R2.1. Our desiderata

We include three desiderata for metabolic knowledge to act as full-blown evaluative or generative considerations in cognitive models. First, the knowledge must be specific to a computation, a cognitive capacity, or a neural system. Specificity in this sense is not restricted to localization in brain areas or systems, as some commentators inferred from our core examples (Williams, Mace et al.). We do not think that including metabolism always entails localizing a cognitive capacity (see R3.1). Second, metabolic knowledge must be relevant to modeling the target, which takes the form of explanatory constraints. Again, some commentators (mistakenly) infer from the mental effort case that metabolism must have direct explanatory relevance (Hendrickx), which is not our view (see R3). Third, metabolic knowledge must be legible, which may require that the model itself is re-formatted, e.g., to include specific information-processing costs (Khoudhary et al.).

Beyond our examples, the desiderata are corroborated by the examples commentators have provided. Zou et al. specify implications of metabolic considerations for cognitive load theory. Adenosine triphosphate (ATP) availability and synaptic maintenance costs are relevant to human working memory, as they constrain “how many novel elements can be processed and integrated at once.” Metabolic knowledge is also legible to experimental measures like an instruction-efficiency metric, although further work is needed to translate it into modeling parameters. Zenon & Solopchuk specify metabolic considerations for cognitive fatigue, arguing that passive fatigue paradigms rule out global resource or prefrontal-only models of this capacity. Their generative hypothesis shows that aerobic glycolysis is legible to cognitive fatigue models, resulting in testable quantitative predictions about cognitive parameters like degree of representational degradation. The commentaries on metabolic considerations for models of mind-wandering and the default mode network (Wong), alternatives to stimulus–response views of neural function (Alvarez-Hernandez & Theriault), and neurotypical and neurodivergent development (Siddiqui & Jones) are also consistent with our desiderata.

Yet, our desiderata are not trivially met in all cases. Historically, metabolism at the microscale of ATP use was hard to measure (Kondrakiewicz & Gonçalves), leading to an absence of data to explicitly fit cognitive models (Khoudary et al.). Yet there are “work-arounds,” such as using coding efficiency schemes (Pohl), indirect constraint by matching models to neural dynamics (Kondrakiewicz & Gonçalves), calculating ATP using basic molecular biological knowledge (Sengupta, Stemmler, & Friston, Reference Sengupta, Stemmler and Friston2013), and measuring metabolism at different scales (Deery et al.). Thus, despite challenges, the examples and strategies suggest that satisfying the desiderata is possible and, in fact, a substantial achievement when developing cognitive models. They also highlight that a model in which the challenge is not met is deficient, even if it is understandable why it has not been met.

R2.2. Realities of evaluating and generating scientific models

With these desiderata reiterated, we defend their use in evaluation and generation. Shahmoradi argues that our distinction collapses because generation is simply evaluation at an earlier modeling stage. This criticism rightly points out that there is no sharp distinction between the context of discovery and of justification (Bechtel & Richardson, Reference Bechtel and Richardson2010; Kästner & Haueis, Reference Kästner and Haueis2021), so generating new models may require evaluation. However, it would be wrong to conclude that we cannot articulate the difference between evaluating and generating models, nor that there are no clear cases of the two. In clear cases of generating a model, we are not testing a model’s adequacy, explanatory power, or predictive success regarding the cognitive capacity or neural system that is targeted. In clear cases of evaluation, we explicitly assess two (or more) models along these dimensions. We think that our core examples in Sections 3.1.2 and 4.1.2 provide clear cases. Hendrickx picks up predictive success as an evaluation criterion to argue that in hybrid mental effort models (Section 4.2), “the metabolic component explains the same behavioral change that motivational/computational models already predict.” At least currently, metabolic considerations do not seem to uniquely favor hybrid models. We agree with Hendrickx’s conclusion, but for different reasons. First, evaluation is a method for assessing the relative strength of models. It need not be a success condition, where a model is shown to be strongest qua the evidence in question. Second, we did not present hybrid models as evaluative metabolic considerations, but to argue that computational and resource models are not in opposition. In fact, hybrid models like Holroyd’s (Reference Holroyd2025) better match generative metabolic considerations. They provide a proof-of-concept (“controllosphere”), potential explanations (fatigue may signal neurotoxic waste buildup) and help us reassess the target system (link between computational mechanisms and metabolic processes). Third, pace Hendrickx, metabolism can be used in evaluation even when it is not directly explanatory of features of mental effort but instead acts as an explanatory constraint (Section 3.1.1). Metabolic considerations, when they meet the desiderata, can be used to evaluate any model of mental effort, including resource models like the ego depletion model. Hendrickx juxtaposes computational and motivational approaches to modeling with metabolic ones, whereas we challenge this juxtaposition. We think that metabolism can be used to decide which of the many motivational or computational model versions are the most feasible. But we do not think that simply adding metabolic details to models makes them better in all cases. Rather, adding such details needs to fulfill the functions that we highlight in our desiderata and guidelines.

