History-Dependent Decision Architecture
A Decision-Science Framework for How Learning History Shapes the Parameters of Choice
ABSTRACT
Decision science contains substantial accounts of information acquisition, representation, prediction, valuation, and learning. Yet a recognized but incompletely resolved problem remains: how prior learning configures the decision parameters that shape subsequent information sampling, and how the resulting evidence environment in turn contributes to future decision formation.
This article proposes History-Dependent Decision-Parameter Architecture (HDPA), an empirically testable decomposition into five analytically distinguishable classes of history-sensitive decision parameter: state and schema representations Z, source weights W, subjective outcome valuations V, predictive action–outcome expectations Q, and information-selection policies X. Through these policies, current decision architecture can influence which evidence is subsequently encountered, making part of the future evidence environment endogenous to the current architecture and allowing that environment to contribute to its subsequent updating.
HDPA does not propose a new universal learning algorithm, a newly discovered neural module, or a unitary psychological trait, nor does it claim to have solved the underlying problem empirically or to supersede existing computational frameworks. Its proposed scientific contribution is an empirically testable decomposition and a set of falsifiable hypotheses concerning longitudinal and recursive coupling. The architecture earns independent explanatory status only if those hypotheses improve prediction, longitudinal explanation, generalization, or intervention performance beyond simpler component models and broader computational formalisms.
The framework distinguishes constraint from determination and persistence from immutability. Stable choice tendencies may therefore be dynamically regenerated without requiring immutable neural representations. After establishing this core, the article introduces several deliberately non-load-bearing extensions: a frontier hypothesis concerning possible early-emerging explanatory orientations; a distinction between individual decision architecture and institutional selection; a bounded game-theoretic extension; and a normative-operational governance proposal. Each extension may fail without invalidating HDPA.
CLAIM DISCIPLINE
The article distinguishes seven claim classes:
E — established external science.
R — prior RATIUM.AI synthesis.
S — new synthesis.
H — new empirical hypothesis.
D — hypotheses inherited specifically from the Duality of Innate Cognition framework.
M — model-internal CEP propositions.
N — normative-operational governance propositions.
These classes are intentionally non-convertible. A useful synthesis is not a causal finding. A formal equation is not validation. An empirical association is not a normative permission.
RATIUM.AI sources are used to establish conceptual provenance, construct definitions, and continuity with prior stages of the framework. They do not constitute independent empirical validation of those constructs. External empirical or theoretical support is cited separately.
1. DECISIONS HAVE HISTORIES
A choice made at time t depends not only on the information currently available but also on the architecture through which that information is represented and evaluated. HDPA therefore distinguishes the encountered observation or context, denoted o_i,t, from the internally constructed decision state, denoted z_i,t. The former is input to the decision process; the latter is partly a product of the decision maker's current representational architecture.
At the highest level:
P(a_i,t | o_i,t, Θ_i,t) = F(o_i,t, Θ_i,t)
The internal state construction through Z is specified later as:
z_i,t = Φ(o_i,t, m̃_i,t, Z_i,t)
This distinction matters because two decision makers can encounter materially similar information while constructing different decision states from it. HDPA does not treat such divergence as arbitrary: one of its empirical questions is whether previous learning history contributes systematically to that difference.
The problem is recognized across adjacent decision-science literatures rather than created by HDPA. Decision theories have often simplified the acquisition problem by treating relevant information as given rather than modeling how decision makers learn and implement active sampling policies (Gottlieb, 2018). Reviews of information sampling note that the field does not yet fully understand how sampling choices guide future decisions and, in turn, how past decisions affect subsequent information sampling (Kaanders et al., 2021). Representation-learning research likewise identifies the dynamic interaction between information selection and learning as a central problem, emphasizing that the representations guiding behavior are themselves shaped by experience (Radulescu, Shin, & Niv, 2021). HDPA addresses a narrower architectural question within this recognized problem: how prior learning can be decomposed into decision-relevant parameter classes that shape subsequent information acquisition and are themselves updated through the resulting evidence environment.
The central problem is not simply why one option is chosen rather than another. It is how the system within which those options are interpreted acquired its present structure.
A person does not approach a decision with a raw archive of previous experiences. Experience is forgotten, abstracted, categorized, integrated, and reorganized. Nevertheless, some consequences of prior learning remain operative. Long-term knowledge shapes immediate goal-directed processing, adult neural systems retain meaningful experience-dependent plasticity, and stable memories can coexist with flexible or drifting neural representations rather than requiring immutable physical traces (Miller & Constantinidis, 2024; Lövdén et al., 2013; Zaki & Cai, 2025).
The relevant question is therefore:
Which consequences of learning history remain operative as parameters of later decision?
HDPA imposes no null-state assumption on Θ_i,0. Initial parameter configurations may differ across individuals, and their developmental origin remains an independent empirical question.
The core model concerns history-dependent change in decision parameters. Whether initial differences themselves arise from earlier learning, developmental organization, biological variation, or other sources is a separate question.
The canonical update equation is:
Θ_i,t+1 = U(Θ_i,t, o_i,t, m_i,t, a_i,t, y_i,t)
where o_t is the encountered observation or context, m_t acquired information, a_t the action taken, and y_t the experienced outcome.
Identical present evidence therefore need not yield identical choice distributions:
P(a | o, Θ_i) ≠ P(a | o, Θ_j)
when relevant prior configurations differ.
This is a probabilistic claim. HDPA changes the distribution of subsequent decision computation; it does not transform biography into destiny.
2. FROM LEARNING HISTORY TO DECISION ARCHITECTURE
The broader Prior Structure Principle identifies a recurrent explanatory form:
Input → Structure → Intelligibility
and explicitly does not reduce prior structure to innateness: organizing structure may be learned, developmental, cultural, computational, or otherwise acquired (Dunavich, 2026g).
HDPA addresses a narrower problem:
Learning History → Decision Parameters → Choice
The distinction is:
Structure for intelligibility ≠ Parameter for decision.
