Some degree of resilience is putatively accepted to be innate in the human response to adversity (Bonanno, 2021). Early work defined resilience as the absence of psychopathology, often PTSD (Agaibi and Wilson, 2005) after trauma. Traumatic events are common worldwide with exposure estimates ranging as high as 70% (Benjet et al., 2016). Disentangling the prevalence of long-term effects on mental and physical health is not yet feasible but lifetime estimates of PTSD are around 8% (Gradus, 2017). However, the classic 32-year longitudinal study of Kauaian children who had experienced adverse developmental conditions found one out of three grew into competent, confident, and caring adults (Werner, 1999) suggesting a profound effect of developmental chronic stress, along with remarkable resilience. This work led to a broader notion of resilience as “good survival,” or adaptation to and recovery from stress (Bonanno and Mancini, 2012; Masten, 2001). In good survival, some may even be positively changed by this adaptive process (Boden et al., 2014), through successful negotiation and management of the stressors experienced across the lifespan (Windle, 2011) to result in what has been described as post-traumatic growth and psychosocial gains from adversity (Johnson and Boals, 2015; Mancini, 2019).

Biopsychosocial systems research has demonstrated that resilient individuals have identifiable traits and practices, or resilience factors that are shaped by genetics (Feder et al., 2009), epigenetics (Smeeth et al., 2021; Zannas and West, 2014), culture, social resources (Hobföll, 1989, 2001), and life circumstances (Armbruster et al., 2012). Some argue that stress resilience is distinct from recovery from trauma (Richter-Levin and Sandi, 2021), although others argue that the same practices that help resilient individuals overcome significant adversity provide transferrable protection to coping with stress (Fletcher and Sarkar, 2013). We argue that enactive allostasis can explain resilient processes and outcomes associated with both trauma and with chronic stress.

While relatively static trait and malleable contextual resilience factors have been combined to reliably predict the likelihood of resilience outcomes at the group level, resilience outcomes are very difficult to predict in individuals. This contextual dependence has been labeled the “resilience paradox” (Bonanno, 2021).

Similar to the resilience trajectories introduced above (Bonanno, 2004; Bonanno et al., 2011, 2023; Galatzer-Levy et al., 2018), Kalisch et al. (2024) provide a longitudinal model to enhance the valid measurement of resilient processes and resilient outcomes (Kalisch et al., 2021). Dubbed the Frequent Stressor and Mental Health Monitoring (FRESHMO) paradigm, the ratio of mental health reactivity to stress exposure, termed “stress reactivity,” is calculated across time.

Bonanno’s earlier work identifying four prototypical trajectories of adjustment following potentially traumatic events (Bonanno, 2004; Bonanno et al., 2011) also comports with the notion of allostatic phenotypes. A meta-analysis (Galatzer-Levy et al., 2018) confirmed those trajectories: (1) no dysfunction following the event; (2) immediate dysfunction with gradual recovery; (3) a delayed trajectory with dysfunction increasing over time; and (4) an emerging chronic level of dysfunction. Bridging work on trajectories and FRESHMO, we propose the resilient outcome of individuals reflects the accuracy of the brain’s predictive models that undergird regulatory control during and after stress and/or adversity. i.e., enactive allostasis. Prediction includes all the biopsychosocial systems associated with resilience in a complex, dynamic model from which phenotypes emerge. These phenotypes allow for the application of our theoretical approach to individuals.

Fundamentally, prediction both anticipates events that may occur and prepares the system to maintain regulation before the need arises. Predictive allostasis reduces error, matches the capacities of different response components, shares resources among systems to reserve overall capacity, and integrates past errors to improve future predictions (Sterling, 2012). In prediction, the brain develops models based on prior experience and traits to allocate resources needed to react to stress. Successful allostatic predictions not only indicate an ability to predict the internal demands needed to respond to an environmental change, e.g., fight or flight, but it also suggests the subsequent allostatic accommodation, e.g., rest and recovery. Allostatic predictions can result in allostatic responses that ultimately lead to chronic homeostatic disruption—e.g., the association of diabetes with AL (Steptoe et al., 2014). It is equally plausible to argue predictive allostasis can maintain and potentially enhance the efficiency of the neuropsychobiological systems that maintain homeostasis across time and environmental change.

As elegantly evidenced in the regulation of cortisol (de Kloet and Joëls, 2023), stress responses demonstrate that with every activation there is a possible recovery mechanism. Stress accommodation is situational, determined by interaction with the broad physical and affiliative environment, shaped by factors occurring across variable time courses and realized across all levels of neurobiological expression, genetic to behavioral. Thus, we argue that allostatic predictions are enactive, informed by both the person and their environment, where we use the terms enactive and predictive allostasis interchangeably. Within this context, a broad range of individual differences in predictive allostasis is observed, which are presented as phenotypes.

