \ [àyálásílè ] 'he/she has hit them' [B07:294] As seen, the Perfect is exponed with H 2-F if the SM is high-toned (16a and 16c), but exponed with H 2 if the SM is toneless (16b and 16d). In both of these example sets, we see the same first-last tone harmony generalization: if the first TBU is toneless then the last one is also toneless.

Non-local conditioning
All three of these SM-conditioned tone allomorphy pairings apply regardless of the phonological or morphological locality of the trigger and target. The trigger is always the value of the SM at the left edge, and the target is always grammatical tone which appears towards the right edge. Intervening morphemes and/or intervening tonemes are transparent and do not disrupt this allomorphy relationship, as shown in (17)

Other TAMs are unaffected by tone of SM
Other TAMs which are co-exponed with grammatical tone do not show allomorphy triggered by the tone of the subject marker. For these TAMs, no first-last tone harmony is observed. For example, the Remote Perfect /a-. . . -a H 2-F / is consistently realized with a H 2-F grammatical tone in all SM contexts, shown in (18).
'he/she has already buried' The same facts hold for other non-allomorphic TAM designations with H 2-F in Table 4, such as the Far Past /a-. . . -il-e H 2-F / and the Narrative Past /-a H 2-F /. Further, Table 4 mentioned several TAMs which have an idiosyncratic tonal effect on the subject marker. For example, the Potential / Ø ngá-. . . -a H F / has a special property of making all SMs toneless regardless of their underlying value (see Ta-ble 2), while the Past Inceptive / H aa-. . . -a Ø/ makes all SMs high tone. Despite the neutralization of SM tone, neither TAMs show any tonal allomorphy.

Local summary
We summarize the allomorphic TAM patterns of focus in Table 6. Allomorphy in grammatical tone is conditioned by whether the subject marker is high-toned or toneless. Allomorphy only affects GT, and the segmental morphs in the same TAM designation never change.
Based on these allomorphy patterns we established a number of observations, summarized in (21).

(21)
Core observations of the allomorphic patterns: a. Restricted trigger set: it is the phonological identity of SMs and only SMs which trigger allomorphy. b. Restricted target set: the GT of certain TAMs (and only these TAMs) are the target of allomorphy.
Morphosyntactic insensitivity: allomorphy applies regardless of morphosyntactic context (e.g. matrix vs. embedded clauses, with and without object markers, presence or absence of derivation such as passive, etc.). d. Non-locality: allomorphy is insensitive to intervening tones; these do not trigger nor block allomorphy. e. First-last tone harmony: in allomorphic TAMs, the first and last TBUs have the same tonal value (either both are H or both are toneless).
A major focus of the next chapter is on the interpretation of first-last tone harmony. Is it an incidental aspect of the allomorphic pairings or is it a constraint driving the patterns in the first place? Our discussion of the allomorphic patterns was necessarily brief. At the end of this paper, we provide two appendices which detail some further complications involving this allomorphy. Appendix 1 discusses the one morphosyntactic context which does systematically disrupt regular allomorphy, namely negation. Further, Appendix 2 includes additional complicating data for full transparency, particularly with very short roots.

Morphology or Phonology?
We can now begin to address the central question of this paper: should these patterns be treated as morphology or as phonology? And what are the implications upon choosing one over the other? As stated in the introduction, we argue that this is morphology, i.e. that the patterns emerge based on a claim we make about the morphological component of Cilungu grammar. In this section, we wish to neutrally present the case for both morphology and phonology, to the best of our ability. In §5 we will turn to our arguments in favor of morphology.

Interpretation as morphology: Phonologically-conditioned suppletive allomorphy (PCSA)
Let us first establish the basics of the morphological solution. As our starting point, we assume a realizational model of morphology whereby exponents are made up of phonological substance which realizes the already present morphosyntactic features, rather than introducing these features themselves (Stump, 2001;Trommer, 2012). We model this with simple realizational rules of the type familiar to item-based realizational models such as Distributed Morphology (Halle & Marantz, 1993;Embick, 2015). For example, the Narrative Past /-a H 2-F / consists of the final vowel /-a/ and a grammatical tone H 2-F which associates from the second to final TBUs of the stem, shown in (22). No TAM prefixes or suffixes occur in this TAM designation.
'and then I accompanied' We can write a realizational rule for this TAM as in (23) (24), where [v] is realized as /-ile/ in the context of remote past [T:RPAST], but is realized as /-a/ elsewhere.
(25) Cilungu final vowel exponence: The contexts which trigger /-e/ are after the perfect suffix /-il/, in affirmative subjunctives, and in imperatives with an object marker (OM) or andative marker. Because these do not form a natural class, we write three disjunctive statements for the exponence rule. The elsewhere form is /-a/. We follow Bickmore (2007) who analyzes /-il-e/ in Cilungu as constituting separate morphs (unlike in Ndebele), since they can be interrupted by the passive morph /-u/.
With this analysis of the Bantu final vowel in mind, let us now examine a subset of complex TAM designations in Table 7. We contrast two sets of TAMs here, one denoting more remote situations and one denoting more recent situations. Only the second set shows allomorphy.
We can observe that certain TAM affixes co-vary with specific grammatical tones, e.g. the toneless prefix /a-/ always co-varies with the grammatical tone H 2-F . We capture this covariation directly by positing that they are 'co-exponents' of a morphosyntactic feature [T:REMOTE], written with a single rule in (26).
Segmental exponents are linearized at the left edge, while the grammatical tone targets TBUs at the right edge and sensitive to the stem domain. In this way the coexponents are circumfixal. Under this morphological approach, we can now define the exponence involved in the allomorphic TAMs. All but one of these TAMs involve the prefix /á-/, which denote a more recent temporal situation (the counterpart to remote /a-/). The prefix /á-/ co-varies with H F ∼Ø allomorphic grammatical tone, and we therefore interpret them as suppletive co-exponents of [T:RECENT]. The suppletive exponence rules are in (27).  [RECENT] are in a paradigmatic relationship with a consistent meaning difference. Table 8 organizes the data with /á-/ vs. /a-/ based around the presence or absence of other affixes. Observe that the /a-/ form is used to indicate Remote Perfect without a suffix, and Far Past with /-il/ and Far Past Progressive with /-ang/. Compare the equivalent forms with /á-/ and /á-cí-/, which denote the equivalent meanings but with more recent temporal reference. 7 Let us now analyze the TAM suffixes found in this table, i.e. /-il/ and /-ang/. In TAM designations introduced thus far (Table 4), /-ang/ expresses imperfective aspect. It can also be used to modify certain other TAM designations to indicate continuative meaning, e.g. with the Imperative in (28).
(28) Imperfective /-ang/ used in the Imperative to indicate continuative meaning Based on its consistent meaning contribution, we analyze /-ang/ as directly exponing the aspectual category [IMPERFECTIVE]. Its realizational rule is in (29). (29) [ASP:IMPERFECTIVE] ↔ -ang Unlike the temporal prefixes /a-/ and /á-/, aspectual /-ang/ does not consistently covary with a specific grammatical tone, and we therefore do not include tone in this realizational rule. This aspectual suffix is in a paradigmatic relationship with the aspectual suffix /-il/, the only other inflectional morph found in this position. While no TAM designation exists where /-ang/ appears without a TAM prefix, /-il/ may occur without one. In contexts on its own, /-il/ expresses perfect aspect (also called 'anterior' aspect, e.g. throughout Nurse, 2008). 8 An example set is in (30). When /-il/ [PERFECT] appears on its own, it is co-exponed with H 2-F if there is a high-toned SM (30a), but with H 2 if the SM is toneless (30b). 7 Their paradigmatic relationship can also be seen in certain compound TAM designations, such as the Obligative formed with a construction [SM-. . . -lí na INFINITIVE] with the auxiliary verb /lí/. See Bickmore (2007:397-399) for data. We do not discuss compound TAMs in this paper. 8 Cognates of /-il/ and /-ang/ are widespread across Bantu, with meanings identical or similar to Cilungu. For the morph /-il/ (often referred to as /-ile/), Nurse (2008:24) notes that "at least 66 percent of Bantu languages have an anterior ('perfect') in suffixal -ile", which "refers predominantly to anterior (aspect) or to various degrees of past (tense) perfective" (Nurse & Philippson, 2006:181).
A non-trivial issue in interpreting the TAM system of Cilungu is the semantic compositionality of morphs within a single TAM designation. We have proposed thus far that prefixes such as /á-/ and /a-/ express tense, while suffixes /-ang/ and /-il/ express aspect. However, in several cases the meaning of their combination is not transparently composed from their individual meanings. Some examples are provided in Table 9. On its own, /-il/ indicates Perfect, while /a-/ indicates Remote Perfect, but together they indicate the Far Past. Likewise, /á-/ indicates Recent Perfect, but with /-il/ is Yesterday Past.
A full semantic account of these idiosyncrasies is outside of the scope of this paper. The solution must involve a combination of default temporal and aspectual states, and the Perfect marker /-il/ having some kind of temporal meaning in some contexts. Across Bantu, issues of tense and aspect compositionality are notoriously difficult with cognates of /-il/. Nurse (2008) notes the "difficulty of distinguishing anterior (aspect) [Perfect] and near past (perfective), in which the perfective is also an aspect, but typically unmarked in Bantu", remarking that "in some cases it is clear, in others not" (p. 94). Specifically regarding the sequence /a-. . . -ile/ across Bantu compared to /-ile/, Nurse states the following: . . -ile has often has been recategorized from anterior to middle or far past perfective." (Nurse, 2008:157) From the perspective of Bantu, it is not surprising that Cilungu exhibits these complications in compositionality.
There are several responses one could pursue to account for these complexities. One is to say that /-il/ occupies a structurally low grammatical category, but it is some hybrid category fusing tense and aspect. Under this alternative, individual TAM designations realize more complex categories like [FAR PAST PERFECTIVE] and are co-exponed with prefixes, suffixes, and grammatical tones directly. A realizational rule would look as in (32).