Experience-dependent representation provides a concrete bridge. Research on task-state learning shows that humans construct task representations by identifying relevant features and inferring latent causes, and that such representations can support later learning and decision-making (Niv, 2019). Recent work also connects long-term learning to persistent changes in prefrontal population activity and treats schemas as structured knowledge potentially learnable through prediction errors, hierarchical organization, and dimensionality reduction—while appropriately presenting that schema–RL connection as a synthesis rather than a settled universal mechanism (Miller & Constantinidis, 2024; Bein & Niv, 2025).
These findings justify three conservative propositions:
Experience can change representation.
Some consequences of that change can persist.
Persistence ≠ immutability.
They do not establish HDPA as an integrated architecture. That is the new hypothesis.
3. THE FIVE PARAMETERS OF HDPA
HDPA proposes:
Θ_i,t = {Z_i,t, W_i,t, V_i,t, Q_i,t, X_i,t}
The components are analytically distinguishable classes, not yet demonstrated independent latent constructs.
3.1 Why these five parameter classes?
The five classes are not proposed as an exhaustive taxonomy of every variable capable of influencing choice. Affect, effort, attention, uncertainty, executive control, action cost, habit, and other variables may matter in particular models or domains. HDPA instead proposes a bounded decision-relevant decomposition for the problem addressed here: how prior learning can remain operative at several functionally different points in subsequent decision formation.
The five classes occupy distinguishable explanatory roles:
Z: What decision state is being represented?
W: How much influence does encountered information receive as a function of its source?
Q: What outcomes are expected from available actions in that represented state?
V: How much do those possible outcomes matter to the decision maker?
X: Which information or sources are sampled before subsequent choice?
The distinctions are consequential rather than merely terminological. Z and Q differ because representing the current state is not the same as predicting what an action will produce from that state. W and X differ because weighting information already encountered is not the same as determining which information will be encountered. Q and V differ because prediction is not valuation. Z and V differ because representing what situation one is in is not equivalent to determining how much its possible outcomes matter.
HDPA therefore does not assume in advance that the five classes will emerge as five statistically independent latent variables. Their empirical separability, internal dimensionality, and possible reduction to a smaller architecture are themselves part of the research programme.
3.2 State and schema representation — Z
Z concerns how the current decision state is represented:
z_i,t = Φ(o_t, m_t, Z_i,t)
The same observation can support different generalization or action when different latent states are inferred. Representation-learning research makes this problem explicit: learning involves determining which features define the current task state and which latent causes should organize incoming information (Niv, 2019).
3.3 Source weighting — W
Information does not enter belief and choice independently of its source.
Let W_i,j,d,t represent the influence assigned by decision maker i to source j in domain d.
A meta-analysis covering 346 effect sizes from 129 independent datasets and 17,296 participants found that information indicating advisor quality was the only unique predictor of pooled advice weight among the principal factors examined; advice use increased substantially when advisors were represented as higher quality (Bailey et al., 2023).
Thus:
W_j,d1 ≠ W_j,d2
may be reasonable, and:
Perceived Reliability ≠ Actual Reliability
must remain explicit.
Source weighting can also affect subsequent source selection. Across nine experiments, Jaquiery and Yeung (2024) first exposed participants to advisors differing in objective accuracy, agreement with the participant, or both, and subsequently allowed participants to choose their advisors. With objective feedback available, participants strongly preferred the more accurate source; without feedback, agreement exerted greater influence and substantial individual differences remained. This provides direct evidence that previous experience with a source can influence later source choice, while leaving the broader HDPA recursion to be tested independently.
Accordingly:
Past Source Experience_t → Source Choice_t+1
is empirically grounded as a component relation, but W_t → X_t+1 is not treated here as an already established general law.
W is not equivalent to an epistemic policy. It concerns a particular source or source class.
3.4 Subjective outcome valuation — V
Let:
V_i,t(y, z)
denote the subjective value assigned to outcome y in represented state z.
The Identity-Value Model proposes that identity-relevant behavior can acquire greater subjective value than comparable identity-irrelevant behavior, thereby changing choice probability without determining behavior (Berkman, Livingston, & Kahn, 2017).
Accordingly:
Identity Relevance → ΔV
is used here as an externally grounded theoretical antecedent, not as a general causal law already established for all identity-sensitive decisions.
3.5 Predictive action–outcome expectations — Q
In HDPA:
Q_i,t(y | a, z)
denotes the predictive expectation that action a in state z will yield outcome y.
Q denotes predictive action–outcome expectations and should not be confused with the scalar action-value Q(s,a) convention used in standard reinforcement-learning notation.
This separation prevents prediction from being double-counted as valuation:
Q = What is expected to happen?
V = How much does that outcome matter?
A schematic decision value becomes:
DV_i,t(a) = Σ_y Q_i,t(y | a, z) V_i,t(y, z)
3.6 Information-selection policy — X
X concerns which evidence or source is sampled before subsequent choice.
This is not a cosmetic addition. Information seeking is increasingly treated as a decision process in its own right; the information people elect to acquire affects subsequent decision quality, and information itself may possess subjective value related to uncertainty, curiosity, or expected utility (Kobayashi & Kable, 2024).
Thus:
X_t → Exposure_t+1
makes the future evidence environment partly endogenous.
3.7 Canonical within-episode sequence
The five components can be placed in one schematic episode:
k_t ~ X_i,t
m_t = Acquire(k_t)
m̃_t = Γ(m_t, W_i,t)
z_i,t = Φ(o_t, m̃_t, Z_i,t)
Q_i,t(y | a, z_i,t)
DV_i,t(a) = Σ_y Q_i,t(y | a, z_i,t)V_i,t(y, z_i,t)
a_i,t ~ π(DV_i,t, Context_t)
followed by:
Θ_i,t+1 = U(Θ_i,t, o_t, m_t, a_t, y_t)
This is an analytical sequence, not a claim that the brain executes eight discrete serial modules.
