Enactive Drift Regulation and the Emergence Machine: Sustaining Coherent Adaptation in Chaotic and Non-Stationary Time Series



Abstract

Adaptive systems increasingly operate in environments characterized by persistent non-stationarity, where patterns reorganize rather than merely vary. While existing approaches such as online learning, continual learning, and adaptive filtering address performance degradation under changing data distributions, they typically treat drift as noise, error, or distribution shift to be corrected. This paper argues that such framings miss a more fundamental challenge: the loss of organizational coherence over time.

We introduce Enactive Drift Regulation (EDR) as a general adaptive principle that treats drift as a regulatory signal indicating breakdowns in coherence between a system’s internal organization and its environment. Rather than optimizing predictions or retraining models, EDR reframes adaptation as the regulation of structure—maintaining, reorganizing, or transitioning internal dynamics to sustain viable operation under change. We present the Emergence Machine as an architectural instantiation of EDR, organized around regimes, attractors, coherence measures, reorganization dynamics, and memory across regimes. By shifting the focus from error minimization to coherence regulation, this work provides a principled framework for long-duration adaptation under non-stationarity and offers a bridge between adaptive control and enactive accounts of cognition.

1. Introduction: Adaptation Under Drift

Adaptive artificial systems are increasingly deployed in environments characterized by continuous change. From neural and physiological signals to climate dynamics, financial markets, and interactive human–machine settings, the conditions under which systems operate are rarely stationary (Benjamini, 2008; Gama et al., 2014; Widmer & Kubat, 1996). Patterns shift, relationships reorganize, and the relevance of past experience decays unevenly over time. Yet much of contemporary machine learning and adaptive systems research remains grounded in assumptions of stationarity, or treats non-stationarity as an exceptional condition to be corrected rather than as a fundamental property of real-world environments (Quiñonero-Candela et al., 2009).

This mismatch has practical consequences. Systems that perform well under controlled or short-term conditions often degrade during prolonged deployment. Predictive accuracy declines, internal models become misaligned with their environments, and ad hoc solutions—such as retraining, recalibration, or sliding windows—are repeatedly applied to restore performance (Žliobaitė et al., 2015). These interventions are costly, brittle, and frequently insufficient. More importantly, they obscure a deeper issue: many adaptive systems are designed to learn under stable assumptions, not to regulate themselves under drift.

This paper argues that drift is not an anomaly but the norm, and that effective adaptation under such conditions requires a different computational framing. Rather than treating drift as noise, error, or distribution shift to be eliminated, we propose Enactive Drift Regulation (EDR) as a principled approach that treats drift as a first-class signal indicating breakdowns of coherence between a system and its environment. From this perspective, adaptation is not primarily a matter of updating parameters or optimizing predictions, but of regulating the organization of internal dynamics to maintain coherence over time (Ashby, 1956; Di Paolo, 2005).


1.1 Why Stationarity Assumptions Fail

Many dominant approaches to learning and prediction assume that the statistical properties of the data-generating process are stable, or at least piecewise stationary. Even online and continual learning methods often rely on implicit stationarity assumptions within bounded windows (Bottou et al., 2018; Parisi et al., 2019). When these assumptions hold approximately, incremental updates can track gradual change and maintain performance.

However, in many real-world domains, stationarity fails in more fundamental ways. Signals may exhibit regime changes, phase transitions, or shifts in underlying generative structure that cannot be captured by smooth parameter drift (Peters, Janzing, & Schölkopf, 2016). In such cases, updating a single model—even continuously—forces incompatible patterns into a shared representational space. The result is not graceful adaptation but instability, interference, or loss of interpretability (French, 1999; Kirkpatrick et al., 2017).

Examples of this failure appear across domains. In neural and physiological modeling, signal characteristics vary with fatigue, attention, learning, and context (Makeig et al., 2004; Millán et al., 2010). In environmental and economic systems, external shocks and long-term trends interact in non-linear ways (Farmer & Foley, 2009). In interactive systems, human behavior changes in response to the system itself, producing feedback loops that violate assumptions of independent and identically distributed data (Suchman, 1987; Licklider, 1960). In all of these settings, the environment does not merely drift—it reorganizes.

Stationarity assumptions fail because they conflate two distinct phenomena: variation within a regime and transitions between regimes. While the former can often be handled through incremental learning, the latter require structural reorganization. Treating regime change as noise or gradual drift obscures the need for mechanisms that can detect and respond to deeper breakdowns of coherence (Gershman, Blei, & Niv, 2010).


1.2 Why Retraining Is Insufficient

A common response to non-stationarity is retraining. When performance degrades, models are retrained on recent data, sometimes from scratch and sometimes through fine-tuning. While retraining can temporarily restore accuracy, it does not address the underlying problem and often introduces new failure modes (Goodfellow et al., 2016).

First, retraining is reactive rather than regulatory. It responds to degraded performance after the fact, rather than maintaining coherence proactively as conditions change. Second, retraining often leads to catastrophic forgetting, erasing previously learned structure that may become relevant again (McCloskey & Cohen, 1989; Parisi et al., 2019). Third, retraining assumes that there exists a single model configuration that can adequately capture the current environment, an assumption that fails in settings where multiple regimes coexist or recur.

More fundamentally, retraining treats adaptation as a problem of model content rather than model organization. It assumes that the primary challenge lies in fitting the right parameters, rather than in managing the structure of representations across time. As a result, retraining addresses symptoms—declining accuracy or increasing error—without providing insight into why the system’s organization has become mismatched to its environment.