Having a generative model that entails agency, necessarily, postulates a model of the consequences of action in any domain (i.e., motoric or autonomic). In a minimal sense, this kind of generative model—under which the degree of surprise or free energy is defined—is a model of the self as agent.

To reframe, the brain essentially acts as a prediction-processing network, where prior beliefs are updated with new information to form posterior beliefs, in the spirit of Bayesian belief updating ³ (Friston, 2010; Friston et al., 2006). Surprise cannot be quantified but is analogous to the variational free energy that systems seek to minimize. Beliefs are updated based on their accuracy, or the degree to which they correctly predict current sensory information, and their precision or reliability over time ⁴ (Friston, 2009). Precision is encoded by neurons reporting prediction errors being given higher synaptic gain—encoded at fast timescales through neuromodulation of synaptic efficacy, and through to neuroendocrine-mediated plasticity and learning, and slower structure learning and reconfiguration of generative models associated with changes in immune responses (Arnaldo et al., 2022).

Technically, EFE, i.e., expected surprise, is reduced through action via policy selection (Arnaldo et al., 2022). Acting to optimize preferred sensory inputs will pre-emptively minimize the divergence between anticipated and sampled sensory input, thereby minimizing expected surprise (Friston et al., 2016). Expected surprise can also be reduced by updating relevant beliefs to reduce uncertainty. These two aspects of minimizing EFE can be read as epistemic and instrumental or pragmatic affordances; in exactly the same way that distinguishes between exploration and exploitation. In other words, the single imperative to minimize EFE manifests as curious, information-seeking behavior that is constrained by the prior preferences that shape goal-seeking behavior. The relative precision of epistemic and instrumental affordances translates into preferences to explore or exploit the environment (Caddick and Rottman, 2021; Friston et al., 2017), and may rest on genetic, epigenetic, and developmental factors.

It is also helpful to decompose EFE into risk, the difference between predicted and a priori preferred outcomes in the future, and ambiguity, which is the uncertainty associated with future observations, given current states (Da Costa et al., 2020). This further clarifies the role of exploitative, i.e., risk minimizing (goal-seeking), and explorative (information-seeking), ambiguity minimizing, policy selection. Da Costa et al. (2020) highlight the crucial role of awareness as such: “planning and decision-making, respectively, correspond to evaluating the expected free energy of different policies, which scores their goodness in relation to prior preferences and forming approximate posterior beliefs about policies.” This EFE formulation of active inference has also been related to interoceptive control under allostasis (Tschantz et al., 2022), supporting the idea of allostatic growth, using the same model as AL.

The enactive process described by Friston et al. (2021) as sophisticated inference is also essential to our model. Using the economic definition of sophisticated as ‘beliefs about beliefs’, or meta-beliefs, of either one’s own or others, meta-beliefs are stable, high-level beliefs that constrain lower-level beliefs—are foundational for the enactive niche construction. Meta-beliefs begin with the assumption that the intrinsic value of every action is its epistemic value or affordance (Friston et al., 2015). Under sophisticated inference, planning becomes a belief generation strategy. This link introduces conditional dependencies between the past for the selection of actions in the present and for similar selection of future paths, e.g., mental time-travel (Friston et al., 2021). In effect, these meta beliefs reflect beliefs states or “what I would believe about what would happen if I did that,” as compared to “what would happen if I did that” as is the case when modeling a single belief (Friston et al., 2021).

Impacting the successful formation of predictions, we consider three interacting levels of awareness as consistent with the multidisciplinary literature on cognition and psychology and easily integrated into active inference: (1) sensory, (2) interoceptive, and (3) enactive, or meta, awareness. We observe overlap in these levels similar to the heterarchy, Arnaldo et al. (2022) propose to account for the coding of precision of beliefs across timescales. These levels of awareness similarly occur across timescales and can feed forward and backward to impact awareness in the other levels.

Sensory awareness and processing relate to the rapid neuromodulatory and neurotransmitter effects of sensory inputs; in subjective terms, what we are hearing, seeing, feeling, smelling, and tasting at each given moment. Sensory differences are implicated in numerous clinical conditions (Harrison et al., 2019). Sensory awareness predominates in certain enactive strategies, where sensory sensitivity and processing disruptions interrupt the development of enactive awareness as in obsessive-compulsive disorder (OCD) (Hulle et al., 2019). Sensory awareness precedes appraisal in our formulation. Conversely, we argue that when enactive awareness is developed through optimization of the enactive niche, sensory awareness provides input into all predictive models including epistemic models of self as well deep temporal “planning as inference” models.

With heightened sensory sensitivity, overly precise beliefs can be developed that override prior beliefs that underwrite intentions and agency, resulting in false perceptual inference, e.g., delusions in schizophrenia (Friston et al., 2013) or an awareness-enhanced level of precision associated with maintenance of negative emotional states (Schwartenbeck et al., 2015).

