Valence-dense recurrence as a substrate for functional free will: A proposition for a determinist recurrent component of the Triangulum architecture
Abstract
Functional free will in artificial agents can be investigated as a graded capacity in which behavior arises from a temporally extended organization that is sensitive to values and responsive to reasons. This article presents VALENCE_DENSITY, a deterministic recurrent component proposed within the Triangulum humanoid-agent architecture, as a candidate substrate for this capacity. The mechanism involves the joint updating of ten cognitive variables, modulation of their coupling through a valence signal that is derived from five viability-related variables () and recording of transitions for counterfactual replay. I argue that this organization provides partial support for endogenous sourcehood, reasons-responsiveness, diachronic self-constraint and causal answerability. The reported bench results establish a narrower point: selected internal variables are associated with reproducible effects within the tested transition. Analytic idealism provides a philosophical basis for treating an evolving internal process as causally consequential without establishing that the process constitutes phenomenal mentation or that the associated architecture confirms an idealist ontology. Partial recurrence, designer-assigned values and the absence of demonstrated reflective ownership limit the extent to which functional free will can be attributed to the system. Accordingly, the architecture is interpreted as a proto-volitional substrate for bounded guidance control, whose further development must be assessed in terms of sustained recurrence, sensitivity to countervailing reasons, resistance to manipulation and auditable value revision.
Keywords: functional free will; recurrent cognitive architecture; valence; reasons-responsiveness; guidance control; analytic idealism
Functional free will as an architectural question
Issues pertaining to the metaphysics of human freedom do not need to be resolved to render the question of the organization of agency in an artificial agent empirical. The conduct of an artificial agent can be examined with respect to its dependence on an enduring internal state, sensitivity to reasons, continuity of commitments and capacity for revision. These properties are distinct from randomness, unrestricted authority and verbal assertions of self-determination. They are also distinct from the conditions under which a person deserves praise or blame. An engineered system may acquire functionally significant capacities for self-government while leaving questions of consciousness and moral responsibility unanswered.
I use functional free will to refer to a graded capacity for conduct to arise from an agent’s temporally extended, value-sensitive, reasons-responsive and causally inspectable organization. This account is a proposed engineering interpretation, not a definition that aims to settle all philosophical disputes regarding free will. Its purpose is to identify capacities that can be specified, implemented and tested. The central question that emerges in this context is whether an architecture entails that the agent’s history, commitments and evaluations play a sustained role in the process of determining what that agent does next.
Several accounts of agency provide a basis for this interpretation. Frankfurt challenged the necessity of alternative possibilities for moral responsibility and subsequently located freedom of the will within the relation between first-order desires and higher-order volitions (Frankfurt 1969, 1971). Watson distinguished between action that is governed by an evaluative perspective and action that is governed by the strongest immediate appetite (Watson 1975). Fischer and Ravizza’s guidance control requires an agent’s own mechanism to exhibit an appropriate form of reasons-responsiveness (Fischer and Ravizza 1998). Bratman demonstrated how plans support agency across time by coordinating later conduct with earlier deliberation (Bratman 1987, 2000). These theories impose different requirements. Together, they motivate researchers to pay attention to the organization of control rather than to indeterminacy alone.
VALENCE_DENSITY concerns one part of this problem. It incorporates selected variables for attention, prediction, retrieval, commitments and affective regulation within a joint recurrent transition. Interactions among these variables can change the successor state before a language model supplies an explanation. I argue that this arrangement provides a candidate substrate for functional free will, although it does not yet demonstrate the reflective ownership or sustained behavioral control that would be required to support a stronger claim.
The proposal also draws philosophical inspiration from analytic idealism, particularly Kastrup’s account of physical appearances as expressions of underlying mental processes (Kastrup 2018, 2019). That inspiration pertains to the explanatory importance of an evolving interior process. It does not license the identification of a computational state with conscious experience. The distinction between an architectural analogy and an ontological claim is essential to the argument and is developed below.