In contrast, adaptive biological systems rarely respond to change by erasing and rebuilding internal structure wholesale. Instead, they regulate stability and plasticity through mechanisms that preserve coherence across time, enabling reorganization without collapse (Ashby, 1956; Di Paolo & De Jaegher, 2012). This suggests that effective adaptation under drift requires regulatory mechanisms that operate at the level of system organization rather than parameter optimization alone.


1.3 Why Drift Is the Norm, Not the Exception

Treating drift as an exceptional condition reflects a narrow view of adaptation. In open-ended, long-duration systems, drift is not a deviation from normal operation—it is the default state. Environments change, agents act within them, and those actions alter the conditions to which the system must respond (Clark, 1997; Di Paolo, 2005). Even in seemingly static domains, long-term dynamics introduce shifts that invalidate fixed assumptions.

From this perspective, drift should be understood not merely as distribution shift, but as a breakdown of coherence between a system’s internal organization and the structure of its environment. Such breakdowns manifest as increasing instability, loss of predictive consistency, or fragmentation of internal dynamics. Crucially, these signals contain information: they indicate that the system’s current mode of operation no longer fits its context (Ashby, 1956).

Recognizing drift as normative reframes the adaptive challenge. The goal is no longer to eliminate drift, but to use it as a signal for regulation. This requires mechanisms that can detect when coherence is degrading, identify the scale at which reorganization is needed, and restructure internal dynamics accordingly. Adaptation becomes an ongoing process of balancing stability and change, rather than a sequence of corrections applied to an otherwise fixed model.


1.4 Enactive Drift Regulation and the Emergence Machine

In response to these challenges, this paper introduces Enactive Drift Regulation (EDR) as a computational principle for adaptation under non-stationarity. EDR treats drift as a first-class regulatory signal and reframes adaptation as the maintenance of coherence through structural reorganization rather than error minimization alone (Di Paolo, 2005; Ashby, 1956).

We present the Emergence Machine as a concrete instantiation of this principle. Rather than operating as a monolithic predictor, the Emergence Machine maintains multiple regimes, stabilizes attractors within those regimes, and reorganizes its internal structure when coherence degrades. In doing so, it supports long-duration operation under drift without assuming stationarity or relying on continual retraining.

The remainder of this paper develops EDR as a general adaptive framework, details the architecture and dynamics of the Emergence Machine, and situates this approach in relation to existing learning and control paradigms. Together, they offer an alternative foundation for adaptive systems designed to operate where drift is not an exception, but the rule.

2. Drift Reframed: From Error to Breakdown of Coherence

The term drift is widely used across machine learning, signal processing, and adaptive systems, yet it is rarely given a precise or unified interpretation. Depending on context, drift may refer to noise, error accumulation, distribution shift, or changes in task structure (Gama et al., 2014; Quiñonero-Candela et al., 2009). This section argues that these interpretations obscure a more fundamental phenomenon. Rather than treating drift as a statistical anomaly or performance degradation, we reframe it as a breakdown of coherence between a system’s internal organization and the structure of its environment. This reframing is central to Enactive Drift Regulation (EDR), as it shifts the adaptive problem from correction to regulation.


2.1 Drift Versus Noise

Drift is often conflated with noise. In signal processing and statistical modeling, noise refers to stochastic variability superimposed on an underlying signal. Noise is typically assumed to be independent, zero-mean, and uninformative with respect to the system’s structure (Kay, 1993). From this perspective, robustness to noise is achieved through filtering, smoothing, or averaging, all of which aim to recover a stable signal by suppressing irrelevant variation.

Drift differs fundamentally from noise. Whereas noise is structureless, drift is structured and directional. Drift reflects systematic changes in the relationship between signals, latent variables, or environmental conditions (Widmer & Kubat, 1996; Žliobaitė et al., 2015). Importantly, these changes are not random perturbations around a stable mean, but alterations to the mean itself—or to the structure that defines what the mean represents.

Treating drift as noise leads to inappropriate responses. Filtering strategies that suppress variability may delay detection of drift, while smoothing can actively mask the onset of structural change (Basseville & Nikiforov, 1993). From an enactive perspective, drift is not something to be eliminated, but something to be interpreted. It signals that the system’s current organization is losing its fit with the environment. Suppressing this signal prevents timely reorganization and accelerates longer-term breakdown.


2.2 Drift Versus Error

Drift is also frequently understood through the lens of error. In predictive systems, rising error is often taken as evidence that a model has become miscalibrated or outdated. The standard response is to reduce error through retraining, parameter adjustment, or model replacement (Goodfellow et al., 2016).

While error and drift are related, they are not equivalent. Error is a local, outcome-based measure: it quantifies deviation between predicted and observed values at specific points in time. Drift, by contrast, is a temporal, relational phenomenon. It reflects the accumulation of mismatches across time that indicate a deeper loss of coherence (Ashby, 1956).

A system may experience increasing error without drift—for example, due to transient perturbations that do not alter underlying structure. Conversely, a system may exhibit minimal short-term error while drifting structurally, particularly if it compensates locally while its internal organization becomes increasingly fragmented (Parisi et al., 2019). In such cases, error-based triggers delay adaptation until breakdown becomes severe.

EDR therefore treats error as an insufficient signal for adaptation. While error may indicate that something is wrong, it does not reveal what kind of reorganization is needed or at what scale. Drift, understood as a breakdown of coherence, provides richer information: it indicates that the system’s current regime no longer organizes experience effectively.


