Our earlier article, Evolutionary Psychological Structures: Socialization, Meaning, and Value-Grounded AI, introduced EPS through Belonging, Identity, and Power, then connected those social motivations to a computational account of context, values, and emotion. The latest manuscript develops that foundation into a first-principles argument about group coordination, a proposed account of neural execution, and an implemented recurrent model that learns from consequences.
Mvuselelo Houston Khanyile's Evolutionary Psychological Structures and the Neural Execution of Human Centered Reasoning, subtitled A first principles account of group coordination and adaptive social behavior, brings these strands together in version 0.2. Its central question is how an event becomes significant to a person, organizes competing motives, and changes subsequent behavior. Sociology describes participation, roles, and authority; psychology describes their meaning and motivational force; the neural account asks how the required operations could be executed.
Why coordination needs a psychological account
The paper begins with continued existence as an organizing objective: maintaining viability, dependable relationships, meaningful commitments, and opportunities for improvement. This is a way of structuring the inquiry, not a claim that evolution pursues a conscious purpose or that every person explicitly calculates survival in every interaction.
Individuals act with limited time, resources, and information. Collective action can extend what they can accomplish, but it also creates communication costs, conflicts, and dependence. Cooperation is useful when its benefits exceed those costs. A group therefore needs more than proximity: its members must sustain participation, establish meaningful roles, and influence consequential decisions. Their interests need not coincide with one another or with the group's interests.
The manuscript uses Tools plus Constraints equals Objective as a first-principles reasoning heuristic. It asks which available capacities could support an objective under specified constraints. The expression is not a mathematical identity or an empirical law. Likewise, the coordination problem motivates EPS without uniquely proving that BIP, or its five operations, must be the only possible architecture.
Belonging, Identity, and Power as coordination functions
EPS organizes these demands through three interacting motivational dimensions. They describe what matters in participation, rather than assigning a single motive to each observable behavior.
- Belonging
- Inclusion, connection, relational reliability, and continuing participation. Belonging can support disagreement or a move to another group; it does not require compliance with the present one.
- Identity
- Continuity, recognition, and expression of self-relevant roles, commitments, and memberships. Being assigned a role is different from recognizing that role as meaningful to oneself.
- Power
- Influence over consequential outcomes and the management of dependence. Influence can be shared, distributed, or negotiated. A fixed hierarchy is one possible arrangement, not a requirement of EPS.
The same response may serve several dimensions, and priorities can change with context. BIP is a proposed organization of social motivation, not an exhaustive catalogue of human needs or evidence that each dimension is a separate evolutionary adaptation.
Survival, Stability, and Quality answer a different question: what regulatory condition is a person trying to maintain or improve? Survival concerns immediate viability and threat; Stability concerns dependable functioning, resources, and relationships; Quality concerns meaning, capability, and future improvement. These orientations can coexist. They are neither a compulsory ladder nor a one-to-one mapping onto BIP: belonging, for example, may matter to immediate safety, reliable support, or a more meaningful life.
Five operations connect interpretation to adaptation
The functional pipeline explains how an event acquires significance and informs a response. The stages distinguish operations, not five isolated brain regions or a rigid, once-only sequence.
Contextual filtering
Select and regulate the influence of cues, memories, and interpretations. Filtering is not simply discarding information: it can retain ambiguity and competing readings while deciding what deserves attention in the present situation.
Value mapping
Relate the interpreted situation to needs, commitments, learned expectations, and possible consequences. The Cultural Matrix is the proposed learned mapping from values to Belonging, Identity, and Power at the individual level, not a fixed numerical description of an entire culture.
Motivational allocation
Translate valuations into relative BIP priorities. Multiple motives can remain active, compete, or support one another. Organizing motivation is distinct from choosing an action, which also depends on the responses available.
Emotional modulation
Alter the urgency, influence, and persistence of processing in light of bodily conditions, learned associations, and expected consequences. Emotion can change earlier interpretation and valuation; it is not merely a final label attached to a decision.