R2.3. Holding onto “implausible” models

The converse of the previous worry is that evaluative metabolic considerations are too stringent. Embrey worries that following guidelines (E1)–(E5) would lead us to prematurely discard potentially productive models. There may be no way to currently incorporate metabolic considerations into opportunity cost models of mental effort, for example, which we assume worries Embrey for evaluation against other, more developed models.

We think that Embrey helpfully points toward the context of pursuit (Laudan, Reference Laudan1978), which exists alongside the context of discovery (akin to generative considerations) and of justification (akin to evaluative considerations). New models initially may be less explanatory or predictive than older alternatives but nevertheless be pursuitworthy because of their potential to solve outstanding problems (Haueis & Kästner, Reference Haueis and Kästner2022; Šešelja & Straβer, Reference Šešelja and Straßer2014). That a model is metabolically implausible initially does not mean it must be discarded immediately. Rather, it is a limitation that must be addressed in its development. We embrace Embrey’s point to pursue nascent cognitive models but insist that it does not remove concerns about their plausibility. Having no idea how something is implemented is still a deficiency, but it alone need not utterly undermine pursuit of an algorithmic model, for instance.

This appeal to pursuit cuts both ways. The opportunity models Embrey discusses are now developed, and pursuit should support linking them with metabolic considerations. We can (and should) ask questions about their metabolic feasibility. This, we think, aligns with the heart of Marrian analysis (Marr, Reference Marr1982). If Marr’s “levels” are a mode of analyzing systems from the top down, then evaluation of models in terms of how well they fit metabolic constraints seems consistent with it. We do not think anything in Marr’s analysis precludes researchers from using implementational-level details in assessing a computation that is likely implemented in this system, much less in using implementational constraints in this way. Marrian analysis presupposes some implementational story as part of this analysis. Put in slogan form, if a top-down analysis gets stymied before it reaches the bottom, there is a problem.

R3.1. Connections between mechanistic and computational modeling

Our response to Najenson may suggest that we favor a mechanistic approach, in which systems are decomposed and localized to entities and activities, over other modeling approaches. Williams contends that nonlocalizationist modeling speaks against a putatively “monist” mechanistic modeling approach.

For our desiderata, specificity need not entail localization in the classic mechanistic sense (Bechtel & Richardson, Reference Bechtel and Richardson2010). Our guidelines do not only apply to mechanistic modeling, and many of the commentator’s examples show how metabolism plays a role in other approaches (Kong et al., Jamalabadi & Hahn, Starita et al., Penacchio & Clemente, Deery et al.). Contra Williams, we do not think of cognitive models as a specific type of (nonmechanistic) model, nor of all cognitive models as (ideally) mechanistic. Rather, cognitive models convey information about particular cognitive capacities or neural systems relevant to these capacities. Likewise, we think that different modeling approaches can contribute to explanatory understanding of a given target (Hochstein, Reference Hochstein2017). Our account thus provides no defense of monist over pluralist stances to modeling.

We agree with Williams that our core examples fit the mechanistic approach. But we reject that including metabolic considerations in any manner that satisfy our desiderata automatically makes a model mechanistic. We worry that this reading implies that any metabolic details count as localization. Likewise, our guidelines (E1)–(E5) at no point require explicitly representing metabolic considerations, much less representing them as system components; explicitly represented considerations are neither necessary nor sufficient for successful evaluation. Mere appeal to metabolism counting as localization or componential is too low of a bar for distinguishing cognition or computation from mechanism, especially when other mechanistic criteria are not satisfied. Further, it underestimates the helpful methods that are employed in cognitive modeling. For instance, these systems can be approached via appeal to Simon’s near-decomposability of complex systems (Simon, Reference Simon1962; Weinberger, Reference Weinberger2020), allowing us to assess how a system’s operations relate to an ascribed computational activity. Modeling methods like this supply a defense for our position without limiting us to any one framework, such as neuro-computationalism (Piccinini, Reference Piccinini2020), from which some concerns about the relation between mechanism and computation may stem. Nonetheless, any approach must answer when a model contributes to explanatory understanding, and metabolism is part of the evidential constraints by which we can prune, partition, and populate the space of possible models.