2.3 Drift Versus Concept Shift

In machine learning, drift is often formalized as concept drift: changes in the mapping between inputs and outputs over time (Gama et al., 2014). Concept drift frameworks distinguish between gradual, abrupt, and recurring shifts, and propose strategies such as windowing, ensemble models, or change-point detection to manage them (Bifet & Gavaldà, 2007).

While concept drift captures an important class of non-stationary phenomena, it remains limited in two respects. First, it presupposes a fixed task structure—typically input–output mappings—within which drift occurs. Second, it frames adaptation primarily as model selection or replacement, rather than as regulation of internal dynamics.

EDR generalizes beyond concept drift by decoupling drift from predefined tasks. In many real-world systems, especially those involving interaction, embodiment, or complex environments, there is no stable concept to drift from (Clark, 1997; Di Paolo, 2005). Instead, the system’s engagement with the environment continuously reshapes what counts as relevant structure. Drift in such settings reflects not just changes in mapping, but changes in the organization through which experience is interpreted.

Moreover, concept drift approaches often assume that once a shift is detected, the appropriate response is to switch to a new model. EDR, in contrast, emphasizes continuity: preserving useful structure while reorganizing where necessary. Drift does not demand wholesale replacement, but selective reconfiguration guided by coherence signals.


2.4 Drift as Loss of Regime Coherence

The core claim of EDR is that drift should be understood as a loss of regime coherence. A regime, in this sense, is a temporally extended organization of internal dynamics that stabilizes perception, prediction, or action under particular conditions (Gershman et al., 2010). Regimes are not static models; they are patterns of coordination that remain viable as long as the environment supports them.

Drift occurs when the coherence that sustains a regime degrades. This degradation may manifest as increasing instability, reduced predictive consistency, conflict between internal processes, or fragmentation across temporal scales. Importantly, these signals are not arbitrary—they reflect the system’s ongoing engagement with an environment that no longer conforms to the assumptions embedded in the current regime (Ashby, 1956; Di Paolo & De Jaegher, 2012).

By framing drift as loss of regime coherence, EDR shifts the adaptive focus from correcting outputs to regulating structure. The task is not to force a regime to fit indefinitely, but to detect when its coherence is breaking down and to reorganize accordingly. Drift thus becomes a regulatory signal that guides transitions between regimes, stabilizes new attractors, and preserves continuity across change.

This reframing has significant implications. It suggests that adaptation under drift requires mechanisms for monitoring coherence, maintaining multiple regimes, and supporting transitions without collapse. It also implies that successful long-term adaptation depends less on predictive optimality and more on the system’s ability to remain organized in the face of structural change.

Drift is not noise to be filtered, error to be minimized, or concept shift to be patched. It is a signal that a system’s current organization is losing coherence with its environment. By reframing drift in this way, Enactive Drift Regulation provides a foundation for adaptive systems that can reorganize themselves under non-stationarity rather than merely reacting to its symptoms. The next section formalizes EDR as a regulatory principle and articulates how coherence, rather than accuracy alone, becomes the central variable of adaptation.

3. Enactive Drift Regulation (EDR)

This section introduces Enactive Drift Regulation (EDR) as a general computational principle for adaptive systems operating under non-stationary conditions. EDR reframes adaptation not as the optimization of predictions or the correction of error, but as the regulation of internal organization in response to drift. In doing so, it provides a principled alternative to learning-centric accounts of adaptation and establishes the theoretical foundation for the Emergence Machine described in subsequent sections (Ashby, 1956; Di Paolo, 2005).


3.1 Definition of Enactive Drift Regulation

We define Enactive Drift Regulation (EDR) as follows:

Enactive Drift Regulation is the process by which an adaptive system maintains coherence with its environment by monitoring drift as a regulatory signal and reorganizing its internal dynamics accordingly.

Several aspects of this definition are essential.

First, EDR treats regulation, rather than learning, as the primary adaptive act. While learning updates parameters or representations within a given organizational structure, regulation operates at the level of that structure itself (Ashby, 1956; Conant & Ashby, 1970). EDR is therefore concerned with when existing organization should be preserved, when it should be modified, and when it should be replaced or reconfigured.

Second, EDR is drift-driven. Drift is not treated as a residual error or performance artifact, but as an endogenous signal indicating that the system’s current mode of organization is losing its coherence. In EDR, adaptation is triggered not by failure at the output level alone, but by degradation in the internal consistency that sustains viable operation (Di Paolo, 2005; Gershman et al., 2010).

Third, EDR is coherence-oriented rather than objective-oriented. The goal of regulation is not to optimize a predefined cost function, but to preserve the system’s capacity to remain meaningfully coupled to its environment over time. This orientation distinguishes EDR from control-theoretic and reinforcement learning approaches that assume stable objectives or reward structures (Sutton & Barto, 2018; Friston, 2010).


3.2 Drift as a Regulatory Signal

A central departure of EDR from conventional adaptive paradigms lies in how it interprets drift. Rather than viewing drift as a problem to be corrected, EDR treats it as information about the adequacy of the system’s current organization.

In predictive and learning-based systems, rising error is often the primary trigger for adaptation. However, error is a local and outcome-dependent signal: it indicates mismatch at specific points in time but provides little guidance about how the system should reorganize (Goodfellow et al., 2016). Drift, as understood in EDR, is a higher-order signal. It reflects the loss of coherence across time, manifested as instability, fragmentation, or inconsistency in the system’s internal dynamics (Ashby, 1956; Parisi et al., 2019).

