Brett Reynolds

Why do humans alone develop language? I argue that evolution selected a binding bias, not specialized syntactic machinery: a hyperprior treating addressed, multi-timescale communicative events (infant-directed speech, with eye contact and rhythmic prosody) as reliable evidence. Under active inference, the same precision profile biases both perceptual uptake and vocal convergence; no separate imitation system is required. Human echolocation illustrates the logic: trained blind individuals navigate by self-generated clicks, a complex skill from generic neural systems; complexity doesn’t entail selection. The account explains language’s universality and the combinatorial limits of other vocal learners. Three falsifiable tests: a neonatal preference study, a computational replay varying the gain profile, and a newborn functional near-infrared spectroscopy (fNIRS) assay. Selection targeted what we notice, not what we compute.

Amie Thomasson’s metaphysics treats ontological and modal disputes through rules of use, discourse roles, and conceptual revision. I accept this function-first turn and detach it from the systemic functional linguistics used to motivate it in Rethinking Metaphysics. The needed premise is role pluralism: categories do different jobs in different practices. Natural-kind cases show why role pluralism needs a projectibility constraint. The cases of fish, jade, and race don’t separate conceptual engineering from projectibility realism, since Thomasson can accommodate their empirical corrections. The harder case arises when a cat- egory keeps coordinating a field after losing the projection that gave it kind-standing. A pro- jectibility realist account adds the missing demotion rule: once a field and projection are fixed, stable projection failure can defeat a useful category profile.

Thomasson’s (2025) Rethinking Metaphysics brings a grammatical debate into metaphysics. Thomasson uses systemic functional linguistics to explain why noun-term recastings such as the property of redness are useful without behaving like observational object talk. This paper accepts the discourse-role insight but detaches it from Halliday’s grammatical architecture. SFL explains why the recast formulations are worth having; easy ontology licenses the inferences. For linguists, the lesson is broader than syntax: some inferences and expectations follow from linguistic description, but not every expectation imported from familiar models of reference, objecthood, or category.

Anti-essentialist theories explain how categories can be real without essences. Homeostatic property cluster theory gives a Boydian answer, but homeostatic now covers achievements that should be kept apart: stable projectible profiles, network order, maintenance, and corrective control. This paper gives anti-essentialist kind theory a stricter support-grade and demotion vocabulary. A category earns standing by tracking a projectibility profile: a worldly pattern whose partial observation licenses inferences under specified conditions. Profile analysis asks what the category projects, what supports the projection, what orders and stabilizes it, and where the claim should be demoted or rejected. Homeostatic control is a subcase, not the default.

Anti-essentialist theories explain how categories can be real without essences. Homeostatic property cluster theory gives a Boydian answer, but homeostatic now covers achievements that should be kept apart: stable projectible profiles, network order, maintenance, and corrective control. This paper gives anti-essentialist kind theory a stricter support-grade and demotion vocabulary. A category earns standing by tracking a projectibility profile: a worldly pattern whose partial observation licenses inferences under specified conditions. Profile analysis asks what the category projects, what supports the projection, what orders and stabilizes it, and where the claim should be demoted or rejected. Homeostatic control is a subcase, not the default.

Recent work suggests that some large language model representations have content or reference. Grounding can secure either without supplying live routes for correction. This paper asks what follows from that gap. An output is answerable when discrepancies can affect what a target- and task-specific arrangement produces, accepts, or withdraws. The arrangement has corrective control only when live, sufficiently independent routes can detect and repair fresh discrepancies. A route profile records which routes constrain the arrangement and how they are related. Those profiles support analysis of truth-tracking: patterned support for representational success. Language models are the pressure case; text-only arrangements provide a task-relative limiting case. Text-trained models inherit patterns of testimony, coherence, and prior correction. Where target-sensitive correction survives training, these can supply derivative answerability (inherited constraint); live answerability is the relation supplied by a current route for fresh discrepancies. Fluent failures should follow when a task requires independently informative access to the facts. Self-consistency, retrieval, tools, code execution, multimodal input, and feedback should help selectively. Route-by-task interactions test the distinctions. The decomposition's empirical burden is to predict held-out route–task combinations or improve intervention choice without conceptual refitting. Surface improvement and truth-tracking improvement can come apart.