(32)
Alternative realizational rule with hybrid category: Such rules circumvent compositionality in that [FAR PAST PERFECTIVE] would not be derived from two separate exponence rules, one for /a-/ and one for /-il/. However, this sacrifices the transparent semantics as seen e.g. in Table 8. It is important to note that regardless of interpretation, the core properties of allomorphy remain unchanged. The trigger of allomorphy is the tone of the subject marker, and the target of allomorphy is the grammatical tone portion of a TAM designation. Accepting for our purposes that TAM is indeed compositional, what decides the winning tone pattern when two or more TAM components with GT co-occur? For example, the Yesterday Past combines [RECENT] /á-H F ∼ Ø/ and [PERFECT] /-il H 2-F ∼ H 2 /, each with their own GT. As shown in (33) We can say that tense /á-/ is 'dominant' over aspect /-il/, following standard interpretation of stress/accent/tone competition in the prosodic literature (Kiparsky & Halle, 1977;Kiparsky, 1984;Inkelas, 1998, inter alia).
With these aspects of a morphological account established, we turn to the final and perhaps thorniest issue, namely idiosyncratic effects from the TAM prefix /cí-/ which appears in several TAM designations (see Table 4). Bickmore (2007:287) states /cí-/ is used to emphasize "the recent nature of the action in question", often translated as 'just'. This morph does not co-vary with any specific grammatical tone. We posit the following realization rule: ['just'] ↔ cí-This morph /cí-/ has a peculiar tonal effect in one context, shown in Table 10. The allomorphic pairing for /-il/ and /á-/ are provided in rows a-b. Row c shows that /cí-/ has no effect on the tonal realization with /á-/. Further, as we stated above when /á-/ and /-il/ co-occur, /á-/ is dominant and its pattern prevails. The quirk comes in row e: when all three morphs co-occur -/á-/, /cí-/, and /-il/ -the allomorph in the context of a high-toned SM is as expected (i.e. H F ) but its counterpart allomorph is unexpectedly H 2 . Row e appears to be a 'blending' of the two other allomorphic pairings.
There are multiple ways to adjust our analysis to accommodate the last pattern where the GT is unexpected H 2 in the context of a toneless SM. Since all three Table 10 Recent Past with /cí-/ 'just' as a unique GT allomorph pairing morphs appear in this construction, we may associate this special GT pattern with either /cí-/ or /-il/; we would not with /á-/ because the patterns are as expected when co-occurring with only /cí-/ (row c) or only /-il/ (d). To capture this pattern, we posit another allomorph of /-il/ [PERFECT] which occurs in the context of /cí-/ 'just', shown in (35) (an update of 31 above).
The first allomorph, which appears in the context of a high-toned SM, remains unchanged. Below that, the new special allomorph H 2 appears in the context of /cí-/ ['just']. The box denotes that it is a special grammatical tone which has exceptional faithfulness and cannot be overridden by dominant tone from the prefix /á-/. This is unlike the other two GTs -H 2-F and the elsewhere H 2 -which are overridden.
Notice that the first two allomorphs appear in non-overlapping environments, and therefore based on these realizational rules in (35) we cannot appeal to a subset principle to decide which should be inserted. We can tell that the one inserted is the H 2-F GT (in the context of H-τ-[SM]), because the first form in row e from Table 10 is not H 2 , which we would expect if the one inserted were H 2 (in the context of /cí-/ ['just']). In Appendix 1, we discuss a principled reason why the first realizational rule wins over the second when the two conflict. Based on additional data from negative clauses, the first rule actually involves two environment conditions: the clause must be both affirmative and have a high-toned SM. Because this rule has two conditions but the second has only one (i.e. /cí-/ ['just']), the former is more specific and is therefore inserted.
This concludes the morphological interpretation of the GT patterns as suppletive allomorphy. We further address several aspects of this analysis when we directly compare the morphological versus phonological interpretations in §5, but this serves to establish the basic logic.