Crucially, drift signals are not binary. They vary in magnitude, scale, and persistence, allowing regulation to occur incrementally rather than catastrophically. Small coherence losses may call for minor adjustments or increased plasticity, while sustained or large-scale drift may require regime transition or structural reorganization (Gama et al., 2014). By operating on drift rather than error alone, EDR enables adaptation that is anticipatory and graded rather than reactive and disruptive.

This interpretation aligns with how adaptive biological systems respond to change. Rather than waiting for failure, such systems continuously monitor their own stability and reorganize proactively when coherence begins to degrade (Di Paolo & De Jaegher, 2012; Ashby, 1956). EDR abstracts this principle into a computational framework suitable for artificial systems.


3.3 Regulation as Structural Reorganization

In EDR, regulation is not implemented through continuous parameter tuning alone. Instead, it involves structural reorganization of internal dynamics. Structure here refers to the patterns of coordination—regimes, attractors, coupling relations—that govern how the system interprets and responds to its environment (Gershman et al., 2010).

Structural reorganization may include:

Importantly, EDR does not assume that reorganization entails discarding prior structure. On the contrary, effective regulation preserves useful organization whenever possible, enabling reactivation or reuse when conditions recur. This contrasts with retraining-based approaches, which often erase prior structure and thereby lose long-term continuity (McCloskey & Cohen, 1989; Parisi et al., 2019).

By separating regulation from learning, EDR clarifies a distinction that is often blurred in adaptive systems. Learning modifies content within a given structure; regulation modifies the structure that makes learning meaningful. Both processes may coexist, but they operate at different levels and on different timescales. EDR foregrounds the latter as essential for long-duration adaptation under drift.


3.4 Relation to Enaction

EDR is inspired by enactive theories of cognition, but it does not attempt to reproduce enactivism in full philosophical detail. Instead, it adopts a disciplined enactive stance focused on organizational principles relevant to adaptive systems.

From an enactive perspective, cognition is not the manipulation of internal representations but the maintenance of viable coupling between an organism and its environment. Meaning arises through action, and breakdowns in coupling signal the need for reorganization (Varela, Thompson, & Rosch, 1991; Di Paolo, 2005). EDR operationalizes this insight by treating drift as a signal of lost viability and regulation as the means by which coupling is restored.

However, EDR differs from some enactive accounts in its scope and intent. It does not make claims about subjective experience, embodiment, or consciousness. Nor does it require that artificial systems enact meaning in the human sense. Instead, EDR translates enactive ideas into a computationally tractable framework concerned with coherence, organization, and adaptation over time.

In this sense, EDR occupies a middle ground. It avoids the representational commitments of predictive and optimization-centered models (Clark, 2013), while also avoiding metaphysical claims about artificial agency. Enaction functions here as a design constraint, not a philosophical endpoint: adaptive systems must regulate their organization to remain coherently engaged with their environments.


3.5 Summary and Implications

Enactive Drift Regulation reframes adaptation as the regulation of coherence under drift rather than as the minimization of error under stationary assumptions. By treating drift as a first-class regulatory signal and emphasizing structural reorganization, EDR provides a principled account of how adaptive systems can operate effectively in environments that change in fundamental and unpredictable ways.

This reframing has several implications. It suggests that long-term adaptability depends less on predictive accuracy than on organizational resilience. It highlights the importance of regime management and multi-scale dynamics. And it provides a conceptual foundation for architectures—such as the Emergence Machine—that are designed not merely to learn from data, but to remain coherently organized as conditions evolve.

The next section builds on this theoretical foundation by describing the Emergence Machine as a concrete instantiation of Enactive Drift Regulation, demonstrating how these principles can be realized in an adaptive computational system.


4. The Emergence Machine Architecture

The Emergence Machine is a computational architecture designed to instantiate Enactive Drift Regulation (EDR). Rather than operating as a single predictive model optimized for accuracy, it functions as a regulator of internal organization under drift. Its purpose is not merely to produce predictions, but to maintain coherence across time by organizing, monitoring, and reorganizing its internal dynamics in response to changing conditions (Ashby, 1956; Conant & Ashby, 1970).

This section describes the architectural principles of the Emergence Machine. The description is intentionally abstract: the goal is to articulate the organizational logic that supports EDR rather than to prescribe a specific implementation. The architecture is structured around five interrelated components: regimes, attractors, coherence measures, reorganization dynamics, and memory across regimes (see Figure 1).

4.1 Regimes: Structured Modes of Operation

At the highest level, the Emergence Machine organizes its operation into regimes. A regime is a temporally extended mode of internal organization that stabilizes perception, prediction, or action under a particular set of conditions (Gershman et al., 2010). Regimes are not static models, nor are they tied to predefined tasks. Instead, they are patterns of coordination that remain viable as long as they maintain coherence with incoming signals (Di Paolo, 2005).

Each regime encodes a distinct way of structuring experience. Different regimes may correspond to different signal characteristics, behavioral contexts, or environmental conditions. Importantly, regimes can coexist: the Emergence Machine does not assume that only one regime is valid at a time. Instead, it maintains a repertoire of regimes and regulates which ones are active, dominant, or latent (Grossberg, 1980).

This regime-based organization enables the system to avoid forcing incompatible patterns into a single representational space. Rather than continuously adapting one model to fit all conditions, the Emergence Machine can stabilize multiple modes of operation and transition between them when coherence degrades (Parisi et al., 2019).