Homeostatic property cluster theory gave natural-kinds realism a powerful anti-essentialist form, on which a kind can be real because its properties cluster through mechanisms rather than because it has an essence. In the HPC-descended debate, the homeostatic component can be made to carry too much weight when asked to cover several achievements that aren't the same: reliable co-occurrence, network order, maintenance, and feedback-sensitive control. This paper treats projectibility as the target of kind claims and offers a diagnostic ladder: identify the projection target, then ask whether a stable profile supports it, whether network order makes the profile non-accidental, what if anything maintains the profile, and whether that maintenance is corrective control. On the disciplined use proposed here, homeostasis belongs at the last step, where perturbation-sensitive organization preserves a higher-scale relation through lower-scale variation. The Watt governor calibrates control, glycemic regulation tests the strict biological version of the control diagnostic, interjection tokens supply a contrast linguistic case, and monitored agreement production supplies a candidate positive one. The result is a disciplined way of using the framework's insight without turning homeostatic into a loose synonym for stable or projectible.

Social explanation faces a forced choice: treat social statuses as causal patterns with moral labels, or as normative facts whose social uptake is incidental. Neither fits a person effectively hounded for a promise they never made, or a standing that a practice successfully confers while morality rejects it. Status-assignments, I argue, are projectible: observing one licenses defeasible expectations about duty, recognition, and response. From a promise we project duty and reliance; from release, lapse even while demand continues; from judgment, records, enforceability, and preclusion; from discriminatory standing, credibility loss, exclusion, and anticipatory adjustment. Extending Khalidi (2018) on projectible kinds, I model these projections as travelling through a causal-normative network: a typed structure joining causal uptake, recognitional representation, directed status-transition, and non-directed correlative ties within a practice or field. Interventionist tests type the edges, and misrecognition supplies the diagnostic case: the framework separates what a false recognition can still cause from what a wrongfully conferred standing has become. The payoff is an account of how status-assignments can be socially effective, institutionally valid or communally conferred, enforceable or final, recognitionally mistaken, and morally unwarranted in different combinations.

The word language names several distinct objects across the language sciences: capacity, idiolect, communicative practice, population convention, named sociohistorical object, and formal system. Pluralist accounts (especially Santana's distinction among psychological, social, and abstract ontologies) show why monism fails but don't say when a proposed sense licenses inference or when evidence about one sense can travel to another. This paper adds a homeostatic property cluster (HPC) discipline: a sense earns realist standing where its properties are stabilized by identifiable mechanisms and project for a field's purposes; where the mechanisms diverge, the shared label doesn't license transfer. The map distinguishes three categories proper (population convention, named sociohistorical object, abstract formal system) from three neighbourhoods of substantively coupled constituent categories (capacity, idiolect, communicative practice). The payoff is diagnostic. The Casten / Pinker exchange over genomic evidence for spoken-language ability shows how developmental verbal-performance evidence gets read as evidence about a uniquely human language faculty; similar slippages arise when census recognition is read as convention, formal models as cognitive architecture, or analyst diagnostics as the object. The umbrella label remains useful for coordination, but the load-bearing scientific units are the categories and neighbourhoods, not the umbrella.