Interpretation as phonology: First-last tone harmony
Let us now compare this morphological account to a phonological one which attributes the allomorphy to a phonological constraint to derive the surface differences. As with the morphological approach we assume realizational rules which map morphosyntactic features to exponents, shown in (36). The important difference is that there is no suppletive allomorphy; exponents are uniform in all phonological/morpho-logical contexts and only later modified. Compare these rules to the previous section above which exhibited grammatical tone allomorphs (Ø and H 2-F respectively).
'he/she has just untied' [B07:269] Under the phonological account, we interpret the driving force behind the delinking of the final tone as the phonological rule of 'first-last tone harmony', which applies in all of the allomorphic contexts which have been introduced. We define this rule in more formal terms in (38).
(38) First-last tone harmony: For a domain D, the first/leftmost TBU τ FIRST (τ F ) has the same tone value as the last/rightmost TBU τ LAST (τ L ) Tone rules of this nature have been lingering in the Bantu literature at least since Meeussen (1967), under the name 'tonal harmony' or the 'Law of Initials and Finals'. As summarized in Hyman (2012:109), such rules state that "certain verb forms end H if the subject prefix is /H/, but L if the subject prefix is /L/". To exemplify, Hyman states that "this rather unusual tonal agreement [is] often found in [non-subject relative clauses]", e.g. in relative clauses in the Konda variety of Mongo [lol] (Nsuka Nkutsi, 1982:189). An example is in (39). The relative clause is in square brackets, and the subject agreement marker on the verb agrees with the relativized noun (denoted by subscripts), not with the embedded subject mí 'I' in post-verbal position. Here, if the agreement marker has no inherent tone then the final vowel will also have no tone (39a), but if it has an inherent high like bá-(39b) then the final vowel also bears high tone. Parallel patterns are reconstructed for Proto-Bantu (Meeussen, 1967:113-114). Hyman in fact cites Meeussen (1971) in linking these patterns to the loss of a grammatical marker between the final vowel of the verb and post-verbal argument which may have left a floating tonal effect.
A rule of this type was originally proposed for the Cilungu facts (Bickmore, 2007:250,271), called 'Final Vowel Delinking' (FV Delinking). This rule stated that tone on the final mora of the (phonological) word delinks from a H toneme, illustrated in (40). This rule must be qualified to applying "only when the TAM prefix /á-/ is present, as well as a toneless SM" (p. 250), /á-/ being the familiar [RECENT] morph. Because such a phonological rule cannot be attributed to the general tonology of Cilungu, it must be understood as 'morphologically-conditioned phonology' (Inkelas, 2014, for an overview). The representations below illustrate sequences which are allowed (41a-b) and those which are not allowed (41c-d) according to the first-last tone harmony constraint from (38) above stating that the first TBU (τ F ) must have the same tone value as the last (τ L ). Crucially, these representations must also make reference to the morphological contexts where it applies, i.e. with subject markers (i.e. [SM]) and the two relevant TAM morphs [T:RECENT] and [ASP:PERFECT]. In (41), we capture this morphological conditioning directly by placing these features within the representations themselves. 9 (41) Let us continue to formalize a phonological account using a prominent model developed to capture long-distance phonological agreement between phonological units, namely Agreement by Correspondence (ABC - Rose & Walker, 2004), couched within Optimality Theory (Prince & Smolensky 2004). Under this model, individual phonological units are in correspondence with each other if they are sufficiently similar. The basic ABC configuration is in (42), illustrating two consonants in a correspondence relation (C x ) which consequently enter an agreement relationship along some additional dimension ([αF]). Data from Kikongo exemplifies this configuration (Rose & Walker, 2004:510). Here, underlying /n/ and /l/ are sufficiently similar in both being alveolar and sonorant, and therefore must be in correspondence, as indicated by a subscripted '1'. Parasitic on this, units in correspondence must additionally agree in nasality which causes changes to the input form, i.e. /l/→[n]. Other segments are not similar enough to be in correspondence (e.g. /k/), and therefore do not show nasal agreement. For these Kikongo data, Rose & Walker propose a grammar with highly ranked correspondence constraint CORR-N↔L and agreement constraint enforcing identity ID-C L C R (NAS), ranked above a faithfulness constraint for nasality. We choose to model the phonological account via ABC because it is a leading model in the theoretical literature on harmony. Further, general tonal processes have recently been incorporated into a unified theory of ABC in Shih and Inkelas (2018), extending the model to capture tone spreading, tone plateauing, tone absorption, tone polarity, contour tone copying, among other phenomena. For example, for the unbounded tone spreading rule in Cilungu, Shih and Inkelas (2018) posit that (adjacent) vowels must be in correspondence with each other (using a constraint CORR-VV). Parasitic on this correspondence, all vowels which are in correspondence must also agree for their tone value -in Cilungu H or toneless (using a constraint IDENT-VV[Tone]). This ensures that an input like /kú-fulumy-<a>/ 'to boil over' (where final <a> is extraprosodic) is mapped to an output [kú 1 -fú 1,2 lú 2 my-<à>], where vowels exist in correspondence chains and therefore must all agree for tone value.
Let us now apply Shih & Inkelas' proposal to first-last tone harmony, using the constraints in (43). The first two (43a-b) are parallel to ones discussed above, while the latter two (43c-d) involve faithfulness to underlying tone structure. Under this analysis, we must assume that edge TBUs, by virtue of being at a domain edge, are sufficiently similar to one another, to the exclusion of intermediate TBUs. Domain edges are cross-linguistically prominent positions, and it is therefore reasonable that they could be in correspondence parasitic on their similarity in prominence. We illustrate these constraints with the H F /Ø alternation of the Recent Perfect /á-. . . -a/. Example (44a) has a high-toned subject marker and a toneless root, while (44b) has a toneless subject marker but with a high-toned root. The first and last TBU's in the word are in correspondence (the subscripted numeral). TBUs and tones between the first and last positions are ignored by these correspondence relations.
'he/she has just untied' [B07:269] We can use the constraints from (43) to derive the mapping between the two intermediate forms for each example in (44). This mapping is shown in the tableaux in Table 11, illustrated with a sample of candidates. The input contains both phonological structure and the morphological conditioners (but for space reasons, they are only implied in the output candidates). The tableau on the left matches (44a) rendered in τ notation, while the tableau to the right matches (44b). Structure in output candidates in bold indicates inserted structure, while that structure in gray indicates structure deleted from the input.
In the tableau at the left, high-ranked CORR-τ F τ L eliminates candidates d-h because the first and last TBUs are not in correspondence (i.e. they are not co-indexed). While candidate c shows correspondence, it violates ID-τ F τ L (T) because the relevant TBUs do not agree in their tonal value (the second H is deleted, indicated in light gray). Candidate b shows both correspondence and agreement, but superfluously: intermediate TBUs also agree resulting in violations of DEP-IO(H). Candidate a is op-timal. In this case, no changes take place to the input other than establishing first-last correspondence.
Consider now the tableau at the right. Candidates g-h do not show first-last correspondence and are therefore eliminated. Next, candidate f shows correspondence but not agreement, and so is also eliminated. All remaining candidates show both correspondence and agreement. Candidates d-e accomplish agreement by adding H tonemes to the outputs, in violation of DEP-IO(H). Finally, candidates a-c all violate MAX-IO(H) by deleting H's in the input. Candidate a does so only for the final TBU, whereas b-c also gratuitously eliminate the root's H tone. This derives the fact that first-last harmony is a non-local process which leaves intermediate tone structure unaffected.
In total, the tableaux in Table 11 which use ABC mechanics constitute a straightforward way to derive delinking of the H tone on the final TBU, and could be equally applied to the other allomorphic GT patterns.