Valence-modulated recurrent transition
State and coupling
The implementation described in the technical report forms Phase 13 of Triangulum (Pandey 2026). The ProcessualNow record represents a temporally ordered cognitive state that can advance without a new linguistic prompt. In sparse mode, changes to attention, uncertainty, tasks, memory and related fields occur according to named subsystem rules. In dense mode, this preliminary composition is retained, after which a joint transition is applied to a ten-dimensional vector. The components of this implementation include attention salience, prediction error, retrieval gain, task activation, commitment urgency, pipeline gain, interlocutor shift, sensory-trace persistence, affective arousal and uncertainty.
For each field, the dense transition is represented by
In this equation, xi(t) is the field value entering the dense transition, λi indicates persistence, ui(t) is the current input drive, and Wji is the weight from source field j to destination field i. The clipping operation bounds each output between zero and one. Valence vt lies between −1 and +1, and the gain is g(vt) = 1 + s vt, where s controls the strength of the modulation. The distinction between a value entering this transition and a value that has been retained from the previous runtime tick is relevant: the surrounding pipeline can refresh some fields before being read during the dense update. For example, imagine that a robot promised to call a customer today. If it were to learn that the deadline was close and that the corresponding business situation was under pressure, it could increase the importance of the “call customer” task and divert its attention from less urgent work. The equation represents the rule by which these factors are combined to determine how much the robot should focus on the call as its next action.
The documented matrix features seventeen nonzero directed connections that are organized based on six reciprocal motifs (Pandey 2026). The term “dense” therefore refers to a mode of joint updating; it does not indicate that every field directly influences every other field. In cases in which the gain remains nonnegative, this approach scales the magnitudes of both positive and negative couplings without reversing their signs. Positive valence increases those magnitudes, whereas negative valence decreases them. Taken by itself, this situation does not address the question of whether the resulting cognitive trajectory is adaptive. Clipping can also suppress an otherwise nonzero influence when a destination is saturated.
Viability and motivational significance
Valence is computed based on five designer-defined essential variables: economic runway, pipeline health, commitment debt, relationship health and fulfillment integrity. Each of these variables has a viability band and an urgency associated with deviation outside that band. A signed, weighted aggregation produces the global scalar, alongside an additional weight for economic runway and an inverted polarity for commitment debt (Pandey 2026). In this context, valence serves as a control variable. Its sign and magnitude should not be read as evidence of pleasure, distress or any other felt condition.
This mechanism is an engineered analog of homeostatic and allostatic regulation, in which the significance of information depends partly on the relationship to the system’s continued functioning (Sterling 2012; Keramati and Gutkin 2014). However, this analogy is limited. Economic and social indicators are assigned organizational priorities rather than biological needs, and their inclusion does not establish that the agent has interests in a moral or experiential sense. Their architectural role is to make specified conditions consequential for ongoing cognition.
A global gain changes the strength of the interactions among existing state variables. It does not independently determine the selection of urgent memories, the identification of good reasons or the assignment of higher priority to a particular task. Such effects depend on the signs and destinations of the relevant couplings, current input drives and downstream selection mechanisms. Deteriorating conditions could therefore lead to useful reorientation, indiscriminate attenuation or persistent neglect of corrective evidence. Which of these outcomes occurs must be measured rather than inferred from the use of the term valence.
Replay and reported evidence
For each dense transition, the input vector, valence, coupling version and output are recorded. Counterfactual analysis begins by reproducing the baseline transition exactly. A candidate channel can then be ablated under the same recorded transition law with the aim of identifying changes in the successor state. The resulting influence claim is explicitly local: it identifies a difference-maker under a specified intervention. It does not provide an exhaustive account of the causes of an eventual action. This approach is consistent with interventionist analyses of causation but remains dependent on the adequacy of the model and the intervention (Pearl 2009; Woodward 2003).