Although computational modeling is often associated with Marrian analysis, controversies over what “computational level” or “computation” means abound (Egan, Reference Egan2025; Shargrir, Reference Shagrir2010). On our ecumenical approach, computational modeling provides mathematical descriptions of operations, which process information in the brain, and there is a specific relation (mapping, narrative, or other), which relates this attributed computational function to features of cognitive and neural systems. Our view of evaluative metabolic constraints also follows mainstream computational neuroscience, which takes metabolism to be external to the operations or mechanisms that implement the computational function (see Section 3.2.1).

For some commentators the abstract nature of computation is in tension with metabolic considerations (Dengsø), whereas others argue that it makes metabolism irrelevant for this modeling approach (Embrey, Mace et al.). There are too many presuppositions for us to definitively address this issue here. What is often overlooked in these discussions, however, is that in complex biological and artificial neural networks, the computational function that modelers attribute approximates the operations of the actual system (cf. Cao & Yamins, Reference Cao and Yamins2024, p. 3). Like other constraints (e.g., tractability van Roij, Reference Van Rooij2008), metabolic considerations can help us adjudicate which candidate best approximates an actual system operation. We agree with Williams and Embrey that not every model will currently allow such comparison directly. They may have to be re-expressed before they can be connected to metabolic considerations (Khoudary et al.).

R3.2. Dynamical and multiscale modeling

Some commentators emphasize the importance of metabolism for modeling scale and dynamics of cognitive systems. Kong et al., for instance, highlight that greater appreciation of meso- and macroscale neural networks can offer a better venue for integrating metabolic considerations in cognitive modeling. We agree that the mesoscale is important (Colaço, Reference Colaço, Bickle, Craver and Barwich2021; Haueis, Reference Haueis2014 Reference Haueis, Reference Haueis2021), yet what counts as the “mesoscale” can vary amongst models, and these models can also span multiple levels of neural organization. This speaks to Penacchio & Clemente, where system-level considerations are recommended in addition to research at the neuronal level, and Deery et al., in which a scaling gap is overcome through imaging methods to relate metabolic activities to larger-scale networks. Our desiderata and other commentaries highlight that many cognitive phenomena are on one reading or another “multilevel” or “multiscale” (Haueis, Reference Haueis2022; Mok & Love, Reference Mok and Love2023).

Combining scale and dynamics, Favela also encourages us to incorporate multiscale dynamics into our framework. We agree that incorporating metabolism can also accommodate and benefit from bringing in temporal dynamics, as metabolic processes, including some from our examples, highlight temporal dynamics, sometimes as an implication of their multiscale spatial character. This is not something one gets for free from including metabolism; it must be addressed. However, which dynamics matter will be as variable as which considerations are incorporated. There is not a one-size-fits-all option for incorporating time. Likewise, we should distinguish incorporating brain dynamics from dynamic approaches to cognitive modeling. The latter often speak of “energy functions” or “energy landscapes” without being explicitly constrained by metabolic considerations (Beer, Reference Beer2003; Hopfield, Reference Hopfield1982; Rabinovich, Huerta, & Laurent, Reference Rabinovich, Huerta and Laurent2008). De Wit makes a similar point about ecological approaches, which are theoretically sympathetic to, but background, metabolism. These examples illustrate why we do not think that dynamical approaches are the favored, let alone only, options available for incorporating metabolic considerations. Other options are live, and whether they can incorporate metabolic considerations depends on the hard currency of the descriptive, explanatory, and predictive success of the models.

R3.3. Predictive processing approaches

Predictive processing (PP) and free energy principle (FEP) combine tenets of the previously discussed approaches. We read PP and FEP as presupposing that (a) information is physical, (b) information processing is subject to efficiency constraints, and (c) cognitive functions are subject to metabolic constraints because they minimize prediction error or free energy (Friston, Reference Friston2010; Sengupta, Stemmler, & Friston, Reference Sengupta, Stemmler and Friston2013; Sprevak & Smith, Reference Sprevak and Smith2023) These assumptions make PP and FEP models of cognitive capacities suitable candidates for metabolic considerations because any neural system that regulates metabolism “must embody a model of the variables it regulates” (Jamalbadi & Hahn). Likewise, minimizing prediction error reduces redundant signals, which could reduce “metabolic costs within a global computing scheme” (Pennacchio & Clemente, and see Hechler, de Lange, & Riedl, Reference Hechler, de Lange and Riedl2023 for experimental evidence).

