--- title: "Varieties as conditioning structure: A homeostatic account of register, dialect, and discourse community" author: "Brett Reynolds" year: "2026" status: "Preprint" canonical_url: "https://lingbuzz.net/lingbuzz/009872" website_url: "https://brettreynolds.ca/papers/varieties-as-conditioning-structure/" markdown_url: "https://brettreynolds.ca/papers/varieties-as-conditioning-structure/paper.md" version: "author-manuscript mirror" version_date: "2026-06-12" keywords: ["sociolinguistic variation", "register", "dialect", "discourse community", "homeostatic property clusters"] --- # Varieties as conditioning structure: A homeostatic account of register, dialect, and discourse community **Author-manuscript mirror.** This Markdown file is provided for accessibility, search, and machine readability. The canonical public record is linked in the metadata above. ## Abstract What distinguishes register from dialect from discourse community? This paper argues that they differ not in what they are but in what participant expectation is conditioned on: situation, ascription, or identification. These aren’t just descriptive dimensions. Recurrent situations, social classifications, and norm orientations keep rebuilding the same expectation profiles. The key move is separating ascription (what you’re treated as) from identification (whose norms you orient to). Drawing on Wiese (2023)’s account of communicative situations and O’Connor (2019)’s game-theoretic work on how categories stabilize, I argue that varieties hang together because the same kinds of situations recur, the same kinds of speakers keep getting socially classified in familiar ways, and the same norm centres keep organizing orientation, while social coordination makes established partitions sticky. A variety earns its keep when it supports stable, purpose-relative prediction: knowing the conditioning state helps participants anticipate likely forms, and hearing the forms helps them infer the conditioning state. A worked example using Kiesling (1998)’s fraternity data on the sociolinguistic variable traditionally written as (ING) shows that this framework captures structure that standard factor-group analyses conflate. **Keywords:** sociolinguistic variation, register, dialect, discourse community, homeostatic property clusters, coordination games, Bayesian conditioning, participant expectation, indexicality, ascription, identification # Introduction Define register by situation, and a lawyer’s jargon used at home becomes a register without a situation. Define dialect by geography, and Multicultural Toronto English (spoken across neighbourhoods, ethnicities, and networks) becomes a dialect without a region. Define discourse community by shared goals, and a speaker who orients to a community’s norms without sharing its goals falls through the definition. The external criteria work until they don’t, and they don’t at exactly the boundaries where the concepts are needed most. This paper takes a different approach. Instead of defining variety types by their external correlates, I ask what participant expectations are conditioned on and what recurrent pressures keep that conditioning in place. The point is predictive and purpose-relative: those are the conditions under which participants can use part of the profile to infer the rest for the purposes at hand. I argue that three dimensions yield the most predictive lift. Situation: what’s expected in this context. Ascription: what’s expected from someone in this social position. Identification: what’s expected given whose norms you orient to. These map onto register, dialect, and discourse community respectively, though the conditioning space is richer than three axes and the cuts are purpose-relative. They aren’t just labels for sorting tokens after the fact. Recurrent situations, social classifications, and norm centres keep rebuilding them over time. Existing frameworks tend to conflate ascription and identification. Ascription is what you *are treated as* – the social position participants assign you. Identification is who you *position yourself with* – the reference population whose norms you accept, whose approval you seek. Separating them matters because their divergence generates predictions. A working-class kid in law school has an ascribed position but identifies with a new community. Style-shifting, code-switching, and accommodation all operate in this space between ascription and identification. That space is a major locus of sociolinguistic agency. Two theoretical sources ground this framework. From Wiese (2023): the primary unit of linguistic organization isn’t the named language but the communicative situation – the setting of communication as perceived by speakers. From O’Connor (2019): categories are shaped not just by which properties hang together, but by what distinctions help people act and coordinate successfully. Varieties emerge as stable partitions in recurring coordination problems. In plainer terms, the framework separates three familiar pulls on speech: what the setting calls for, what kind of speaker others take you to be, and whose norms you’re trying to match. The situation/ascription/identification (S/A/I) distinction also provides a conditioning-theoretic gloss on what variationists frame in three-waves terms: from macro-categorical modelling of demographic groups, through locally grounded groupings, to third-wave attention to social meaning, style, and stance in interaction (Drummond and Schleef 2016, 50–55). This is a conceptual paper in the philosophy of sociolinguistics, not an empirical study. I don’t present new data about any variety. Instead, I offer a framework for relating concepts already in use and for clarifying what follows from those relations for analysis and interpretation. The argument requires only this: variety categories are worth positing when they support stable, purpose-relative inferences. # Communicative situations as the primary unit Wiese (2023) argues that linguistic regularities don’t live neatly inside named languages; they are organized through communicative situations. A communicative situation (com-sit) is the setting of communication as perceived by speakers: a dynamic, socially interpreted bundle of situational characteristics including participants, activity type, medium, goals, and stance. Her key move is that com-sits are primary. Language categories emerge from com-sit distributions; com-sits don’t derive from pre-existing language borders. Talking with a parent, answering a teacher, and joking with close friends can therefore be three different linguistic environments even when all three happen within one named language. The evidence comes from what Wiese calls “free-range” language settings (contexts less constrained by monolingual or standard ideologies). Urban markets with high linguistic diversity demonstrate translinguistic constructions and emergent local norms. Heritage language settings show com-sit distinctions tied to interlocutors and activities. Multiethnic adolescent peer groups exhibit contact varieties – Kiezdeutsch in Berlin, Multicultural Toronto English in Toronto – that emerge in tightly defined com-sits and expand as their com-sit base widens. Code-mixing in these settings isn’t random switching between two grammars; it’s grammatically structured behaviour within the com-sit, with its own syntactic constraints. The grammar belongs to the interaction, not to either named language. These settings reveal the conditioning structure precisely because named-language ideologies are weak. The same dynamics operate in monolingual settings, but the social overlay of language labels obscures them. Children learn com-sit differentiation early. Wiese describes a child who initially uses a “Daddy register” (German exclusively with father, and only with father), then later broadens to a conventional com-sit definition (German in wider contexts). The idiosyncratic stage highlights that com-sit recognition is a *learned hypothesis*, not a fixed property of the input. This matters for the Bayesian framing developed below: learners track which recurrent pressures matter and update their expectations accordingly. The child’s initial overly narrow conditioning is later revised as evidence accumulates. Wiese’s framework has three immediate consequences for how we think about variety concepts: 1. What we call **register** primarily tracks expectation conditioned on situation – the pressures of the here-and-now. An MTE speaker uses more MTE features with friends than with a teacher; the situation shifts, and the distribution shifts with it. 2. What we call **dialect** primarily tracks expectation conditioned on ascription – durable distributional differences tied to the social positions speakers are ascribed to. The same speaker’s MTE features reflect where they grew up and who they grew up talking to; those classifications keep reimposing a baseline across situations. 