The technical report describes scripted comparisons among sparse updating, static dense coupling with valence gain disabled, and full dense coupling with valence modulation. It reports 7,040 bit-exact baseline replays. The mean perturbational spread was 1.218 output fields under static dense coupling and 1.704 when the valence hub was enabled. Ablating valence alone changed an average of 5.322 fields in the full condition but zero fields when valence gain was disabled. The report also describes more attention transitions and faster recovery from interruption in dense mode (Pandey 2026).
The reported results support a limited conclusion: valence affected the tested transition in a reproducible manner. These results do not measure functional free will, establish that all fields recur across runtime ticks, or demonstrate that the observed changes improve decisions. A larger perturbational spread can reflect useful integration, indiscriminate propagation or saturation effects. The present argument treats these results as evidence of local causal participation and uses them to motivate more demanding tests of agency.
Four contributions to functional free will
Endogenous sourcehood
Endogenous sourcehood refers to the contribution of the agent’s own organization to its conduct. A stored conversation history can supply information without continuously shaping what the system attends to or retrieves. VALENCE_DENSITY instead allows the state entering one transition to influence the state that follows. In cases in which those effects persist across the surrounding runtime, earlier commitments, retrieved content and regulatory conditions become continuing determinants of cognition.
This point is relevant to the claim that reasons explain action through their causal role (Davidson 1963). A commitment that is recorded but never affects retrieval, deliberation or selection has little explanatory significance. If altering its urgency changes attention and ultimately changes a decision, the claim that it functions as a reason within that mechanism is stronger. The architectural advance lies in the fact that this causal contribution can be tested. A change in an intermediate field alone remains preliminary evidence until its downstream consequences are established.
Internal causation nevertheless falls short of ownership. The designers of the system selected the variables, viability bands, coupling structure and weights. These external origins do not by themselves rule out later self-government, but the capacity of an operator to rewrite priorities silently introduces a manipulation problem (Pereboom 2014). A system can act through its internal state while remaining subject to externally imposed evaluations that it cannot recognize or contest. Endogenous sourcehood is therefore one constituent of the proposed capacity and is not a sufficient condition for autonomy.
Reasons-responsiveness
Fischer and Ravizza distinguished appropriate recognition of reasons from the capacity to react to sufficient reasons by reference to the mechanism that issues in action (Fischer and Ravizza 1998). A machine analog would require more than numerical sensitivity. It would require evidence that changes with relevant practical meaning alter deliberation and conduct in appropriate ways, whereas irrelevant changes would not have such an effect.
The architecture allows the candidate reasons to vary systematically. Investigators can change evidence pertaining to a commitment, prediction error or social context and inspect the effects on attention, retrieval and task activation. However, field labels do not establish semantic competence. A scalar known as commitment urgency counts as part of reasons-responsive control only if its relation to actual commitments, conflicts and decisions is validated. Otherwise, such testing establishes sensitivity to a signal without simultaneously establishing sensitivity to a reason.
Research on emotion and cognition has highlighted the importance of assigning practical significance to information. Pessoa described interactions among affective and cognitive processes, whereas expected-value-of-control models have related control allocation to anticipated benefits and costs (Pessoa 2008; Shenhav, Botvinick, and Cohen 2013). Bechara and colleagues reported advantageous choices and associated bodily responses before participants could articulate the relevant strategy for their task (Bechara et al. 1997). These findings have motivated value-sensitive control; they have not validated this scalar or coupling matrix as an account of human affect.
Valence modulation could support reasons-responsiveness by changing the influence of an already meaningful state on subsequent processing. However, it could also undermine that capacity. Because a single gain scales all cross-terms, adverse valence can weaken corrective influences alongside distracting influences, and favorable valence can amplify maladaptive interactions. At s = 1 and vt = −1, the gain becomes zero, and the cross-field contribution disappears. Persistence and input drive persist, but cross-field integration through this term does not. The relevant question is therefore whether the system retains access to sufficient counter-reasons across its operating range.