Local summary
To summarize this section, we have compared two ways to account for the grammatical tone allomorphy seen in Cilungu, one morphological, the other phonological. This is summarized in Table 12. Under the first account, we appeal to irregular morphology in the form of contextual realizational rules with multiple suppletive allomorphs, i.e. different exponents for the tense feature [RECENT] are inserted depending on its environment (e.g. the presence of absence of a high-toned subject marker). In this account, phonology is maximally regular and general, without appeal to morphologically-conditioned statements.
Under the second account, morphology is maximally regular by having no stored suppletive allomorphs. Instead, it is phonology which is irregular in that it has morphologically-conditioned phonology in the context of subject markers [SM] and [RECENT]/[PERFECT]. We called the rule deriving the surface alternations first-last tone harmony, and formalized it via Agreement by Correspondence with constraints CORR-τ F τ L requiring correspondence between the first (τ F ) and last (τ L ) TBUs, and IDENT-τ F τ L (T) requiring TBUs in correspondence to have the same tone value. This resulted in the deletion/delinking of final H tones in these contexts, e.g. mapping an input (τ . . . τ´. . . τ´) D to an output (τ 1 . . . τ´. . . τ 1 ) D but leaving intermediate tone structure unaltered.

In support of the morphological account
In this section, we argue for the morphological account (involving suppletion) over the phonological account. Our support for the morphological account is based on three considerations: (i) empirical complications from relative clauses, (ii) the morphological restrictedness of GT allomorphy, and (iii) the phonological unnaturalness of first-last tone harmony.

Empirical complications from relative clauses
One empirical complication for the phonological account comes from relative clauses. Recall that under first-last tone harmony, the first TBU and the last TBU of the word must agree in tone value. In the data we have seen, the first TBU is always occupied by a subject marker (SM). In general, SM's are the initial morph of the word, but one exception are relative clauses.
Relative clauses have the basic structure of matrix clauses, but before the SM there is a relativizer prefix, which is a H-toned copy of the subject marker's vowel. An example of relative clause structure is in (45), where we notate the relativizer prefix as /V-/ REL.

AUG-CL7-caterpillar REL-SM7-TAM-fall-FV
'the caterpillar which is falling' [B07:187] What happens to the allomorphic TAMs in this context? A relativized version of the Yesterday Past TAM designation (/á-. . . -il-e/ with H F /Ø allomorphy) is in (46). This example shows that it is still the tone value of the SM which conditions GT allomorphy; the high tone of the relativizer prefix plays no role. Here, the high-toned SM /yá-/ triggers the H F form (46a), while the toneless SM /u-/ does not (46b).

(46)
Relative form in Yesterday Past -SM still triggers the allomorphy REL-SM-TAM-OM-hit-TAM-FV TAM 'those who hit them' [B07:254] REL-SM-TAM-OM-hit-TAM-FV TAM 'one who hit them' [B07:257] Equivalent data for the other allomorphic TAMs show parallel patterns. Such data reveal that it is not strictly speaking the left-edge value which is important but rather it is truly what the value of the SM is. Therefore, the domain for first-last tone harmony is not an independently-supported prosodic constituent, e.g. the phonological word. 10

Morphological restrictedness of GT allomorphy
The relative clause examples just presented illustrate our next consideration: it is specifically the tone of a single morpheme class (the subject marker) which conditions GT allomorphy, and not a class which could be defined purely phonologically (e.g. the initial morpheme of the word). This illustrates how the GT patterns are morphologically restricted in two ways. First, as stated the trigger of allomorphy is restricted to tone of the SM, and not other morphs in similar positions. Second, the target of allomorphy is also restricted to the GT of TAM designations with /á-/ RECENT and /-il/ PERFECT only. We saw explicitly that other comparable TAMs are not subject to allomorphy -e.g. the Remote Perfect /a-. . .  Inkelas (2014). Consider the case of the Mayan language Mam (England, 1983;Willard, 2004) in Table 13. Mam contrasts short vs. long vowels. By default, suffixes do not alter the length of the stem to which they attach (classified as 'recessive' by Willard), e.g. the instrumental suffix /-b'il/ in this table. In contrast, another set of suffixes systematically shorten preceding long vowels ('dominant' in Willard), e.g. /-na/ an adjectival participle also (b).
Vowel shortening constitutes morphologically-conditioned phonology because its application cannot be attributed to any regular phonological rule of Mam. The number of suffixes like /-na/ which trigger vowel shortening is restricted, and hence we can consider the trigger as restricted, as we do in the Cilungu case. However, within the context of the trigger /-na/, vowel shortening is a regular process and we expect it to apply regularly; in other words the target of allomorphy should not be restricted. This is indeed the case, with both England (p. 126) and Willard (p. 10) describing /-na/ as quite productive.
Such regularity is not a quirk of Mam. Rather, the bulk of cases in Inkelas (2014), as in many other works, illustrate that the target of morphologically-conditioned phonology by default is not a restricted set. If the Cilungu grammatical tone alternations were morphologically-conditioned phonology, we would expect it to apply in a much larger set of contexts than the restricted set of those TAMs involving [RECENT] /á-/ and [PERFECT] /-il/. The fact that this is not the case supports suppletion. 11