4. RECURSIVE PERSISTENCE WITHOUT NEURAL LOCK-IN
The strongest integrative claim of HDPA concerns recursion.
Past learning can affect future choice in two different ways.
The first is familiar:
Past Learning → Current Parameter → Current Choice.
The second is more consequential:
Current Architecture → Information Selection → Future Exposure → Future Parameter State.
Thus:
Θ_t → Information Selection_t → Exposure_t+1 → Θ_t+1
where Information Selection_t is generated by the information-selection policy X_t contained within Θ_t:
Information Selection_t ~ X_t.
The loop is recursive only if three analytically distinct relations are present.
First, the current information-selection policy must influence subsequent sampling:
X_t → Information Selection_t.
Second, that selection must materially alter the evidence environment encountered next:
Information Selection_t → Exposure_t+1.
Third, the altered exposure must contribute to the subsequent parameter state:
Exposure_t+1 → Θ_t+1.
If current architecture does not influence sampling, if sampling does not alter actual exposure, or if later exposure leaves the architecture unchanged, the process may still involve ordinary learning across time but does not instantiate the endogenous recursive mechanism proposed here.
The consequence is partial endogeneity of the future evidence environment: some later evidence is encountered not independently of the current architecture, but because information-selection behavior generated under that architecture altered what was sampled. The endogeneity is only partial. External events, institutional availability, other agents, and environmental constraints continue to determine much of what can be encountered.
Recursion also does not imply self-confirmation. The same architecture can produce a stabilizing loop:
Θ_t → Congruent Search → Congruent Evidence → Parameter Reinforcement
or a corrective loop:
Θ_t → Diagnostic Search → Disconfirming Evidence → Parameter Revision.
Thus:
Recursion ≠ Self-Confirmation.
The empirical issue is not whether feedback exists, but whether present decision architecture systematically changes the evidence from which future decision architecture is formed.
Selective-exposure research supports one component of this possibility but also shows why the loop must not be described as intrinsically self-confirming. A large meta-analysis found a moderate average preference for congenial information, but the effect depended on defensive and accuracy motives; when uncongenial information was more useful for a current goal, participants could preferentially seek that information instead (Hart et al., 2009).
The resulting distinction is:
Parameter Persistence
Behavioral Persistence
Recursive Persistence.
The third is the HDPA synthesis.
A behavior can remain stable even while its neural representation changes. Engram research explicitly documents this tension between stability and flexibility, including representational drift under stable behavioral performance (Zaki & Cai, 2025).
Hence:
Plasticity > 0
is entirely compatible with:
Behavioral Persistence >> 0.
The core proposition is:
Persistence can be dynamically regenerated rather than physically locked in.
5. LIFE STAGE, ROLE, AND UNEVEN REALIZATION
History-dependent architecture changes across a life course, but chronological age is not itself a decision architecture.
The relevant decomposition is closer to:
Age ⇒ {Developmental State, Knowledge, Experience, Future Horizon, Health, Resources, Social Position}.
Explore–exploit research illustrates why this decomposition matters. A four-task study comparing early adolescents with adults found distinguishable random and directed exploration components in both groups; younger participants were more exploratory and flexible overall but less strategically directed in their exploration (Harms et al., 2024). Research on healthy older adults likewise found lower average information seeking and lower reliance on random variability while retaining adaptive increases in exploration when exploration was useful; substantial heterogeneity remained within the older sample (Mizell et al., 2024).
Therefore:
Younger ≠ Better Explorer
Older ≠ Rigid Exploiter.
Institutional role must be separated from age as well.
A role can alter payoff, responsibility, reputation, investment, and switching cost without any necessary decline in cognitive capacity.
Institutional persistence should also not be reduced to switching cost. Job embeddedness was originally formulated through three broader classes of relation: links to other people, teams, and groups; perceived fit with the job, organization, and community; and sacrifices associated with leaving (Mitchell et al., 2001). Subsequent review work has developed embeddedness as a broader account of why people remain situated within jobs and organizations and of its consequences beyond turnover (Lee, Burch, & Mitchell, 2014).
Thus:
Embeddedness ≠ Switching Cost.
HDPA may also provide a partial microfoundation for the previously proposed Reason–Realization Gap, in which shared rational capacities coexist with uneven realization under differences in knowledge, evidence, language, education, incentives, institutional position, and access to problem structure (Dunavich, 2026j).
But:
HDPA ≠ RRG.
A person may possess adequate decision architecture yet lack evidence, time, standing, or authority. RRG therefore remains broader than the individual parameter model.
6. MEASURING AND FALSIFYING HDPA
An architecture becomes scientifically useful only if its proposed components and couplings can lose.
The first requirement is independent measurement.
Z should be assessed through latent-state inference, structural generalization, contextual representation, or transfer.
W should be assessed by varying source accuracy, expertise, independence, or domain and measuring resulting informational influence.
V should separate subjective or self-relevant valuation from predictive expectation.
Q should estimate P(y | a,z) rather than incorporate outcome utility.
X requires genuine choice over information or sources.
Longitudinal comparison requires evidence that apparent change is not an artifact of changing measurement. For questionnaire or latent psychometric constructs, conventional measurement-invariance testing provides one relevant framework (Luong & Flake, 2023). Computational task parameters create a different validation burden: researchers should establish parameter recovery, determine whether rival models can be reliably discriminated, validate the selected model, and test whether the experimental design contains sufficient information to estimate the parameters being interpreted (Wilson & Collins, 2019). No single measurement procedure is therefore assumed to apply to every HDPA component.
Temporal analysis must also distinguish stable between-person differences from within-person dynamics. Conventional cross-lagged panel models can produce misleading lagged effects when stable person-level heterogeneity is ignored (Lucas, 2023).
6.1 From component validity to architectural validity
Establishing that each parameter class can be measured does not establish HDPA. The architectural claim concerns whether relations among the components add explanatory value beyond their separate effects.