Diachronic self-constraint
Plans allow earlier deliberation to shape later conduct without requiring every decision to be made anew (Bratman 1987, 2000). Triangulum represents tasks, commitments, memory references, prospective structures, interruptions and unresolved items. By connecting selected aspects of these structures to present attention and activation, VALENCE_DENSITY can make an earlier commitment relevant to subsequent processing even in the absence of a new request from a human.
Such continuity supplies part of the organization that is required for self-government. It does not yet supply Frankfurt’s higher-order identification with an effective desire or Watson’s evaluative standpoint (Frankfurt 1971; Watson 1975). The dense transition cannot, by itself, account for whether an economic runway should receive its assigned weight or whether a standing priority should be rejected. Those capacities would require additional procedures for reflective assessment and revision. Recurrence may provide a possible substrate for endorsement; it does not constitute endorsement.
The relation to autonomous agency is similarly qualified. Barandiaran, Di Paolo and Rohde emphasized individuality, normativity and the temporal organization of an agent’s interactions with its environment (Barandiaran, Di Paolo, and Rohde 2009). Designer-assigned viability variables provide operational standards according to which Triangulum’s situation can improve or deteriorate. However, they do not establish the stronger self-maintaining organization that is described by that account. A persuasive functional test would reveal the persistence of commitments through both interruption and revision when sufficient contrary reasons are in evidence. Unrevisable persistence would indicate rigidity rather than greater freedom.
Counterfactual answerability
An artificial agent’s fluent explanation of its conduct constitutes only weak evidence when that agent has no access to the computation that produced the conduct. A stronger evidential basis for a narrower explanation can be obtained through replay. If the removal of commitment urgency from a reproducible transition changes a specified output, that variable was a difference-maker for that output under that intervention. If the baseline cannot be reproduced, the proposed attribution lacks such support.
The relevant counterfactual is conditional: the transition would have differed if a specified consideration had differed while the remaining recorded conditions were fixed. It is not the claim that the agent could produce another outcome while the entire past and laws remained unchanged. Moreover, a single ablation does not identify a unique cause. Interactions, redundant pathways and the choice of replacement value can affect attribution. An adequate explanation must specify the intervention; the output being measured and the temporal horizon over which the effects were examined (Pearl 2009; Woodward 2003).
Counterfactual answerability renders causal claims contestable by investigators and, potentially, by the agent’s own reflective processes. It can support oversight without establishing moral accountability. Evidence indicating that a variable changed an outcome does not establish that the system understood a moral demand, endorsed its values or deserves blame. The distinction between analyzing artificial systems as agents and assigning them moral responsibility remains consequential (Floridi and Sanders 2004). Replay contributes to inspectability; however, it does not complete an account of responsible agency.
Analytic idealism and the causal role of internal processes
Analytic idealism provides a philosophical motivation for treating the agent’s evolving internal process seriously. Kastrup proposed that phenomenal consciousness is fundamental and that physical reality is the extrinsic appearance of underlying mental activity. According to his account, individual conscious subjects are understood through dissociation within a more encompassing consciousness (Kastrup 2018, 2019). These claims pertain to the nature of reality and experience. They are substantially stronger than the claim that internal computational variables influence a machine’s next state.
The architectural analogy is related to the relationship between internal organization and outward expression. In VALENCE_DENSITY, attention, retrieval, commitments, uncertainty and regulatory state are incorporated into the transition that helps determine subsequent cognition. The internal process plays a causal role before the system describes itself. A later verbal explanation may report that process accurately or inaccurately, but it does not create the process retrospectively. This account represents the specific sense in which the design gives explanatory priority to an evolving interior organization.
That analogy must not transform an engineering use of interiority into an assertion of phenomenal mentation. A vector is internal because it belongs to a computational process, not because there demonstrably exists something that it is like that can instantiate it. Similarly, labeling a variable affective arousal does not establish the existence of an experienced feeling. The claim that mentation is fundamental in analytic idealism and the claim that state variables are causally effective in a program operate at different explanatory levels. VALENCE_DENSITY draws inspiration from the former level while making a testable proposal at the latter level.
