Abstract
Predictive Processing (PP) is commonly described as a mechanism sketch—an incomplete, primarily mechanistic representation of cognitive processes. While this characterization rightly emphasizes PP’s structural and causal explanatory aspects, I argue that it tends to overlook important functional and normative dimensions that are essential for a comprehensive account of cognitive processes. In response, I claim that PP, while providing mechanism sketches in the standard sense, simultaneously performs a second, normative-functional epistemic role. This expanded view on PP preserves the value of its mechanistic explanation while clarifying the essential role of functional constraints in shaping predictive models. Specifically, I show how PP models postulate explanatory constraints—formal, structural, and functional—that are inherently normative, as they specify the conditions viable mechanisms must satisfy to be biologically plausible and adaptive. Drawing on Levenstein et al. (2023), I situate this account within a pragmatic framework that recognizes the legitimacy of mechanistic, normative, and descriptive theories without reducing one to another. This pluralistic approach enables genuine integration across explanatory levels and points toward more adequate future models in cognitive neuroscience.
Similar content being viewed by others
1 Introduction
Predictive Processing (hereafter PP) is a prominent theoretical framework in contemporary cognitive neuroscience that draws from two traditions:Footnote 1 the Bayesian view of the brain as a probabilistic inference machine (cf. Knill & Pouget, 2004; Oaksford & Chater, 2007) and the idea of unconscious perceptual inference (cf. Gregory, 1980; Helmholtz, 1867). At its core, PP posits that the brain implements a hierarchical generative model that continuously predicts sensory input and updates its states by minimizing prediction errors—differences between expected and actual input (Clark, 2013, 2016; Friston, 2009; Hohwy, 2013). The mechanism by which likelihoods and priors in a model are updated in light of prediction errors (which refer to the discrepancy between the model’s expectations and incoming sensory data) is typically characterized through Bayesian inference, wherein hypotheses are revised as new evidence becomes available. Rather than performing exact Bayesian inference, which is computationally intractable, PP employs variational approximations to optimize an upper bound on the true posterior (Buckley et al., 2017). This results in hierarchical message passing: top-down predictions attempt to suppress bottom-up sensory discrepancies. Prediction errors can be reduced either by updating internal models (perceptual inference) or by acting on the world to fulfill expectations (active inference) (cf. Friston, 2009).
Because the minimization of prediction errors via approximate Bayesian inference can be interpreted as the means by which neural information-processing mechanisms perform variational inference (cf. Badcock et al., 2019), this raises an important question concerning the explanatory status of PP as a theory of cognition—specifically, whether it can account for (at least some) cognitive, perceptual, and motor functions (cf. Clark, 2013, 2016; Hohwy, 2013). In response, many researchers maintain that PP provides a mechanism sketch—an incomplete or elliptical representation of the phenomena to be explained (explananda)—in which certain structural features of a full mechanistic explanation are left unspecified (cf. Piccinini & Craver, 2011). As such, explanations derived from PP are generally regarded as mechanistic in character (cf. Badcock et al., 2019; Förster, 2023; Gładziejewski, 2019; Gordon et al., 2019; Harkness, 2015; Harkness & Keshava, 2017; Hohwy, 2015; Keller & Mrsic-Flogel, 2018; Venter, 2021).
However, ongoing debate surrounds the precise interpretation of this mechanism sketch in the context of predictive architecture: does it represent a single, hierarchically organized causal structure encompassing a broad range of cognitive functions, or rather a collection of partially independent mechanisms that can be explained through the common inferential schema posited by PP? (Gładziejewski, 2019, pp. 665–667; Hohwy, 2015, pp. 9–10). Regardless of how, this issue is ultimately resolved, some researchers argue that PP, even in its current mechanistic form, offers “a very powerful explanatory mechanism for the mind” (Hohwy, 2015, p. 6), capable of “guiding researchers in finding mechanistic explanations of a target cognitive phenomenon” (Harkness & Keshava, 2017, pp. 1–2).
However, this mechanistic reading is contested. Critics argue that PP lacks empirical specificity (Cao, 2020), bridges theoretical levels loosely (Litwin & Miłkowski, 2020), and may be too flexible to be falsifiable (Miłkowski & Litwin, 2022; cf. Bowers & Davis, 2012). Others point out that PP’s reliance on formal principles distances it from the causal-mechanical style of explanation typical in neuroscience (Colombo & Hartmann, 2017; Klein, 2018).
These concerns raise a central question for this paper: what kind of explanation does PP provide? Given these debates, I propose the following view: I argue that the mechanism sketches employed within the PP framework should be understood as performing two complementary roles:
-
(i)
A mechanistic role, capturing the structural and causal organization of the system, and
-
(ii)
A normative-functional role, which introduces explanatory constraints that govern cognitive performance under adaptive pressure, uncertainty and resource limitations.
These two roles are irreducible because mechanistic explanation without norms doesn’t explain why given mechanism matter (vs. other possible mechanisms) and normative principle without mechanism doesn’t explain how norms are physically realized. This is why PP is a hybrid framework: it essentially involves both perspectives in complementary relationship. For this reason and developing the ideas proposed by Levenstein et al. (2023), I shall distinguish between mechanistic and normative explanations. The former account for phenomena by identifying, decomposing, and describing the components of the mechanism responsible for producing a given phenomenon, along with their operations and organization (cf. Craver, 2007; Bechtel, 2008; see Sect. 3). Such explanations answer the question “How does it work?”.
Normative explanations, by contrast, seek to answer the question “Why is it organized this way?” by showing that certain features or functions are optimal or adaptive relative to specific constraints. In this paper, I argue that these two forms of explanation are complementary and mutually irreducible, and that their integration is characteristic of the neurosciences.Footnote 2
In this paper I distinguish between three types of explanatory constraints: structural constraints, which arise from the biological organization of neural systems; formal constraints, which derive from the mathematical and computational principles underlying information processing; and functional constraints, which reflect the adaptive requirements of cognitive systems. Together, these constraints delimit the space of viable mechanisms and specify the conditions under which predictive models are considered biologically plausible. While structural and formal constraints are largely intrinsic to the biological and mathematical properties of the system, functional constraints are partly attributed—they reflect the explanatory goals and modeling choices of researchers operating within a given scientific context.
Following Ross (2023), explanatory constraints are boundaries that delimit which mechanisms are plausible and explain why certain outcomes can occur while others are ruled out. As Ross (2023, p. 56) writes, they explain why the outcome can take on some states and not others and, in some cases, why particular states are impossible”. In PP’s case, such constraints arise from the biological architecture of the nervous system, its adaptive functions, and from the mathematical and physical principles governing information processing. These sources are not unique to PP; they constrain any theoretical framework in neuroscience. What is distinctive to PP, as I argue below (Sects. 4–5), is the specific formal coupling of these sources through variational inference and prediction error minimization, which simultaneously functions as a normative criterion and as a guide for mechanistic discovery.
In this analysis, I adopt the pragmatic perspective on neuroscience articulated by Levenstein et al. (2023), in which there is room for descriptive, mechanistic, and normative models. These models and theories are mutually interrelated, as they are embedded in scientific practice and do not exist in isolation.Footnote 3 Moreover, their differentiation depends on this practice, as well as on the context, goal, and research problem at hand (Levenstein et al., 2023, 1078).
In accordance with this view, I argue that:
-
(1)
The primary role of PP is to provide explanatory constraints that guide the construction of mechanistic explanations in neuroscience, and for this reason,
-
(2)
PP can be understood as comprising both mechanistic and normative-functional roles, which emerge depending on the scientific context and modeling purpose.
When interpreted mechanistically, PP contributes to the identification of mechanisms by specifying the constraints under which they operate. When interpreted normatively, PP highlights constraints grounded in broader biological and environmental factors, enabling researchers to ask questions about the functions and purposes of these mechanisms.
It is worth noting that several researchers have recognized PP’s integration of mechanistic and normative-functional roles. Clark (2013, p. 52) describes PP as an “intermediate-level model” that bridges computational principles and neural implementation. Hohwy (2013, p. 8) emphasizes that PP combines top-down functional analysis with bottom-up mechanistic modeling. Colombo and Seriès (2012) critically discuss how PP places constraints on mechanism discovery through its normative principles.
However, these discussions typically treat PP’s dual character as self-evident without systematically analyzing how mechanistic and normative-functional roles interact or what role explanatory constraints play in this integration. My contribution builds on these insights but goes further in three ways. First, I provide a systematic account of how mechanistic and normative-functional roles are complementary through explanatory constraints—formal, structural, and functional boundaries that delimit viable mechanisms. Second, I show that these constraints are not merely heuristic but constitutive of PP’s explanatory power. Third, I demonstrate through an empirical case study how constraints function in practice to shape mechanistic inquiry. This moves beyond acknowledging PP’s “hybrid nature” to explaining the mechanics of integration.
Recognizing PP’s normative-functional role has direct practical payoffs: it provides a principled basis for rejecting mechanistic proposals on normative grounds before exhausting experimental tests, clarifies which explanatory gaps require normative rather than mechanistic solutions, and explains PP’s integrative power across levels of neuroscientific explanation. In each case, the result is a more disciplined and productive research practice—one better equipped to identify what cognitive systems actually do and why they are organized the way they are.
The paper proceeds as follows. I first establish the theoretical framework by reconstructing Levenstein et al. (2023) tripartite distinction between descriptive, mechanistic, and normative theories (Sect. 2), and by clarifying how constraints function within mechanistic explanation (Sect. 3). Building on this, I argue that PP integrates both mechanistic and normative-functional roles through explanatory constraints (Sect. 4), distinguish three types of such constraints (Sect. 5), and illustrate their operation through a case study of Mismatch Negativity research (Sect. 6).