Existential state confirmation
Provisionally estimate how a situation or response supports Survival, Stability, and Quality. This assessment can occur before action and be revised after an outcome. Confirmation means an updated estimate, not a guarantee that the response actually secures those conditions.
Feedback makes this architecture recurrent. An emotional change can redirect attention; an unexpected outcome can revise a value association; an updated sense of security can change which possibilities matter next. Adaptive behavior here means revising interpretations and responses relative to objectives and constraints. It does not automatically mean that the resulting behavior is correct, ethical, or beneficial to everyone affected.
From neuronal properties to candidate computations
The neural argument starts with the operation that must be performed, then asks which cellular properties and patterns of connectivity could support it. This gives the proposal a mechanistic direction: contextual filtering requires control over which signals influence a circuit, while integration requires combining signals whose significance depends on their sources and timing.
Inhibitory and disinhibitory interneuron circuits provide candidate mechanisms for regulating input access, gain, and timing. Inhibiting an inhibitory neuron can release a target from inhibition, so filtering can open a pathway as well as suppress it. Pyramidal neurons provide candidate integration mechanisms through their inputs, dendritic compartments, and sensitivity to coincident activity. Whether a signal is transmitted can depend on how these inputs interact, not simply on an undifferentiated sum.
These roles operate together in recurrent circuits. The proposal does not require interneurons to finish all filtering before pyramidal neurons begin integrating. Nor does it posit dedicated Belonging, Identity, or Power neurons. The same classes of cells can participate in different functions, depending on their connectivity and the information available to them.
The manuscript extends this reasoning to valuation, learned value-to-motive connections, affective modulation, and estimates of bodily and environmental conditions. The studies it discusses supply candidate ingredients for those operations. They do not, by themselves, identify an EPS circuit or establish a neural Survival-Stability-Quality triad. The resulting account is a biologically motivated hypothesis with specific computational roles to investigate. The detailed anatomical argument and its supporting literature appear on pages 8-11 of the manuscript.
What version 0.2 implements
Version 0.2 gives selected functions an executable form. Its recurrent population model represents four contextual cues, three value channels, three BIP driver activities, two affective components (valence and arousal), and three regulatory support estimates. Filtering changes cue influence; recurrent excitation and competition organize driver activity; affective gain changes responsiveness; and regulatory estimates feed back into processing. These are abstract computational populations, not recorded neurons, and the simulation steps are not measurements of biological time.
The learning mechanism is especially important. A three-by-three Cultural Matrix starts with equal weights of 0.2. After a response, an outcome prediction error updates the eligible associations between the values active before the action and the response pathway actually used. Different pathways can therefore strengthen or weaken differently. This response-specific rule replaces the earlier equal-row broadcast approach, which could change weights without differentiating the driver rows from their equal starting point.
Actions are selected from driver activity and declared response affordances, with exploration. Learned associations, experienced regulatory conditions, and the preceding outcome persist between episodes, while within-event activity is reset. This makes it possible to test whether the implemented system acquires an association, loses its advantage when consequences change, and learns a new one.
Learning when the successful response changes
The reported study runs ten seeds, eight model or control conditions, and 600 episodes per condition and seed: 48,000 synthetic interactions in total. Each situation contains one of three relevant cues and an independent distractor. The model chooses among coordinating participation, asserting a role, and directing a plan. Their one-to-one BIP affordances, cue reliability, and outcome effects are supplied by the experiment; the model does not discover the social meanings of those labels.
A hidden mapping determines which response succeeds for each cue. After episode 300, all successful responses change. This reversal tests adaptation to changed consequences, rather than continued exploitation of the original mapping. A simpler linear action-value learner receives the same supplied information and provides a useful comparison without EPS recurrence.