Interoceptive sensations constitute the afferent physiological information from the body to the brain and allow the organism to be aware of its regulatory status. Interoceptive awareness provides feedback on the success of enactive allostasis as inputs provided by the viscera (Craig, 2003; Quadt et al., 2022) and, in combination with sensory awareness and ANS and HPA activation (Schulz and Vögele, 2015). Providing mental percepts for afferent internal inputs is critical to the construction of affective and social experience (Barrett et al., 2016; Quigley et al., 2021; Seth, 2013). Interoception may influence “the dynamic basis to the concept of self” (Critchley and Harrison, 2013). These levels of interoceptive awareness also operate under active inference to minimize unexpected, energy-consuming surprises (Paulus et al., 2019). Thus, interoception may not only impact, or possibly be, awareness (Seth, 2013); it is also critical in the ability to understand others’ intentional and belief states, i.e., theory of mind (Ondobaka et al., 2017). Social allostasis extends similar mechanisms to group dynamics (Saxbe et al., 2020).

Finally, enactive awareness encompasses what others have described as self-awareness (Friston K., 2018) or meta-cognition (Fleming and Dolan, 2012), but also includes a nested awareness of how one’s actions impact the experience—and construction—of one’s environment, again part of the enactive niche. This brings to the fore the importance of belief-based behavioral policy selection within an enactive niche. Behavioral policies are selected to optimize EFE, which requires epistemic (exploration) and pragmatic (exploitation) behaviors to reduce uncertainty and to enable reward-seeking, respectively (Friston et al., 2016) within the enactive niche. At the highest level, enactive behavioral policy selection requires sophisticated inference, or nested beliefs about beliefs (Friston et al., 2021), which can each be optimized. Because sensory and interoceptive errors feed into enactive awareness, the latter includes perception and emotion, but also includes higher-order “top-down” processes. This allows for direction of attention that impact those processes, including cognition, mental time travel, the construction of concepts of self and self-esteem, and the construction of social representations necessary for empathy and compassion. Deane et al. (2024) refer to adaptive narrative control or the ability to model one’s own attentional states and how they can be controlled as the mental action that allows for affective and physiological regulation. It is sophisticated inference that allows for the proactive development of enactive awareness and policy selection capable of reducing EFE in one’s niches, potentially extending to novel, even chaotic, enactive niches.

Within enactive awareness, we include default states or biases which stem from a combination of genetics, epigenetics, and experience, which can influence one’s beliefs without overt awareness. In addition to experience, each organism is imbued with some degree of evolutionary determined preferences that underwrite precise constraints on enactive beliefs and policies. Be it explorative or exploitive, inhibitory or appetitive, there are core preferences satisfying evolutionary demands that influence phenotypic predictive processes. Although many predictive models operate independent of awareness, e.g., osmotic adjustments and circadian hormonal fluctuations, there are also beliefs and policies that through a variety of genetic, epigenetic, and learned mechanisms underwrite beliefs that inferentially select policies that may or may not use awareness. When such beliefs select behavioral policies without the influence of awareness, it reflects bias, or the default state of predictive processing.

When sensory attention (i.e., awareness) is decreased, we are more likely to rely on prior beliefs (i.e., inductive biases) for policy selection. These include such innate policy selections as those imbued by personality preferences or learned habits, such as placing one’s keys in the same place each time—although this may remain aspirational for some. Regardless, biases can be adaptive or maladaptive. If maladaptive, as so often is the ubiquitous confirmation bias (Tversky and Kahneman, 1974), they can only be altered with effortful, top-down attention.

Predictive processing in the absence of awareness can reduce expected surprise in many situations and informs how habits develop and are maintained. Such biases can also serve to perpetuate uncertainty, a process that may explain associations between personality and mental health. Traits associated with low extraversion and high neuroticism (Hakulinen et al., 2015) can be argued to reflect a lack of emotional regulation and a sense of self indicative of a bias that tolerates expected uncertainty (Clark et al., 2018).

Stress directs cognitive resources to limbic vigilance, decreasing cognitive attention and flexibility. Stress amplifies bias leading to preference for habitual, overlearned solutions without exploring new options. We refer to this as belief bias, at least partially distinct from the more common emotional and cognitive biases. Belief biases are rarely context specific and reflect meta-beliefs that are not fully enactive, i.e., do not minimize EFE. Belief biases act as the lens through which we process the world, e.g., evolved control parameters associated with personality (Safron and Sheikhbahaee, 2023). When used in model generation, (meta) belief bias fails to select policies that consider the full repertoire of coping or responses—and restrict optimally enactive policy selection. Such biases as personality develop from characteristic adaptations and evolutionary selection to play a role in long-term phenotypic plasticity and adaptation and may also play a role in canalization.