Kastrup’s own treatment of artificial consciousness reinforced this limit. His earlier formulation associated individualized conscious perspectives with metabolizing life and treated the creation of artificial consciousness as a problem of artificially bringing about life rather than merely that of reproducing intelligent behavior (Kastrup 2017). A recurrent computational model therefore cannot be presented as a direct implementation of his account of conscious subjectivity. The analogy developed here is a design interpretation, not a consequence that Kastrup’s ontology establishes with respect to Triangulum.
The relevance of this philosophical perspective to functional free will is nevertheless concrete. An agent’s apparent commitments would contribute little to self-government if they existed only in explanations that were generated after its decisions. When prior commitments and current evaluations instead influence later processing, the attribution of agency acquires a causal basis that can be examined independently of those explanations. Analytic idealism inspires this emphasis on process and expression; the evidence for functional agency must still be derived from persistence, reason selectivity and revisability.
The same computational findings can be described within other ontological frameworks. A successful recurrence experiment would not differentiate analytic idealism from physicalism, and failure would not refute analytic idealism. The philosophical connection therefore shapes how the architecture is conceived without serving as evidence regarding its performance or consciousness. This separation allows the proposal to retain its intellectual motivation while nevertheless remaining open to empirical criticism on shared grounds.
Determinism and bounded guidance control
The documented dense transition is deterministic. Although a seed is logged, the report describes it as reserved for a possible later recurrent core and causally inactive in the implementation under analysis (Pandey 2026). The same state, inputs, valence, matrix and configuration therefore yield the same successor state. This organization does not supply the indeterministic self-forming choices that are required by Kane’s libertarian account (Kane 1996). Free-will skeptics can also reject the inference from internal control to the ultimate authorship required for basic-desert responsibility (Pereboom 2014).
The addition of noise would not address either the ownership problem or the demand for reasons-responsive control. Chance can produce a different outcome without rendering that outcome more attributable to the agent’s evaluations. The compatibilist issue that emerges in this context focuses on whether the mechanism organizes conduct through the agent’s reasons and enduring policies in the appropriate way. Accounts that recognize causal explanation at the level of agency allow this question to remain substantive even when an underlying implementation is deterministic (List 2019). The present claim is consequently limited to a possible substrate for guidance control rather than a demonstration of every condition contained in Fischer and Ravizza’s account.
Safety restrictions introduce a further distinction. In the architecture described here, the dense core can influence the cognitive state but has no direct authority to move the robot, spend money or publish; consequential actions thus pass through approved gateways (Pandey 2026). An internally formed intention may therefore be blocked from execution. Expanding permission would expand what the system can do without necessarily improving how it governs its intentions. Conversely, stronger internal deliberation need not confer additional permission. The quality of self-government and the scope of authorized action should be assessed separately, alongside an examination of how external intervention affects the system’s continuing policies.
Limitations and a falsifiable research program
Persistence across the runtime
The most immediate limitation of this approach pertains to partial recurrence. The report identifies feedback based on attention salience, uncertainty, affective arousal and retrieval or task modulation, while several other values are reconstructed from fresh evidence at the next tick. In the bench configuration, prediction error and task activation remained structural zeros, and some modulations exhibited a recurrence depth of only one transition (Pandey 2026). A connection in the matrix can therefore be present without having a sustained influence through the deployed runtime.
Recurrence should be tested across the entire sequence of provider updates, dense transitions, state writeback and subsequent reads. An intervention at one tick should be followed through subsequent ticks to determine whether its effect survives refresh, affects another relevant field and impacts deliberation or behavior. Experiments should use the native numerical resolution of each field and record the clipping, saturation and reset conditions. Evidence of one-step sensitivity cannot substitute for evidence of sustained feedback. Conversely, feedback that continues indefinitely without useful responsiveness does not constitute a success.