4.2 Attractors: Stabilizing Internal Dynamics

Within each regime, the Emergence Machine maintains attractors—stable patterns of internal dynamics that organize ongoing processing. Attractors may take many forms depending on the implementation: recurrent trajectories, stable parameter configurations, or persistent relational patterns. What defines an attractor is not its mathematical form, but its functional role in stabilizing behavior over time (Kelso, 1995; Ashby, 1956).

Attractors provide local coherence. They enable the system to remain consistent in the face of variation and noise, preventing overreaction to transient perturbations. At the same time, attractors are not immutable. Their stability is continually assessed through coherence measures, and they may weaken, shift, or dissolve as conditions change (Haken, 1983).

Crucially, attractors are regime-specific. An attractor that is coherent within one regime may be incoherent in another. This separation allows the Emergence Machine to preserve useful structure without forcing it to generalize beyond its viable domain.


4.3 Coherence Measures: Monitoring Viability

Central to EDR is the ability to assess whether current organization remains viable. The Emergence Machine does this through coherence measures—signals that quantify the internal consistency and stability of ongoing dynamics (Ashby, 1956).

Coherence measures are not limited to prediction error. While error may contribute, coherence is fundamentally relational and temporal. It reflects how well internal processes agree with one another across time, how stable attractors remain under perturbation, and whether regime-level organization continues to organize incoming signals effectively (Di Paolo & De Jaegher, 2012).

Examples of coherence indicators may include:

The precise form of these measures is implementation-dependent. What matters is their role: coherence measures serve as regulatory signals that inform whether the current organization should be maintained, adjusted, or reorganized. In this sense, they operationalize drift as a first-class signal.


4.4 Reorganization Dynamics: Regulating Structure

When coherence degrades beyond tolerable bounds, the Emergence Machine engages reorganization dynamics. Reorganization is not a binary reset or wholesale retraining. Instead, it is a graded process that adjusts internal organization at the appropriate scale (Ashby, 1956; Grossberg, 1980).

Reorganization dynamics may include:

Importantly, reorganization is selective and conservative. The goal is not to maximize change, but to restore coherence with minimal disruption. Where possible, existing structure is preserved and repurposed rather than discarded. This supports continuity and reduces catastrophic forgetting (McCloskey & Cohen, 1989).

Reorganization dynamics embody the core principle of EDR: adaptation through regulation of structure rather than through continuous optimization alone. They enable the system to remain organized under drift without assuming stationarity or relying on external intervention.


4.5 Memory Across Regimes: Continuity Without Freezing

Long-duration adaptation requires memory, but not the kind of memory that freezes a system into outdated configurations. The Emergence Machine therefore maintains memory across regimes—a form of structural memory that preserves the availability of past organizational patterns without forcing their continued dominance (Parisi et al., 2019).

This memory allows regimes to be reactivated when conditions recur, supporting recurrence and cyclic environments. It also enables the system to compare current conditions with prior organizational states, informing reorganization decisions (Gershman et al., 2010).

Crucially, memory across regimes is not a simple archive of past data or parameters. It is a memory of organizational solutions: ways of structuring dynamics that were once coherent. By retaining this memory, the Emergence Machine avoids relearning from scratch and supports adaptive reuse.


4.6 Architectural Summary

Taken together, these components form an architecture that instantiates Enactive Drift Regulation:

The Emergence Machine is therefore not best understood as a predictor, classifier, or controller. It is a regulator of predictive organization—a system designed to remain coherently organized as its environment changes (Ashby, 1956). By separating regulation from learning and treating drift as a signal rather than a defect, the Emergence Machine provides a concrete instantiation of EDR suitable for long-duration, non-stationary domains.

The Emergence Machine has been instantiated and evaluated across multiple chaotic and non-stationary time-series domains, including ETTh1, ETTh2, EEG, and ECG signals. These studies demonstrate the system’s ability to maintain regime coherence and recover from drift over long durations. Detailed empirical results are reported in companion work.

The next section situates this architecture in relation to existing adaptive paradigms, clarifying how EDR and the Emergence Machine address a different class of problems than conventional learning-based approaches.

5. Comparative Analysis: EDR as a Distinct Adaptive Paradigm

Enactive Drift Regulation (EDR) does not compete directly with existing adaptive learning paradigms on performance metrics such as accuracy, convergence speed, or sample efficiency. Instead, it addresses a different class of adaptive problem: how systems can remain coherently organized over time when the structure of their environment changes in ways that violate stationarity assumptions (Ashby, 1956; Di Paolo, 2005). This section situates EDR in relation to several influential approaches—continual learning, adaptive filtering, online learning, and predictive processing—clarifying both points of contact and fundamental differences.


5.1 EDR vs. Continual Learning

Continual learning frameworks aim to enable systems to acquire new knowledge without catastrophically forgetting prior knowledge. Techniques such as rehearsal, regularization, and modularization seek to preserve past task performance while learning new tasks sequentially (McCloskey & Cohen, 1989; Parisi et al., 2019; Kirkpatrick et al., 2017).

While continual learning addresses an important problem, it presupposes a task-centric framing: the environment is decomposed into a sequence of tasks, each of which can be learned, stored, and protected. Drift is treated as a transition between tasks, and success is measured by retention of performance across that sequence.

EDR departs from this framing in two key ways. First, it does not assume that environmental change can be decomposed into discrete tasks. In many real-world domains, especially continuous time-series and interactive settings, change is gradual, overlapping, and recurrent rather than task-bounded (Gama et al., 2014). Second, EDR is not primarily concerned with preserving task performance, but with maintaining organizational coherence.