A minimal comparison should distinguish at least four model families.
Component-specific model M1. Each parameter is used only in the class of decisions for which it is independently relevant. No integrated HDPA architecture is assumed.
Additive model M2. The five measured parameter classes contribute to the observed decision process without cross-component or cross-temporal coupling. Their effects may enter at different points in the within-episode architecture rather than as five interchangeable predictors of the final action.
Partially coupled model M3. Only specific preregistered relations are permitted, for example:
W_t → X_t+1.
Dynamic HDPA model M4. Current parameter configuration affects present choice and future information selection, while resulting exposure contributes to subsequent parameter configuration:
Θ_t → {Choice_t, Information Selection_t} → Exposure_t+1/Outcome_t → Θ_t+1.
The ordering M1–M4 represents increasing model commitment, not an assumption that the most complex model should win. HDPA receives support only to the extent that the additional coupling in M3 or M4 improves explanation beyond the simpler alternatives. Greater in-sample fit alone is insufficient. Relevant evidence may include held-out prediction, longitudinal prediction, intervention response, cross-task or cross-domain generalization, and successful recovery of the parameters and couplings attributed to the model.
A direct intervention test can make a coupling claim sharper. Suppose experimentally generated source-performance history first changes estimated W_t. If a proposed W → X relation is real, that manipulation should subsequently alter which source the participant chooses to consult. If altered consultation then produces systematically different later parameter estimates, the study tests a section of the recursive architecture rather than merely correlating stable individual differences.
Failure at any stage should narrow the corresponding coupling claim without automatically invalidating the remaining components.
The strongest comparison is also not necessarily M4 versus a null model. Broader computational formalisms may reproduce the same behavioral regularities with fewer or differently organized assumptions. Bayesian decision models, model-based reinforcement learning, partially observable decision models, predictive-processing systems, and active inference already combine important subsets of state inference, value, prediction, learning, and information seeking. Active-inference models in particular possess broad explanatory scope, but their own empirical superiority over alternative algorithms remains an open model-comparison problem rather than a settled fact (Hodson, Mehta, & Smith, 2024).
The decisive comparison is therefore:
M_dynamic/coupled > M_best simpler rival
under held-out prediction, longitudinal explanation, intervention, or generalization—not merely in-sample fit.
Accordingly, the canonical novelty claim is:
HDPA does not propose a new universal learning algorithm and does not claim to supersede broader computational frameworks. Its proposed scientific contribution is an empirically testable decomposition into analytically distinguishable classes of history-sensitive decision parameter and a set of falsifiable hypotheses concerning their longitudinal and recursive coupling. Independent explanatory status must be earned through incremental performance relative to simpler component models and broader computational formalisms.
If coupling adds nothing:
Simplify HDPA.
That is not an embarrassment to the framework. It is its primary scientific defeat condition.
EVIDENCE BOUNDARY
Everything above constitutes the empirical and synthetic core.
The remainder of the article changes claim class.
No proposition below is required for HDPA to survive.
7. FRONTIER HYPOTHESES: ONTOLOGICAL PRIOR, EPISTEMIC POLICY, AND EXPLICIT ONTOLOGY
HDPA permits, but does not require, non-identical initial conditions.
This creates a frontier question:
What explains early variance in decision architecture?
The default competitor remains learning and socialization.
A second possibility is that some stable early-emerging explanatory orientation contributes incremental variance.
The provisional name for such a construct is:
ONT^prior.
It denotes a hypothetical pre-explicit orientation toward which forms of causal organization, directionality, constraint, programme, or purpose initially appear plausible.
One candidate dimension concerns the explanatory hierarchy assigned to development relative to entropy. At one pole, genetically and regulatorily organized ontogenesis may be represented as a bounded form of development within a more general entropy-governed physical order: organismic development is real because the organism possesses inherited and regulatory organization, but its existence does not imply that physical reality as a whole constitutes a developmental process. At the other pole, the same ontogenesis may be represented as a local instance of a more general developmental grammar attributed across biological, historical, and potentially physical scales. The relevant contrast is therefore not whether development exists, but whether development is ontologically bounded within entropy or whether organismic development is interpreted as one manifestation of a more general developmental character attributed to reality.
This contrast is proposed only as a candidate dimension of ONT^prior. HDPA does not assume that it is bipolar, universal, developmentally prior, or empirically recoverable. Its structure must be established independently before any mapping to the Duality of Innate Cognition or to CEP ontological categories is permissible.
ONT^prior is therefore not religion, political identity, materialism, idealism, an OPI score, or a CEP category.
Developmental teleology research demonstrates why early explanatory preferences are scientifically investigable, but it does not validate ONT^prior. Chinese children, for example, displayed broad teleological explanation preferences across natural phenomena that declined with grade level, showing that such tendencies cannot be attributed solely to Western Abrahamic cultural environments; the study does not establish their precise developmental source (Schachner et al., 2017).
Accordingly:
Early Emergence ≠ Innateness.
The empirical order must be:
ONT^prior exists?
then:
Structure(ONT^prior) = ?
with possible outcomes including bipolar, multidimensional, domain-specific, or null structure.
A direct operationalization could compare judgments of organized development in systems possessing inherited internal regulatory architecture with judgments of change in systems lacking such architecture, and then test whether developmental grammar generalizes across organismic, population-level, historical, and physical domains. The empirical question is whether a stable dimension governing such cross-domain generalization can be recovered—not whether participants endorse any predetermined philosophical label.
Only afterward is a mapping to Duality of Innate Cognition permissible. The legacy DIC framework proposes two innate orientations and a 50:50 pre-social equilibrium, but those are model claims rather than assumptions of the present article (Dunavich, 2026d). Prior RATIUM.AI comparative work has already provided a methodological warning by failing to recover a simple binary intellectual-profile structure and retaining DIC only as a non-load-bearing hypothesis (Dunavich, 2026a).