Phonological unnaturalness of the alternation
The best evidence for the suppletion view comes from our final consideration: the phonological naturalness of the alternation under the phonological account (i.e. firstlast tone harmony). Logically, if an alternation does not constitute suppletion -i.e. distinct stored allomorphs -then it must be derived via some phonological rule, whether morphologically conditioned or not. As Paster (2016) aptly points out, this requires us to assess "the plausibility of the proposed rule" which serves as the alternative to suppletion, embedded within "a commitment to some formal model for which it is clear what constitutes an allowable operation, trigger, target, and so forth, so that the plausibility of a rule can be assessed" (p. 113). To that end, we assess the phonological naturalness of first-last tone harmony along three dimensions: typological precedence, computational complexity, and (to a lesser extent) learnability in a laboratory setting.
First, we expect a phonological rule of this type to also be attested in some grammar where it would apply generally across that language (i.e. not morphologicallyconditioned). In our experience of tone systems, we know of no fully general tonological rule of this type in any language. Nothing resembles this in any of the surveys of tone (Pike, 1948;Fromkin, 1978;Yip, 2002;Hyman, 2011;Wee, 2019) or in common historical changes affecting tones (Hyman & Schuh, 1974;Hyman, 2007). In the Bantu literature, a phonologically general first-last tone harmony is not found (e.g. not found in Odden & Bickmore, 2014). The literature on computational complexity and learnability in fact accepts that such a pattern is unattested. To quote Heinz (2018:144), "there are no known phonotactic patterns where the last sound in a word depends in some fashion on its first sound".
A morphologically restricted version of first-last tone harmony (not fully general) is found, but is exceedingly rare in the tone/Bantu literature. We introduced a few of these in §4.2 above. We believe the best example other than Cilungu comes from Kikamba [kam] (Roberts-Kohno, 2000Jones & Freyer, 2019). In a minority of TAM contexts, the final vowel of the verb must have the same tone value as the subject marker, either high or low/toneless. TAM designations include the Hodiernal Perfective and Stative (Assertive-nonfinal and Relative, but not Object-Focus), the Immediate Past, among a few others. An example is in (47) (Jones & Freyer,11 Relatedly, we agree with the mounting criticism of 'readjustment rules' (Bermúdez-Otero, 2012:79ff.;Merchant, 2015:282;Haugen & Siddiqi, 2016:349ff.;Paster, 2016:110ff., etc.), which constitute a DM variation on morphologically-conditioned phonology involving both a restricted target set and restricted trigger set. A common example is a I → ae readjustment (e.g. sing → sang), where the trigger is restricted to [PAST] and the target restricted to a small number of irregular roots ( While we do not take a position on Kikamba, in its small literature there is a parallel debate whether to treat this as morphology or phonology. Roberts-Kohno (2014) proposes a morphological account, with a grammatical tone melody / H 2 L PEN H F / for (47a) in the context of a H-toned SM, but / H 2 L F / for (47b). This essentially involves two distinct inputs and as such constitutes suppletion. In contrast, Jones and Freyer (2019) respond with a phonological account involving a rule of 'final lowering' which is morphologically-conditioned, applying in forms "with 3rd singular personal subject agreement" (p. 186), and also requiring the specific TAM environments list above. Like in Cilungu, 3rd singular systematically differs from other subject markers in tone. Under this phonological account, Jones & Freyer posit uniform tonal exponence for these TAMs in all contexts, / H 2 H F / (p. 188). The fact that this is probably the best case of (morphologically restricted) first-last tone harmony speaks to it being a typological anomaly. 12 Second, there is a principled reason why first-last tone harmony should not exist in phonological systems, namely that its computational properties exceed that of the established catalogue of phonological patterns. This can be assessed thanks to a large amount of recent work dedicated to examining phonological patterns computationally (see Heinz, 2018 for a recent overview). To illustrate, let us situate first-last tone harmony within what is known as the Chomsky Hierarchy (Chomsky, 1956, inter alia), which demarcates linguistic patterns into nested regions of complexity (Heinz & Idsardi, 2013:113). This is shown in Fig. 1 (drawn from Lai, 2015;Heinz, 2018;and Avcu & Hestvik, 2020).
One pattern which is specifically characterized by Lai (2015:426) as outside of the known range of subregular patterns is first-last harmony, which has emerged as a consensus within the computational phonology literature (Heinz & Rogers, 2010;Heinz & Idsardi, 2013;Lai, 2015;Graf 2017Graf , 2018Heinz, 2018:146). This is seen within Fig. 1 where it falls outside of other subregular types, e.g. Strictly Local patterns like 12 Other cases of morphologically restricted first-last tone harmony are found in Nyamwezi [nym] (Schadeberg, 1989) and Ekegusii [guz] (Larry Hyman, p.c.). Cilungu is not particularly close to any of these other Bantu languages (Kikamba, Nyamwezi, and Ekegusii). Cilungu is a Zone M language of northern Zambia, while Ekegusii (Zone J/E transition) and Kikamba (Zone E) are in Southern Kenya; between them is Nyamwezi (Zone F) in central Tanzania. Genealogically, Cilungu is also in a separate East Bantu branch from all of them, as well (Nurse, 1999). standard nasal place assimilation where a "string's well-formedness can be determined solely by examining . . . contiguous substrings of bounded length" (Chandlee, 2014:25), or Strictly Piecewise patterns like sibilant harmony, which may operate at a distance. To quote Heinz and Idsardi (2013:119), "even if humans were exposed to words conforming to this [first-last harmony] pattern, they would fail to detect it, as it lies outside the hypothesis spaces of humans' phonological 'pattern detectors"'.
To illustrate in more concrete terms how first-last tone harmony is different, consider the following discussion taken from Jardine (2020) looking at another Bantu language Karanga Shona [sna] (Odden 1984(Odden , 2014. Jardine summarizes that nonassertive verb stems are limited to a two-way H-toned vs. L-toned contrast. The surface patterns of this H vs. L contrast is predictable based on the number of TBUs. This is shown in (48).

(48)
Karanga Shona 1σ 2σ 3σ 4σ 5σ 6σ 7σ a. H-toned roots H HL HLH HHLH HHHLH HHHLLH HHHLLLH b. L-toned roots -LH LHL LHHL LHHLL LHHLLL LHHLLLL In (48a), a H-toned root generally shows a HLH pattern if there are enough TBUs. The first TBU is always H and this spreads over a maximum of three TBUs. This is followed by a L (again, spreading maximally over three TBUs), and a H is final (except in 2σ cases). In (48b) with a L-toned root, the initial TBU is always L (which does not spread), followed by a H maximally over two TBUs, and then the remaining TBUs are L. Notice that the 3-7σ stems of both root types superficially show a firstlast tone pattern, as both positions are tonally identical.
Examining these patterns computationally, Jardine shows that these Karanga Shona data can be analyzed in strictly local terms using contiguous forbidden substrings on the melody tier. For example, a forbidden melody set {#HL#, HLHL, LHLH} bans instances with not enough contours at the word-level but equally in-stances with too many contours, as well. 13 These melody constraints can be used in conjunction with another set of constraints at the TBU tier, and what emerges is that these surface patterns and only these patterns are permissible strings. Pattern like Karanga Shona do not involve any literal long-distance first-last tone harmony, and the list of banned strings always constitute a finite set.
Jardine explicitly contrasts this with a hypothetical first-last harmony constraint, i.e. forbidden melodies of the type {#H. . . L#, #L. . . H#}. Because such forbidden melodies do not constitute a contiguous string (the '. . . ' expressing the variable structure in between), they are necessarily non-local. Why this is crucially different from other subregular patterns is that the banned structures do not form a finite set of forbidden substring constraints, as the "melodies of these can strings can be arbitrarily long. . . : mldy(HLHL) = HLHL, mldy(HLHLHL) = HLHLHL, mldy(HLHLHLHL) = HLHLHLHL, ad infinitum" (Jardine, 2020(Jardine, :1170. If we were to capture phonologically non-local interactions (like first-last tone harmony) literally via non-local constraints, this would drastically increase the power of the phonological module and consequently lessen its restrictiveness significantly.
The third and final point regarding the phonological unnaturalness of first-last tone harmony comes from experimental evidence from artificial language designs (Lai 2012(Lai , 2015Finley 2012Finley , 2017Avcu & Hestvik, 2020). To discuss a few of these, Lai (2012Lai ( , 2015  . The results match the predictions given its unattestedness: "intensive First-Last participants definitely failed to internalize the First-Last Assimilation grammar that was intended in this study", showing that "First-Last Assimilation is harder to learn than Sibilant Harmony" (Lai, 2015:445). A recent experiment in Avcu and Hestvik (2020) largely replicated Lai's findings, though they observe "a residual sensitivity to the [First-Last Assimilation] rule in the [First-Last] and [Intensive First-Last] groups, which contradicts Lai's previous conclusion" (p. 15). The authors interpret this as an artifact of the laboratory learning situation rather than reflecting linguistic-specific learnability. As an anonymous reviewer points out, however, all of the studies above show complications undermining a straightforward interpretation, and none of them directly probe first-last harmony with tone itself. Further studies are required.