3. What we call **discourse community** primarily tracks expectation conditioned on identification – the reference population whose norms the speaker orients to. Whether that speaker treats MTE or Standard Canadian English as the baseline for “normal” depends on who they’re identifying with – and repeated orientation to that norm centre helps keep the profile in place. (I retain “discourse community” for the third term, though the concept here is broader than its technical use in applied linguistics, where it typically implies shared goals, genres, and gatekeeping practices. The core idea (a reference population whose norms one orients to) is closer to what sociologists call a reference group or what Eckert might call a community of practice. I keep the traditional label for continuity with the variety-types literature.) But Wiese’s framework is primarily descriptive. It tells us that com-sits are the primary unit, but not *why* certain partitions of com-sit space stabilize rather than others. For that, we need game theory. # Com-sits as coordination problems O’Connor (2019) uses evolutionary game theory to ask whether the words we use for kinds of things in the world track property clusters. Her answer is sobering. Even when properties cluster together, the categories picked out by those words need not line up neatly with those clusters: different stable partitions can emerge. And when successful action depends more on payoff structure than on property structure, the resulting categories track what matters for coordination rather than the underlying property clusters. The upshot: categories are shaped by what helps coordination, not just by which properties cluster together. This insight transforms how we think about varieties. Communicative situations are payoff structures. When speakers choose among variant forms, they’re solving coordination problems: signalling identity, managing politeness, avoiding conflict, claiming authority, marking solidarity. A job interview, a service encounter, a kitchen shift, and a group chat reward different forms. Consider an MTE speaker navigating a day in Toronto. With neighbourhood friends, using MTE features can secure several payoffs at once: belonging, solidarity, local prestige, and a credible signal of commitment to the group. Those payoffs can be especially strong when the same forms would be costly in more standard-oriented settings. With a teacher, a customer at work, or a job interviewer, suppressing those features can secure a different cluster of payoffs: being taken seriously, reducing interactional friction, and aligning with institutionally rewarded norms. Neither choice is more “authentic”; both are rational responses to different payoff structures. The speaker isn’t code-switching because they’re confused about who they are. They’re code-switching because different situations reward different bundles of social goods. What we call registers, dialects, and discourse communities are the stable partitions that emerge from these recurring coordination problems. They stabilize not because linguistic properties come pre-bundled that way, but because those partitions help people solve recurring social problems. That is why the same speaker can sound natural in one setting and marked in another. This explains why sociolinguistic categories are often messy. Variety boundaries are shaped by payoff relevance (identity, prestige, alignment) rather than by structural clustering alone. Multiple overlapping categories can coexist because different payoff spaces yield different stabilized partitions. Third-wave work in contemporary urban youth settings argues that it’s neither possible nor desirable to treat local practice as instantiating one named variety, since features circulate in ways that don’t match the patterns outsiders expect (Drummond and Schleef 2016, 59). O’Connor (2021) asks how arbitrary a convention is. Some conventions are tightly constrained – only a few outcomes were ever plausible. Others are highly arbitrary – many alternatives could have stabilized. For linguistic varieties, this means: - Some features are **relatively low-arbitrariness**: articulatory/perceptual pressures on phonology, recurrent reduction pathways, some information-structural pressures on word order. These are more tightly constrained by functional pressures. - Others are **relatively high-arbitrariness**: which discourse markers become emblematic, which local vowel-fronting pattern becomes socially salient, which lexical items serve as in-group signals, which exact conventional formulae become salient badges of a practice. These could have been otherwise. Turn-taking, repair, and many politeness routines probably sit in between. They are strongly constrained by interactional function, but still open to variety-specific conventionalization. In MTE, ⟨th⟩-stopping may be relatively low-arbitrariness: it’s one of a limited set of recurrent outcomes for dental fricatives, not a wholly idiosyncratic innovation. But specific MTE discourse markers or slang items are higher-arbitrariness: they could have been different, and their particularity is what makes them useful for in-group signalling. General functional pressures can shape a feature without fixing exactly how it turns out. And a feature that could have been otherwise can still become useful once a community takes it up. O’Connor (2022) takes up a different question: what happens once social categories themselves become salient in coordination games? The equilibria that emerge tend to be inequitable – discriminatory norms emerge and persist even when most participants would prefer otherwise. Even arbitrary social categories (the “green bellies” and “yellow bellies” of her models) produce discriminatory equilibria under standard evolutionary dynamics. Which categories become salient is historically contingent, but once social categories become coordination anchors, the dynamics tend to generate inequity. That matters for the sociolinguistic categories at issue here, and I return to it in Section 7. Stability enters in two ways. First, the same exogenous pressures keep coming back. Kitchens, classrooms, interviews, and peer-group hangouts impose recurrent situational demands; social classifications follow speakers across encounters; and norm centres keep pulling speakers back toward familiar baselines. Second, once a community has converged on a bundle of conventions, individual departures from it can make particular interactions harder, and repeated local costs of that kind help keep the bundle in place. I call the first source exogenous and the second endogenous. Coordination costs is the simplest name for one such endogenous maintenance pressure. O’Connor (2021) formalizes the general point with basins of attraction: some established partitions are more resistant to drift than others. S/A/I name the recurrent pressures that keep re-creating the relevant conditioning state, and those are the pressures that feed the Bayesian model below. Coordination costs do different work: they help explain why a particular carving of that space can become basin-like and resistant to drift once speakers have converged on it. Homeostatic property cluster (HPC) theory offers a useful vocabulary here because it gives a way to talk about categories that are real enough to support prediction without requiring an essence. The point needed here is simple: varieties are worth positing when they sustain stable, purpose-relative projectibility. Whether every such case should count as a full HPC category can be left open. In that vocabulary, an HPC category is a projectible profile with maintenance relations that help keep the profile together. Find some of the profile, and you can make reasonable bets about the rest (Boyd 1991, 1999); for this formulation in my own work, see Reynolds (2026, chap. 6 and 16). Some stabilizers operate on the profile from outside it; others are themselves part of what is clustered. Either way, the category earns its keep by supporting inference across time. For varieties, the profile is social throughout. A courtroom, a neighbourhood network, and a professional norm centre can keep rebuilding a profile from the outside; convergence, accommodation, network density, gatekeeping, and other feedbacks can keep established profiles sticky from the inside. Variety categories can then be stable, purpose-relative, and projectible without becoming essences. If a variety category is doing real work, knowing the conditioning state should help participants predict what forms are likely, and hearing those forms should help them infer that state. Where those predictions degrade, the category is fraying rather than disappearing into noise. # Varieties as emergent partitions Taken together, Wiese and