2 Three Explanatory Approaches in Neuroscience: Descriptive, Mechanistic, and Normative
Before analyzing PP’s explanatory structure, I establish the theoretical framework that motivates my account. Levenstein et al. (2023) defend a pragmatic approach to theories in neuroscience that recognizes multiple, non-reducible explanatory modes. This framework is crucial for my argument because it provides conceptual space for PP to be both mechanistic and normative without collapsing one into the other. The key insight is that explanatory adequacy depends on context and purpose, not on reduction to a single “correct” level. This allows me to argue (in Sects. 4–5) that PP’s mechanistic and normative-functional roles are complementary rather than competing, and that their integration occurs through explanatory constraints. Now I briefly outline Levenstein et al.’s framework below, focusing on aspects directly relevant to my analysis of PP.
Scientific practice compels researchers to formulate theories and explanations at different levels of abstraction. For example, at the biophysical level, neurons transmit signals via action potentials, the trajectory of which depends on the interaction between ion channels and the cell membrane. At the level of neural networks, individual neurons connect to form circuits, enabling signal integration and sensory information processing. At the computational level, biophysical processes implement specific computational functions, such as minimizing information redundancy or predicting stimuli. From a pragmatic perspective, the existence of multiple levels of abstraction assumes the coexistence of different levels of explanation, none of which is privileged, as the level at which we work depends on the research goal, available tools, and scientific context. This means that theories, and the models and explanations based upon them, are primarily assessed based on their utility—specifically, what empirical problems they can solve, how easily they can be used to solve them, and the quality of their solutions (Levenstein et al., 2023, p. 1075)—as well as their accuracy, simplicity, falsifiability, generalizability, and reproducibility (cf. Laudan, 1977; Van Fraassen, 1980).
This is why Levenstein et al. (2023) distinguish three irreducible explanatory approaches in neuroscience: descriptive, mechanistic, and normative. Descriptive theories answer the question “What is the phenomenon?”. They characterize phenomena at the descriptive level by identifying their recurring features. Mechanistic theories pose the question “How does the phenomenon arise?”. They explain the processes and mechanisms underlying phenomena by analyzing them in terms of parts and interactions of other phenomena at lower levels of abstraction. Normative theories, on the other hand, seek answers to the question “Why does the phenomenon exist?”. In practice, this means determining the goal or function of a given biological or cognitive process (Levenstein et al., 2023, 1077). Thus, rather than analyzing the structure or mechanisms of brain function, the normative approach focuses on the principles governing its operation and the evolutionary benefits of a particular method of information processing, i.e., why a given mechanism is adaptive or optimal in biological and functional contexts.Footnote 4
Normative explanation, as defined by Levenstein et al. (2023), does not exclusively point to the causal mechanism generating a phenomenon. Instead, it justifies why a particular realization of the mechanism serves a specific function—such as minimizing uncertainty, adaptive fitting, or resource conservation. This type of justification is normative because it refers to the conditions for the effective or correct functioning of the system, which supports its adaptation, rather than just its structural composition.
An example of the functioning and relationship between these three approaches is presented in the following table:
Level of explanation | Typical question | Level of abstraction | Example |
|---|---|---|---|
Descriptive | Which brain regions activate during vision? | Functional description of neural correlations | fMRI shows that V1 activates in response to visual stimuli |
Mechanistic | How is the visual signal processed in the cortex? | Biophysical level: Neurons transmit signals via action potentials, which depend on interactions between ion channels and the cell membrane. Network level: Individual neurons connect into circuits that enable signal integration | V1 analyzes contrast and orientation of objects, as described by Hubel and Wiesel’s feature detection theory |
Normative | Why is visual processing organized this way? | Computational level: Biophysical processes implement specific computational functions | The efficient coding theory suggests that the visual system reduces redundancy to conserve energy |
According to Levenstein et al., when mechanistic models at the biophysical level fail to account for behavioral-level regularities, researchers are motivated to introduce normative explanations at a higher level of abstraction (2023, p. 1077). Just as mechanistic theories or explanations analyze phenomena in quantitative terms, normative theories formalize the goal of a phenomenon in an objective function (sometimes referred to as a utility or cost function), which defines how the system functions “well” or “correctly” (cf. Körding, 2007). These models rely on an assumed definition of the system’s goal and constraints within which it operates. For example, efficient coding theory (Barlow, 1961) formalizes the goal of visual processing as maximizing informational efficiency while minimizing metabolic cost—an objective function that specifies what the system should achieve. Empirical research shows that retinal receptor distribution is adapted to natural scene statistics, enabling efficient encoding of common features like edges and contrasts (Geisler & Diehl, 2002). This illustrates how normative theories identify functional goals that mechanistic theories must explain how to achieve.
In the perspective of Levenstein et al. (2023), the choice of theory or explanation is not arbitrary but depends on the context in which they are used and the purpose for which they are formulated. Different models pertain to different levels of abstraction, with normative models linking descriptive and mechanistic models by explaining how the less abstract features serve more abstract goals. For this reason, Levenstein et al. argue that we should speak of a tripartite division of labor in the neurosciences, which implies from researchers’ perspective, recognizing the importance of each type of description or theory (2023, p. 1080).
We may observe that the pragmatic pluralism defended by Levenstein et al. (2023) resonates with broader discussions in philosophy of science. Sandra Mitchell’s (2003) integrative pluralism argues that complex phenomena require multiple, non-reducible models that jointly constitute adequate explanation. Similarly, Michela Massimi’s (2018) perspectival realism holds that different theoretical perspectives can each track real features of the world without reducing to a single correct representation. These frameworks provide philosophical grounding for Levenstein et al.’s neuroscientific proposal and contextualize my argument that PP’s mechanistic and normative dimensions are genuinely complementary rather than competing.
Yet Mitchell’s and Massimi’s positions, like Levenstein et al.’s own framework, establish the legitimacy of explanatory pluralism without specifying when a given explanatory constraint is epistemically warranted—that is, when there are good grounds for believing it tracks something real about the system, rather than being an artifact of a particular model or method (cf. Eronen, 2015)—or why constraints that hold in one modeling context need not generalize to others. Two further positions help address these questions. Wimsatt’s (2007) robustness analysis suggests that explanatory constructs gain warrant when they remain invariant across independent methods, perspectives, and background assumptions—a criterion PP’s explanatory constraints arguably satisfy by linking structural, formal, and functional sources of evidence, as I show in Sect. 5. Cartwright’s (1999) account of nomological machines suggests, conversely, that explanatory generalizations hold only locally, within the specific arrangements that license them, rather than universally—a qualification consonant with the limited scope of the MMN case study discussed in Sect. 6. Together, these two positions supply criteria of warrant and scope that the tripartite division of Levenstein et al. leaves implicit.Footnote 5
I claim that this pragmatic framework has direct implications for understanding PP’s explanatory character. If we accept that: (1) neuroscience legitimately employs multiple explanatory modes; (2) these modes are non-reducible but interdependent and (3) choice between modes depends on research context and goals, then we should expect that successful frameworks like PP might integrate multiple explanatory modes rather than exemplifying only one. Indeed, I will argue (Sect. 4) that PP does exactly this: it provides both mechanistic sketches and normative constraints. The mechanistic role addresses “how” questions (how does PP work?), while the normative role addresses “why” questions (why must mechanisms have these features?). Crucially, these roles are linked through explanatory constraints—formal, structural, and functional conditions that delimit and enable the space of viable mechanisms. This is where PP exemplifies Levenstein et al.’s pragmatic pluralism: normative constraints and mechanistic details mutually shape each other in research practice.
This framework directly motivates my analysis of PP. If neuroscience legitimately employs multiple non-reducible explanatory modes, we should expect successful frameworks like PP to integrate rather than exemplify only one mode. In Sect. 4, I argue PP does exactly this through explanatory constraints. But first, I must clarify how constraints function in mechanistic explanation (Sect. 3).
3 Mechanistic Explanations and Constraints
The goal of neuroscience is to explain the functioning of the brain, the nature of perceptual mechanisms, motor control, and decision-making from infancy to adulthood. This explanation aims to enable control, which is linked to the need for diagnosing and treating various diseases, dementia, or brain damage, and more generally, to understanding how and why our brains function. For these reasons, the question of what constitutes explanation in neuroscience proves to be crucial. Proponents of the mechanistic approach to explanations (“New mechanism”) argue that scientific explanation can be characterized in terms of discovering, describing, and decomposing the mechanisms responsible for specific phenomena (cf. Bechtel, 2008; Craver, 2007; Kaplan & Craver, 2011; Miłkowski, 2016a, 2016b). To explain a given phenomenon, one must identify its mechanism, meaning the system of components and actions that together realize or account for it. A mechanism is understood as a system composed of components, their operations, and their organization, which collectively perform a specific function (Bechtel, 2008, p. 13). As Craver (2007, p. 6) notes, “mechanisms are entities and activities organized such that they exhibit the explanandum phenomenon.” Such mechanisms lead to regular changes from an initial to a final state (Machamer et al., 2000, p. 3).
Mechanisms consist of objects (components), their actions and operations (activities), and the relationships between them. They form multi-level hierarchies, which are tied to their organization. The objects are the elements of mechanisms that participate in actions. The condition for their functioning is the specific set of properties of these objects. The operation performed by a component of a given mechanism can serve as an explanandum and be explained by a mechanism at a lower level in the hierarchy.