Local summary
In this section, we provided evidence in favor of a morphological account for the Cilungu allomorphy patterns over a phonological account. The morphological account entails that the Cilungu patterns constitute suppletion, whereby one allomorph cannot be derived from the other and both need to be stored in the lexicon (Embick, 2010:43;Paster, 2016;Bermúdez-Otero, 2018). The Cilungu patterns meet the threshold for suppletion assessed against three commonly used criteria for suppletion found in the literature. These are summarized in (49). Note that our use of 'suppletion' here applies to any allomorphic pairing and is explicitly not restricted to lexical roots.
(49) Criteria for assessing suppletion 14 a. Phonological distance of forms: two forms F 1 and F 2 are suppletive if they exhibit phonological distance past a threshold T, measured with respect to phonological structure b. Uniqueness of alternation: two forms F 1 and F 2 are suppletive if the alternation is not found in comparable morpho-phonological contexts c. Phonological naturalness of alternation: two forms F 1 and F 2 are suppletive if the alternation cannot be derived via a phonologically natural rule The first criterion involves the phonological distance of two allomorphic forms and is the weakest criterion. At first glance, Cilungu allomorphs do not exhibit much phonological distance, e.g.
[PERFECT] /-il H 2-F / vs. /-il H 2 / where the difference is solely located in how the GT associates. However, complicating any straightforward interpretation is the fact that it is not settled how different kinds of floating tones should be represented. For example, if one interpreted the difference in tonal patterns as due to different combinations of floating H and L -/-il H 2 / would be rendered something like / L H L il L / -this would impact the degree of distance between morphs, tipping the scales towards suppletion. 15 The two more informative criteria from (49) involve the uniqueness of the alternation and the phonological naturalness of the alternation, and both support the Cilungu patterns as suppletive. As this section has showed, these GT allomorphic patterns are restricted both with respect to the set of triggers and to the set of targets. Furthermore, we considered a potential phonological rule -first-last tone harmony -which could derive the patterns without suppletion, and firmly concluded that it is both cross-linguistically rare if not non-existent and phonologically unnatural from the view of computational phonology (and likely laboratory phonology, as well). The result is that a morphological account with suppletion emerges as a superior analysis compared to a purely phonological account without it.

Implication: Outward-sensitive PCSA is possible
If the Cilungu allomorphic GT patterns are bona fide examples of suppletion, their inclusion in the wider typology of suppletion has significant implications for morphological theory. Specifically, they constitute a case of outward-sensitive phonologically-conditioned suppletive allomorphy (PCSA), which has been claimed to be unattested/impossible in several different strands of the morphological literature. To demonstrate this, in this section we take as our starting point the verbal syntactic structure of Cilungu, based on the extensive syntactic literature on Bantu. Based on this structure, we demonstrate the outward-sensitive nature of these suppletion patterns and argue that this supports a view of exponence as simultaneous rather than strictly inside-out. Finally, we discuss a prediction made by simultaneous exponence, namely that reciprocal phonologically-conditioned allomorphy is possible and show that this prediction is borne out by the Cilungu data.

Directionality and allomorphy
Directionality with PCSA is a recurrent issue across the different schools of morphology (Carstairs, 1987), and has played a particularly pronounced role in Distributed Morphology (Bobaljik, 2000(Bobaljik, , 2012Embick 2010Embick , 2015. Directionality involves the relative position of the trigger of allomorphy and its target. One type is 'inwardsensitive' allomorphy, whereby the trigger of allomorphy is in a structurally inner position compared to the target of allomorphy, i.e. it is closer to the lexical head of the construction. This type is also referred to as 'outward-conditioning' in the literature. Its counterpart is 'outward-sensitive' allomorphy (a.k.a. 'inward-conditioning'), where it is the trigger which is in an outer position and the target in an inner position.
To illustrate inward-sensitivity with PCSA, consider English indefinite allomorphy where a is used before consonants but an is used before vowels, shown in (50). The trigger of allomorphy (bold, underlined) is a phonological property of the inner element (the head noun); in contrast, the target of allomorphy (boxed) is an outer element (the indefinite article). Allomorphy here is sensitive to an inner phonological property. Here, there would be two hypothetical allomorphs a and na, conditioned by the preceding segment. The trigger would be a phonological property of the outer element (the preposition) and the target would be an inner element (the article). Prepositions which end in consonants (at, with, in, during) would condition one allomorph (a), while those which end in a vowel (to, by, below, via) would condition another (na). Allomorphy here would be sensitive to an outer phonological property. While inward-sensitive PCSA is very common cross-linguistically, outwardsensitive PCSA is considered rare at best and essentially unattested by many morphologists. Paster (2009) presents a typological survey of 137 instances of PCSA, and finds that "135 indisputably have 'inside-out' conditioning" (see also Paster 2006Paster , 2015. Separately, work within DM uniformly adopts the position that outwardsensitive PCSA is impossible, e.g. Bobaljik (2000), Embick (2010Embick ( , 2015, among others. The proposed empirical gap is insightfully linked to the procedure of exponence itself: terminal nodes in the syntax are spelled out cyclically from the most deeply embedded node, moving outward. Inner vocabulary items (i.e. exponents) are exponed earlier, and as such would have no access to the phonological features of outer vocabulary items. Few cases of outward-sensitive PCSA exist, and none has been universally accepted as a bona fide example. Potential examples are found in Fula (Carstairs, 1987:185-188, 205-206, citing Arnott, 1970, Welsh (Hannahs & Tallerman, 2006), Ndyuka (Bye, 2007:77, citing Huttar, 1996, the Surmiran dialect of Swiss Rumantsch (Anderson, 2008), Italian (Wolf, 2008:193ff.) Icelandic (Svenonius, 2012), Armenian (Wolf, 2013), and Nez Perce (Deal & Wolf, 2017;however, cf. Kiparsky, 2021).