Success percentages reported in Table 5 of the manuscript. Values are means, with sample standard deviations in parentheses, across ten seeds. Each column covers 60 episodes; variation is across simulated runs, not human participants. On small screens, scroll the table horizontally.
| Model or control | First blockEpisodes 1-60 | Before reversalEpisodes 241-300 | After reversalEpisodes 301-360 | Final blockEpisodes 541-600 |
|---|---|---|---|---|
| Recurrent EPS | 84.3 (4.3) | 92.0 (4.2) | 39.3 (9.7) | 91.5 (3.4) |
| Frozen weights | 31.2 (4.8) | 30.2 (5.6) | 33.5 (6.6) | 30.8 (10.3) |
| Broadcast learning | 31.2 (4.8) | 30.2 (5.6) | 33.5 (6.6) | 30.8 (10.3) |
| Unrelated feedback | 31.5 (10.6) | 35.0 (14.1) | 37.2 (11.0) | 26.3 (12.0) |
| Filtering removed | 66.0 (11.8) | 92.0 (4.2) | 26.7 (4.4) | 91.5 (3.4) |
| Affective gain removed | 82.3 (5.2) | 91.8 (4.2) | 35.3 (7.9) | 91.3 (3.1) |
| Regulatory feedback removed | 84.7 (4.4) | 92.0 (4.2) | 39.3 (10.2) | 91.5 (3.4) |
| Linear learner | 81.2 (6.0) | 93.2 (3.4) | 46.7 (11.9) | 90.8 (3.4) |
Source: version 0.2, page 17. Uniform random choice has an expected success rate of one third.
Full EPS reaches 92.0% before reversal, drops to 39.3% in the first post-reversal block, and recovers to 91.5% in the final block. Frozen weights, broadcast learning, and feedback unrelated to the response do not show comparable recovery. These controls support the contribution of differentiated, outcome-contingent learning to success in this task.
The linear learner finishes at 90.8%. The mean paired final difference is only 0.67 percentage points, with a sample standard deviation of 4.39 points. The reported comparison therefore does not establish an EPS performance advantage. Both architectures can solve the constructed contingency using the supplied representations and experienced outcomes.
Filtering has a more specific effect: removing it reduces initial success from 84.3% to 66.0% and first post-reversal success from 39.3% to 26.7%, although both conditions finish at 91.5%. Removing affective gain or regulatory feedback also produces similar final performance. The task reveals differences in learning dynamics without establishing the necessity of every EPS component. It does not imply that emotion or regulatory assessment is unnecessary in human behavior.
A more explicit, testable EPS research programme
The advance over the earlier exposition is a tighter connection between the coordination problem, the required psychological operations, candidate neural mechanisms, and an executable learning system. The model develops differentiated associations from equal starting weights and revises them after reversal. The manuscript also reports 16 passed functional and regression checks, verification of all 48,000 episode records, and an identical replay of the full model for seed 0. These are reported implementation checks, not an independent replication or a test of human psychology.
The next questions concern what those computations explain beyond the constructed task. Psychological studies would need to test whether BIP and the regulatory orientations capture distinct, useful features of motivation. Richer social settings would need to introduce negotiated roles, competing commitments, other people's perspectives, and meanings that are not supplied in advance. Neural studies would need to test the proposed circuit operations and their causal contribution. Comparisons with simpler models remain essential throughout.
Read alongside the earlier EPS article, this edition moves the research from a broad account of socialization and value-grounded cognition toward a more explicit account of execution and adaptation. Its computational result is concrete but bounded: selected EPS functions can be implemented and can learn changing response contingencies. Whether the larger architecture explains human social reasoning remains the research programme that this implementation helps make testable.
Read the source manuscript
Mvuselelo Houston Khanyile. Evolutionary Psychological Structures and the Neural Execution of Human Centered Reasoning: A first principles account of group coordination and adaptive social behavior. CVEST, South Africa. Version 0.2, 2 October 2026. This article summarizes the supplied manuscript; the experimental results above are reported by its author.