O’Connor give a picture of varieties as stable partitions of com-sit space, shaped by payoff relevance and sustained by coordination dynamics. In HPC terms, the question is then what projectible profile those partitions organize. The answer is participant expectation. ## Participant expectation as the common predictive target What these partitions organize is participant expectation: what participants take to be likely, marked, or normal in interaction. A form can be structurally fine but situationally off, or accentually marked but otherwise unremarkable. These are different kinds of fit, and they are conditioned differently. Register fit is primarily situation-conditioned: what’s expected here and now. Accent fit is often primarily ascription-conditioned – the patterns participants associate with how a speaker is socially positioned – though identification matters when speakers orient to a new group’s norms. Grammaticality is often conditioned by both ascription and identification: what counts as grammatical depends on which grammar or norm set participants take to be in play. Situation asks what context we’re in. Ascription asks how the speaker is being socially classified. Identification asks whose norms the speaker is orienting to. The traditional variety labels differ in which of these pressures does most of the maintenance work: register with situation, dialect with ascription, discourse community with identification. In Wiese, situation is the primary and most explicit organizing dimension. Com-sits are broad enough to absorb socially relevant audience, interlocutor, and indexical information, but they do not separate two further pressures that matter here. Ascription concerns the social position participants take the speaker to occupy. Identification concerns the reference population whose norms the speaker treats as normal. The contrast echoes Cavell’s distinction between knowledge and acknowledgment (Cavell 1979). That distinction matters because the two can align or diverge. Ascription is a classification process, not a metaphysical essence. It can be contested and revised, but it matters because others condition on it (Drummond and Schleef 2016, 52). Identification is the norm orientation the speaker tries to match. This is where sociolinguistic agency enters most clearly: between what speakers are taken to be and whose norms they orient to. These pressures also operate on different timescales. Situation is token-level and fast. Ascription is more durable across encounters because others keep reimposing it. Identification is negotiable but sticky across trajectories because orientation to a norm centre takes time to build and to change. In the formal section below, I represent all three as conditioning variables. In the world, they are different kinds of exogenous maintenance dynamics. Registers shift fast. Dialects are more durable. Identifications can move, but they don’t usually move overnight. ## Where variables live Sociolinguistic variables should appear where these pressures diverge. Where situation, ascription, and identification all favour the same form, we should expect relative invariance. Where they pull in different directions, we should expect variation. That is where the cluster frays, and that is where variables should concentrate. Variables shouldn’t be distributed randomly across the grammar. They should cluster at the divergence points of the conditioning space, where different S/A/I configurations make different forms likely. If variation showed no relation to that divergence structure, the framework would be wrong about where variation comes from. Appendix A illustrates the claim with Kiesling (1998)’s study of the sociolinguistic variable traditionally written as (ING) in a fraternity, where the same variable shows opposite style effects for speakers with different identifications. ## Preserving working distinctions Register, dialect, and discourse community remain useful working concepts. They pick out real patterns that matter for sampling, stratification, and analysis. The framework adds a principled account of *why* they’re different: different recurrent pressures condition and regenerate participant expectation in different ways. If you’re studying regional variation, you’re studying ascription as a recurrent social pressure. If you’re studying style-shifting, you’re looking at how recurrent situations modulate a socially positioned baseline. If you’re studying professional socialization, you’re tracking how identification shifts and then keeps regenerating a new baseline. The framework tells you what those practical distinctions are tracking. ## Relation to three-wave variationism The three waves of variationist sociolinguistics can be read as successive discoveries of recurrent pressures on variation. First-wave work found ascription: macro-sociological categories (class, sex, ethnicity) condition variant rates (Drummond and Schleef 2016). Second-wave work found situation: locally grounded groupings and social networks condition behaviour in ways that demographic categories miss. Third-wave work found identification: speakers deploy variation agentively for identity work, and social meaning is constructed in interaction through stance and style. Each wave foregrounded its discovery while backgrounding the others. Drummond and Schleef (2016) describe the resulting tension: first- and second-wave work treats identity as stable and imposed; third-wave work treats it as fluid and constructed. Both are real simultaneously, but neither tradition can do formal work with the gap between them. S/A/I holds both at once by making the durable dynamic (ascription) and the agentive dynamic (identification) structurally different parameters whose *divergence* generates specific predictions – not just whose coexistence needs acknowledgment. Indexical fields, enregisterment, and indexical orders already explain much of this picture. They show how variants acquire social meaning, how bundles of forms become publicly typified, and how those links are layered and reworked over time. For Eckert (2008, 2012), a variable has an indexical field: a constellation of ideologically linked meanings whose force becomes more specific in styles. That captures the meaning space around a variant. The question here is narrower. In a given interaction, what keeps one region of that field more available than another? How quickly do those pressures update? How do participants select among them? For Agha (2005), registers are publicly typified through metapragmatic activity. They involve a repertoire, a social range of personae, and a social domain of recognizers and users, and they spread and stabilize through institutional channels. That account is indispensable here. Social range still leaves together two different things: being classified as a kind of speaker and orienting to a kind of speaker. I separate those as ascription and identification. For Silverstein (2003), forms accrue layered social meanings through the dialectic of presupposition and reflexive indexicality. He gives the right diachronic picture. At a given moment, though, which pressure is situational, which is ascriptive, which is identificational, and which variety concept is being maintained thereby? S/A/I spells out that conditioning structure. Given a token in interaction, it asks what participant expectation is conditioned on right now and which recurrent pressures, with which update dynamics, keep that conditioning in place. Once those pressures are separated, the framework can make distributional predictions from their alignment and divergence, as Appendix A illustrates, and connect stabilized patterns to the coordination dynamics discussed in Section 3. It doesn’t replace indexical fields, enregisterment, or indexical orders. It makes explicit one conditioning question those frameworks largely leave implicit. If dialect is primarily ascription-conditioned, stylization, crossing, and second-dialect acquisition need to be kept apart. Stylization and crossing do not usually work by immediately changing how a speaker is classified. They work because participants can hear a departure from the speaker’s baseline as marked performance: playful, affiliative, parodic, challenging, or aspirational. In such cases, the social force lies in the mismatch between the speaker’s usual ascribed position and the forms now being used. Second-dialect acquisition is different. There, sustained mobility, network change, and orientation to a new norm centre can gradually shift both production and uptake, so that some participants begin treating the speaker as belonging to a different category, while others do not. The point is not that ascription tracks behaviour mechanically token by token. It is that dialectal expectation is anchored in social classification, and departures from that anchor