Determining directionality based on Cilungu verbal syntax
In order to determine the directionality exhibited by Cilungu grammatical tone allomorphy, we must articulate the underlying morphosyntactic structure which the individual allomorphs are mapped from. Relevant for our purposes are the positions of three types of morphemes: the subject marker (the trigger), tense before the root (e.g. /á-/ [RECENT]) and aspect after the root (e.g. /-il/ [PERFECT]), the latter two being the targets of allomorphy. We repeat the Cilungu verbal template in (52) (slightly modified from §2.1), where we mark in bold the relevant positions. As reflected in the bracketing in this template, the consensus within Bantu literature is that prefixes are treated as further away from the verbal root than suffixes (e.g. Downing, 1999;Nurse, 2008, and literature cited therein). This inner complex is referred to as the stem across Bantu (and the macro-stem if including the prefixal object markers), a constituent which is morpho-phonologically demarcated across Bantu languages, e.g. being the domain of nasal harmony, reduplication, and vowel coalescence, among other phenomena (Nurse, 2008:14). As we have seen in Cilungu, evidence for the stem comes from the alignment of grammatical tone. In general, this verbal complex is formalized in the generative syntactic work on Bantu by positing a verb root and a series of higher functional heads which express verbal grammatical categories, e.g. subject agreement, polarity, tense, aspect, mood, etc. We analyze the syntactic positions of Cilungu verbal categories as in the tree in Fig. 2, which synthesizes various accounts across Bantu while simultaneously being faithful to Cilungu-specific patterns. At the bottom of this tree is the lexical root, which we simply denote as V 0 . This is merged with a categorizing little-v head v 0 , which raises to form a complex head with an aspectual head Asp 0 (if present). As we stated in §4.1 above, we follow Pietraszko (2018) in positing that the final vowel in Cilungu is the realization of this v 0 head. The realizational rules are repeated in (53).

(53)
Cilungu final vowel exponence: Before the final vowel, the two aspectual suffixes /-il/ for [PERFECT] and /-ang/ for [IMPERFECTIVE] may appear. We assume that v 0 raises to Asp 0 in order to account for this linear order. 16 Figure 2 diverges in a few ways from the general Bantu syntactic literature. One way is that only v 0 raises, but V 0 does not. Summarizing this literature, Pietraszko (2018:305) states that "it is typically assumed that the verb in Bantu languages does not move all the way to [Tense] T, but that it stops in a lower position-the head hosting the final suffix" (supporting literature include Myers, 1990;Julien, 2002;Carstens, 2005;Buell, 2005;Harford, 2008;Van der Wal 2008Cheng & Downing, 2012, inter alia). However, if we follow this literature and assume V 0 also raises, we would additionally have to account for why the linearized order is [V-Asp-v], rather than [V-v-Asp] which would be expected. While this perhaps could be attributed to a type of head-lowering, or morphotactics which rearrange heads post-syntactically (e.g. CARP templates à la Hyman, 2003), we assume the simpler analysis whereby the linear order of suffixes simply reflects their syntactic position (Baker, 1985). Regardless, nothing hinges on this decision with respect to the relevant directionality relations.
Further, the order of subject markers and tense markers reflect their syntactic position, as well. Subject markers realize an AgrS 0 head, while tense realizes lower T 0 heads. Subject agreement being outside of tense is well-supported in both the typological literature (Bybee, 1985:35) and generative literature (Chomsky, 1989:68ff.;Belletti, 1990;Speas, 1991:183ff;Harford, 2008), and cross-linguistically tense predominantly scopes over aspect (Cinque, 2014). 17 With this established, we may now examine the directionality relations of the allomorphy patterns. We will diagram the syntactic structure and subsequent morphological realization of the examples in (54-55), taken from (16) and (13) above respectively. The trigger (bold and underlined) conditions the GT co-exponing either tense or aspect.
→ \ a-ful-íl -e \ [ àfùzílè ] 'he/she has washed' 'he/she has just accompanied' Figure 3 illustrates the spell-out of these examples, going from the syntactic structure to the linearized sequence of exponents. The spell-out of AgrS 0 and V 0 as subject markers and lexical verb roots (respectively) happens straightforwardly, and the spell-out of v 0 as the final vowel proceeds according to the realizational rules in (53) Fig. 3, because the subject markers /tú-/ and /yá-/ have a pre-linked lexical high tone, this triggers the more specific suppletive allomorph to be inserted in these contexts.

Fig. 3 Cilungu illustration of Spell-out and allomorphy
The directionality relations in these diagrams are clear: the trigger (agreement) is structurally higher and in an outer position while the target (tense/aspect) is structurally lower and in an inner position. Because (i) the trigger is outward compared to the target, and (ii) a phonological property of the trigger is part of the conditioning environment, this thus constitutes the aforementioned rare type of allomorphy, namely outward-sensitive PCSA.

Simultaneous exponence
If Cilungu constitutes outward-sensitive PCSA and therefore shows that this type of allomorphy is rare but possible, what ramifications does this have for morphological theory? We contend here that the Cilungu facts support the notion of 'simultaneous exponence'.
To explain, recall that within the DM literature it is generally assumed that exponence proceeds cyclically moving inside-out. We can call this model 'serial exponence'. For example, for the leftmost tree in Fig. 3, this would involve four steps of exponence, starting with the most deeply embedded (V 0 ) and ending with the least embedded (AgrS 0 ). One can identify several variations on serial exponence in the literature (e.g. exponence takes place one at a time but the steps may not necessarily reflect linear order -Myler, 2017), but all generally share the idea that exponence is staggered. In contrast is a model of simultaneous exponence, whereby the realizational rules all apply simultaneously within the relevant spell-out domain. The Cilungu facts are compatible with simultaneous exponence but not with strict insideout serial exponence. Under the latter model, inner [PERF] and [REC] should be already realized before phonological properties of the outer subject marker become available.
While simultaneous exponence is less prominent than serial exponence, there is some precedence for it outside of Cilungu. For example, under Nanosyntactic approaches involving 'spanning' (Starke, 2009;Caha, 2009;Svenonius, 2016, inter alia), larger contiguous constituents can be realized as a single exponent (e.g. portmanteaux), which entails that inner heads cannot always be exponed before and to the exclusion of outer ones. Other works have even pointed out that two heads may not be structurally local but still condition portmanteaux or allomorphy, e.g. bracketing paradoxes like [under][went] from [[UNDER [GO]]PAST] (see Haugen & Siddiqi, 2016). If many-to-one exponence is permitted, it is a short move to also allow manyto-many exponence as well. The latter is simultaneous exponence.