A second frontier construct is Epistemic Policy:
EPI = {ES, EIS, UC, RP, AT}
representing candidate policies concerning evidence sensitivity, evidence-independence sensitivity, uncertainty calibration, revision, and alternative testing.
This is not source weight W. It asks not how much do I trust this source? but what counts as sufficient warrant for revision?
Contemporary theory distinguishes belief updating proper from upstream evidence search, evidence evaluation, hypothesis specification, integration, and reasoning, supporting the legitimacy of studying higher-order evidence-processing policies while not validating this particular EPI decomposition (Sommer, Musolino, & Hemmer, 2024).
The frontier hypotheses are therefore:
ONT^prior → HDPA
and potentially:
ONT^prior → EPI.
The strongest competitor is simpler:
HDPA → EPI
without a separate ontological prior.
Finally, explicit ontology:
ONT^explicit
must be distinguished from ONT^prior.
The canonical CEP relation:
EPI —M:warrant→ ONT^explicit
is a model-internal relation about warrant, not an established developmental law. The empirical psychological hypothesis:
EPI_t —H→ ONT^explicit_t+1
requires independent longitudinal evidence.
Any failure in this section leaves HDPA intact.
8. FROM INDIVIDUALS TO INSTITUTIONS
Institutions are not individuals enlarged.
Therefore:
Individual State ≠> Institutional State.
Individual architecture can matter institutionally only through mediating processes such as observable behavior, selection, role allocation, gatekeeping, incentives, authority, and permission:
HDPA_i → Behavior_i → Selection/Role/Gate → Institutional Outcome.
Permission denotes the operative institutional state that allows, conditions, blocks, suspends, withdraws, or reauthorizes a consequential action, decision, policy, or deployment. It is not synonymous with evaluation, recommendation, or authority. An evaluator may judge an action defective without possessing authority to alter its permission state; an authority may possess formal decision rights yet decline to change that state.
Accordingly:
Evaluation ≠ Authority ≠ Permission.
The institutional question is not merely whether a problem can be recognized, but whether recognition can reach a decision right capable of changing what the system is actually permitted to do.
This prevents cognitive explanation from silently becoming institutional explanation.
Organizational-routine research provides a useful non-cognitive analogue. Routines can be understood as repetitive, recognizable patterns of interdependent action carried out by multiple actors. The persistence of such a pattern therefore does not require persistence of one individual, one shared psychological state, or one homogeneous belief system (Pentland & Hærem, 2015).
Accordingly:
Institutional Persistence ≠> Cognitive Homogeneity.
Employee voice research likewise makes clear that the presence or absence of visible criticism cannot be equated with underlying agreement. Because silence is a distinct and context-sensitive organizational behavior, lower visible dissent is insufficient evidence of higher underlying agreement (Morrison, 2023).
Observed Quiet ≠> Demonstrated Consensus.
The institutional construct Governing Ontology (GO) refers here not to the average worldview of participants but to the operational categories through which a decision regime represents relevant entities, relations, harms, objectives, failures, and permissible interventions.
Thus:
GO ≠ mean(ONT^explicit_i).
A GO claim should be based on observable institutional traces—rules, rubrics, categories, authorized models, gate criteria, or permission effects—not on an analyst's attribution of a hidden worldview.
The hard evidential rule is:
No Operational Trace → No GO Claim.
Correction must likewise be analyzed as a chain rather than inferred from the existence of a complaints procedure:
Criticism → Review → Judgment → Gate Change → Implementation → Restoration.
A proposed Correction Conversion Profile therefore measures each conversion rather than collapsing correction into one score:
CCP = {
P(Review | Eligible Signal),
P(Adverse Judgment | Review),
P(Gate Change | Adverse Judgment),
P(Implementation | Gate Change),
P(Restoration | Implementation)
}.
Soft Closure is reserved for cases in which criticism remains formally expressible while consequential correction is structurally impaired. It is an article-defined institutional condition, not an established universal organizational diagnosis.
9. STRATEGIC PERSISTENCE AND CEP AS A BOUNDED EXTENSION
Institutional persistence and game-theoretic equilibrium are different analytical objects.
Two axes should be kept separate.
The institutional-diagnostic axis is:
Persistence → Path Dependence? → Correction Asymmetry/Soft Closure?
The strategic axis is:
Strategic Interdependence? → Repeated Game? → Equilibrium Concept?
Path dependence itself requires more than recurrence. Pierson's influential formulation emphasizes increasing returns, timing, sequence, coordination effects, adaptive expectations, and rising reversal costs as mechanisms through which earlier institutional choices alter later feasible paths (Pierson, 2000).
Repeated interaction is still not enough for equilibrium diagnosis. Repeated-game research shows that future interaction can sustain behavior unavailable in a one-shot environment, while also showing that an equilibrium-supportable pattern need not be the pattern actors actually select (Dal Bó & Fréchette, 2018).
An empirical strategic analysis requires identifiable actors, action sets, information or monitoring structure, payoff-relevant consequences, and history-dependent strategies:
σ_i: H_t → Δ(A_i).
A fixed point of an arbitrary institutional transition:
I* = R(I*)
does not establish Nash equilibrium.
CEP enters only after these distinctions.
Its canonical core identifies ontological and epistemological strategy dimensions and treats S4 = D × B as a Pareto-inefficient Nash equilibrium inside the model; the foundational CEP framework explicitly treats such equilibrium claims as model-internal or interpretive unless separately supported empirically (Dunavich, 2026c).
Hence:
CEP Core ≠ Empirical Carrier Game.
A real institutional application must separately identify actors, strategies, information, payoffs, deviations, and a reliable mapping to CEP categories.
Likewise:
Pareto_CEP-Core ≠ Pareto_Empirical-Game.
Strategic payoff is not moral welfare.
CEP is therefore a bounded downstream model, not a new name for institutional persistence.