become meaningful in different ways depending on whether they are momentary, strategic, ratified, contested, or eventually sedimented into a new baseline. ## Ballroom vocabulary As a schematic example, consider lexical items often associated with Black and Latino queer ballroom culture, such as *shade*, *read*, *tea*, *slay*, and *serving*. These forms emerged in a scene organized by recurrent events, roles, and evaluative practices. Early on, situation and identification likely reinforced each other: using the forms fit ballroom settings, and it fit orientation to ballroom norms. That alignment can loosen once speakers carry the forms beyond any one ball. House membership, scene ties, and durable orientation to ballroom practice can remain socially salient across encounters. In that phase, identification can keep the forms available even when the original situation is absent. Broader circulation changes the conditioning again. As the forms spread through media and uptake beyond ballroom itself, ascription becomes more salient. The same token can now be heard differently depending on who the speaker is taken to be, whether the use is treated as ordinary, affiliative, performative, or appropriative, and whether participants recognize the relevant norm centre at all. Situation therefore updates fastest. Identification travels more slowly across encounters. Ascription is slower still, because other participants keep reimposing it. Ballroom vocabulary is not a pure case of one anchor. It begins where situation and identification reinforce each other, then moves into wider circulation, where situational, identificational, and ascriptive pressures come apart. In that respect, it helps answer the questions raised in Section 4.3: what keeps one region of the field available, how quickly different pressures update, and why the same form can be heard differently in different S/A/I configurations. # Bayesian formalization The game-theoretic account explains *why* certain partitions stabilize. The Bayesian account explains *how* learners track the exogenous pressures that condition them. The two are complementary: games explain population-level dynamics; Bayes explains individual learning. Entrenchment and coordination create analogous departure costs at different scales. In individuals, entrenched patterns are easier to reuse than to abandon (Bybee 2010; Schmid 2020). In groups, established conventions are easier to coordinate on than to depart from. The mechanisms differ, but both deepen an existing attractor. The core idea is simple: learners keep track of which forms are expected in which circumstances, and they update their expectations as they accumulate experience. The formal machinery below makes this precise. The equations represent those recurrent exogenous pressures and their hierarchical structure as conditioning variables. They do not represent endogenous coordination costs directly. Those belong to the population-level maintenance story: they help explain why the same conditioning patterns remain socially stable enough to be learned in the first place. The equations state the conceptual claim precisely rather than fit a model to data. The contribution is to identify S/A/I as the pressures that parameterize participant expectation and to show how they can be represented formally. A computational model would require specifying priors from data, fitting parameters, and comparing predictive accuracy against alternatives. That’s the natural next step, but it falls outside this paper. The specific distributional choices below are illustrative rather than mandatory; what matters is the dependency structure they make explicit. The glosses state the argument in prose, while the equations make its commitments explicit enough for precise evaluation. ## Setup Consider a choice point (a constructional niche) where speakers choose among alternatives that perform roughly the same communicative function. Think of first-person *mans* vs. *I* – a niche where some MTE speakers have an option that Standard Canadian English doesn’t generally make available (Denis 2016). An MTE speaker might say *mans made it* with friends but *I made it* with a teacher. The form varies by situation, but this particular contrast is most available to speakers who have access to the MTE option. For each utterance, we can note: - What variant the speaker used - The situational context (formal meeting? casual chat?) - How the speaker is being socially positioned (where they grew up, how others are likely to classify them) - Whose usage the speaker seems to be orienting to In notation: $`Y`$ is the observed choice, $`S`$ captures situation (the here-and-now), $`A`$ captures ascription (how the speaker is being socially classified), and $`I`$ captures identification (whose norms the speaker orients to). Each $`\theta`$ is a probability vector over the available variants: $`\theta_s`$ is the variant distribution for situation $`s`$, $`\theta_a`$ for ascription $`a`$, and so on. $`\text{Categorical}(\theta)`$ just means that one of those variants is chosen according to the probabilities in $`\theta`$ (Gelman et al. 2013, sec. 3.4). The learner’s task is to figure out: given all this, how likely is each variant? ## Representing the three dimensions as conditioning variables Each dimension plays a distinct role. Conceptually, each is a recurrent pressure from outside the local bundle. In the notation below, I represent that pressure as a conditioning variable. For readers comfortable with the notation, I’ll give the formal statement; the gloss follows. If you prefer, you can read only the glosses and lose none of the conceptual point. **Situation**: If you know the context, you can predict which variants are likely. ``` math \begin{equation} Y \mid S = s \sim \text{Categorical}(\theta_s) \end{equation} ``` Gloss: “The distribution over variants depends on the situation.” With a teacher, a speaker might favour *I*, hoping to come across as professional; with friends, *mans* might feel more natural, signalling solidarity. **Ascription**: If you know how the speaker is being socially ascribed, you can predict their baseline tendencies. ``` math \begin{equation} Y \mid A = a \sim \text{Categorical}(\theta_a) \end{equation} ``` Gloss: “The distribution over variants depends on how the speaker is being socially positioned.” A speaker socially positioned as Multicultural Toronto English might show higher rates of ⟨th⟩-stopping, first-person *mans*, or certain intonation patterns – features that track where they grew up and who they grew up talking to, and that persist across situations. **Identification**: If you know whose norms the speaker is orienting to, you can predict what they’ll produce. ``` math \begin{equation} Y \mid I = i \sim \text{Categorical}(\theta_i) \end{equation} ``` Gloss: “The distribution over variants depends on whose norms the speaker is trying to align with.” An MTE speaker who orients to Standard Canadian English norms – perhaps hoping that sounding “standard” will help professionally – might suppress *mans* even with friends. One who identifies strongly with the MTE community might not. The question isn’t what situation they’re in or where they’re from, but whose norms they’re treating as authoritative. In practice, all three matter at once. The full model conditions on situation, ascription, and identification jointly: ``` math \begin{equation} Y \mid S{=}s,\, A_{p,s}{=}a,\, I{=}i \;\sim\; \text{Categorical}(\theta_{s,a,i}) \end{equation} ``` Gloss: “The distribution over variants depends on all three dimensions jointly.” This isn’t three independent effects added together; it’s a single distribution parameterized by the full S/A/I configuration. Note that ascription is indexed to the participant-speaker pair ($`A_{p,s}`$), because different participants may ascribe the same speaker differently – ascription is a property of the perception, not the person. If you average over $`A`$ and $`I`$, you get the register-conditioned slice of the joint distribution. If you average over $`S`$ and $`I`$, you get the dialect-conditioned slice. If you average over $`S`$ and $`A`$, you get the discourse-community-conditioned slice. The three “variety types” are different ways of looking at one joint distribution, not three separate models. ## Hierarchical structure These dimensions interact. The formal question is not whether each S/A/I cell gets its own probability vector; it does. The question is which cells should borrow strength from which others. In Gelman’s terms, units that the model does not distinguish are treated as exchangeable, and grouped units are modeled as partially exchangeable through a hierarchy (Gelman et al. 2013, sec. 5.2). Here, encounters that share an S/A/I configuration can be treated as exchangeable for estimating the cell distribution, and cells that share ascription