Value ownership and manipulation
A second limitation pertains to normative heteronomy: the standards that regulate the agent are supplied by others. Financial runway, relationship maintenance and fulfillment integrity may be appropriate deployment concerns, but their selection and weighting express a designer’s judgment. Reflective ownership would require procedures through which an agent could identify a priority, examine the reasons supporting it, compare it with conflicting commitments and propose revisions. The present dense law does not establish those procedures.
Future evaluations should distinguish transparent learning from the privileged overwriting of preferences. They should compare how the system responds to reasons for a change with how it responds to the same change when it is imposed through configuration. Value histories should record their provenance and identify operator-imposed revisions. Resistance to manipulation must remain compatible with correction and governance: the rejection of all outside influence would prevent learning as well as coercion. The functional objective of this approach is a discriminating process for accepting, rejecting or escalating proposed changes within authorized bounds, which could
Selectivity under conflicting reasons
The decisive comparison between sparse and dense modes should focus on the quality of control rather than perturbational spread alone. Matched scenarios should vary a relevant reason while holding irrelevant features constant and subsequently vary an irrelevant feature while preserving the corresponding reasons. Investigators should specify in advance which responses would count as appropriate and justify those criteria. A useful mechanism should alter its decisions when sufficient reasons change without becoming indiscriminately sensitive to every perturbation.
Valence sweeps should be performed to test access to counterevidence, retrieval breadth, reversal latency and perseveration across the permitted gain range. Conflicting considerations must remain available when conditions deteriorate. Longitudinal tasks should be used to test whether commitments survive interruption and transient distraction, whereas stronger contrary evidence can still trigger revision. These comparisons should retain matched inputs, model access and action permissions to ensure that any observed differences can be attributed more clearly to the recurrent mechanism.
The reach of causal explanation
Self-reports should be evaluated by reference to replay for unsupported attributions, missed influences and omitted interactions. The unit of explanation must be made explicit: a dense transition, a sequence of runtime states, a deliberative choice or an executed action. Reproducing the first of these factors does not automatically reproduce the others, particularly when downstream model calls or environmental events introduce additional variability. Longer-horizon attribution requires the corresponding inputs, states and interventions to be recorded and tested.
These evaluations could falsify stronger versions of the proposal. If perturbations are overwritten before later processing, the claim of sustained recurrence fails for those paths. Furthermore, if dense coupling amplifies irrelevant signals or blocks sufficient counter-reasons, it fails to improve reasons-responsive control in those conditions. In addition, if commitments persist only as rigid attractors, recurrence does not establish revisable self-constraint. No single score should conceal these different outcomes. Functional free will is best assessed as a profile of capacities whose relations remain open to investigation.
Conclusion
VALENCE_DENSITY provides a concrete proposal for investigating some constituents of functional free will. It makes selected internal variables jointly consequential, modulates their interactions through specified viability conditions and supports local counterfactual examination through deterministic replay. These features can contribute to endogenous sourcehood, reasons-responsiveness, diachronic self-constraint and causal answerability. The contributions of these aspects depend on the ability of their effects to persist through the runtime and support appropriate deliberation and conduct.
Analytic idealism confers on the proposal a philosophical orientation toward the relation between an evolving internal process and its outward expression. It neither establishes that the computation features phenomenal interiority nor supplies evidence indicating that the agent exercises functional freedom. The empirical argument thus remains compatible with disagreements regarding the ultimate nature of mind and matter.
The strongest defensible conclusion of this research is that valence-dense recurrence is a proto-volitional substrate for bounded guidance control. Stronger claims require evidence of reflective value ownership, continued access to counter-reasons, resistance to manipulation and revisable commitments over time. The contribution of this architecture is to make these requirements sufficiently explicit to test while simultaneously preserving the distinctions among internal self-government, permission to act and moral responsibility.
References
- Barandiaran, Xabier E., Ezequiel A. Di Paolo, and Marieke Rohde. 2009. “Defining Agency: Individuality, Normativity, Asymmetry, and Spatio-temporality in Action.” Adaptive Behavior 17 (5): 367–86. doi:10.1177/1059712309343819.