In the mechanistic framework, the brain is composed of hierarchically organized causal mechanisms that perform a range of functions. The aim of neuroscience is to discover and describe these mechanisms, thereby providing explanations of their operation.Footnote 6 Carl Craver (2007) argues that neuroscience should not be viewed as a unified theory, but rather as a mosaic of mutually connected and coherent theories and models that integrate a mechanistic approach to studying the brain at various levels. Neuroscience consists of many complementary, yet not necessarily reducible, theories that examine different levels of brain mechanisms. The integration of models and theories in neuroscience, therefore, occurs through the discovery of mechanisms at various levels, rather than through reduction to a single underlying theory (cf. Miłkowski, 2016a). According to Craver, the process of theoretical integration in neuroscience proceeds through the fragmentary identification and gathering of constraints on the space of possible mechanisms—i.e., the set of all mechanisms that could potentially explain a given phenomenon.Footnote 7 Points within this space represent possible explanations. Craver refers to these as “how-possibly mechanisms.” These are preliminary, hypothetical explanations for a phenomenon, as opposed to more developed and empirically grounded “how-actually mechanisms” (Craver, 2007, pp. 112–113). Scientists never explore the entire space; rather, by identifying constraints, they refer to selected points within it, i.e., mechanisms. From a modeling perspective, this implies that a given theory or model may provide constraints on the space of possible mechanisms that cannot be identified within the framework of an alternative theory or model.
I claim that mechanistic explanations directly correspond to what Levenstein et al. refer to as mechanistic theories. In both approaches—illustrated here through Craver’s framework—mechanisms are understood as systems of units and actions that are organized in a way that realizes a given phenomenon, with explanations being the identification, decomposition, and description of these systems. What further unites Craver’s approach with that of Levenstein et al. is the emphasis on the non-autonomy of mechanistic theories. However, while Craver asserts that integration in neuroscience relies on the collaboration between successive mechanistic theories or models that provide constraints on the space of possible mechanisms, Levenstein et al. take a broader view: mechanistic theories are not autonomous—their explanatory power derives directly from their relationship with normative and descriptive theories. In this context, the key issue for my analysis is whether integration can also encompass non-mechanistic models, in the sense outlined by Craver, namely those that Levenstein characterizes as normative. I argue that it can.
Although mechanistic explanations occupy a central place in cognitive science—their proponents argue that identifying and decomposing mechanisms constitutes the gold standard of scientific explanation in this domain (Kaplan & Craver, 2011), many authors see the possibility of cross-integration between mechanistic explanations and dynamic or psychological explanations (Taylor, 2021; cf. Bechtel, 2011; Bechtel & Abrahamsen, 2010). Is the situation analogous in the case of normative theories? Not entirely. If we accept the premise that (at least some) psychological explanations are sketches of mechanisms (cf. Piccinini & Craver, 2011), and dynamic explanations elucidate patterns of dynamic organization that characterize the overall state of the cognitive system as a whole (cf. Bechtel, 2011), these are mechanistic in the sense described by Levenstein et al., rather than normative. Therefore, the possibility of cross-integration between these explanations and mechanistic ones does not automatically resolve the issue of integrating the latter with normative explanations. It is important to recall that, firstly, the distinction between normative and mechanistic theories depends on the questions they aim to address: while mechanistic theories inquire into the processes and mechanisms underlying phenomena, normative theories seek to identify the goal or function of a given process and its evolutionary benefits. Secondly, the relationship between these theories is meant to be complementary. The existence of one does not preclude the existence of the other, but rather assumes it (Levenstein et al., 2023, p. 1080). This means that neuroscience should encompass both mechanistic and normative-functional explanatory strategies, which suggests that an explanation formulated at one, lower level of abstraction may be mechanistic, while at another, higher level, it may contribute to a more abstract normative explanation.
In summary, mechanistic explanations in neuroscience require the identification of both structural and dynamic components of cognitive processes. However, for these explanations to be considered complete, it is also necessary to account for the functions these mechanisms are designed to perform and why a particular structure is biologically plausible. In the following section, I will demonstrate that PP provides a mechanism sketch which offers a normative framework that governs the structure of explanations in accordance with functional adaptive requirements.
4 Predictive Processing as Hybrid Normative-Mechanistic Theory of Explanations
In this section, I argue that PP, while providing mechanism sketches in the standard sense, simultaneously performs a second, normative-functional epistemic role. This role is twofold: it formulates the constraints that mechanistic models must satisfy, and it constitutes the criteria by which such models are evaluated as biologically realistic and functionally adequate. In other words, PP does not merely reconstruct how cognitive systems operate. It specifies the conditions a viable mechanism must meet.
Before proceeding, it is necessary to clarify three key terms used throughout this paper. By functional, I mean relating to the causal role a component plays within a larger system—in line with Cummins (1975) functional analysis. By adaptive, I mean contributing to the organism’s survival and fitness under environmental constraints. By normative, I mean subject to standards of correct or incorrect operation. These three terms are related but not synonymous: a process can be functional without being adaptive (e.g., a maladaptive habit), and adaptive without being normative in the evaluative sense—as in purely reflexive responses that increase fitness without involving any standard of correct functioning.
Below I distinguish two senses in which this applies to PP. Biological normativity refers to whether a mechanism operates according to its biological function—contributing to the organism’s viability and maintenance of far-from-equilibrium conditions (Bickhard, 2009; cf. Piekarski, 2026). Epistemic normativity refers to the researcher’s criteria for evaluating which mechanistic models count as adequate explanations—for instance, whether a proposed model satisfies the constraint of precision-weighted prediction error processing. These two senses are related: biological norms, when formalized in models (e.g., as the imperative to minimize prediction errors), become epistemic norms that constrain theory construction. This is the bridge between the adaptive and the methodological dimensions of PP’s normative character.
As noted in Sect. 1, numerous authors argue that explanations formulated within the PP framework are mechanistic, with PP itself providing a mechanism sketch (cf. Gładziejewski, 2019; Harkness, 2015; Harkness & Keshava, 2017; Hohwy, 2015, among others). Unlike a schematic representation, a mechanism sketch allows for the omission of certain components of the causal structure underlying the phenomenon in question and thus constitutes an incomplete representation of the phenomenon (Piccinini & Craver, 2011). This means that within such a sketch, functional concepts (e.g., precision) may operate alongside mechanistic terms referring to structural properties (e.g., dopamine) (Harkness, 2015, p. 6).
The difference between the mechanistic and normative-functional roles that PP performs is not a matter of descriptive specificity or detail, but of explanatory role. Consider precision-weighting as an illustrative example. From a mechanistic perspective, we may ask: “What neural processes implement precision-weighting?”. Candidate answers identify specific physical substrates: dopaminergic modulation, synaptic gain control, attentional selection mechanisms, or combinations thereof. These are mechanistic specifications because they name concrete neural structures and processes. From a functional perspective, we ask a different question: “Why must any adequate PP implementation include precision-weighting?”. The answer derives from the normative principle of Bayesian optimality: a system performing inference under uncertainty should weight prediction errors by their estimated reliability to approximate optimal probabilistic inference. (cf. Feldman & Friston, 2010). Precision-weighting is thus a necessary condition for approximating Bayesian optimal inference: without it, a system cannot adequately weight evidence by its reliability. Precision-weighting is thus a functional requirement that any adequate implementation must satisfy. Critically, the same feature—precision-weighting—appears in both perspectives but with radically different epistemic roles. Mechanistically, precision is a target for empirical discovery: we must identify which specific neural mechanisms realize it in actual brains. Functionally, precision is an explanatory constraint: any proposed mechanism must include some means of achieving it to count as an adequate realization of PP.
As more components of the underlying causal mechanism are identified, these functional components may be progressively integrated with mechanistic ones. Accordingly, proponents of PP offer a mechanism sketch that integrates two elements: a functional analysis of the brain as an adaptive prediction machine, and a mechanistic decomposition of the hierarchical structure of the neocortex (cf. Kanai et al., 2015; Keller & Mrsic-Flogel, 2018).
This draws on a tradition that situates the brain within evolutionary and environmental context, rather than treating it as an isolated set of causal mechanisms (cf. Bickhard, 2009; Pezzulo et al., 2021). I claim that this functional perspective does more than merely gesture toward the brain’s adaptive role within the organism as a whole; by identifying relevant explanatory constraints, it also introduces a set of commitments that ought to guide future mechanistic explanations—commitments that are not entailed by the mechanistic framework or its associated theories alone.
I claim that PP should be understood as a dynamic, multi-level explanatory framework that integrates mechanistic and normative-functional roles. On the one hand, the mechanistic aspect of PP characterizes specific neural and computational processes that enable predictive coding (cf. Friston, 2005; Rao & Ballard, 1999), thereby providing the basis for mechanistic explanation in terms of system components (cf. Gordon et al., 2019).Footnote 8 On the other hand, the normative-functional role of PP confers a normative interpretation on these processes by highlighting that their organisation and operation are directed towards the minimisation of prediction error. This functional orientation supports the view that PP constitutes a theory with a distinctly adaptive—and potentially even teleological—character (cf. Gong & Wei, 2025; Pezzulo et al., 2021). In this light, PP thus extends beyond a purely mechanistic account: it represents cognitive systems as dynamically organised, sensitive to interlevel dependencies, and embedded in environmental context. Predictive mechanisms consist of specific neural and computational processes and function within a broader framework of adaptive engagement with the environment. Thus, PP integrates mechanistic and normative dimensions, and this reveals their complementarity in the explanation of cognitive systems.Footnote 9
This complementarity can be further illuminated by Tinbergen’s (1963) four questions: What is the mechanism underlying the behavior? How does it develop over the organism’s lifetime (ontogeny)? What is its function? And how did it evolve (phylogeny)? Mechanistic dimensions of PP address the first two; the normative role addresses the third. The fourth—phylogeny—raises a distinct challenge. This challenge is particularly salient for PP: formal implementations under the FEP often assume ergodicity, meaning that the bounds on variability can be translated between time-averaged statistics of individual systems and population-level statistics (cf. Colombo & Palacios, 2021). This limits PP’s ability to address historically contingent adaptation (cf. Bolhuis et al., 2011; Richardson, 2007; van Elk, 2021). Moreover, PP’s normative claims about biological optimality must be treated with caution: not every feature of a biological system is an adaptation in the strict sense, and attributing functional significance to structural traits risks committing the adaptationist fallacy identified by Gould and Lewontin (1979). Some features of predictive systems may be phylogenetic by-products or constraints inherited from ancestral architectures rather than selected-for solutions to adaptive problems. Addressing this phylogenetic dimension fully would require a separate analysis.