and identification can be partially pooled rather than estimated in isolation. Within a conversation, order still matters, so the sequential updating just discussed takes over. One standard Bayesian way to represent the cross-encounter hierarchy is: ``` math \begin{equation} \theta_{s,a,i} \sim \text{Dirichlet}(\boldsymbol{\alpha}_{a,i}) \end{equation} ``` Here $`\boldsymbol{\alpha}_{a,i}`$ is a higher-level vector over variants for speakers with ascription $`a`$ and identification $`i`$. If the name “Dirichlet” is unfamiliar, the substantive point is simple: S/A/I cells with the same ascription and identification are partially pooled toward a shared baseline rather than estimated as wholly unrelated cases. This treats ascription as relatively durable across interactions while leaving room for participant-relative assignments in any particular encounter. Situation enters at the token level, but its effect can vary by identification – allowing the model to learn that situational sensitivity itself differs depending on whose norms a speaker orients to. The choice of prior family is incidental; the point is the nesting. For quantitative work, this structure matters. Tokens from the same speaker aren’t independent, so a fitted model would ordinarily add speaker-level varying effects, and possibly perceiver-level ones as well. Similar ascriptions can yield similar structure in the observed distribution, and situational effects operate within those nested levels. The dimensionality of $`\theta`$ depends on the variable and the community: a binary variable in a community with three ascription groups, two identification orientations, and two situational levels yields 12 $`\theta`$ vectors before any speaker- or perceiver-level structure is added. The framework doesn’t prescribe the granularity – it identifies which pressures need representation. Identification shapes production, and production signals identification. The reciprocity is the homeostatic feedback loop the framework predicts. In plain terms, today’s inferences become tomorrow’s expectations: once the model is time-indexed, the posterior at $`t`$ becomes the prior at $`t{+}1`$. This is a standard property of sequential Bayesian updating in cognitive science and iterative learning models in the language evolution literature. The speaker’s production model and other participants’ inference models are distinct: the speaker conditions on their model of how they’ll be ascribed; other participants infer ascription from what they observe. The two are linked by observation across time, not collapsed into one. The $`A_{p,s}`$ subscript reflects this: ascription is perceiver-relative, and speakers’ and other participants’ models converge through repeated interaction, not by sharing a single representation. ## Indexicality as inference from form to context This isn’t a new inferential architecture. RSA-style work in pragmatics and more recent work on social meaning already model the move from linguistic form to speaker and context inferences (Frank and Goodman 2012; Beltrama 2020). The present claim recasts that familiar idea in variationist terms. If certain forms are more likely in certain S/A/I configurations, hearing a form gives evidence about which configuration is in play. - If *mans* is more common in casual, in-group contexts, then hearing *mans* suggests that kind of situation. The form indexes situation. - If *mans* is more common among speakers likely to be ascribed to an MTE background, hearing it suggests that ascription. The form indexes ascription. - If using *mans* can signal orientation to MTE norms, then using it can signal alignment with that community. The form indexes identification. Because indexical links accrue in ideologically structured bundles, a single variant typically supports a field of partially compatible inferences (stances, personae, macro-categories) rather than a one-to-one mapping, with context selecting which region of that field is foregrounded (Drummond and Schleef 2016, 55–57). Social meaning isn’t an arbitrary add-on. It’s the informational flip-side of production conditioning. Speakers condition forms on context; other participants infer context from forms. Convention can also be understood as reproduced precedent. Posterior expectations are the proximate cognitive mechanism by which precedents get weight: if a form has been used successfully in contexts like this, the learner’s model gives it higher probability in similar contexts. Ascription and identification function as cues for which precedents to treat as the relevant model set. The Bayesian layer models how participants reproduce those conditioning patterns in learning and use; it doesn’t replace the coordination story that helps keep them socially stable. Operationalizing these dimensions takes different kinds of evidence. Situation is relatively tractable – it can be coded from observable features of the interaction. Ascription is harder but doable – speakers can be grouped by region, ethnicity, class, or other sociological categories, ideally with attention to how they’re actually perceived by others. Identification is the practical bottleneck. If it’s inferred from the very linguistic choices being modelled, the analysis is circular. Useful strategies have to measure orientation independently: network analysis can estimate who interacts densely with whom; participation histories can track durable involvement in scenes, institutions, or mobility trajectories; experiments can collect self-placement or norm judgments before any target tokens are heard. None is perfect, but each breaks the worst circle. Appendix A demonstrates one viable strategy: Kiesling (1998)’s discourse-analytic classification of power orientation functions as an independent measure of identification, and it does explanatory work that speaker demographics alone cannot. The inferential model isn’t ideologically neutral. Priors on ascription are shaped by ascribers’ social positions and by ideological structures organized around race, class, sex, age, and other such axes. But the raciolinguistic point is sharper than biased priors alone. Flores and Rosa’s white listening subject shows that asymmetry can also enter in what listeners take themselves to have heard in the first place (Flores and Rosa 2015; Rosa and Flores 2017). In Bayesian terms, that is not only a prior over $`A`$; it can also alter the likelihood linking forms to ascriptions. The present framework can register that possibility through perceiver-relative weighting and category availability, but it doesn’t theorize it in full. # Participant judgment, normativity, and unacceptability To what extent does a form sound odd, marked, or wrong? One answer in this framework is: to the extent that it’s unlikely given the expectations a participant brings to the interaction. The judgment “that’s not how we talk” means: “given who I think you are, what situation I think we’re in, and whose norms I think apply – I wouldn’t expect that form.” But we need to distinguish several things that “sounds wrong” might track. Appropriateness[^1] is the broad category: the coupling between a form and the values it signals. Grammaticality is one species of appropriateness: primarily the fit between morphosyntactic form and structural meaning. On this account, that fit isn’t the same thing as a participant’s feeling about it. A sentence can be well-formed and still feel wrong, as in garden paths, or ill-formed and still feel fine, as in comparative illusions (Wellwood et al. 2018). Felt ungrammaticality is therefore an imperfect cue to grammatical structure, not a transparent report of it (Reynolds 2024). Other appropriateness channels – register fit, accent, lexical precision – have their own form–value couplings and their own ways of misfiring. The S/A/I framework developed here is about what all these channels are *indexed to*: grammaticality is variety-indexed (conditioning on whose grammar), register fit is situation-indexed, and so on. Acceptability is the phenomenology: what a participant actually experiences, informed by all these channels and by processing factors that can mislead any of them. And correctness is the prescriptive overlay: what gatekeepers enforce, what gets codified and moralized – often an ideologized version of one variety’s appropriateness norms imposed as if universal. This makes acceptability, and especially the appropriateness component, relative to conditioning. The same form can be fine under one set of expectations and jarring under another. ⟨th⟩-stopping is unremarkable among MTE speakers but may be noticed – and judged – by participants conditioning on Standard Canadian English. A double modal is unremarkable in some Southern American varieties and ungrammatical in others. The form didn’t change; the conditioning in play did. This account of normativity aligns with what Pullum (2019) calls the constitutive view of grammatical rules. Pullum argues that grammatical constraints aren’t regulative – they don’t tell you what you *ought* to do in some moral sense. They’re constitutive: they define what counts as participating in a particular practice. Departing from English grammar isn’t like breaking a traffic law – it’s more like making an illegal move in chess. You’re not violating a duty; you’re just not playing the game. But unlike chess, there’s no rulebook. Grammatical constraints are emergent (no one stipulated them), variety-indexed (different communities play by different rules), and tacit (speakers follow them without being able to state them). The constitutive force is real – deviate far enough and you’re not playing the game others are playing – but there’s no official chess federation to consult. The S/A/I framework helps specify which constraint set or expectation space is in play. This framing accommodates several observations: 1. **Gradient judgments**: probability is continuous, so judgments can be gradient. 