Our target article points out that PP and FEP approaches can generate novel cognitive models (Section 3.3.2). Yet, metabolic considerations do not uniquely favor these approaches, as metabolism can also be integrated into other types of modeling. There is also a further question that merits careful examination. Because PP and FEP incorporate metabolism and energy-information efficiency as theoretical axioms, models derived from these axioms are already metabolically constrained (e.g., Ali et al., Reference Ali, Ahmad, de Groot, van Gerven and Kietzmann2022; Zhang et al., Reference Zhang, Chitic and Bohté2025). Can evaluative metabolic considerations also be used to select which of multiple PP schemes is the most biologically plausible for a given cognitive capacity or neural system?

R5.1. Functionalism and multiple realization/realizability

In Section 4.1 we argue that there is a substantive reading of multiple realization that is compatible with incorporating metabolic considerations into cognitive modeling. Metabolic factors themselves might be multiply realized across the biological realm, as their life-sustaining function is something they have in common. Guest et al. support this reading because they argue that the same functional capacity is multiply realized if there are multiple mechanisms that underlie that capacity.

Although not necessarily unclear, scholars do not always mean the same thing by “multiple realizability.” For example, Thagard argues against strong multiple realizability but favors a scant version, where brains and certain AI systems realize intelligence because they mutually satisfy energy, time, space, and history constraints. Is this in tension with Guest et al.? The answer not only depends on the content, but also our epistemic stance toward multiple realization (or realizability). If Guest et al. and Thagard’s positions are read as working hypotheses, for instance, then they are not at odds with one another. Both are revisable means of exploring relations between cognition and its physical basis. However, if they are core metaphysical commitments, then facts about metabolism challenge them (Cao, Reference Cao2022), and both are in tension with general worries about multiple realization (Polger & Shapiro, Reference Polger and Shapiro2016). Nonetheless, neither Guest et al. nor Thagard are in principle at odds with our thesis. It is just that incorporating metabolic considerations into cognitive modeling does not require any one metaphysical outlook, nor does it alone prove any one outlook correct.

R5.2. Medium (in)dependence of computation

Several commentaries shed light on how “medium independence” should be understood as connecting with our account. According to Williams (see also Williams, Reference Williams2025) there are two senses of medium independence: as a formal system that is realizable in many physical systems (Haugeland, Reference Haugeland1985), and as an abstract, less detailed description of a physical system that has enough degrees of freedom to execute a computational rule (Piccinini, Reference Piccinini2015). Mace et al. criticize that our view is too focused on Piccinini’s account. We take seriously that it might be idiosyncratic qua medium independence positions and perhaps different throughout his existing works. What Piccinini gets right, however, is that medium independence specifies a relation between a physical system and a cognitive activity that permits variability between systems that underwrite (or realize) the activities. When we engage in cognitive modeling, we have expectations about how the model relates to the brain (or the brain plus, for 4E proponents), including expectations about the media that realize computational operations. Those media may vary without corresponding variance in the operations.

Taking medium independence as an epistemic stance toward the system (see R5.1), we wonder whether implications of the two senses manifest in the context of cognitive modeling. We do not read Haugeland’s (Reference Haugeland1985, p. 58) sense of medium independence as entailing that any arbitrary medium can materialize the same formal system. And we leave open the idea that metabolic considerations can materialize in different media yet have a unity to be captured in our modeling practices. This parallels our points about Marrian levels (R2.3). As cognitive modelers we operate with a working but revisable delineation of media of interest, e.g., spiking time differences for neural computation (Piccinini, Reference Piccinini2015). These media can share constraints, such as availability of metabolic resources. It seems entirely consistent with the spirit of medium independence that we can use these constraints to assess models of how these systems operate. Thagard’s position, we think, is a litmus test: we take it as consistent with medium independence as a working hypothesis, but we imagine other commentators would not. Generally, how metabolism relates to medium independence requires us to explicate the position, our epistemic stance toward it and careful reflection upon scholarship opposed to medium independence.