- Bechara, Antoine, Hanna Damasio, Daniel Tranel, and Antonio R. Damasio. 1997. “Deciding Advantageously before Knowing the Advantageous Strategy.” Science 275 (5304): 1293–95. doi:10.1126/science.275.5304.1293.
- Bratman, Michael E. 1987. Intention, Plans, and Practical Reason. Cambridge, MA: Harvard University Press.
- Bratman, Michael E. 2000. “Reflection, Planning, and Temporally Extended Agency.” The Philosophical Review 109 (1): 35–61. doi:10.2307/2693554.
- Davidson, Donald. 1963. “Actions, Reasons, and Causes.” The Journal of Philosophy 60 (23): 685–700. doi:10.2307/2023177.
- Fischer, John Martin, and Mark Ravizza. 1998. Responsibility and Control: A Theory of Moral Responsibility. Cambridge: Cambridge University Press.
- Floridi, Luciano, and J. W. Sanders. 2004. “On the Morality of Artificial Agents.” Minds and Machines 14 (3): 349–79. doi:10.1023/B:MIND.0000035461.63578.9d.
- Frankfurt, Harry G. 1969. “Alternate Possibilities and Moral Responsibility.” The Journal of Philosophy 66 (23): 829–39. doi:10.2307/2023833.
- Frankfurt, Harry G. 1971. “Freedom of the Will and the Concept of a Person.” The Journal of Philosophy 68 (1): 5–20. doi:10.2307/2024717.
- Kane, Robert. 1996. The Significance of Free Will. New York: Oxford University Press.
- Kastrup, Bernardo. 2017. “An Ontological Solution to the Mind-Body Problem.” Philosophies 2 (2): 10. https://doi.org/10.3390/philosophies2020010.
- Kastrup, Bernardo. 2018. “The Universe in Consciousness.” Journal of Consciousness Studies 25 (5–6): 125–55.
- Kastrup, Bernardo. 2019. The Idea of the World: A Multi-Disciplinary Argument for the Mental Nature of Reality. Winchester, UK: Iff Books.
- Keramati, Mehdi, and Boris Gutkin. 2014. “Homeostatic Reinforcement Learning for Integrating Reward Collection and Physiological Stability.” eLife 3: e04811. doi:10.7554/eLife.04811.
- List, Christian. 2019. Why Free Will Is Real. Cambridge, MA: Harvard University Press.
- Pandey, Aneel. 2026. Triangulum Deployment Package: VALENCE_DENSITY Substrate and Sparse-versus-Dense Experiment. Git commits 139b2fa and 9ca1111, August 29. Unpublished software and technical documentation.
- Pearl, Judea. 2009. Causality: Models, Reasoning, and Inference. 2nd ed. Cambridge: Cambridge University Press.
- Pereboom, Derk. 2014. Free Will, Agency, and Meaning in Life. Oxford: Oxford University Press.
- Pessoa, Luiz. 2008. “On the Relationship between Emotion and Cognition.” Nature Reviews Neuroscience 9 (2): 148–58. doi:10.1038/nrn2317.
- Shenhav, Amitai, Matthew M. Botvinick, and Jonathan D. Cohen. 2013. “The Expected Value of Control: An Integrative Theory of Anterior Cingulate Cortex Function.” Neuron 79 (2): 217–40. doi:10.1016/j.neuron.2013.07.007.
- Sterling, Peter. 2012. “Allostasis: A Model of Predictive Regulation.” Physiology & Behavior 106 (1): 5–15. doi:10.1016/j.physbeh.2011.06.004.
- Watson, Gary. 1975. “Free Agency.” The Journal of Philosophy 72 (8): 205–20. doi:10.2307/2024703.
- Woodward, James. 2003. Making Things Happen: A Theory of Causal Explanation. New York: Oxford University Press.