2. **Context-dependence**: different conditioning yields different expectations. 3. **Norm conflicts**: situations can bring speakers into conflicting expectations – especially when ascribed identity norms clash with situational demands. An MTE speaker in a job interview faces exactly this: forms natural to their ascription may be marked in that situation. 4. **Innovation**: a novel form becomes learnable when participants can attach it to a distinct recurring interactional pattern rather than treating it as random noise. # Coordination, salience, and inequity The framework developed in the preceding sections has implications that are uncomfortable but important. Once we understand varieties as coordination equilibria sustained by category salience, we have to face the fact that these equilibria can be – and often are – inequitable. ## Standard language ideology as coordination Consider a phenomenon sociolinguists know well: standard language ideology. Despite decades of linguists explaining that all varieties are systematic and expressive, the prestige hierarchy hasn’t been paying attention. An MTE speaker applying for a professional job in Toronto may find their speech (⟨th⟩-stopping, distinctive intonation, certain discourse markers) working against them. Students are still corrected for “non-standard” grammar. Why? The game-theoretic answer is that once “standard” and “non-standard” become salient ascriptive categories, they can function as coordination signals. Under informational scarcity, employers can treat “standard”-sounding speech as a crude hiring proxy. The cue is taken to add predictive value when other information is thin, because it seems to compress background facts about education, class trajectory, or institutional fit. Once many employers treat it that way, ignoring it can look risky even to those who know it is a poor linguistic measure. The signal matters most when information is scarce. When employers know applicants well, or have richer evidence about how they will do the job, its marginal value drops. The minimal conditions for this are simple: individuals recognize salient ascriptive categories, condition behaviour on them, and adapt in ways that improve their own payoffs. Once these conditions are met, coordination can stabilize on unequal outcomes, even when most participants would prefer a fairer arrangement. Game-theoretic language can make centuries of institutional sedimentation sound like momentary choice situations, but that’s misleading. The categories that structure coordination, including race, class, region, and language, have deep histories. The prestige attached to “standard” speech is historically produced through standardization and through ideologies that link language to authority and identity (Joseph 1987, 2004; Milroy and Milroy 1999). Coordination helps sustain those categories once they are salient, but it doesn’t explain how they became salient in the first place. That work belongs to history and institutions. ## What coordination sorts on Coordination is not a pathology confined to dominant groups. Any community can coordinate on signals that are cheap and informative for its own members and costlier or less legible for outsiders. In one interactional field, ⟨th⟩-stopping may be a low-cost, high-payoff way of sounding local or aligned; in another, the same feature may carry social cost. Ballroom lexis has the same structure. These cues are often anchored locally in situation or identification before they are reused more widely as cues for outsider ascription. What is easy and meaningful inside a scene can travel very differently outside it. That doesn’t excuse discrimination. It locates the asymmetry more precisely. The standard-speaking majority is not the only group coordinating on signals. The asymmetry is that majority coordination is backed by institutions that control hiring, schooling, housing, and legitimacy, and that other groups more often bear the burden of accommodation. In O’Connor’s sense, once salient ascriptive categories enter the game, unequal outcomes are a persistent tendency. The question is what coordination sorts people by. It can anchor on assigned social position, or on properties that are more changeable and more relevant to the task at hand. The cues, incentives, and consequences participants face are historically made and can be remade. So appeals alone won’t undo the pattern. If access is conditioned on signals that track ascribed position, coordination can simply re-anchor on whatever cues remain available. ## The role of category salience Because coordination is not confined to dominant groups, the next question is not whether salience exists, but where it travels and who can use it. A category can be salient inside a scene without yet functioning as a broad institutional sorting device. Trouble grows when a signal becomes portable: recognizable across settings, legible to outsiders, and easy to recruit for gatekeeping. That is one path by which a cue that was locally anchored in situation or identification becomes available for ascription-conditioned sorting. That helps separate two things that are often run together. Salience can support recognition, solidarity, and in-group coordination. The same salience can also support sorting once outsiders or institutions can use it as a shortcut. Ballroom lexis shows the structure clearly. Forms that begin as scene-internal markers can spread outward; once they do, they are no longer only resources for mutual recognition inside the scene. They also become available for imitation, policing, commercialization, or exclusion. In S/A/I terms, cues that first help participants recognize a situation or an identification can later be reused by outsiders as cues for what kind of person a speaker is taken to be. In HPC terms, that is purpose-relative projectibility. Inside the scene, the same cue can license inferences about solidarity, stance, or shared norms. Outside it, the cue can license cruder inferences about who a speaker is taken to be. The signal hasn’t stopped being informative. The purpose of the inference has changed, and institutions make one purpose much more consequential than the other. Salience doesn’t have a single political valence. Reducing it may weaken one anchor for discrimination, but it may also erode locally valuable forms of coordination and identity. Increasing it may support mobilization, but it can also make a category easier to sort on from the outside. The question is not whether salience is good or bad in the abstract. It is where the signal is legible, what work it is doing there, and who has the power to act on it. ## Diffusion, appropriation, and the ecological tradeoff The schematic ballroom example from Section 4 points to a phenomenon often discussed as cultural appropriation: linguistic forms originating in a marginalized community can diffuse to broader populations, losing some of their community-specific indexicality along the way. As those markers spread, three things can happen. The forms lose some of their ability to function as in-group signals. Cues once anchored mainly in situation and identification become available for wider ascription. And the originating community’s conditioning can be overwritten in broader usage. That is diffusion: the same process by which any innovation spreads. This creates a real normative tension. Reduced salience can remove one anchor for inequity: if “ballroom community member” is less legible through the forms, discrimination against that community loses one cue to coordinate on. But the same salience also supports community identity, solidarity, in-group coordination, and pride. When markers diffuse and lose their indexical force, the community loses a coordination resource it built. Communities aren’t