Reflecting the predictive link shared between allostasis and active inference, efforts to frame allostatic resilience were previewed by Feldman-Barrett and colleagues who presented depression as an allostatic disruption (Barrett et al., 2016). This work identified interoception as the signal of allostatic change, arguing that if allostasis is central to brain architecture, affect is better considered as an aspect of consciousness, not emotion per se, and that all perception is a consequence of predictive allostatic change represented through interoception. In our terminology, these are beliefs, which in the case of depression, result in prediction error causing repeated negative interoceptive signals. Over time, the net result is a metabolically and statistically inefficient internal (i.e., generative) model.

Miller et al. (2022) suggest three concepts of resilience; (1) inertia or the state of being resistant to change, (2) elasticity or bouncing back to its setpoints, and (3) plasticity or the ability to expand one’s repertoire of good states. Under an active inference model, inertia is argued to reflect high precision of prior beliefs, although the potential for inflexibility to develop is recognized. Elasticity is related to homeostasis and allostasis, and the use of temporal models to anticipate the consequences of future actions allows for recovery from environmental or prosocial perturbations. The notion of optimality is discussed in terms of balance between epistemic and instrumental actions, with resilience requiring the cognitive flexibility needed to explore new hypotheses and then update models in terms of updating the relative precision of prior preferences (i.e., biases) that, in turn, affect the balance between epistemic and instrumental affordances. Finally, plasticity is discussed in terms of degeneracy, or useful redundancy. While redundant systems are less efficient, degenerate models afford the opportunity to seek out surprises that are likely to provide maximal information gain. The argument is made that the best strategy is not to attempt to avoid inevitable expected surprises but to become a system that thrives amongst varying degrees of risk and uncertainty.

Recent work (Waugh and Sali, 2023) explicitly placed resilience and allostasis into an active inference framework. Linking resilience with emotional intelligence, they argue active inference selects for the maintenance of well-being. Defining resilience as a pre-determined ability that can be learned across life, a hierarchical tradeoff between belief development and updating is proposed such that a positive sense of well-being is at the top level of beliefs. Model stability is argued to reduce variational free energy as an explanation for why resilience persists through both minor and major prediction errors, with an exception being made for consistent minor errors.

The fragile phenotype stems from genetic predispositions, developmental/epigenetic disruptions or overwhelming trauma, alone or in combination and presents as susceptible to psychological and physiological dysregulation. The commonality across these diverse etiologies and symptom clusters is ineffective allostatic predictions. Over time—and without change to the enactive niche—phenotypic canalization, or the opposite of phenotypic plasticity (Belsky and Pluess, 2013; Carhart-Harris et al., 2023), becomes increasingly entrenched, dominating the person’s cognitive, emotional, and behavioral experiences (Deane et al., 2024) and making allostatic predictions more difficult. Rigid and recurrent patterns persist despite allostatic errors that indicate a need for adaptation and change (Lewis and Todd, 2007). This suggests what has been termed “motivated inattention” or “avoidant mental action” (Deane et al., 2024) where reduced interoceptive awareness. i.e., lessened emotional recognition likely stemming from canalization, is a central mechanism underlying the fragile phenotype.

Many of the errors the fragile phenotype displays involve sensory processing disruptions, notably sensitivity to external and/or internal, i.e., interoceptive, sensory inputs, with frequent allostatic responses. Within a predictive framework, sensory precision and processing differences and stress reactivity impact awareness. With sensory disruptions the effect on awareness is to limit the ability of the person to use active inference to update their predictive models or enactive niche. The association of sensory reactivity and obsessive compulsive disorder (Ben-Sasson and Podoly, 2017), along with anxiety (Podoly and Ben-Sasson, 2020) supports the suggestion that sensory disruptions of the fragile phenotype may result in niche limitation.

Sensory differences, often overlooked in clinical research, are related to several mental health conditions (Harrison et al., 2019). Within a predictive processing framework, individual differences in sensory processing modulate attention given to sensory inputs (i.e., sensory sensitivity) and, crucially, the ability to ignore certain — usually self-generated — sensations (i.e., sensory attenuation). Regarding sensory awareness, assigning high precision to prior beliefs reduces the impact of incoming sensory prediction error signals, and vice versa (Clark et al., 2018).

The fragile phenotype may over-attend to specific external stimuli, thus updating beliefs with limited consideration of prior beliefs (Lawson et al., 2014). Failure to integrate sensory modalities is evident. Sensory over-reactivity has been associated with OCD and may reflect a high degree of uncertainty stemming from the persistently low precision of sensory predictions, with the obsessive need for order being a policy preference to mitigate sensory over reactivity (Poletti et al., 2023).