10. CORRECTIVE ARCHITECTURE AS A GOVERNANCE IMPLICATION
This section changes from descriptive explanation to normative-operational proposal.
Define Governance Reliability independently as:
GR = Stable, consequential, pressure-resistant correction.
Epistemological Reopening (ER) denotes the capacity of relevant evidence to trigger authorized reconsideration not only of a decision made under an existing rule, but—where the evidence warrants it—of the categories, assumptions, evidentiary standards, or governing representations through which that rule or decision was generated.
The distinction is therefore:
Appeal Within Rule ≠ Reopening of Rule.
A system may permit case-level appeals while treating the categories producing those cases as categorically unavailable for reconsideration. ER concerns the second possibility.
Because the proposition introduced here is normative-operational rather than an established empirical law, it is denoted:
N_GR: ER is non-substitutable for GR.
The claim is deliberately stronger than saying that reopening is merely useful, but weaker than claiming that reopening is sufficient. A system may preserve epistemological reopening and still fail because standing, competent review, decision authority, implementation, remediation, or verification are defective.
The proposed sequence is a normative-operational synthesis rather than an existing regulatory standard:
Traceability + Appropriate Transparency
→ Contestability
→ Correctability
→ Reversibility/Remediation.
Its components nevertheless have substantial antecedents in established governance frameworks. The National Institute of Standards and Technology (NIST) AI RMF 1.0 incorporates post-deployment monitoring, relevant-actor input, appeal and override, decommissioning, incident response, recovery, and change management. The revised OECD AI Principles connect transparency and traceability to the ability of adversely affected persons to challenge AI outputs and call for mechanisms capable of override, repair, or decommissioning where appropriate. Article 20 of Regulation (EU) 2024/1689, in its consolidated text through 27 July 2026, requires corrective action for non-conforming high-risk AI systems, including conformity correction, withdrawal, disabling, or recall. These frameworks support the relevance of the component capabilities; they do not establish the particular cumulative architecture proposed here (NIST, 2023; OECD, 2024; European Union, 2024).
The synthesis proposed here goes one level deeper:
Correction Depth ≥ Failure Generation Depth.
A defective output may require output correction.
A defective model may require model correction.
A defective gate cannot necessarily be corrected by changing one case.
A defective governing ontology cannot necessarily be corrected merely by applying its categories more carefully.
HDPA adds a specific prediction:
In some cases, effective correction will require intervention on the conditions generating or updating decision parameters rather than the addition of corrective information alone.
Thus:
Information Availability ≠ Correction.
Where literal reversal is impossible, remediation and restoration replace an impossible demand for complete undoing.
This governance proposition is related to HDPA. It is not an empirical conclusion logically entailed by HDPA.
11. FAILURE CONDITIONS AND RESEARCH PROGRAMME
The framework should not survive by reinterpretation.
Acceptable outcomes include:
CONFIRM, NARROW, SIMPLIFY, RECLASSIFY, SPLIT, REJECT.
These outcomes describe different empirical consequences for the architecture.
CONFIRM at the tested scope applies when the proposed decomposition and preregistered couplings add robust explanatory or predictive value beyond strong rivals.
NARROW applies when the architecture is supported only under specified domains, populations, developmental periods, or task structures rather than at the scope initially proposed.
SIMPLIFY applies when fewer parameter classes or fewer couplings perform as well as the fuller architecture.
RECLASSIFY applies when a proposed HDPA component is empirically useful but is better interpreted as an instance of an already established construct or as a parameter internal to another computational framework.
SPLIT applies when a proposed parameter class proves to contain multiple processes that cannot defensibly be represented as one class.
REJECT applies when the architecture fails to provide meaningful explanatory, predictive, or discriminative value beyond its constituent mechanisms and stronger rivals.
These outcomes can occur at different levels. A coupling can be rejected while a component survives; the integrated architecture can be simplified while retaining a useful taxonomy; an upstream frontier hypothesis can fail while the HDPA core remains unaffected.
HDPA loses independent explanatory privilege if:
M_dynamic/coupled ≤ M_best simpler rival
under robust held-out prediction, longitudinal comparison, intervention, or generalization.
The research programme should therefore proceed in three stages.
Stage I — Core HDPA validation. Validate the component measures; establish their discriminability; compare component-specific, additive, partially coupled, and dynamic models; test recursive exposure; and use intervention where feasible to identify directional relations.
Stage II — Optional individual frontier. Only after the HDPA core is independently tractable should research test ONT^prior—including any recoverable development–entropy hierarchy—EPI, developmental ordering, and any possible DIC mapping.
Stage III — Institutional and strategic extension. Only after individual-level effects can be linked to observable behavior should research test selection, gates, authority, permission, correction conversion, strategic interdependence, and—under still stricter conditions—CEP mapping.
Downstream Fit ≠ Upstream Validation.
Evidence at a later stage cannot retroactively validate an earlier one.
No single study can validate:
HDPA + ONT^prior + EPI + DIC + Institutional Conversion + CEP + Governance Architecture.
The complete architecture gains credibility only through independent module convergence.
Its defeat relations are deliberately asymmetric:
DIC fails → HDPA survives.
ONT^prior fails → HDPA survives.
EPI fails → HDPA may survive.
Institutional extension fails → individual HDPA survives.
CEP fails → HDPA survives.
Governance proposal fails → descriptive HDPA survives.
Only one class of negative result defeats the central scientific claim: failure of the HDPA decomposition and its couplings to add meaningful explanatory or discriminative value beyond its constituent mechanisms and stronger computational alternatives.
Even that result need not erase all utility. It may leave a descriptive taxonomy.
It would, however, defeat the claim that HDPA deserves independent explanatory status.
CONCLUSION
Decisions have histories, but history need not be represented as destiny.
Prior experience can matter not only because remembered propositions are retrieved at the moment of choice, but because learning can alter how later situations are represented, which sources are credited, how outcomes are valued, what consequences their actions are expected to produce, and which information is obtained before the next decision.