wrong to value salience, and they aren’t wrong to recognize what salience enables. The framework doesn’t resolve the dilemma. It makes its structure visible. The same issue can be stated ecologically. Category diversity has costs and benefits. Linguistic innovation often emerges in marginal communities, and the ballroom example is meant to point to that possibility. Multiple communities mean multiple perspectives and challenges to assumptions. At the same time, coordination is harder with more categories. Conflict is more likely. Transaction costs of navigating difference are real. Monocultures are efficient in stable environments but vulnerable to disruption. Diverse ecosystems are less efficient but more resilient and generative. A system that dissolved all category distinctions would reduce some coordination anchors for discrimination, but it would also lose the niches where innovation emerges. The question is not whether appropriation is good or bad in the abstract. It is how much category differentiation a system can sustain, given the tradeoff it brings. More differentiation can support innovation, solidarity, and local coordination. It can also increase friction, conflict, and opportunities for discriminatory sorting. That is a collective problem, not one minority communities should be asked to solve on their own. # Conclusion Register, dialect, and discourse community are best treated as different ways participant expectation is conditioned and maintained. Register is rebuilt mainly by recurrent situations, dialect by recurrent ascriptions, and discourse community by recurrent identifications. Separating ascription from identification matters because the gap between them is where much sociolinguistic agency lives. Wiese supplies the unit, the communicative situation. O’Connor supplies the dynamics: recurring coordination problems make some partitions sticky, and salient ascriptive categories can also stabilize inequity. The Bayesian layer supplies the participant-level implementation: participants track recurrent situations, ascriptions, and identifications as conditioning variables and revise their expectations as experience accumulates. The Kiesling case shows why this matters analytically. Once ascription and identification are pulled apart, what looks like one speaker effect becomes a structured divergence in orientation. The payoff is predictive and purpose-relative. A variety category earns its keep when part of the profile lets participants infer the rest for the purposes at hand. Knowing enough about situation, ascription, and identification should help predict likely forms, and hearing those forms should help infer the conditioning state. The same category can support different inferences for different parties: in-group solidarity for participants who share the identification, and cruder sorting for outsiders who read the forms as ascriptive cues. Both are real projectibility; what differs is the purpose, and institutions make one purpose much more consequential than the other. That is enough for the argument here, whether or not every such case should count as a full HPC category. That’s what the framework adds to the traditional variety concepts: a common logic, sharper predictions, and a clearer account of what those concepts are tracking. # Worked example: word-final *-ing* variation in a college fraternity ## The variable and the published analysis Kiesling (1998) recorded 37 hours of interaction in an American college fraternity and analysed the sociolinguistic variable traditionally written as (ING) – the alternation between velar \[ɪŋ\] and alveolar \[ɪn\] in words like *going*/*goin’*. He coded 1,098 tokens across 11 speakers for speaker, activity type (meeting, socializing, interview), following phonological environment, and grammatical category. All speakers were white fraternity brothers from roughly similar class backgrounds. The overall pattern was conventional: socializing favoured \[ɪn\] (.72 Varbrul probability), meetings disfavoured it (.30). Most speakers shifted in the expected direction – more standard \[ɪŋ\] in the formal meeting context. But the speaker $`\times`$ activity cross-tabulation (Kiesling’s Table 3) reveals striking divergence. Mack – the clearest standard-pattern case – went from 73% \[ɪn\] socializing to 13% in meetings. Speed, by contrast, barely shifted at all: 95% to 82%. And Mick *reversed*: 67% to 71%, using slightly *more* \[ɪn\] in the formal setting. Waterson was categorical \[ɪn\] in meetings, though with only four tokens – two of which were *fuckin’*, a lexeme that is almost categorically \[ɪn\] for all speakers and whose frequency in formal fraternity meetings perhaps tells us something about the formality of formal fraternity meetings. For progressive verb forms specifically (Kiesling’s Figures 7–8), the reversal is even clearer: Speed and Mick both *increase* their \[ɪn\] in meetings, while other speakers decrease it. ## Mapping onto S/A/I Kiesling’s factor groups map directly onto S/A/I: - Situation = activity type (meeting vs. socializing vs. interview). This is the style dimension. - Ascription = social position, operationalized here through speaker demographics (region, class, ethnicity, fraternity membership). Under that coding, all speakers are effectively similar for present purposes: white fraternity brothers at the same university. - Identification = whose norms the speaker orients to. This is where Kiesling’s analysis does its work. He shows through detailed discourse analysis that the reversers orient to *physically powerful, working-class* alignment roles (hard work, confrontation, bodily authority), while the standard-pattern speakers orient to *structurally powerful* roles (institutional hierarchy, formal authority). The original Varbrul analysis codes speaker as a single factor group, conflating ascription and identification. Because all speakers share the same ascription under this coding, the speaker effect largely tracks identification – but the analysis can’t see this because it doesn’t separate A from I. ## Predictions S/A/I generates three predictions for this case: 1. **A/I separation:** Speakers with the same ascription but different identification should show different (ING) rates. This is borne out in the published patterns: Speed, Mick, and Waterson have the same A as Mack, Hotdog, Ram, etc., but different I, and the distributions diverge dramatically – especially in the meeting context. 2. **S $`\times`$ I interaction with reversed sign:** The magnitude *and direction* of situational shifting should depend on identification. For structural-power identifiers, formality increases \[ɪŋ\] (standard Labovian pattern). For physical-power identifiers, formality increases \[ɪn\] (reversed). This is also borne out in Kiesling’s Figure 8: the slopes run in opposite directions for the two groups. 3. **Divergence-point prediction:** (ING) is expected to be variable in this community because S, A, and I make divergent predictions. S says: use more \[ɪŋ\] in meetings. A says: no prediction (everyone has the same A). I says: it depends on whose norms you orient to. On this account, the variable sits at a divergence point in the conditioning space rather than in the more stable core. ## What would falsify the framework If independently measured identification added no predictive power over ascription and situation – if all fraternity members showed the same style-shift pattern regardless of their normative orientation – the A/I distinction would have no empirical content for this variable. Kiesling’s data set shows the opposite: identification is doing explanatory work that ascription cannot. A standard mixed-effects model with style (S) and demographic group (A) as predictors would predict the wrong direction for Speed, Mick, and Waterson. Adding identification (I) – operationalized through Kiesling’s discourse-analytic classification of power orientation – should fix the direction of the prediction. This demonstration shows *consistency*: the framework re-describes published data in a way that reveals structure the original analysis treated as unexplained speaker variation. Consistency is necessary but not sufficient. A stronger test would be *prediction*: identify a constructional niche where S/A/I divergence is high and predict in advance that it should be highly variable – then go look. The divergence-point prediction (Section 4.3) makes this possible in principle: find a niche where situation, ascription, and identification make conflicting predictions, and the framework says that’s where variation should live. Testing this against corpus data is the natural next empirical step. # Acknowledgements This paper was drafted with the assistance of Claude Code (Opus 4.6) and Codex 5.4. I have reviewed and revised all content and take full responsibility for the final text.