From an EDR perspective, catastrophic forgetting is a symptom rather than the core problem. The deeper issue is that forcing incompatible regimes into a single organizational structure creates internal incoherence. By explicitly maintaining multiple regimes and regulating transitions between them, EDR avoids the need to freeze parameters or protect representations. Forgetting becomes a controlled reallocation of dominance rather than a failure to retain content.


5.2 EDR vs. Adaptive Filtering

Adaptive filtering techniques, such as Kalman filters and their variants, are designed to track changing signals by continuously updating estimates in response to new data (Kalman, 1960; Haykin, 2002). These methods excel in environments where changes are smooth and can be captured by incremental parameter updates.

However, adaptive filtering assumes that the form of the model remains valid as parameters change. Drift is treated as gradual variation around a stable generative process, and adaptation consists of tracking that variation as efficiently as possible.

EDR addresses a different scenario: environments in which the structure of the generative process itself changes. In such cases, no amount of parameter tuning within a fixed filter structure is sufficient. The problem is not tracking, but reorganization.

Where adaptive filtering emphasizes responsiveness and smoothness, EDR emphasizes regime viability. A regime may be locally stable and still globally incoherent. EDR therefore introduces coherence monitoring and regime transition as first-class mechanisms, enabling adaptation through structural change rather than continuous estimation alone.


5.3 EDR vs. Online Learning

Online learning frameworks update models incrementally as data arrives, often under assumptions of bounded regret or slowly changing distributions (Cesa-Bianchi & Lugosi, 2006; Shalev-Shwartz, 2012). These approaches are well suited to streaming data and non-stationary environments where adaptation must occur continuously.

Despite their flexibility, online learning methods typically optimize a single objective function over time. Drift is handled by adjusting learning rates, discounting older data, or sliding windows to emphasize recent observations. While these techniques can improve responsiveness, they do not fundamentally alter the organization of the model.

EDR differs by decoupling adaptation from optimization. Rather than continuously chasing an objective that itself may be drifting, EDR treats coherence as the primary variable to regulate. When coherence degrades, the appropriate response may not be faster learning, but structural reconfiguration—activating a different regime, stabilizing a new attractor, or reorganizing memory across timescales.

In this sense, EDR complements rather than replaces online learning. Learning may occur within regimes, but regulation determines which regime is active and when learning should be emphasized or suppressed.


5.4 EDR vs. Predictive Processing

Predictive processing and active inference frameworks model cognition as the minimization of prediction error or variational free energy (Friston, 2010; Clark, 2013). These approaches have been influential in both cognitive science and machine learning, emphasizing hierarchical generative models and continual updating.

EDR shares with predictive processing an emphasis on coherence and stability rather than static correctness. However, it diverges in how adaptation is conceptualized. Predictive processing typically treats error minimization as the primary driver of change, with reorganization occurring implicitly through parameter updates or model selection.

In contrast, EDR treats drift—not error—as the central regulatory signal. Error may increase for many reasons, including noise or transient perturbations, without indicating a need for structural change. Drift, as loss of regime coherence, directly signals when the system’s current organization is no longer viable (Ashby, 1956).

Moreover, predictive processing frameworks often assume a single hierarchical model that must be continually refined. EDR explicitly allows for multiple coexisting regimes and does not require convergence toward a unified internal model. This plurality enables EDR to handle recurrent and cyclic environments without forcing reconciliation of incompatible structures.


Across these comparisons, a consistent distinction emerges. Continual learning, adaptive filtering, online learning, and predictive processing all ask some version of the question:

How can a system improve or maintain performance as conditions change?

EDR asks a different question:

How can a system remain coherently organized as the structure of its environment changes?

This shift in focus—from performance to coherence, from learning to regulation, from error to drift—defines the contribution of Enactive Drift Regulation. EDR does not replace existing adaptive methods; it reframes the problem they implicitly assume. In domains where non-stationarity is the norm and structural change is unavoidable, EDR offers a principled foundation for building systems that can endure.

7. Discussion: What Enactive Drift Regulation Enables

The preceding sections have argued that drift is not an exceptional failure mode but a defining condition of real-world adaptive systems, and that Enactive Drift Regulation (EDR) offers a principled response by reframing adaptation as the regulation of coherence rather than the optimization of performance (Ashby, 1956; Di Paolo, 2005). This discussion elaborates what this reframing enables in practice and why it matters for the design of adaptive systems intended to operate over extended durations in non-stationary environments.


7.1 Long-Duration Operation Under Drift

A central consequence of EDR is its support for long-duration operation. Many adaptive systems perform well in short deployments yet degrade over time as accumulated drift undermines their internal organization (Žliobaitė et al., 2015). Traditional responses—retraining, recalibration, or manual intervention—reset performance at the cost of continuity and stability.

EDR enables long-duration operation by introducing mechanisms that continuously monitor coherence and reorganize internal structure before breakdown becomes catastrophic. Because drift is treated as a regulatory signal rather than as a terminal failure, adaptation occurs incrementally and proactively (Ashby, 1956). Regimes can weaken, transition, or be reactivated without forcing the system to relearn from scratch (Parisi et al., 2019).

This capacity is especially important in domains where continuous operation is essential, such as physiological monitoring, environmental modeling, or interactive systems embedded in everyday contexts (Millán et al., 2010; Farmer & Foley, 2009). In these settings, adaptation must preserve continuity across hours, days, or months rather than optimizing for short-term accuracy. EDR directly targets this requirement by designing for endurance rather than episodic performance.