The core architecture is:
Θ_t = {Z_t, W_t, V_t, Q_t, X_t}
with:
Θ_t+1 = U(Θ_t, o_t, m_t, a_t, y_t)
and the proposed recursive extension:
Θ_t → Information Selection_t → Exposure_t+1 → Θ_t+1.
These equations organize hypotheses. They do not validate them.
HDPA does not claim that its five parameter classes are already established as one empirical architecture. It creates a way to test whether this parameter-level decomposition and its recursive couplings provide incremental explanatory value beyond simpler component accounts and broader computational formalisms.
That gives the framework a simple scientific burden:
If the coupling adds nothing, simplify the theory.
The bounded scientific conclusion is:
Learning history can configure the representations, source weights, valuations, action–outcome expectations, and information-selection policies through which later decisions are made. Their configuration can constrain subsequent choice without determining it, and their recursive interaction may dynamically reproduce stable choice tendencies without requiring neural immutability.
The governance framework adds a separate proposition:
Where historically accumulated constraints become consequential, reliable governance should preserve a pressure-resistant path by which evidence can reopen the level generating the consequence and alter, reverse, or remediate the operative state when correction is warranted.
The two propositions are connected.
They are not the same proposition.
REFERENCES
Bailey, P. E., Leon, T., Ebner, N. C., Moustafa, A. A., & Weidemann, G. (2023). A meta-analysis of the weight of advice in decision-making. Current Psychology, 42(28), 24516–24541. https://doi.org/10.1007/s12144-022-03573-2
Bein, O., & Niv, Y. (2025). Schemas, reinforcement learning and the medial prefrontal cortex. Nature Reviews Neuroscience, 26, 141–157. https://doi.org/10.1038/s41583-024-00893-z
Berkman, E. T., Livingston, J. L., & Kahn, L. E. (2017). Finding the “self” in self-regulation: The identity-value model. Psychological Inquiry, 28(2–3), 77–98. https://doi.org/10.1080/1047840X.2017.1323463
Dal Bó, P., & Fréchette, G. R. (2018). On the determinants of cooperation in infinitely repeated games: A survey. Journal of Economic Literature, 56(1), 60–114. https://doi.org/10.1257/jel.20160980
Gottlieb, J. (2018). Understanding active sampling strategies: Empirical approaches and implications for attention and decision research. Cortex, 102, 150–160. https://doi.org/10.1016/j.cortex.2017.08.019
Harms, M. B., Xu, Y., Green, C. S., Woodard, K., Wilson, R., & Pollak, S. D. (2024). The structure and development of explore-exploit decision making. Cognitive Psychology, 150, 101650. https://doi.org/10.1016/j.cogpsych.2024.101650
Hart, W., Albarracín, D., Eagly, A. H., Brechan, I., Lindberg, M. J., & Merrill, L. (2009). Feeling validated versus being correct: A meta-analysis of selective exposure to information. Psychological Bulletin, 135(4), 555–588. https://doi.org/10.1037/a0015701
Hodson, R., Mehta, M., & Smith, R. (2024). The empirical status of predictive coding and active inference. Neuroscience & Biobehavioral Reviews, 157, 105473. https://doi.org/10.1016/j.neubiorev.2023.105473
Jaquiery, M., & Yeung, N. (2024). Preferences for advisor agreement and accuracy. PLOS ONE, 19(9), e0311211. https://doi.org/10.1371/journal.pone.0311211
Kaanders, P., Juechems, K., O'Reilly, J., & Hunt, L. (2021). Dissociable mechanisms of information sampling in prefrontal cortex and the dopaminergic system. Current Opinion in Behavioral Sciences, 41, 63–70. https://doi.org/10.1016/j.cobeha.2021.04.005
Kobayashi, K., & Kable, J. W. (2024). Neural mechanisms of information seeking. Neuron, 112, 1741–1756. https://doi.org/10.1016/j.neuron.2024.04.008
Lee, T. W., Burch, T. C., & Mitchell, T. R. (2014). The story of why we stay: A review of job embeddedness. Annual Review of Organizational Psychology and Organizational Behavior, 1, 199–216. https://doi.org/10.1146/annurev-orgpsych-031413-091244
Lövdén, M., Wenger, E., Mårtensson, J., Lindenberger, U., & Bäckman, L. (2013). Structural brain plasticity in adult learning and development. Neuroscience & Biobehavioral Reviews, 37(9), 2296–2310. https://doi.org/10.1016/j.neubiorev.2013.02.014
Lucas, R. E. (2023). Why the cross-lagged panel model is almost never the right choice. Advances in Methods and Practices in Psychological Science, 6(1), 1–22. https://doi.org/10.1177/25152459231158378
Luong, R., & Flake, J. K. (2023). Measurement invariance testing using confirmatory factor analysis and alignment optimization: A tutorial for transparent analysis planning and reporting. Psychological Methods, 28(4), 905–924. https://doi.org/10.1037/met0000441
Miller, J. A., & Constantinidis, C. (2024). Timescales of learning in prefrontal cortex. Nature Reviews Neuroscience, 25, 597–610. https://doi.org/10.1038/s41583-024-00836-8
Mitchell, T. R., Holtom, B. C., Lee, T. W., Sablynski, C. J., & Erez, M. (2001). Why people stay: Using job embeddedness to predict voluntary turnover. Academy of Management Journal, 44(6), 1102–1121. https://doi.org/10.5465/3069391
Mizell, J.-M., Wang, S., Frisvold, A., Alvarado, L., Farrell-Skupny, A., Keung, W., Phelps, C. E., Sundman, M. H., Franchetti, M.-K., Chou, Y.-H., Alexander, G. E., & Wilson, R. C. (2024). Differential impacts of healthy cognitive aging on directed and random exploration. Psychology and Aging, 39(1), 88–101. https://doi.org/10.1037/pag0000791