A prediction borne out: Reciprocal PCSA
Simultaneous exponence makes different predictions from serial exponence. As pointed out in Deal and Wolf (2017), under simultaneous exponence theories "we expect to find cases where allomorphic selection for two adjacent morphemes is resolved in a way that involves mutual phonological dependence" (p. 57). Thus, simultaneous exponence uniquely predicts that two exponents can trigger 'reciprocal PCSA', where both have a phonological property which triggers suppletive allomorphy of the other. Serial exponence predicts this to be impossible.
Importantly, this prediction is actually borne out by Cilungu when we consider other facts of the language. Specifically, tonal properties of subject markers can condition allomorphy of tense, while at the same time segmental properties of tense can condition allomorphy of subject markers. The relevant data involve the class 1 (CL1) subject marker (expressing third singular reference). In Table 2 from §2.3, we listed the CL1 subject marker with two allomorphs, /u-/ and /a-/ (both toneless). Bickmore (2007:29) details how the /u-/ allomorph is only used next to exponents of the shape /a/, while the /a-/ allomorph is used in all other contexts. In (57), the /u-/ CL1 allomorph appears before the tense prefix /a-/ [REMOTE] (e.g. in Far Past /a-. . . -il-e H 2-F /) as well as before the tense prefix /á-/ [RECENT] (e.g. in the Yesterday Past /á-. . . -il-e H F ∼ Ø/). All TAM designations which involve these prefixes /a-/ or /á-/ equally condition this CL1 allomorphy. In (57), the tense prefix (the trigger) is in bold and underlined while the subject marker (the target) is boxed. This can straightforwardly be understood as a dissimilation effect, whereby two exponents of the shape /a/ cannot be adjacent, a kind of 'repeated morph constraint' (Menn & MacWhinney, 1984). In all other contexts, the /a-/ CL1 allomorph is used instead. For example, if the tense prefix (the trigger) and the subject marker (the target) are interrupted by the negation prefix /tá-/, the default /a-/ form surfaces, shown in (58).
(67) / u-á-sukilil-a Ø / \ u-á-sukilil-a \ [ wààsúkílìlà ] 'he/she has just accompanied' In general, if tense were realized first, then the relevant phonological features triggering allomorphy (the tone of the subject marker) would not be available to condition its allomorphy. Equally, if agreement were realized first, then the relevant phonological features triggering allomorphy (the segments of tense) would not be available. This mutual dependency is naturally captured if exponence is simultaneous and exponents have access to the phonological shapes of their neighbors in both directions. It is not naturally captured under purely serial models of exponence (whether inside-out or outside-in).

Conclusion
This paper has sought to incorporate a case of grammatical tone suppletive allomorphy into ongoing debates on allomorphy within the morphological literature. We showed that in the Bantu language Cilungu, the grammatical tone (GT) which is co-exponed with the tense/aspect/mood (TAM) morphs /á-/ RECENT and /-il/ PER-FECT shows allomorphy depending on the tone of prefixed subject agreement markers (SMs). We posited the following realizational rules to account for these allomorphic patterns: If the SM is high-toned, then one type of GT is selected, but if it is toneless then an elsewhere GT is selected. We presented several aspects of the data which support an interpretation as suppletive allomorphy: the patterns are not subject to phonological locality, other TAMs involving similar tone patterns are not subject to this alternation, and other high-toned prefixes do not trigger it. We contrasted our account to an alternative phonological account which would involve no suppletive allomorphy and instead a morphologically-restricted phonological operation of first-last tone harmony, where the first and last positions of a word must agree in tone value. We dismissed this as a highly phonologically unnatural rule: it has little to no cross-linguistic support and it exceeds the computational properties of all well-known and established phonological operations (e.g. with respect to the Chomsky Hierarchy and the 'Subregular Hypothesis'). The Cilungu suppletive GT patterns have important implications for morphological theory. Specifically, they constitute a rare type of outward-sensitive phonologically-conditioned suppletive allomorphy (PCSA), which has been claimed to be unattested/impossible in several places in the morphological literature. We showed that across the Bantu literature, it is generally accepted that subject agreement (the trigger of allomorphy) is in an outer position compared to tense and aspect (the target of allomorphy). This therefore entails that this pattern is outward-sensitive, i.e. more inwardly located exponents are sensitive to features of more outwardly located exponents. If we accept that Cilungu shows outward-sensitive PCSA, this supports a view of exponence as simultaneous rather than strictly inside-out. A prediction of simultaneous exponence is that that reciprocal PCSA should be possible, and prediction which is borne out by PCSA exhibited by class 1 subject agreement conditioned by the phonological shape of an inner exponent.
We conclude with a few general comments for future work. First, if we accept outward-looking PCSA as possible and empirically attested, we must address why it is so vanishingly rare. If we cannot localize its rarity to factors of grammatical architecture (i.e. within a Universal Grammar), then what are the potential usagebased, diachronic, or cognitive pressures which counteract its development? Second, a hope of this research is to encourage closer examination of grammatical tone systems through the lens of morphological theory, systems which proliferate across sub-Saharan Africa (amongst other places). Many of these languages demonstrate complex, nuanced, and consistent patterns, collected by careful fieldwork and textual analysis. We have no doubt that many of these are well-described enough at the present moment to be valuable for testing morphological hypotheses.

Appendix 1: Special GT allomorphy in negative clauses
As alluded to in §3.1, one morphological context systematically disrupts normal GT allomorphy, namely negative clauses. Let us first establish the basic facts about verbal negation in Cilungu. In most TAM designations, negative morphology is marked by a prefix /tá-/∼/táa-/ NEG (see Bickmore, 2007:240 for details on the distribution of short vs. long variants). The prefix appears after the SM but before TAM prefixes. Before this negative prefix, SMs become toneless, which we notate as / Ø tá-/ with a superscript Ø. An example of the negative is in (69), in the Future Progressive TAM designation. The first plural SM /tú-/ which has lexically-high tone surfaces as low in this negative context. This tonal effect is found across the full range of TAMs and SMs whenever they are in a negative context. We must also adjust the H 2-F allomorph to applying only in affirmative clauses (marked [POS]). This accounts for something we alluded to in §4.1, namely a principled reason why the second allomorph (with H 2-F ) triggered by a high-toned SM should win over the third allomorph (with H 2 ) triggered by the presence of /cí-/ 'just', if both conditions exist in the environment at the same time. The former actually has two conditions and is therefore more specific, while the latter has only one condition and is less specific. The most specific allomorph takes priority during insertion.

Appendix 2: Unexpected tone patterns of TAMs with H F /Ø allomorphy
This final appendix presents a small set of recalcitrant data, included for full transparency. All of these involve those TAMs with H F /Ø allomorphy, where it appears that the wrong allomorph is selected. While these present additional complications for our analysis (and likely any analysis), we stress that they constitute a small portion of the overall paradigm. The overwhelming majority of data support the generalizations laid out in this paper. To exemplify, consider the Yesterday Past TAM designation /á-. . . -il-e/, which we showed is co-exponed with H F GT if the SM is high-toned, but with Ø GT if the SM is toneless. In a limited number of contexts, however, it appears as if the H F is chosen with a toneless SM. An example is in (76). Here, the verb forms end in a high tone in the intermediate and surface forms, rather than the expected form with a final toneless TBU in the intermediate form (low on the surface).
(78) / u-á-el-a H F / \ u-á-el-á \ [ wèélá ] SM-TAM-fish-FV GT 'he/she has just fished' [B07:271] (Cf. expected: */ u-á-el-a Ø / *\ u-á-el-a \ *[ wèélà ] ) The unexpected form is also found with short roots extended with the derivational suffix /-il/ APPLICATIVE, where the unexpected H F GT is also selected, shown in (79). This happens both with toneless roots (79a) and high-toned roots (79b), demonstrating that the irregularity cannot be attributed to some kind of interface from the root tone. As seen here and the other data points, many of the exceptional surface patterns end in a downstepped high, but it is not known at this point whether this is a coincidence. Further study of these exceptional patterns is required.