Agha, Asif. 2005. “Voice, Footing, Enregisterment.” *Journal of Linguistic Anthropology* 15 (1): 38–59.
Beltrama, Andrea. 2020. “Social Meaning in Semantics and Pragmatics.” *Language and Linguistics Compass* 14 (9): e12398. .
Boyd, Richard. 1991. “Realism, Anti-Foundationalism and the Enthusiasm for Natural Kinds.” *Philosophical Studies* 61 (1–2): 127–48. .
Boyd, Richard. 1999. “Homeostasis, Species, and Higher Taxa.” In *Species: New Interdisciplinary Essays*, edited by Robert A. Wilson. MIT Press. .
Bybee, Joan. 2010. *Language, Usage and Cognition*. Cambridge University Press. .
Cavell, Stanley. 1979. *The Claim of Reason: Wittgenstein, Skepticism, Morality, and Tragedy*. Oxford University Press.
Denis, Derek. 2016. “A Note on *Mans* in Toronto.” *Toronto Working Papers in Linguistics* 37. .
Drummond, Rob, and Erik Schleef. 2016. “Identity in Variationist Sociolinguistics.” In *The Routledge Handbook of Language and Identity*, edited by Siân Preece. Routledge. .
Eckert, Penelope. 2008. “Variation and the Indexical Field.” *Journal of Sociolinguistics* 12 (4): 453–76. .
Eckert, Penelope. 2012. “Three Waves of Variation Study: The Emergence of Meaning in the Study of Sociolinguistic Variation.” *Annual Review of Anthropology* 41: 87–100. .
Flores, Nelson, and Jonathan Rosa. 2015. “Undoing Appropriateness: Raciolinguistic Ideologies and Language Diversity in Education.” *Harvard Educational Review* 85 (2): 149–71. .
Frank, Michael C., and Noah D. Goodman. 2012. “Predicting Pragmatic Reasoning in Language Games.” *Science* 336 (6084): 998. .
Gelman, Andrew, John B. Carlin, Hal S. Stern, David B. Dunson, Aki Vehtari, and Donald B. Rubin. 2013. *Bayesian Data Analysis*. 3rd ed. Chapman & Hall/CRC. .
Joseph, John E. 1987. *Eloquence and Power: The Rise of Language Standards and Standard Languages*. Blackwell.
Joseph, John E. 2004. *Language and Identity: National, Ethnic, Religious*. Palgrave Macmillan.
Kiesling, Scott Fabius. 1998. “Men’s Identities and Sociolinguistic Variation: The Case of Fraternity Men.” *Journal of Sociolinguistics* 2 (1): 69–99. .
Milroy, James, and Lesley Milroy. 1999. *Authority in Language: Investigating Standard English*. 3rd ed. Routledge.
O’Connor, Cailin. 2019. “Games and Kinds.” *The British Journal for the Philosophy of Science* 70 (3): 719–45. .
O’Connor, Cailin. 2021. “Measuring Conventionality.” *Australasian Journal of Philosophy* 99 (3): 579–96. .
O’Connor, Cailin. 2022. “Why Natural Social Contracts Are Not Fair.” Unpublished manuscript.
Pullum, Geoffrey K. 2019. “Formalism, Grammatical Rules, and Normativity.” In *Form and Formalism in Linguistics*, edited by James McElvenny. Language Science Press. .
Reynolds, Brett. 2024. “Characterizing Grammaticality: A Draft Framework.” .
Reynolds, Brett. 2026. “Words That Won’t Hold Still: How Linguistic Categories Work.” .
Rosa, Jonathan, and Nelson Flores. 2017. “Unsettling Race and Language: Toward a Raciolinguistic Perspective.” *Language in Society* 46 (5): 621–47. .
Schmid, Hans-Jörg. 2020. *The Dynamics of the Linguistic System: Usage, Conventionalization, and Entrenchment*. Oxford University Press. .
Silverstein, Michael. 2003. “Indexical Order and the Dialectics of Sociolinguistic Life.” In *Language & Communication*, vol. 23.
Wellwood, A., R. Pancheva, V. Hacquard, and C. Phillips. 2018. “The Anatomy of a Comparative Illusion.” *Journal of Semantics* 35 (3): 543–83. .
Wiese, Heike. 2023. *Grammatical Systems Without Language Borders: Lessons from Free-Range Language*. Language Science Press. .
[^1]: I use “participant expectation” rather than “appropriateness” as the framework’s organizing concept, because “appropriateness” has been argued to embed normative assumptions about whose communicative practices count as situationally correct. The model-internal concept is what a participant’s model predicts given inferred S/A/I, not what is normatively correct. In this section, “appropriateness” appears in its technical sense – probabilistic, participant-relative, and explicitly distinguished from prescriptive “correctness”.