Like sensory awareness, interoceptive awareness may be altered, with a focus on a limited number of systems. This focus may at times play a role in the development of advanced specific interoceptive systems, e.g., high interoception is associated with an ability to detect others’ emotion (Dobrushina et al., 2020), and also with higher empathy (Fukushima et al., 2011). Impaired interoceptive ability has been reported in depression (Eggart et al., 2019) and autism (DuBois et al., 2016), while hyper-interoceptive sensitivity has been reported in anxiety (Domschke et al., 2010). Decreased interoceptive ability is related to elevated alexithymia (Brewer et al., 2016), which is associated with a range of disorders (Kojima, 2012).

The fragile phenotype demonstrates repeated allostatic responses, which promote even more frequent responses of longer duration and greater intensity (McEwen, 1998). Repeated exposure to rapid and strong allostatic responses without the foreseeable opportunity of recovery may result in allostatic overload (Fava et al., 2019). These conditions are colloquially referred to as burnout or exhaustion. This process of allostatic overload can be further understood as the top-down collapse of the “highest goals” (Goekoop and de Kleijn, 2021) similar to what is seen in hierarchical Bayesian control networks. As in sophisticated inference discussed above (Friston et al., 2021), organisms optimize to reduce the dimensions needed to most parsimoniously represent the niche. These higher level models interact with those lower on the hierarchy to use planning as inference to reach solutions and control behavior with a minimal amount of information, a formalization of Occam’s razor (Maisto et al., 2015). This hierarchical structure allows for the anticipation of more complex inter- and extra-personal events into deeper realms of the future using predictive modeling (Goekoop and de Kleijn, 2021). When an organism is confronted with conditions that interfere with the anticipated state, a useful definition of stress (Peters et al., 2017), prediction errors accumulate throughout the hierarchy, leading to the collapse of the highest level models, or allostatic overload. This top-down failure degrades integrative function at lower levels, consistent with the sensory and emotional focus of those suffering from states of overload such as burnout.

Exposure to severe and/or chronic allostatic activation during critical developmental periods has the potential to alter allostatic predictions across the lifespan (Wilkinson and Goodyer, 2011). Developmental exposure to trauma increases allostatic reactivity and sensitivity across the lifespan, increasing risk for not only stress-related disorders such as PTSD, but also for the entire spectrum of disorders of allostatic systems, e.g., metabolic, cardiovascular, immune (Danese and McEwen, 2012; Finlay et al., 2022).

Awareness in the fragile phenotype is aligned closely with sensory inputs. An inability to integrate sensory information is associated with AL (Azizi et al., 2024) and conforms with this phenotype. This has been alluded to as being “a slave to one’s senses” (Lawson et al., 2014; Peters et al., 2017; Quattrocki and Friston, 2014). With increased uncertainty, allostatic predictions are more closely linked to immediate neuromodulation of sensory inputs (Arnaldo et al., 2022). Given sensory over-reactivity, and focal interoceptive awareness, beliefs emphasize prediction of limited aspects of the environment. This circumscribes the ability to incorporate new experiences into beliefs. Beliefs in general are not updated quickly or with accuracy or precision. The inferential development of higher levels of awareness such as flexibility, theory of mind, and gratitude is inconsistent. Associated clinical characterizations include autism spectrum, OCD, affective conditions, Cluster B personalities, a history of ACES, and PTSD.

While this phenotype may be explained as allostatic over-reactivity stemming from childhood trauma and stress, it can also emerge from highly protected early environments. Notably, this phenotype can be expressed among individuals who were not previously fragile through a combination of prior experiences and current stressors developing an overly sensitive biological system, as can be the case in allostatic (over) load (Pfaltz and Schnyder, 2023).

When considered from the perspective of enactive allostasis, all aspects of the niche can be contributors to plasticity and to canalization. While we have focused on the maladaptive predictions characteristic of the fragile phenotype, it is important to consider interventions using this model. Specific interventions have been developed for a variety of conditions we understand to be part of this phenotype.

Havening, a technique using touch to develop adaptive processing of distressing thoughts and memories (Sumich et al., 2022), can be seen as using sensory inputs (soothing touch) to minimize sensory prediction errors to alter maladaptive prediction errors stemming from trauma and stress. Eye movement desensitization and reprocessing (EMDR) therapy assists patients through bilateral stimulation (Hase, 2021), usually guided eye movements, to reprocess traumatic memories by arguably reducing the prediction error associated with the trauma. The Safe and Sound Protocol, based on Polyvagal Theory (Porges, 2022), employs music tuned to the frequency of human speech to reduce auditory sensitivity and improve speech processing and social awareness in autism spectrum disorder (Kawai et al., 2023). The success of these and other sensory-based therapeutic approaches in improving adaptive functioning by focusing on sensory predictions which in turn enhance predictions across the enactive niche—support the existence of the fragile phenotype and the potential value of our model in identifying and developing interventions to enhance phenotypic plasticity.