R5.3. Ontological perspectives

Our discussions of realization and medium, paired with Section 4, return us to an exploration of ontological perspectives when incorporating metabolic considerations in cognitive modeling. No doubt, ontological positions can influence modeling. As Dengsø illustrates, if you are ontologically committed to metabolism being constitutive of cognition – which we are not – then you build different models. More minimally, if one commits to metabolism being relevant to cognition, which is closer to what we hold, then this commitment can influence what system characteristics are deemed plausible during evaluation. This is shown in commentaries that address architectural commitments of cognitive systems and, by consequence, how they should be modeled. Van der Velde argues that metabolic considerations favor architectures that allow compositional processing and access to long-distance information (e.g., small-world networks). Percy uses energy efficiency to hypothesize that evolution favors anatomically specialized structures, whereas Siddiqui & Jones show how such specialization can arise via connection strengthening on a developmental timescale. Moreau & Wiebels alternatively argue that because cognitive systems draw on common resource pools, “sharp and stable boundaries between cognitive components are unlikely to hold across all contexts.”

We think these takes on the ontological implications of metabolism are all worthy of further investigation. Yet they are not all the same. The position of Moreau & Wiebels includes metabolism in the specification of cognitive modules as local optima, making them medium-dependent. This aligns with criticisms of medium independence which argue that metabolism challenges the distinction between computationally relevant and irrelevant properties (Cao, Reference Cao2022; Chirimuuta, Reference Chirimuuta2022). By contrast, Percy and Siddiqui & Jones posit energy-saving mechanisms on evolutionary or developmental timescales that could be implemented in different media, akin to energy-saving learning rules (cf. Section 3.2.3). Looking at metabolism through an ontological lens, we are skeptical that there is a singular perspective on cognitive systems and how we model them.

Shifting to a potentially different sense of “ontology,” Mace et al. challenge our account by highlighting how questions of cognitive ontology affect the role metabolic considerations may play in cognitive modeling. Having hopefully allayed worries about monism, we do not commit to any idea that metabolic properties of a physical system are needed for identifying, much less individuating, a cognitive capacity or process. We think this matter lies in the domain of conceptualization (R4). Identification is a complex topic, both when identifying target phenomena (Colaço, Reference Colaço2020) and domains for concepts (Colaço, Reference Colaço2022; Haueis, Reference Haueis2024, Haueis & Margulies, Reference Haueis and Margulies2026). In these works, we address general and discipline-specific processes by which categorization of the constructs of a so-called “cognitive ontology” might operate, provisionally, iteratively, and non-monistically. This means that reconceptualizing cognitive capacities will happen alongside, and can inform (or be informed by), metabolically inspired cognitive modeling.

The idea, however, that models cannot be compared or generated, even provisionally, without resolving conceptual issues first is too strong for productive, realistic science. If guidelines E1 and G1 cannot be satisfied, then metabolic incorporation does not get off the ground. For this reason, Mace et al.’s discussion of vehicle indeterminacy points to issues in cognitive modeling more generally. If modelers cannot specify the capacity that the model targets, a whole swath of comparisons is undercut, not just evaluative metabolic considerations. If it is not possible to compare models on any dimension, then metabolism will not inform evaluation between the two. Metabolic considerations are no “silver bullet.” Indeed, we think that no one factor, be it about brains, the body, or the environment, deserves this label in any realistic depiction of the sciences of the mind.

R6.1. Efficiency

Section 2.1.2 discusses several senses of “efficiency,” which are not equivalent but nonetheless can be relevant to cognitive modeling. Comparing systems in terms of “efficiency” risks conflation and requires us to specify the target and scale at which we compare it. Karbowski suggests that at the scale of neurons and synapses, the brain uses more energy per operation than digital and neuromorphic supercomputers, even though the brain is more efficient at other spatial and temporal scales. This highlights how efficiency comparisons become misleading when one ignores scale. Fleig-Goldstein similarly points out how comparing brains and Large Language Models (LLMs) like Chat-GPT must consider time and task specificity. Both commentaries point to our first tidbit in the target article. The brain uses a lot of energy, and we are not surprised that on some calculations its operations are more expensive than those in silico. This is why principles of neural design (Section 2.1.1) are valuable; they delimit how energy costs can be minimized because otherwise these costs add up quickly.