To fully justify this interpretation, it is necessary to demonstrate that PP’s functional role is normative in nature. In the following section, I examine the constraints that PP articulates and illustrate their application through a case study (Sect. 6). These constraints shape the space of possible mechanisms and determine their specific structure and function. PP thus provides a framework for explaining neuronal processes and articulates critical constraints that delineate the boundaries of possible mechanisms—thereby serving an integrative role within neuroscience.
5 Predictive Processing in Search of Explanatory Constraints
The preceding section established that PP’s normative-functional role operates through explanatory constraints—formal, structural, and functional conditions that delimit the space of viable mechanisms. What remains to be clarified is the internal structure of this constraint space: not all constraints operate in the same way, derive from the same sources, or carry the same explanatory weight. This section therefore introduces a threefold distinction between structural, formal, and functional constraints, and—cutting across this typology—a distinction between ontic constraints (the actual biological and physical limits on neural systems) and epistemic constraints (the modeling principles and functional criteria that guide which mechanistic hypotheses are considered plausible). A further distinction, developed below, separates constraints that are merely restrictive—ruling out proposals on structural or logical grounds alone—from those that are genuinely normative. Understanding how these types interact is essential for explaining why PP can reject mechanistic proposals on normative grounds, independently of direct experimental disconfirmation.
What makes PP’s explanatory constraints normative rather than merely restrictive? A constraint is merely restrictive when it rules out mechanistic proposals on structural or logical grounds alone—as any coherent theory does. PP’s constraints are normative because they derive from the telos of the system: the imperative to minimize prediction error is not a neutral formal requirement but an adaptive goal, grounded in the organism’s biological imperative to maintain viability (cf. Piekarski, 2026). In Bickhard’s terms (2009), such constraints reflect the system’s far-from-equilibrium dynamics. When these ontic constraints—the actual biological and physical limits on neural systems—are formalized in predictive models, they become epistemic norms: they specify not merely what mechanisms exist, but what any adequate mechanistic model must represent in order to count as biologically viable. Thus, a mechanism that violates PP’s functional constraints does not merely fail formally—it fails to represent a biologically viable system. This is the normative force of PP’s constraints. Thus mechanisms are normative in virtue of their counterfactual capacity to realize specific functions (cf. Piekarski, 2022).
What, then, are explanatory constraints? These are formal, structural, and functional limitations and enablings that influence how theoretical models can represent, simulate, and explain neuronal and cognitive processes and explain why a given outcome may adopt one state rather than another (Ross, 2023; cf. Hooker, 2013). These constraints arise both from the biological properties of biological systems (particularly the nervous system) and from the mathematical and physical principles governing information processing. Therefore, it is important to distinguish between:
-
Structural (ontic) constraints: these may refer to biological aspects of specific structures or mechanisms. An example of this is the structural organization of neural networks (interpreted in different models as modular, hierarchical, or heterarchical) or the biophysical properties of neurons (such as their temporal delays or energy costs);
-
Formal (ontic/epistemic) constraints: these arise from the computational methodology related to mathematical models or simulations. For example, constraints related to the need for approximation (variational inference) in Bayesian inference, which is directly connected to the impossibility of performing exact Bayesian inference (cf. Turner & Van Zandt, 2018), or constraints related to difficulties in modeling the dynamics of large neural networks, such as the scalability problem (cf. Chen & Pesaran, 2021);
-
Functional (epistemic) constraints: these arise from the adaptive requirements of the cognitive system, such as the need to optimize information processing. They are related, for example, to Barlow’s (1961) principle of efficient coding, which specifies that sensory systems must be evolutionarily and functionally adapted to encode information in the most efficient manner—i.e., with minimal resource use while maximizing the retention of essential data (Abbott & Dayan, 2005, 123–150).Footnote 10
PP brings together and systematically integrates a range of explanatory constraints in neuroscience. Among the constraints postulated by PP is the constraint related to precision-weighted prediction error processing, which specifies that not all prediction errors, or information, should be treated equally. The brain dynamically adjusts their precision to determine which predictions should be updated and which should be ignored, i.e., which information should be accepted and which should be discarded (cf. Feldman & Friston, 2010). In practice, this means that PP imposes the requirement for a mechanism of dynamic regulation of prediction error weights, which imposes additional conditions on the model’s structure. Another constraint postulated by PP is the requirement to build a hierarchical model of information processing, where lower levels encode detailed sensory data, while higher levels represent more abstract predictions (cf. Keller & Mrsic-Flogel, 2018).Footnote 11 Yet another constraint is related to the energy and metabolic efficiency of predictive coding: the brain’s drive to minimize computational and metabolic costs is achieved through the optimization of processing by reducing redundancy and minimizing uncertainty (cf. Sengupta et al., 2013). In practice, this means that a mechanistic model must account for the trade-off between accuracy and processing cost, which implies the need to adopt efficient coding strategies (cf. Ali et al., 2022).
Thus reconceptualizing PP as a hybrid normative-mechanistic framework has direct practical implications for neuroscientific modeling. First, it allows researchers to reject mechanistic proposals that satisfy structural constraints but fail functional ones—without awaiting costly experimental confirmation. A model that specifies neural substrates for precision-weighting but assigns uniform weights to all prediction errors violates a functional constraint independently of its structural plausibility. Second, as argued in Piekarski (2022), normative constraints provide counterfactual conditions for evaluating mechanistic models: a mechanism is adequate not merely because it implements a given computation, but because it does so in a way that satisfies the conditions for adaptive functioning. This is the practical upshot of PP’s hybrid character.
Considered individually, each constraint type excludes a different class of mechanistic proposals: structural constraints rule out architectures that lack the required biological organization; formal constraints rule out implementations that are computationally intractable; functional constraints rule out mechanisms that fail to satisfy adaptive requirements regardless of their structural plausibility. Their explanatory force, however, is best seen when all three operate together—as they do in empirical research practice. The following section illustrates this through a case study of Mismatch Negativity research, where structural, formal, and functional constraints jointly shape both the construction and the evaluation of competing mechanistic models.
6 Case Study: The Application of Explanatory Constraints Postulated by Predictive Processing
Research on the Mismatch Negativity (MMN) provides a particularly instructive illustration of how all three constraint types—structural, formal, and functional—operate jointly in scientific practice, and why their integration is constitutive of PP’s explanatory power rather than merely incidental to it. MMN is a well-established electrophysiological response, typically peaking 100–250 ms after the onset of a deviant stimulus, maximal over fronto-central electrode sites, and exhibiting polarity inversion at mastoid/temporal sites—consistent with generators in primary and non-primary auditory cortex, with some accounts positing additional frontal contributions (Garrido et al., 2009a, 2009b; Giard et al., 1990).Footnote 12 MMN is generated when an unexpected or deviant stimulus occurs within a sequence of otherwise regular auditory events, producing a characteristic negative deflection relative to the standard. Crucially, it is a case in which PP functions not as a retrospective interpretive framework imposed on existing data, but as the explicit theoretical basis from which mechanistic models are constructed and evaluated.
The MMN response, first described by Näätänen et al. (1978; cf. Näätänen, 2000), has attracted sustained attention precisely because it resists explanation by purely feedforward or local adaptation models. Classical accounts—most notably the model adjustment hypothesis (cf. Näätänen & Winkler, 1999) but also the adaptation hypothesis (cf. May et al., 1999)—proposed that MMN reflects simple neural fatigue: neurons responding to the standard stimulus become refractory, while deviant stimuli elicit stronger responses from fresh neural populations (May & Tiitinen, 2010). While parsimonious, this account fails to explain several well-documented findings: MMN is elicited by violations of abstract regularities (e.g., violations of learned patterns across stimulus sequences), occurs for deviants that are physically more intense than standards, and is modulated by attention and prior learning in ways that local adaptation cannot accommodate (Garrido et al., 2009a, 2009b).
Garrido et al. (2009a, 2009b; cf. Garrido et al., 2008) address this explanatory gap by developing an explicit predictive coding account of MMN, implemented through Dynamic Causal Modelling (DCM). On their account, the auditory system maintains a hierarchical generative model of the regularities present in the sensory environment. Specifically, the model comprises a three-level hierarchy of five cortical sources—bilateral primary auditory cortex (A1), bilateral superior temporal gyrus (STG), and right inferior frontal gyrus (IFG)—connected by forward (bottom-up), backward (top-down), and intrinsic (within-source) connections (Garrido et al., 2008, 2009a, 2009b). Standard stimuli are predicted with high precision and generate minimal prediction error; deviant stimuli violate these predictions, which produces a burst of prediction error signal that propagates upward through the auditory hierarchy and is reflected in the MMN component. Using DCM, Garrido et al. (2009a, 2009b) specified a family of competing network architectures and used Bayesian model comparison to identify the model that best explained the data; this winning model exhibited a decrease in forward (bottom-up) coupling from STG to IFG for deviants relative to standards, accompanied by an increase in intrinsic connectivity within bilateral A1 and a trend toward increased backward (top-down) coupling from STG to A1—a pattern the authors interpret as reflecting both local adaptation within A1 and changes in long-range coupling among hierarchically organized regions, consistent with the view that MMN generation involves a combination of local and hierarchical mechanisms rather than either in isolation (Fig. 1).