This phenotype is reflective of a degree of canalization, but as opposed to the fragile phenotype, canalization is not necessarily associated with poor allostatic predictions. Carhart-Harris et al. (2023) has suggested four issues be addressed to understand the adaptivity of canalization: (1) the nature of the canalized phenotype; (2) the extent of canalization; (3) the initial context when the process of canalization began; and (4) any changes in context. We propose that the durable phenotype shows a high degree of canalization that can be adaptive under certain circumstances; for example, when a person lives within the strictures of a supportive religious community. The durable phenotype is most concordant with an organism preferring exploitation from genetic, epigenetic, and developmental sources; its nature in effect. The extent of canalization depends on the fit between the phenotypic nature of the organism and on the extent of boundary limitations of the niche. In an initial context—where the niche is well-suited to the nature of the organism—canalization develops efficiently for the specific niche. In this scenario, hierarchical higher order models are consistent across levels allowing for precise predictions and planning as inference within the niche. There is little need for accommodation within such a predictable niche thus canalization is efficient.

The durable phenotype often builds effective predictive allostatic response systems by living in defined cognitive and environmental niches (Constant et al., 2018, 2022). Within a predictable environment, durable persons experience little uncertainty and manage minimal surprise well. Active inference selects epistemic awareness to function within the social/environmental niche of the organism, which comports with the development of “echo-chambers” (Albarracin et al., 2022) or limited enactive niches. If the fragile type were to be described as a slave to one’s senses, the durable type could be considered slave to one’s rigid or narrow enactive niche. Epistemic awareness inferentially selects for the sense of self that reduces surprise within the enactive system. Growing up in Malibu versus rural Oklahoma has marked effects on the sense of self for the durable; “I’m a surfer,” “I’m a cowboy.”

The defined range of sensory inputs allows for the development of precise beliefs with high certainty, making it likely durable individuals will be unaware of the limited range within which the beliefs are accurate. This emphasis on precision, or bad bootstrapping (Miller et al., 2022), can result in predictive allostatic errors related to inflexibility as well as a tendency to develop fixed future positions. Provided that the environment is stable and basic biological and social needs are status quo, there is limited interoceptive awareness. This is a central feature of this phenotype. However, with a change to the niche, or what Carhart-Harris et al. (2023) call context, predictions made from canalization are rendered ineffective and allostatic errors made. This is analogous to the collapse of higher order models that result in the disintegration of lower order models moving the individual towards a more fragile phenotype.

In the durable phenotype, meta-beliefs, or beliefs about beliefs—for example estimates of the precision of prior beliefs—are updated for the niche within which the durable lives. Given the limited cognitive niche of the durable phenotype, awareness is not determined so much by sensory over-reactivity but by the demands of the niche. To maintain homeostasis, beliefs may conform with cultural systems that differ greatly from other niches, or from society in general, yet are functional for the individual’s allostatic system on a short-term basis. Without exposure to a wide range of enactive models, the ability to develop flexible epistemic awareness is hindered.

Even within one’s durable enactive niche, inflexibility can result in unexpected surprise and allostatic errors that accumulate over time, presenting as AL (Peters et al., 2017). We call attention to two conditions where the durable phenotype can be expected to be prone to ineffective allostatic predictions. If a person has genetic/epigenetic/developmental policy preferences (biases) that differ from their niche, uncertainty becomes likely and with time AL is accelerated. For example, if someone with a preference for exploration exists within a niche optimized for exploitation, their predictions will lack precision, making allostatic reactions more likely. Second, in the face of surprise, the repertoire of beliefs, meta-beliefs, and policies available to regain homeostasis is limited. If high levels of precision and certainty were developed within an enactive niche that becomes unviable or if the individual is seriously threatened in any way, allostatic overload is possible. Long-term, and possibly delayed, effects of allostatic overload are also possible, consistent with the delayed trajectory (Bonanno et al., 2011; Galatzer-Levy et al., 2018).

It is also possible that an individual developed beliefs within their enactive niche that allow for resilient responses after a large surprise or allostatic overload. The physical, cultural, and social context of the niche are of notable importance. Given the probabilistic nature of awareness under active inference, a narrow niche does equate with canalization. When a narrow niche underlies a degree of canalization that dominates individual awareness and constantly clashes with new situations and evidence, as well as wider societal norms and values, adversity is likely to increase fragility. Nevertheless, we do empirically observe incredible outcomes in individuals who have also faced incredible adversity—a phenomenon that has been known to make even the most die-hard empiricists feel some degree of awe at the human spirit—demonstrating the meta-awareness needed to explore new hypotheses and update models can emerge from very limited, even dysfunctional, niches.