7.2 Robustness Without Freezing

Another key implication of EDR is a form of robustness without freezing. Many approaches to non-stationarity achieve robustness by reducing plasticity: freezing parameters, constraining updates, or heavily regularizing change to preserve prior knowledge (Kirkpatrick et al., 2017). While these strategies can stabilize behavior temporarily, they ultimately limit the system’s ability to adapt when conditions shift in more fundamental ways.

EDR offers an alternative by separating stability from rigidity. Stability is achieved through regime and attractor organization, which preserves coherence locally, while plasticity is enabled through regulated reorganization at appropriate scales (Grossberg, 1980; Ashby, 1956). This allows the system to remain robust to noise and transient perturbations without becoming brittle in the face of structural change.

In this sense, EDR reframes robustness not as resistance to change, but as the capacity to change without losing organization. Systems that regulate drift can remain stable precisely because they are able to reorganize when necessary, rather than attempting to suppress all variation. This balance mirrors biological adaptation, where robustness emerges from regulated flexibility rather than fixed structure (Di Paolo & De Jaegher, 2012).


7.3 Interpretability Through Regimes

EDR also has implications for interpretability, an increasingly important concern in adaptive and learning systems. Many high-performing models, particularly deep learning architectures, are difficult to interpret because their internal organization evolves continuously and opaquely (Lipton, 2018).

By organizing behavior into regimes and attractors, the Emergence Machine provides a natural unit of interpretability. Regimes correspond to distinct modes of operation that can be inspected, compared, and related to environmental conditions (Gershman et al., 2010). Transitions between regimes signal meaningful changes in how the system is engaging with its environment, rather than arbitrary parameter drift.

This form of interpretability is not explanatory in the sense of feature attribution or causal graphs, but it is organizationally transparent. It allows observers to understand what kind of situation the system believes it is in and how that belief changes over time. Such transparency is particularly valuable in long-duration systems, where understanding why behavior shifts is as important as the behavior itself.


7.4 Alignment with Embodied and Enactive Accounts

Finally, EDR aligns naturally with embodied and enactive accounts of cognition, while remaining disciplined in its claims. Enaction emphasizes that cognition arises through ongoing engagement with the environment, and that breakdowns in this engagement signal the need for reorganization rather than error correction alone (Varela et al., 1991; Di Paolo, 2005).

EDR operationalizes this insight at the level of system design. By treating drift as loss of coherence and regulation as structural reorganization, EDR mirrors the enactive view that cognition is an activity sustained over time, not a static computation performed on fixed representations. Importantly, this alignment does not require claims about consciousness, subjective experience, or full embodiment. Enaction functions here as a constraint on adaptive organization, not as a metaphysical commitment.

This disciplined alignment opens a productive space between symbolic, optimization-driven models and stronger philosophical enactivism (Clark, 2013). It suggests that artificial systems can exhibit enactive organization—regulating their coupling to the environment—without claiming enactive phenomenology. In doing so, EDR provides a bridge between cognitive theory and practical system design.


Taken together, these implications highlight what EDR enables that existing paradigms struggle to support: long-duration operation under drift, robustness without rigidity, interpretability through organizational structure, and principled alignment with enactive cognition. Rather than optimizing performance under assumed stability, EDR shifts the design target toward sustaining coherence in the presence of inevitable change.

This shift has consequences beyond any single architecture. It suggests that the defining challenge for adaptive systems may not be learning faster or predicting better, but remaining organized over time. EDR offers a framework for meeting that challenge by treating drift not as a defect to be eliminated, but as a signal that guides regulation and reorganization.

The following conclusion synthesizes these contributions and outlines directions for future work building on Enactive Drift Regulation.



8. Conclusion

This paper has introduced Enactive Drift Regulation (EDR) as a general adaptive principle for systems operating under persistent non-stationarity. Rather than treating drift as noise, error, or an exceptional condition to be corrected, EDR reframes drift as a regulatory signal indicating loss of coherence between a system’s internal organization and its environment. From this perspective, adaptation is not primarily a matter of optimizing predictions or retraining models, but of regulating structure to sustain viable operation over time.

By articulating EDR at the level of principle rather than algorithm, this work clarifies a gap in existing adaptive paradigms. Approaches such as continual learning, online optimization, and predictive processing address how systems can improve or maintain performance as data distributions change. EDR addresses a different question: how systems can remain coherently organized when the very structure of their environment reorganizes. This shift from performance to coherence, and from learning to regulation, defines the central contribution of the framework.

The Emergence Machine was presented as an architectural instantiation of EDR. Through its use of regimes, attractors, coherence measures, and reorganization dynamics, it demonstrates how regulatory principles can be realized computationally without assuming stationarity or relying on continual retraining. Importantly, the architecture emphasizes continuity across change, enabling long-duration operation while preserving interpretability through organizational structure.

EDR also functions as a conceptual bridge between cognitive theory and adaptive control. Drawing on enactive insights while remaining disciplined in its claims, the framework translates ideas of viability, coupling, and breakdown into design-relevant mechanisms. In doing so, it avoids both purely representational models that obscure organizational change and strong philosophical commitments that resist operationalization. EDR shows how enactive principles can inform the design of artificial systems without anthropomorphism or claims about subjective experience.

Finally, while this paper has focused on time-series and environmental adaptation, the implications of EDR extend beyond these domains. Any system engaged in sustained interaction—whether with dynamic environments, biological signals, or human partners—faces the challenge of maintaining coherence under drift. By foregrounding regulation as the central adaptive act, EDR provides a foundation for interactional systems that must remain intelligible, responsive, and stable across time without freezing or collapse.