The empirical status of this account has evolved unevenly in the years since. Subsequent DCM studies of the MMN network have not uniformly replicated the original connectivity pattern: for instance, a study examining musicianship and melodic predictability found reduced backward (inhibitory) connectivity from STG to A1 and reduced intrinsic connectivity in A1 and STG—consistent with disinhibition and enhanced neural gain—but no significant modulation of forward connections, a result the authors explicitly contrast with most earlier studies, in which both forward and backward connections showed oddball-related effects (Quiroga-Martinez et al., 2021). While such findings remain broadly consistent with a PP interpretation, they indicate that the specific connectivity signature reported by Garrido et al. is not uniformly reproduced across paradigms. More fundamentally, the predictive-coding interpretation of MMN has not gone unchallenged: May (2021) presents a biologically detailed, modernized version of the adaptation model—grounded in short-term synaptic depression within a hierarchically organized auditory cortex—that reproduces a wide range of MMN phenomena previously taken as decisive evidence against adaptation accounts, including responses to stimulus omission and to global pattern violations. May argues that the predictive-coding account remains comparatively underspecified at the mechanistic level, since the generative model it postulates is assumed rather than derived from known cortical architecture. The empirical picture, fifteen years on, is thus one of partial and uneven support rather than settled consensus. This does not undermine the value of the case for present purposes, however: even an unsettled empirical picture can illustrate how PP’s explanatory constraints operate to structure competing mechanistic proposals, which is the question at issue here.
This case illustrates all three types of explanatory constraints described in Sect. 5. Structural constraints arise from the physiological organization of the auditory system. The PP account requires a hierarchically organized network in which higher-level regions generate top-down predictions and lower-level regions compute and transmit prediction errors. Models that lack this hierarchical architecture cannot accommodate the full pattern of MMN findings, particularly the modulation of MMN by abstract sequential regularities and the top-down suppression of responses to predicted stimuli. The structural constraint thus performs genuine explanatory work: it rules out an entire class of mechanistic proposals on structural grounds alone, not merely on grounds of empirical disconfirmation.
Formal constraints arise from the computational methodology underpinning the PP implementation. Garrido et al.’s model employs variational inference because exact Bayesian inference across the full auditory hierarchy is computationally intractable. This is not an arbitrary modeling choice but a formal constraint: any adequate implementation of PP must adopt some form of approximate inference, and the specific approximation chosen imposes conditions on which model architectures are mathematically tractable and which are not. In this way, the formal constraint shapes the space of admissible mechanistic proposals from within the modeling methodology itself. Concretely, because DCM relies on a variational (Laplace) approximation to model evidence rather than exact computation, it cannot search a continuous, unbounded space of possible network architectures; instead, it requires the researcher to pre-specify a finite family of competing models to be compared. This is precisely why Garrido et al. (2009a, 2009b) tested a discrete set of architectures—ranging from a minimal two-source model with no condition-specific connectivity (S2) to increasingly complex hierarchical networks (S4, S5, S6, and their intrinsic-plasticity variants)—rather than estimating a single, maximally flexible model with unconstrained connectivity. The formal constraint thus operates by limiting the very mechanistic hypotheses that count as candidates for evaluation, before any data are considered.
Functional constraints arise from the adaptive requirements of the auditory system. The auditory cortex must efficiently encode environmental regularities while remaining sensitive to surprising or potentially significant deviations—a dual requirement that is formalized in PP as the minimization of variational free energy. A mechanism that failed to weight prediction errors by their estimated precision—treating all auditory mismatches as equally significant regardless of contextual reliability—would violate this functional constraint: it would be unable to distinguish informative deviations from irrelevant noise and adaptive auditory processing would be impossible. This is why the model adjustment hypothesis is not merely empirically inadequate but functionally insufficient: it specifies no mechanism for precision-weighting and therefore cannot represent a biologically viable auditory system in PP’s terms.
Taken together, these three constraint types illustrate how PP functions not merely as a post hoc interpretive device but as a normative framework that actively shapes the construction and evaluation of mechanistic models. The older hypothesis is rejected not because it has been experimentally falsified in every instance, but because it fails PP’s functional and structural constraints: it lacks the hierarchical architecture and precision-weighting mechanism that any adequate model must include to count as a viable realization of predictive auditory processing. This exemplifies the practical upshot of PP’s hybrid character: researchers can eliminate mechanistic proposals that satisfy local structural plausibility conditions but fail the normative requirements imposed by PP’s functional role—without exhausting the full space of possible experimental tests.
The MMN case thus exemplifies the paper’s central thesis. PP operates simultaneously as a mechanistic framework that specifies the neural architecture responsible for generating prediction error signals, and as a normative framework that specifies the functional conditions any adequate implementation must satisfy. Explanatory constraints serve as the link between these two dimensions: they are simultaneously descriptive (specifying what mechanisms are actually present) and prescriptive (specifying what any viable mechanism must achieve). It is this dual character that constitutes PP’s distinctive explanatory contribution to cognitive neuroscience.
In practice, the normative-functional role of PP imparts a normative dimension to the explanatory framework: it delineates the features of the mechanism that are crucial in terms of its adaptive function and biological plausibility (cf. Golkar et al., 2022). Functional normativity in PP can be understood analogously to biological conceptions of function, in which the function of a mechanism depends on its contribution to the self-maintenance and adaptation of the system (cf. Bickhard, 2009; Millikan, 1984).Footnote 13 PP thus provides tools for describing and reconstructing mechanisms, and offers constraints that shape the space of possible explanations. The explanatory constraints postulated within this framework serve as a link between biological realism and the epistemic validity of models.Footnote 14
It should be noted, however, that MMN represents a relatively favorable case for PP: the phenomenon is well-circumscribed, the relevant hierarchy is anatomically tractable, and the adaptation account is sufficiently specific to be evaluated against PP’s functional constraints. In domains where the target phenomenon is less well-defined—such as the PP-based accounts of social cognition or conscious awareness—normative constraints may underdetermine the space of admissible mechanisms to a far greater degree, and the hybrid character of PP argued for here may be correspondingly harder to demonstrate.
7 Conclusion
PP is commonly described as a mechanism sketch—an incomplete, primarily mechanistic representation of the target phenomenon. While this characterization rightly emphasizes PP’s structural and causal explanatory aspects, I argue that it tends to overlook important functional and normative dimensions that are essential for a comprehensive account of cognitive processes. In response, I claim that PP, while providing mechanism sketches in the standard sense, simultaneously performs a second, normative-functional epistemic role. This expanded view on PP preserves the value of mechanistic explanation while clarifying the essential role of functional constraints in shaping predictive models. Specifically, PP models—such as Rao and Ballard (1999) or Friston (2005)—postulate explanatory constraints that regulate information processing efficiency. These constraints also serve as epistemic and adaptive criteria for evaluating neuronal mechanisms under conditions of limited cognitive and environmental resources. I concluded that this perspective allows PP to be treated as a hybrid explanatory strategy, combining the structural organization of mechanisms with a normatively oriented predictive function.
As anticipated in Sect. 1, the sources of PP’s explanatory constraints—biological architecture, adaptive function, mathematical and physical principles—are not unique to PP, since they constrain any theoretical framework in neuroscience. What distinguishes PP from other mechanism sketches is not the mere presence of normative constraints, but their specific formal character. PP’s normative constraints derive explicitly from the mathematics of variational inference and prediction error minimization—formal principles that simultaneously function as normative criteria (specifying what an adequate mechanism must achieve) and as guides for identifying plausible neural architectures. This dual role, grounded in a formal framework, is less prominent in other mechanism sketches that rely on, for instance, lossy compression or modular decomposition strategies.
The normative character of PP derives from three interlocking sources: the mathematics of variational inference, which specifies what an optimal system should do to approximate Bayesian inference; biological constraints, which specify what a system must do to remain viable; and the functional requirements of efficient coding, which specify how a system ought to process information given limited resources. These sources are not reducible to the FEP—the FEP provides a unifying formal language in which they can be expressed, while PP applies them specifically to neural and cognitive mechanisms.
What distinguishes PP from frameworks such as efficient coding theory (Barlow, 1961) is the explicit centrality and formal precision of its normative principles. Efficient coding specifies a goal (informational efficiency) but leaves the mechanistic implementation largely underdetermined. PP, by contrast, couples this normative goal with a specific formal mechanism—variational inference and hierarchical message-passing—that simultaneously constrains which mechanisms are admissible and explains how the normative goal is realized. This dual role is not merely an add-on but constitutive of PP’s explanatory architecture.
Considering the analyses presented here, mechanistic theories lend credibility to normative theories by providing empirical grounding, while normative theories delimit the space of explanatorily relevant mechanisms. Mechanistic theories make normative claims empirically testable; normative theories, in turn, serve as epistemic guides for exploring possible explanations and identify which questions require answers. This bidirectional constraint exemplifies what Levenstein et al. (2023, p. 1080) call the “division of explanatory labor” in neuroscience—different explanatory modes serving distinct but interdependent epistemic functions. Rather than competing frameworks, mechanistic and normative perspectives represent complementary approaches that mutually shape each other in research practice.