This paper argues that grammaticality is best understood as a homeostatic property-cluster (HPC) kind anchored in morphosyntactic form–meaning pairings and regimented by community practice. The framework distinguishes objective grammatical status – whether the community licenses a form–meaning pairing – from the subjective feeling of ungrammaticality, which integrates grammaticality with processing costs, social evaluation, and prescriptive norms. Drawing on contemporary metaphysics of science, we show how grammaticality exhibits the key features of HPC kinds: a cluster of co-occurring properties (morphosyntactic well-formedness, semantic compatibility, community entrenchment), homeostatic mechanisms that maintain cluster stability (language acquisition, repair strategies, sociolinguistic policing), and systematic projectability supporting inductive generalization. Processing constraints shape what patterns can be licensed through acquisition and transmission while also affecting how licensed patterns feel during comprehension. This analysis clarifies why grammaticality supports explanatory and inductive generalization despite surface variation and gradience. The HPC framework predicts which phenomena will pattern with grammaticality (morphosyntactic violations, semantic mismatches at the syntax-semantics interface) versus those that won’t (pure phonotactic violations, processing overload). By grounding grammaticality in both cognitive mechanisms and social practices, the account reconciles the apparent universality of particular constraints with the evident community-relativity of others. The framework has implications for methodology (validating judgment tasks while acknowledging their limits), theory (reconciling generative and usage-based evidence), and typology (predicting which “exotic” grammars can exist as stable linguistic systems).

Coelho Mollo and Millière pose the Vector Grounding Problem: how the internal states of large language models could acquire extra-linguistic reference. They answer it with a two-condition teleosemantic account requiring causal-informational coupling to the world and a selection history that gives states the function of carrying that information. This reply accepts those teleosemantic foundations but argues that what the grounding debate calls “referential grounding” is itself a cluster concept. It bundles at least four distinct relations: teleosemantic selection, causal-informational coupling, communicative calibration, and inferential-role structure. Drawing on homeostatic property cluster theory, I replace the binary grounded/ungrounded verdict with a groundedness profile that records where a system sits on each strand. The payoff is explanatory and predictive: the profile decomposes the broader bundle often circulating under the name “grounding” into stabilizers that predict robustness, corrigibility, and characteristic failure under correction, prompt pressure, and domain shift in ways a binary verdict can't.

The debate over whether large language models “really think” reproduces a familiar pattern: a boundary case meets a binary classification. Nefdt (2026) organizes this debate through a 2 × 2 matrix crossing language with cognition, placing LLMs in a “missing quadrant” of language without cognition. But the binary frame undersells his own insights. His hedging on proxies, partial capacities, and qualified attributions points toward something the table can’t express: that language and cognition are cluster concepts with graded membership. Homeostatic property cluster (HPC) theory, with its projection-purpose analysis, offers a better framework. Under the shared label cognition, different projections track distinct property clusters – one centered on mechanism and substrate, another on functional capacity – that converge in humans and co-vary in non-human animals. LLMs break the co-variation, exposing an ambiguity that convergence in biological systems had kept invisible. The debate persists because disputants track different clusters under a shared label, and resists resolution because they treat cluster kinds as if they had essences. But the categories themselves aren’t fixed: LLMs are reshaping what “cognition” means, not just testing whether they have it.

The Sorites paradox arises from treating tolerance – the principle that tiny changes preserve vague predicates – as absolute rather than scale-dependent. Building on recent non-Archimedean treatments of vagueness (e.g. Itzhaki, 2021), this paper offers a particularly simple hyperreal model of scale-sensitive tolerance: changes that are infinitesimally small relative to the current scale preserve predicate application, while accumulated changes can cross boundaries. The model keeps classical bivalent logic throughout and, in a hyperfinite setting, locates sharp cutoffs at hypernatural indices that remain epistemically inaccessible. The contribution is methodological rather than revolutionary: I make explicit a Relative Tolerance postulate stated in terms of fractional change, combine it with a hyperfinite sorites chain that guarantees a least hypernatural cutoff, and package the resulting picture in a way that is accessible without deep background in nonstandard analysis. This shows that familiar commitments – monotonicity, scale-sensitive tolerance, clear positive and negative cases – are jointly satisfiable in a classical setting. The resulting framework is best seen as a scale-sensitive cousin of epistemicism, though the same machinery is compatible with contextualist, supervaluationist, and other approaches.