In environments where change is inevitable and stationarity is the exception, sustaining coherence may be the defining challenge of adaptive intelligence. Enactive Drift Regulation offers a principled response to that challenge—one that treats drift not as a failure to be eliminated, but as the very signal that makes adaptation possible.

References

Ashby, W. R. (1956). An introduction to cybernetics. Chapman & Hall.

Basseville, M., & Nikiforov, I. V. (1993). Detection of abrupt changes: Theory and application. Prentice Hall.

Benjamini, Y. (2008). Simultaneous and selective inference: Current successes and future challenges. Statistical Science, 23(1), 1–15.

Bifet, A., & Gavaldà, R. (2007). Learning from time-changing data with adaptive windowing. In Proceedings of the SIAM International Conference on Data Mining.

Bottou, L., Curtis, F. E., & Nocedal, J. (2018). Optimization methods for large-scale machine learning. SIAM Review, 60(2), 223–311.

Cesa-Bianchi, N., & Lugosi, G. (2006). Prediction, learning, and games. Cambridge University Press.

Clark, A. (1997). Being there: Putting brain, body, and world together again. MIT Press.

Clark, A. (2013). Whatever next? Predictive brains, situated agents, and the future of cognitive science. Behavioral and Brain Sciences, 36(3), 181–204.

Conant, R. C., & Ashby, W. R. (1970). Every good regulator of a system must be a model of that system. International Journal of Systems Science, 1(2), 89–97.

Di Paolo, E. A. (2005). Autopoiesis, adaptivity, teleology, agency. Phenomenology and the Cognitive Sciences, 4, 429–452.

Di Paolo, E. A., & De Jaegher, H. (2012). The interactive brain hypothesis. Frontiers in Human Neuroscience, 6, 163.

Farmer, J. D., & Foley, D. (2009). The economy needs agent-based modelling. Nature, 460, 685–686.

French, R. M. (1999). Catastrophic forgetting in connectionist networks. Trends in Cognitive Sciences, 3(4), 128–135.

Friston, K. (2010). The free-energy principle: A unified brain theory? Nature Reviews Neuroscience, 11(2), 127–138.

Gama, J., Žliobaitė, I., Bifet, A., Pechenizkiy, M., & Bouchachia, A. (2014). A survey on concept drift adaptation. ACM Computing Surveys, 46(4), Article 44.

Gershman, S. J., Blei, D. M., & Niv, Y. (2010). Context, learning, and extinction. Psychological Review, 117(1), 197–209.

Goodfellow, I., Bengio, Y., & Courville, A. (2016). Deep learning. MIT Press.

Grossberg, S. (1980). How does a brain build a cognitive code? Psychological Review, 87(1), 1–51.

Haken, H. (1983). Synergetics: An introduction. Springer.

Haykin, S. (2002). Adaptive filter theory (4th ed.). Prentice Hall.

Kalman, R. E. (1960). A new approach to linear filtering and prediction problems. Journal of Basic Engineering, 82, 35–45.

Kay, S. M. (1993). Fundamentals of statistical signal processing. Prentice Hall.

Kelso, J. A. S. (1995). Dynamic patterns: The self-organization of brain and behavior. MIT Press.

Kirkpatrick, J., et al. (2017). Overcoming catastrophic forgetting in neural networks. Proceedings of the National Academy of Sciences, 114(13), 3521–3526.

Licklider, J. C. R. (1960). Man–computer symbiosis. IRE Transactions on Human Factors in Electronics, HFE-1, 4–11.

Lipton, Z. C. (2018). The mythos of model interpretability. Communications of the ACM, 61(10), 36–43.

Makeig, S., et al. (2004). Dynamic brain sources of visual evoked responses. Science, 295, 690–694.

McCloskey, M., & Cohen, N. J. (1989). Catastrophic interference in connectionist networks. Psychology of Learning and Motivation, 24, 109–165.

Millán, J. del R., et al. (2010). Combining brain–computer interfaces and assistive technologies. IEEE Signal Processing Magazine, 27(1), 27–37.

Parisi, G. I., et al. (2019). Continual lifelong learning with neural networks. Neural Networks, 113, 54–71.

Peters, J., Janzing, D., & Schölkopf, B. (2016). Causal inference using invariant prediction. Journal of the Royal Statistical Society: Series B, 78(5), 947–1012.

Prokopenko, M., Boschetti, F., & Ryan, A. J. (2009). An information-theoretic primer on complexity, self-organization, and emergence. Complexity, 15(1), 11–28.

Quiñonero-Candela, J., et al. (2009). Dataset shift in machine learning. MIT Press.

Shalev-Shwartz, S. (2012). Online learning and online convex optimization. Foundations and Trends in Machine Learning, 4(2), 107–194.

Suchman, L. A. (1987). Plans and situated actions. Cambridge University Press.

Sutton, R. S., & Barto, A. G. (2018). Reinforcement learning: An introduction (2nd ed.). MIT Press.

Varela, F. J., Thompson, E., & Rosch, E. (1991). The embodied mind. MIT Press.

Widmer, G., & Kubat, M. (1996). Learning in the presence of concept drift. Machine Learning, 22, 23–47.

Žliobaitė, I., et al. (2015). A survey on concept drift adaptation. IEEE Transactions on Neural Networks and Learning Systems, 26(10), 2412–2433.