Recognizing PP’s normative-functional role has concrete payoffs for neuroscientific modeling practice. First, it provides a principled basis for rejecting mechanistic proposals before exhausting experimental tests: a model that lacks a precision-weighting mechanism fails a functional constraint independently of its structural plausibility, and can be set aside without costly empirical disconfirmation (cf. Piccinini & Craver, 2011). Second, it clarifies which explanatory gaps require normative rather than mechanistic solutions: when a model fails not because its components are mis-specified but because it lacks an adequate functional goal, adding more mechanistic detail will not fix the problem (cf. Cao & Yamins, 2024). Third, it explains why PP has genuine integrative power across levels of neuroscientific explanation: by specifying formal, structural, and functional conditions that any adequate mechanism must satisfy, PP provides cross-level constraints that connect computational principles with biophysical implementation. These are not merely philosophical observations—they determine how models are built, evaluated, and discarded in practice, as the MMN case study demonstrated.
Several questions remain for future investigation. How do other computational frameworks—beyond PP—integrate mechanistic and normative constraints? Can we formalize the relationship between different constraint types (structural, formal, functional) and their respective contributions to explanatory adequacy? How do normative constraints evolve as empirical evidence accumulates? Addressing these questions would move the present account from a philosophical framework toward a more operationalizable methodology for constraint-based explanation in cognitive neuroscience.
In line with the pragmatic vision of neuroscience defended in this paper, mechanistic, normative, and descriptive theories are closely interconnected and mutually define one another. They differ in levels of abstraction, but these levels are defined contextually—in relation to system goals, research strategies, and empirical methods. This pragmatic pluralism accepts multiple mutually constraining explanatory modes and offers a productive path forward for philosophy of cognitive neuroscience. PP provides not merely mechanism sketches awaiting completion, but a normative framework that actively shapes mechanistic inquiry. Recognizing both dimensions yields a more accurate understanding of PP’s explanatory power.
Data Availability
Not applicable.
Notes
I use “research tradition” in Laudan’s (1977) sense as a set of shared problem-solving practices, assumptions, and methods within a field. Describing PP this way highlights that it is not a single unified theory but a heuristic framework of interrelated concepts guiding research rather than dictating fixed strategies.
I claim that mechanistic and normative roles of explanation are “irreducible” because neither can be eliminated without an explanatory loss, and neither can be reduced to the other; both are required for an adequate account. I therefore endorse a pragmatically understood irreducibility—concerning explanatory practice—rather than a metaphysical irreducibility, which would pertain to the nature of ultimate reality.
In this paper, I use the term theory to refer to general conceptual or computational frameworks such as PP; model to denote specific implementations of such theories applied to the explanation of phenomena; and explanation to describe the outcome of using a model to answer empirical or conceptual questions. While these terms are used with some variation in the literature, I adopt this usage in line with common conventions in recent discussions of PP.
This tripartite distinction echoes Marr’s (1982) classic levels of analysis in cognitive science—computational (what and why), algorithmic (how), and implementational (physical realization)—although Levenstein et al. emphasize explanatory strategies embedded in scientific practice rather than strictly hierarchical levels of analysis. Also this tripartite distinction maps naturally onto Tinbergen's (1963) four questions in ethology: mechanism, ontogeny, function, and phylogeny. Mechanistic theories address the first two; normative theories address the third — why a given mechanism exists and what adaptive function it serves. The fourth question, phylogeny, points to explanatory demands that neither mechanistic nor normative theories fully satisfy alone.
A full elaboration of how robustness-based warrant and locality-based scope apply across PP's various domains of application would, however, exceed the scope of the present paper.
Salmon (1984) refers to an explanation of a phenomenon through the identification and description of the underlying mechanism as a constitutive explanation. He contrasts this with etiological explanations, which provide a causal history of the phenomenon without describing the mechanism that links the phenomenon to its cause.
This integration of mechanistic explanations within neuroscience entails providing a causal explanation of a cognitive phenomenon (cf. Kaplan & Craver, 2011).
Whether PP requires contentful representations or can be understood in purely dynamical terms remains debated (cf. Gładziejewski, 2016; Williams, 2018). What matters for the normative role of PP is that these constraints regulate which mechanistic models are considered adaptive, not the metaphysical status of the system's internal states.
It is worth noting that the FEP serves as a formal mathematical framework underpinning PP. The FEP makes explicit the conditions under which self-organising biological systems must operate. As such, it provides a normative-mathematical background for PP without itself carrying empirical content — the claim that a self-organising system minimises free energy is analytically true by definition, not an empirical discovery. The FEP can be regarded as a deeper theoretical framework that encompasses both mechanistic and normative aspects, as it provides a formal description of self-organising systems in terms of probabilistic generative models, while also offering a functional perspective on the adaptive significance of free energy minimization for the survival and effective functioning of the organism (cf. Friston, 2012; Friston & Kiebel, 2009). Friston himself observes that the FEP entails the maximum entropy principle (Jaynes, 1957), since free energy is expected energy minus the entropy of predictions (Friston, 2013) — raising the question of whether PP’s normative commitments could in principle be grounded in the more general MaxEnt framework rather than in the full machinery of the FEP. I do not pursue this question here, as the focus of this paper is on the relationship between the mechanistic and normative aspects of PP in the context of explanatory structures in neuroscience, independent of the fundamental assumptions of FEP. Its status as a universal principle for the organization of biological and cognitive systems remains controversial and requires a separate, more detailed analysis. Moreover, it is unclear whether the FEP itself entails any mechanistic commitments regarding target phenomena. Rather, it appears to impose normative constraints on mechanistic-functional theories such as PP, without specifying the structural nature of the systems these theories aim to explain (cf. Andrews, 2021). In this paper, I treat PP as a research framework that can be motivated by the FEP but is conceptually distinct from it.
The conception of explanatory constraints that I propose partially aligns with the typology introduced by Ross (2023). Ross distinguishes three types of explanatory constraints: (1) law-based constraints, which derive from the laws of physics or biology—for example, the law of gravity; (2) mathematical constraints, which stem from mathematical properties, such as the topology of bridges in a given city; and (3) causal constraints, which involve physical structures that guide processes, such as blood vessels directing the flow of blood. I argue that explanatory constraints, as understood in this project incorporate elements of both mathematical and causal constraints in Ross’s sense. While the functional constraints I discuss don’t fully align with Ross’s typology, her framework isn’t meant to be exhaustive (cf. Green & Jones, 2016), but rather to clarify key types of scientific constraints (Ross, 2023, p. 56).
It should be noted, however, that these constraints are general in character—as Spratling (2017) points out, hierarchical prediction error minimization can be realized by multiple different algorithms, which means that PP narrows the space of possible models without fully determining it.
It should be noted that MMN, as an EEG-derived component, is a coarse-grained measure whose relationship to the underlying neural generators is not straightforward, given the synchronization requirements and the ill-posed nature of EEG source localization (cf. Kalogeropoulos et al., 2026). One might accordingly question whether scalp-level event-related potential components should be treated as epiphenomenal traces of lower-level activity rather than as explananda with an independent mechanistic status. I bracket this issue here and treat MMN as a robust, reproducible phenomenon at the level at which Garrido et al. (2009a, 2009b) analyze it, without committing to a stronger claim about whether it constitutes a cognitive phenomenon in its own right.
The reference to function as a criterion of normativity draws on philosophy of biology, especially etiological and systemic accounts. Etiological views (Millikan, 1984; Neander, 1991) link function to evolutionary history, while systemic approaches (Cummins, 1975; cf. Bickhard, 2003) define it by a part’s role in system organization. In this context, constraints in predictive processing act as structural conditions for correct functioning, varying by research aim—whether the focus is on operational norms, semantics, or efficiency (Garson, 2016).
It should be noted that PP can support both reactive (homeostatic) and proactive (allostatic) regulatory strategies, depending on the level of the model hierarchy and the type of information processed, which is directly related to its functional-normative role of PP (cf. Corcoran, Pezzulo & Hohwy, 2020; Kiverstein & Sims, 2021).
References
Abbott, L. F., & Dayan, P. (2005). Theoretical neuroscience: Computational and mathematical modeling of neural systems. MIT Press.
Ali, A., Ahmad, N., de Groot, E., van Gerven, M. A. J., & Kietzmann, T. C. (2022). Predictive coding is a consequence of energy efficiency in recurrent neural networks. Patterns, 3(12), Article 100639. https://doi.org/10.1016/j.patter.2022.100639
Andrews, M. (2021). The math is not the territory: Navigating the free energy principle. Biology & Philosophy, 36(3), 1–19. https://doi.org/10.1007/s10539-021-09807-0
Badcock, P. B., Friston, K. J., & Ramstead, M. J. D. (2019). The hierarchically mechanistic mind: A free-energy formulation of the human psyche. Physics of Life Reviews, 31, 104–121. https://doi.org/10.1016/j.plrev.2018.10.002
Barlow, H. B. (1961). Possible principles underlying the transformation of sensory messages. In W. A. Rosenblith (Ed.), Sensory communication (pp. 217–234). MIT Press.
Bechtel, W. (2008). Mental mechanisms: Philosophical perspectives on cognitive neuroscience. Routledge.