Morrison, E. W. (2023). Employee voice and silence: Taking stock a decade later. Annual Review of Organizational Psychology and Organizational Behavior, 10, 79–107. https://doi.org/10.1146/annurev-orgpsych-120920-054654
Niv, Y. (2019). Learning task-state representations. Nature Neuroscience, 22, 1544–1553. https://doi.org/10.1038/s41593-019-0470-8
Pentland, B. T., & Hærem, T. (2015). Organizational routines as patterns of action: Implications for organizational behavior. Annual Review of Organizational Psychology and Organizational Behavior, 2, 465–487. https://doi.org/10.1146/annurev-orgpsych-032414-111412
Pierson, P. (2000). Increasing returns, path dependence, and the study of politics. American Political Science Review, 94(2), 251–267. https://doi.org/10.2307/2586011
Radulescu, A., Shin, Y. S., & Niv, Y. (2021). Human representation learning. Annual Review of Neuroscience, 44, 253–273. https://doi.org/10.1146/annurev-neuro-092920-120559
Schachner, A., Zhu, L., Li, J., & Kelemen, D. (2017). Is the bias for function-based explanations culturally universal? Children from China endorse teleological explanations of natural phenomena. Journal of Experimental Child Psychology, 157, 29–48. https://doi.org/10.1016/j.jecp.2016.12.006
Sommer, J., Musolino, J., & Hemmer, P. (2024). Updating, evidence evaluation, and operator availability: A theoretical framework for understanding belief. Psychological Review, 131(2), 373–401. https://doi.org/10.1037/rev0000444
Wilson, R. C., & Collins, A. G. E. (2019). Ten simple rules for the computational modeling of behavioral data. eLife, 8, e49547. https://doi.org/10.7554/eLife.49547
Zaki, Y., & Cai, D. J. (2025). Memory engram stability and flexibility. Neuropsychopharmacology, 50(1), 285–293. https://doi.org/10.1038/s41386-024-01979-z
National Institute of Standards and Technology. (2023). Artificial Intelligence Risk Management Framework (AI RMF 1.0). NIST AI 100-1. https://doi.org/10.6028/NIST.AI.100-1
OECD. (2024). Recommendation of the Council on Artificial Intelligence (OECD/LEGAL/0449), revised 2024. https://legalinstruments.oecd.org/en/instruments/OECD-LEGAL-0449
European Union. (2024). Regulation (EU) 2024/1689 (Artificial Intelligence Act), Article 20. Consolidated version: 27 July 2026. https://eur-lex.europa.eu/eli/reg/2024/1689/2026-07-27/eng
RATIUM.AI CONCEPTUAL PROVENANCE
The following RATIUM.AI sources establish conceptual provenance and continuity inside the framework. They are not counted as independent empirical validation of HDPA or of any externally testable claim.
Dunavich, Benny. (2026a). A Hidden Split in Formal Reason. RATIUM.AI. https://www.ratium.ai/articles/hidden-split-in-formal-reason
Dunavich, Benny. (2026b). The AI Configuration Paradox. RATIUM.AI. https://www.ratium.ai/articles/ai-configuration-paradox
Dunavich, Benny. (2026c). The Central Equilibrium Problem — Independent Research Framework. RATIUM.AI. https://www.ratium.ai/articles/cep-doctoral-scale-work
Dunavich, Benny. (2026d). The Duality of Innate Cognition. RATIUM.AI Foundational Source Dossier. https://www.ratium.ai/foundational-source-dossier/duality-of-innate-cognition
Dunavich, Benny. (2026e). The Key to a Stable Governance Layer: Solve the Foundational Decision Problem First. RATIUM.AI. https://www.ratium.ai/articles/the-key-to-a-stable-governance-layer
Dunavich, Benny. (2026f). The Ontogenesis Projection Index: Four Canonical Books and the Civilizational Sublimation of Development. RATIUM.AI. https://www.ratium.ai/foundational-source-dossier/duality-of-innate-cognition/ontogenesis-projection-index
Dunavich, Benny. (2026g). The Prior Structure Principle: Apperception and the Universal Architecture of Cognition. RATIUM.AI. https://www.ratium.ai/articles/prior-structure-principle-apperception-cognition
Dunavich, Benny. (2026h). The Priority of Epistemology: System Convergence and Consensus Ontology. RATIUM.AI. https://www.ratium.ai/articles/priority-of-epistemology-system-convergence-consensus-ontology
Dunavich, Benny. (2026i). The Sublimation of Ontogenesis. RATIUM.AI. https://www.ratium.ai/articles/the-sublimation-of-ontogenesis
Dunavich, Benny. (2026j). Universal Reason, Prior Structure, and the Foundations of Stable AI Governance. RATIUM.AI. https://www.ratium.ai/articles/universalreason
Dunavich, Benny. (2026k). When the Correction Mechanism Fails. RATIUM.AI. https://www.ratium.ai/articles/when-the-correction-mechanism-fails
Related Source and Reference Pages
This article belongs to the public essay layer of RATIUM.AI. For readers who want to move from this article into the broader source, technical, and orientation layers of the project, the following pages provide the relevant entry points.
Articles
The articles page gathers the public essay layer of RATIUM.AI, including arguments on stable AI governance, decision-control architecture, visible governance versus real authority, universal reason, technical competence, purpose governance, and the doctoral-scale framing of CEP.
Foundational Source Dossier
The foundational source dossier presents the deeper intellectual corpus behind CEP, LoopGuard-AI, and the broader RATIUM.AI research structure.
Technical & Reference Dossiers
The technical and reference dossier page collects architecture, visual explanation, methodological context, FAQ material, and technical source pages related to LoopGuard-AI and CEP.
RATIUM.AI / LoopGuard-AI / CEP FAQ
The RATIUM.AI / LoopGuard-AI / CEP FAQ provides a structured orientation to the main concepts behind RATIUM.AI, CEP, and LoopGuard-AI, helping readers navigate the framework through clear questions, definitions, and internal conceptual links.