Cognitive flexibility helps in the application of prior models to new niches, explaining how the durable phenotype may move to successfully occupy a broader niche and/or actively shape the niche to fit their demands following a large surprise. This process could describe what is termed post-traumatic growth (Henson et al., 2021) and includes the development of a stronger sense of self, better social relationships, and a more grounded sense of purpose and meaning. Again, we see parallels to resilience trajectories here, specifically to that of recovery. Further we see the bi-directional nature of enactive niches, even when there are relatively constrained external resources, allowing for the possibility of positive phenotypic plasticity (see Figure 4 for vignettes of positive and negative phenotypic plasticity).

Again, we see therapeutic interventions being developed that support plasticity, in this instance for the durable phenotype. Specifically, those targeting interoception such as mindfulness meditation and the awareness and manipulation of breathing (Weng et al., 2021) have shown promise. Much as interoceptive sensations, e.g., increases in heart rate and respiration alter predictions toward emotional states such as fear and anxiety to reduce interoceptive prediction error, within enactive allostasis, awareness allows for immediate regulation of respiration, which alters interoceptive predictions away from fear. We see interoceptive awareness as a prerequisite for phenotypic plasticity in the durable phenotype.

The resilient phenotype functions as durable within a broader enactive niche. With cognitive flexibility, allostatic systems consistently return to baseline after activation. The “elasticity” of this successful allostatic accommodation allows the enactive allostatic system to recover following exposure to a stressor; the regulatory and cognitive flexibility of the resilient phenotype can “bounce back” from stressors in the traditional sense. We also agree that the affiliation system is essential in the resilient phenotype (Darling Rasmussen et al., 2019; Feldman, 2020).

The defining feature of the resilient phenotype—from an enactive perspective—is the influence of interoceptive inference and a level of meta-awareness allowing for effective allostatic predictions outside of a predictable enactive niche. Unlike the durable phenotype, the resilient phenotype is consistently interoceptively aware, and generally emotionally regulated, enabling effective enactive allostasis. This interoceptive awareness allows for the development of temporal models needed to recover from stress in most situations.

Over time, successful allostatic accommodation engenders a positive sense of self. More specifically, there are cognitive processes that are consistent features of human enactive allostatic systems that return to baseline after challenge (Constant et al., 2022; Miller et al., 2022). A positive sense of self provides a consistent weighting for posterior beliefs and policy selection. The ability to think temporally allows the brain to simulate the outcomes of a variety of predictions prior to policy selection. Simulations are made with the current matrix of priors, allowing for adjustments based on the certainty of simulated outcomes, prior to an active inference.

Cognitive states change enactive allostasis. Cognitive appraisal of events can redefine them from a threat to a challenge, a redefinition that shifts the allostatic response from one with unknown accommodation to one of clear accommodation (Bobba-Alves et al., 2022). There is also evidence that states of awareness such as meditation, mindfulness, and awe have associations with allostasis (Klimes-Dougan et al., 2020). Physiological fitness and fit with the social/environmental niche are also foundational in this phenotype.

The elevated importance of awareness in this phenotype also highlights the role it can play in precluding resilience. Epistemic memories throughout hierarchical inferential processes do not occur to optimize resilience, or well-being as argued by Waugh and Sali (2023); they occur to minimize uncertainty and surprise. Rather than resilience being an ability or even a process, we see it as identifiable enactive allostatic predictions characterized by sensory and interoceptive awareness, cognitive flexibility and meta beliefs capable of accommodating allostatic surprise.

This enactive profile includes optimization of many aspects of the counteractive systems activated by allostasis. Without the benefit of empirical studies using measures that capture the dynamics of the enactive niche, current interventions developed to enhance resilience largely consider resilience traits and factors as outcome measure. Accepting these measurement limitations, a meta-analysis of training interventions including cognitive behavioral therapy (CBT) techniques, e.g., emotion regulation, goal setting, and mindfulness practices, e.g., cognitive flexibility, self-compassion, gratitude, found a moderate benefit over control (Joyce et al., 2018). The inclusion of metrics indicative of reactivity and accommodation—e.g., heart rate, heart rate variability, respiration, available on biometric devices—will expand understanding of the enactive niche and are a logical next step in resilience training.

The resilient phenotype has some ability to alter or migrate niches when the demands of the niche become unsupportive of their allostatic systems. If a resilient person finds their niche collapsing or shifting in ways that cause chronic stress, e.g., a workplace becomes toxic, niche change becomes an important component of this phenotype.

The pro-entropic phenotype emerges from the core principle that the brain reduces surprise through generative cycles of action and perception (Badcock et al., 2019; Friston et al., 2010). This has been rephrased by Badcock et al. (2019), (ibid) as, “every organism seeks to maximize sensory evidence for its own existence,” also termed self-evidencing (Hohwy, 2016). Implicit in this is a drive for self-actualization which in its complete expression defines the pro-entropic phenotype. When awareness and action act in harmony to not only accommodate surprise but also change some element of the niche to preclude future surprise, the pro-entropic phenotype emerges.