Bechtel, W. (2011). Mechanism and biological explanation. Philosophy of Science, 78, 533–557. https://doi.org/10.1086/661513
Bechtel, W., & Abrahamsen, A. (2010). Dynamic mechanistic explanation: Computational modeling of circadian rhythms as an exemplar for cognitive science. Studies in History and Philosophy of Science Part A, 41(3), 321–333. https://doi.org/10.1016/j.shpsa.2010.07.003
Bickhard, M. H. (2003). Process and emergence: Normative function and representation. In J. Seibt (Ed.), Process theories: Cross disciplinary studies in dynamic categories (pp. 121–155). Springer. https://doi.org/10.1007/978-94-007-1044-3_6
Bickhard, M. H. (2009). The interactivist model. Synthese, 166, 547–591. https://doi.org/10.1007/s11229-008-9375-x
Bolhuis, J. J., Brown, G. R., Richardson, R. C., & Laland, K. N. (2011). Darwin in mind: New opportunities for evolutionary psychology. PLoS Biology, 9(7), Article e1001109. https://doi.org/10.1371/journal.pbio.1001109
Bowers, J. S., & Davis, C. J. (2012). Bayesian just-so stories in psychology and neuroscience. Psychological Bulletin, 138(3), 389–414. https://doi.org/10.1037/a0026450
Buckley, C. L., Chang, S. K., McGregor, S., & Seth, A. K. (2017). The free energy principle for action and perception: A mathematical review. Journal of Mathematical Psychology, 81, 55–79. https://doi.org/10.1016/j.jmp.2017.09.004
Cao, R. (2020). New labels for old ideas: Predictive processing and the interpretation of neural signals. Review of Philosophy and Psychology, 11, 517–546. https://doi.org/10.1007/s13164-020-00481-x
Cao, R., & Yamins, D. (2024). Explanatory models in neuroscience, part 2: Functional intelligibility and the contravariance principle. Cognitive Systems Research, 85, Article 101244. https://doi.org/10.1016/j.cogsys.2023.101200
Cartwright, N. (1999). The dappled world: A study of the boundaries of science. Cambridge University Press.
Chen, Z. S., & Pesaran, B. (2021). Improving scalability in systems neuroscience. Neuron, 109(11), 1776–1790. https://doi.org/10.1016/j.neuron.2021.03.025
Clark, A. (2013). Whatever next? Predictive brains, situated agents, and the future of cognitive science. Behavioral and Brain Sciences, 36, 181–204. https://doi.org/10.1017/S0140525X12000477
Clark, A. (2016). Surfing uncertainty: Prediction, action, and the embodied mind. Oxford University Press.
Colombo, M., & Hartmann, S. (2017). Bayesian cognitive science, unification, and explanation. The British Journal for the Philosophy of Science, 68(2), 451–484. https://doi.org/10.1093/bjps/axv036
Colombo, M., & Palacios, P. (2021). Non-equilibrium thermodynamics and the free energy principle in biology. Biology & Philosophy, 36(5), 1–26. https://doi.org/10.1007/s10539-021-09818-x
Colombo, M., & Seriès, P. (2012). Bayes in the brain: On Bayesian modelling in neuroscience. The British Journal for the Philosophy of Science, 63(3), 697–723. https://doi.org/10.1093/bjps/axr006
Corcoran, A. W., Pezzulo, G., & Hohwy, J. (2020). From allostatic agents to counterfactual cognisers: Active inference, biological regulation, and the origins of cognition. Biology & Philosophy, 35, 32. https://doi.org/10.1007/s10539-020-09746-2
Craver, C. F. (2007). Explaining the brain: Mechanisms and the mosaic unity of neuroscience. Oxford University Press.
Cummins, R. (1975). Functional analysis. The Journal of Philosophy, 72, 741–764. https://doi.org/10.2307/2024640
Eronen, M. I. (2015). Robustness and reality. Synthese, 192(12), 3961–3977. https://doi.org/10.1007/s11229-015-0801-6
Feldman, H., & Friston, K. J. (2010). Attention, uncertainty, and free-energy. Frontiers in Human Neuroscience, 4, 215. https://doi.org/10.3389/fnhum.2010.00215
Förster, J. (2023). Predictive processing as a computational mechanism. In M. Curado & S. S. Gouveia (Eds.), Predictive minds: Old problems and new challenges (pp. 229–254). Vernon Press.
Friston, K. J. (2005). A theory of cortical responses. Philosophical Transactions of the Royal Society B: Biological Sciences, 360, 815–836. https://doi.org/10.1098/rstb.2005.1622
Friston, K. J. (2009). The free-energy principle: A rough guide to the brain? Trends in Cognitive Sciences, 13(7), 293–301. https://doi.org/10.1016/j.tics.2009.04.005
Friston, K. J. (2012). A free energy principle for biological systems. Entropy, 14, 2100–2121. https://doi.org/10.3390/e14112100
Friston, K. (2013). Active inference and free energy. Behavioral and Brain Sciences, 36(3), 212–213. https://doi.org/10.1017/S0140525X12002142
Friston, K., & Kiebel, S. (2009). Predictive coding under the free-energy principle. Philosophical Transactions of the Royal Society B: Biological Sciences, 364, 1211–1221. https://doi.org/10.1098/rstb.2008.0300
Garrido, M. I., Friston, K. J., Kiebel, S. J., Stephan, K. E., Baldeweg, T., & Kilner, J. M. (2008). The functional anatomy of the MMN: A DCM study of the roving paradigm. NeuroImage, 42(2), 936–944. https://doi.org/10.1016/j.neuroimage.2008.05.018
Garrido, M. I., Kilner, J. M., Kiebel, S. J., & Friston, K. J. (2009a). Dynamic causal modeling of the response to frequency deviants. Journal of Neurophysiology, 101(5), 2620–2631. https://doi.org/10.1152/jn.90291.2008
Garrido, M. I., Kilner, J. M., Stephan, K. E., & Friston, K. J. (2009b). The mismatch negativity: A review of underlying mechanisms. Clinical Neurophysiology, 120(3), 453–463. https://doi.org/10.1016/j.clinph.2008.11.029
Garson, J. (2016). A critical overview of biological functions. Springer. https://doi.org/10.1007/978-3-319-32020-5
Geisler, W. S., & Diehl, R. L. (2002). Bayesian natural selection and the evolution of perceptual systems. Philosophical Transactions of the Royal Society of London. Series B, Biological Sciences, 357(1420), 419–448. https://doi.org/10.1098/rstb.2001.1055
Giard, M. H., Perrin, F., Pernier, J., & Bouchet, P. (1990). Brain generators implicated in the processing of auditory stimulus deviance: A topographic event-related potential study. Psychophysiology, 27(6), 627–640. https://doi.org/10.1111/j.1469-8986.1990.tb03184.x
Gładziejewski, P. (2016). Predictive coding and representationalism. Synthese, 193(2), 559–582. https://doi.org/10.1007/s11229-015-0762-9
Gładziejewski, P. (2019). Mechanistic unity and the predictive mind. Theory & Psychology, 29(5), 657–675. https://doi.org/10.1177/0959354319866258
Golkar, S., Tesileanu, T., Bahroun, Y., Sengupta, A., & Chklovskii, D. (2022). Constrained predictive coding as a biologically plausible model of the cortical hierarchy. Advances in Neural Information Processing Systems, 35, 14155–14169.
Gong, Z., & Wei, Y. (2025). An adaptive representational account of predictive processing in human cognition. Cultures of Science, 8(1), 3–11. https://doi.org/10.1177/20966083251324230
Gordon, N., Tsuchiya, N., Koenig-Robert, R., & Hohwy, J. (2019). Expectation and attention increase the integration of top-down and bottom-up signals in perception through different pathways. PLoS Biology, 17(4), Article e3000233. https://doi.org/10.1371/journal.pbio.3000233
Gould, S. J., & Lewontin, R. C. (1979). The spandrels of San Marco and the Panglossian paradigm: A critique of the adaptationist programme. Proceedings of the Royal Society of London. Series B, Biological Sciences, 205(1161), 581–598. https://doi.org/10.1098/rspb.1979.0086
Green, S., & Jones, N. (2016). Constraint-based reasoning for search and explanation: Strategies for understanding variation and patterns in biology. Dialectica, 70(3), 343–374. https://doi.org/10.1111/1746-8361.12145
Gregory, R. L. (1980). Perceptions as hypotheses. Philosophical Transactions of the Royal Society of London. B, Biological Sciences, 290, 181–197. https://doi.org/10.1098/rstb.1980.0090
Harkness, D. L. (2015). From explanatory ambition to explanatory power: A commentary on Jakob Hohwy. In T. Metzinger & J. M. Windt (Eds.), Open MIND, 19(C), 1–7. MIND Group. https://doi.org/10.15502/9783958570153
Harkness, D. L., & Keshava, A. (2017). Moving from the what to the how and where: Bayesian models and predictive processing. In T. Metzinger & W. Wiese (Eds.), Philosophy and predictive processing (Vol. 16, pp. 1–10). MIND Group. https://doi.org/10.15502/9783958573178
Helmholtz, H. (1867). Handbuch der physiologischen Optik. Leopold Voss.
Hohwy, J. (2013). The predictive mind. Oxford University Press.