However, we think sympathetic to Karbowski’s general outlook, and one must ask what counts as an “operation” in both biological and artificial systems while calculating energy per operation. Computational power in artificial systems like El Capitan or Hala Point is measured in floating point operations, where an operation is an arithmetic computation (addition, subtraction, multiplication, division). If one takes seriously Karbowski’s claim that “an action potential corresponds to a single neural operation,” it is questionable that one neural operation equates to one arithmetic operation. Likewise, “operation” at the neuronal or synaptic scale is not “task” in the vein of Fleig-Goldstein; if a brain operating at 17W in a human body and AI system operating at 35MW perform the same task, there is a potential time difference in task performance where AI systems will operate more efficiently. It is also worth noting that these wattage estimates do not tell the whole story. On the one hand, the human brain simultaneously performs multiple cognitive functions, so comparing LLM inference to the brain’s whole energy budget (Fleig-Goldstein) is misleading. On the other hand, the human brain’s 17W budget also goes to repair and maintenance, with far less dedicated to cognitive activities (Levy & Calvert Reference Levy and Calvert2021). Maintenance costs for supercomputers, such as humans who maintain and repair data centers, are typically not factored into energy budget estimates of AI systems. These quibbles with calculation reinforce our and Karbowski’s point that precision and specificity are key for informative brain–AI comparisons. Failing to do either leads to a misleading calculation.

This nuance regarding “efficiency” also speaks to Pohl, who argues that abstract coding efficiency considerations can play nongeneric roles in computational neuroscience. His examples succeed precisely because they meet our desiderata; coding efficiency considerations are specific (e.g., bullfrog inner ear), relevant (e.g., ecological relevance, explaining variance), and legible (information theory applied to neural systems). The only difference is that Pohl discusses hypotheses rather than models, so our evaluative guidelines may not apply directly. However, Pohl recognizes that efficient coding need not be linked (via representation or explanation) to metabolic considerations. We thus find abstract considerations promising, yet we emphasize that not all “efficiency” is the same. This also makes precise why “evolution by natural selection need not always produce perfectly optimal solutions” (de Wit; see also Nave, Reference Nave2026). As our core generative example highlights, sometimes different forms of efficiency cannot be mutually maximized, but tradeoff against each other. We also see no issue with expanding efficiency considerations beyond the brain as de Wit proposes, as long as such modeling efforts make metabolism explicit and satisfy our desiderata and guidelines.

R6.2. Systematizing thoughts on energy use

Our account provides a systematic way to incorporate metabolic considerations and to think about energy use. Yet Fleig-Goldstein worries that without a resource-rational analysis (RRA), there is no rationale for uniquely focusing on metabolic considerations over other physical or biological detail of the target system. We repeat that metabolism can be conceptualized in multiple ways, none of which makes it unique or a “silver bullet” for cognitive modeling. Besides, Section 3.1.1 highlights that we are not utterly opposed to RRA, even though we think it is open how RRA’s unifies versus how informative it is for specific cases. Fleig-Goldstein is right that RRA is generative in a top-down manner, but we present systematic guidelines for a bottom-up, exploratory use of metabolic considerations to rigorously generate new cognitive models (Haueis, Reference Haueis2022). In this sense, we could suggest RRA is too top-down just as Fleig-Goldstein suggests we are too bottom-up. No doubt, both together are more beneficial than either alone.

Yet we worry that Fleig-Goldstein mischaracterizes generative considerations for reasons that parallel our reply to Shahmoradi’s collapse worry (R2.2). As our guidelines indicate, model generation need not be ex nihilo. Rather, parameters are constructed or adapted from existing models (G3), and our core example from Theriault shows both construction and adaptation of a model using tenets of predictive processing, which are not monolithic. Choosing and combining existing model components in new ways can result in a new model, even if it fits an existing modeling approach, such as those we discuss in R3. What matters is that the process begins (G1) by specifying the cognitive capacity or neural system that one wants to model, which may well include specifying the task the capacity must solve. Here, we agree with Fleig-Goldstein: determining what one wants to model and why is critical. RRA can achieve this, but we also insist that not every analysis will move from the computational to algorithmic level. We do not always have a computational description of the target (say in G1). Our core generative example shows that researchers equally move from “implementation” to algorithmic or computational descriptions. In these cases, metabolic considerations can serve as starting points for plausible candidates ultimately to be tested. These approaches can synergize, rather than oppose each other.