The PE phenotype assumes the enactive properties of the resilient phenotype with the addition that the allostatic system not only returns to baseline after activation but shows improved allostatic predictions after allostatic responses. At the other end of the spectrum from the fragile phenotype, we propose that sensory awareness is also high in the PE phenotype, but responsivity is appropriate, similar to the adaptive high-sensitivity profiles described for individuals raised in supportive environments (Nave et al., 2022).

Within enactive allostasis, much as certain predictive errors lead to allostatic load, aware engagement with enactive allostasis allows for proactive selection of interoceptive and exteroceptive sensory inputs. This can be thought of as aware enactive niche construction and enables not just allostatic recovery, as seen in the resilient phenotype, but allostatic growth, i.e., accommodation of increasing levels of stress exposure (both frequency and magnitude), both by more adept recovery, as well as heightened thresholds required to elicit an allostatic response. Heightened awareness as well as pro-active enactive shaping of one’s niche work together to increase one’s allostatic response threshold, including in novel social and physical environments.

We see PE as the Bayes optimal selection of meta-beliefs (and the resulting belief states). These belief states are selected in a recursive manner; this allows for the optimal selection of actions across any finite temporal horizon. Sophisticated inference allows for the identification of the end goal state and then through the use of backward planning through time, similar to game theory’s “backward induction,” to determine the course of actions that will optimally reach the goal (Da Costa et al., 2023).

Perhaps the defining feature of the PE is the ability to use sophisticated inference to proactively shape the external aspects of the enactive niche to fit with the regulatory needs of the person. Where the resilient phenotype that can adapt and “bounce back,” the pro-entropic phenotype proactively uses prediction to enact changes in their niche that minimize surprise. In this regard, much of human advancement can be seen as PE: we learned to control fire to survive in more fertile environments; much as we now develop business relationships that may be of future value. In addition to proactively shaping one’s enactive niche, the cognitive flexibility and temporal planning that are part of PE allow for the development of meta-models that can probabilistically infer the demands of new enactive niches. This allows for consideration of novel niches and more effective niche change when needed and/or beneficial.

PE also relies on the integrity of the underlying neurophysiological systems, sense of self, and temporal thinking. PE extends the role for awareness to include an understanding of the probabilistic nature of human life within the environment. Within this context, the PE phenotype proactively uses awareness to optimize active inferences that lead to preferred, unsurprising outcomes. Similar predictive processing models that describe directed awareness within a niche include, “when individual agents restructure their worlds so as to minimize internal processing costs and/or increase reliability,” (Constant et al., 2022) and as a “sense of our own poise over an action space” (Nave et al., 2022). It is this constructed enactive niche that affords PE and allostatic growth. Clinically we see this mapping onto active engagement with novelty, challenge, and creativity, while nourishing one’s niche and regulating well across new niches.

Interventions to enhance phenotypic plasticity may be as much informed by our history as by our present. When Aristotle argued that pain and adversity was needed for the drive for eudemonia, or self-awareness, personal growth, and ultimately contentment, he presaged much of the PE phenotype, as have generations of Buddhists and Hindus seeking enlightenment and dharma. However, we find in the model of enactive allostasis the potential for individualized strategies to facilitate the biopsychosocial operationalization of this wisdom of the ages.

A more pragmatic intervention with clear implications for PE has been receiving wider attention. Episodic future thinking (EFT) is the projection of the self into the future as a means of “pre-experiencing” an event (Atance and O’Neill, 2001; Schacter et al., 2017). This has been most widely studied in the context of delay-discounting, or the tendency to overvalue more immediate rewards. EFT has been repeatedly shown to be associated with a tendency to place higher value on delayed rewards with a concomitant improvement in decision-making (Rösch et al., 2022). This intervention has also been shown to be effective in treating disorders that relate to impulsivity, e.g., alcohol abuse (Myslowski et al., 2024). There are also suggestions that EFT may be an effective intervention to enhance resilient plasticity. Using both pilot work and literature review, Kent et al. (2015) conclude that making and manipulating internal models frees behavior from the present and allows it to become future-oriented. Episodic future thinking has also been associated with performance enhancement in such diverse dimensions as decision-making, emotion regulation, prospective memory and spatial navigation (Schacter et al., 2017). All these findings are consistent with active inference in general and the notion that interventions based on this theoretical model can enhance phenotypic plasticity toward PE. That cognitive rehearsal of future events increases the value placed on the future suggests that the higher order models that allow for inferential planning are strengthened by repetition. Any consistent method of strengthening higher order models is not only likely to improve the functioning of models lower in the hierarchy, such as the behavioral effects noted above, but also seem likely to allow for the higher order models to improve their robustness across novel situations, both of which we find characteristic of PE.