Hohwy, J. (2015). The neural organ explains the mind. In T. Metzinger & J. M. Windt (Eds.), Open MIND, 19(T), 1–22. MIND Group. https://doi.org/10.15502/9783958570016
Hooker, C. A. (2013). On the import of constraints in complex dynamical systems. Foundations of Science, 18(4), 757–780. https://doi.org/10.1007/s10699-012-9304-9
Jaynes, E. T. (1957). Information theory and statistical mechanics. Physical Review, 106(4), 620–630. https://doi.org/10.1103/PhysRev.106.620
Kalogeropoulos, C., Theofilatos, K., & Mavroudi, S. (2026). From neurons to networks: A holistic review of electroencephalography (EEG) from neurophysiological foundations to AI techniques. Signals, 7(1), Article 17. https://doi.org/10.3390/signals7010017
Kanai, R., Komura, Y., Shipp, S., & Friston, K. (2015). Cerebral hierarchies: Predictive processing, precision and the pulvinar. Philosophical Transactions of the Royal Society B: Biological Sciences, 370(1668), Article 20140169. https://doi.org/10.1098/rstb.2014.0169
Kaplan, D. M., & Craver, C. F. (2011). The explanatory force of dynamical and mathematical models in neuroscience: A mechanistic perspective. Philosophy of Science, 78(4), 601–627. https://doi.org/10.1086/661721
Keller, G. B., & Mrsic-Flogel, T. D. (2018). Predictive processing: A canonical cortical computation. Neuron, 100(2), 424–435. https://doi.org/10.1016/j.neuron.2018.10.003
Kiverstein, J., & Sims, M. (2021). Is free-energy minimisation the mark of the cognitive? Biology & Philosophy, 36, Article 25. https://doi.org/10.1007/s10539-021-09788-0
Klein, C. (2018). What do predictive coders want? Synthese, 195, 2541–2557. https://doi.org/10.1007/s11229-016-1250-6
Knill, D., & Pouget, A. (2004). The Bayesian brain: The role of uncertainty in neural coding and computation. Trends in Neurosciences, 27(12), 712–719. https://doi.org/10.1016/j.tins.2004.10.007
Körding, K. P. (2007). Decision theory: What “should” the nervous system do? Science, 318(5850), 606–610. https://doi.org/10.1126/science.1142998
Laudan, L. (1977). Progress and its problems: Toward a theory of scientific growth. University of California Press.
Levenstein, D., Alvarez, V. A., Amarasingham, A., Azab, H., Chen, Z. S., Gerkin, R. C., Hasenstaub, A., Iyer, R., Jolivet, R. B., Marzen, S., Monaco, J. D., Prinz, A. A., Quraishi, S., Santamaria, F., Shivkumar, S., Singh, M. F., Traub, R., Nadim, F., Rotstein, H. G., & Redish, A. D. (2023). On the role of theory and modeling in neuroscience. Journal of Neuroscience, 43(7), 1074–1088. https://doi.org/10.1523/JNEUROSCI.1179-22.2022
Litwin, P., & Miłkowski, M. (2020). Unification by fiat: Arrested development of predictive processing. Cognitive Science, 44(7), 1–27. https://doi.org/10.1111/cogs.12867
Machamer, P. K., Darden, L., & Craver, C. F. (2000). Thinking about mechanisms. Philosophy of Science, 57(1), 1–25. https://doi.org/10.1086/392632
Marr, D. (1982). Vision: A computational approach. Freeman & Co.
Massimi, M. (2018). Perspectival realism. In J. Saatsi (Ed.), The Routledge handbook of scientific realism (pp. 164–175). Routledge.
May, P. J. C. (2021). The adaptation model offers a challenge for the predictive coding account of mismatch negativity. Frontier in Human Neuroscience, 15, 721574. https://doi.org/10.3389/fnhum.2021.721574
May,P. J., & Tiitinen, H. (2010). Mismatch negativity (MMN), the deviance-elicited auditory deflection, explained. Psychophysiology, 47(1), 66–122. https://doi.org/10.1111/j.1469-8986.2009.00856.x
May, P. J. C., Tiitinen, H., Ilmoniemi, R. J., Nyman, G., Taylor, J. G., & Näätänen, R. (1999). Frequency change detection in human auditory cortex. Journal of Computational Neuroscience, 6(2), 99–120. https://doi.org/10.1023/A:1008896417606
Miłkowski, M. (2016a). Integrating cognitive (neuro)science using mechanisms. Avant, 6(2), 45–67. https://doi.org/10.26913/70202016.0112.0003
Miłkowski, M. (2016b). A mechanistic account of computational explanation in cognitive science and computational neuroscience. In V. C. Müller (Ed.), Computing and philosophy (pp. 153–178). Synthese Library, vol 375. Springer. https://doi.org/10.1007/978-3-319-23291-1_13
Miłkowski, M., & Litwin, P. (2022). Testable or bust: Theoretical lessons for predictive processing. Synthese, 200, 462. https://doi.org/10.1007/s11229-022-03891-7
Millikan, R. G. (1984). Language, thought, and other biological categories: New foundations for realism. MIT Press.
Mitchell, S. D. (2003). Biological complexity and integrative pluralism. Cambridge University Press.
Näätänen, R. (2000). Mismatch negativity (MMN): Perspectives for application. International Journal of Psychophysiology, 37(1), 3–10. https://doi.org/10.1016/S0167-8760(00)00091-X
Näätänen, R., Gaillard, A. W. K., & Mäntysalo, S. (1978). Early selective-attention effect on evoked potential reinterpreted. Acta Psychologica, 42(4), 313–329. https://doi.org/10.1016/0001-6918(78)90006-9
Näätänen, R., & Winkler, I. (1999). The concept of auditory stimulus representation in cognitive neuroscience. Psychological Bulletin, 125(6), 826–859. https://doi.org/10.1037/0033-2909.125.6.826
Neander, K. (1991). The teleological notion of ‘function.’ Australasian Journal of Philosophy, 69(4), 454–468. https://doi.org/10.1080/00048409112344881
Oaksford, M., & Chater, N. (2007). Bayesian rationality: The probabilistic approach to human reasoning. Oxford University Press.
Pezzulo, G., Parr, T., & Friston, K. (2021). The evolution of brain architectures for predictive coding and active inference. Philosophical Transactions of the Royal Society B: Biological Sciences, 377(1846), 20200531. https://doi.org/10.1098/rstb.2020.0531
Piccinini, G., & Craver, C. F. (2011). Integrating psychology and neuroscience: Functional analyses as mechanism sketches. Synthese, 183(3), 283–311. https://doi.org/10.1007/s11229-011-9898-4
Piekarski, M. (2022). Motivation, counterfactual predictions and constraints: Normativity of predictive mechanisms. Synthese, 200, 352. https://doi.org/10.1007/s11229-022-03837-1
Piekarski, M. (2026). Rethinking normativity with the free energy principle in light of interactivism. Phenomenology and the Cognitive Sciences. https://doi.org/10.1007/s11097-025-10135-x
Quiroga-Martinez, D. R., Hansen, N. C., Højlund, A., Pearce, M., Brattico, E., Holmes, E., Friston, K., & Vuust, P. (2021). Musicianship and melodic predictability enhance neural gain in auditory cortex during pitch deviance detection. Human Brain Mapping, 42(17), 5595–5608. https://doi.org/10.1002/hbm.25638
Rao, R. P., & Ballard, D. H. (1999). Predictive coding in the visual cortex: A functional interpretation of some extra-classical receptive-field effects. Nature Neuroscience, 2(1), 79–87. https://doi.org/10.1038/4580
Richardson, R. C. (2007). Evolutionary psychology as maladapted psychology. MIT Press.
Ross, L. N. (2023). The explanatory nature of constraints: Law-based, mathematical, and causal. Synthese, 202, Article 56. https://doi.org/10.1007/s11229-023-04281-5
Salmon, W. C. (1984). Scientific explanation and the causal structure of the world. Princeton University Press.
Sengupta, B., Stemmler, M. B., & Friston, K. J. (2013). Information and efficiency in the nervous system: A synthesis. PLoS Computational Biology, 9(7), Article e1003157. https://doi.org/10.1371/journal.pcbi.1003157
Spratling, M. W. (2017). A review of predictive coding algorithms. Brain and Cognition, 112, 92–97. https://doi.org/10.1016/j.bandc.2015.11.003
Taylor, S. D. (2021). Two kinds of explanatory integration in cognitive science. Synthese, 198, 4573–4601. https://doi.org/10.1007/s11229-019-02357-9
Tinbergen, N. (1963). On aims and methods of ethology. Zeitschrift Für Tierpsychologie, 20(4), 410–433. https://doi.org/10.1111/j.1439-0310.1963.tb01161.x
Turner, B. M., & Van Zandt, T. (2018). Approximating Bayesian inference through model simulation. Trends in Cognitive Sciences, 22(9), 826–840. https://doi.org/10.1016/j.tics.2018.06.003
van Elk, M. (2021). A predictive processing framework of tool use. Cortex; A Journal Devoted to the Study of the Nervous System and Behavior, 139, 211–221. https://doi.org/10.1016/j.cortex.2021.03.014
Van Fraassen, B. C. (1980). The scientific image. Oxford University Press.
Venter, E. (2021). Toward an embodied, embedded predictive processing account. Frontiers in Psychology, 12, Article 543076. https://doi.org/10.3389/fpsyg.2021.543076
Williams, D. (2018). Predictive processing and the representation wars. Minds and Machines, 28(1), 141–172. https://doi.org/10.1007/s11023-017-9441-6
Wimsatt, W. C. (2007). Re-engineering philosophy for limited beings: Piecewise approximations to reality. Harvard University Press.
Funding
This research was supported by the National Science Centre, Poland (Narodowe Centrum Nauki), grant no. UMO-2025/57/B/HS6/02566, awarded to the project’s Principal Investigator, Dr Marek Nieznański. The author of this article served as the project’s main investigator.
Author information
Authors and Affiliations
Contributions
Not applicable (single author). AI Assistance (Claude, Anthropic) was used for literature search and to assist in preparing portions of the text. The author reviewed, verified, and takes full responsibility for the final content.
Corresponding author
Ethics declarations
Conflict of interest
The author declares no competing interests.
Additional information
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Rights and permissions
Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/
About this article
Cite this article
Piekarski, M. Where Mechanism Meets Normativity: Predictive Processing in Search of Explanatory Constraints. Minds & Machines 36, 43 (2026). https://doi.org/10.1007/s11023-026-09799-4
Received:
Accepted:
Published:
Version of record:
DOI: https://doi.org/10.1007/s11023-026-09799-4