Before AI Governance: The Prior Formulation of Social Decision Problems

​

Prior Social Decision Structures, Risk-Origin Diagnosis, and the Upstream Conditions of Operational AI Governance

Maturity: Advanced candidate upstream governance theory and measurement framework; conceptually developed, empirically unvalidated.

​

Abstract

AI governance is usually organized around systems that already exist or are already being designed. It asks whether models are safe, fair, robust, explainable, auditable, secure, and appropriately supervised. These questions are necessary, but they may begin after an earlier decision has already been made: which human or institutional problem the system is supposed to represent.

AI systems enter decision regimes whose categories, evidentiary rules, authority relations, incentives, burden distributions, and correction mechanisms may exist independently of the specified integration. If that prior structure remains implicit, governance may treat an inherited institutional failure as a model-local defect, reduce a technically distinctive risk to general social explanation, or miss a failure produced by the coupling of social and technical mechanisms.

This article proposes an upstream diagnostic architecture. A Social Decision Structure (SDS) is a bounded configuration of decision objects, interested parties, need profiles, categories, evidence rules, authority, incentives, benefit–burden distribution, and correction. A Prior Social Decision Structure (PSDS) is an SDS whose relevant causal elements precede or operate independently of a specified AI integration. The primary unit of analysis is the Social Decision Inheritance Instance (SDII), which requires relative causal precedence, a bounded decision structure, a supported inheritance or activation channel, and material governance relevance.

The Risk-Origin Profile (ROP) distinguishes inherited, interaction-emergent, and AI-native causal contributions while preserving unresolved epistemic remainder:

​

ROPj=(Ij,Ej,Nj;Uj)

​

Causal-Level Misclassification (CLM) identifies mismatch between the intervention levels required by the supported causal account and the levels actually addressed. Prior Social Decision Formulation Completeness (PSDFC) is a proposed two-axis adequacy requirement: completeness and epistemic status. It governs transfer of the upstream record into a provisionally admissible AI problem model.

The contribution is limited. Existing scholarship already establishes that problem formulation is negotiated and normatively consequential, technical abstraction may omit essential social relations, affected stakeholders and values matter, harms arise across the machine-learning lifecycle, and context should be mapped before risks are managed. The residual proposal is the integration of relative causal precedence, bounded inheritance-instance classification, explicit channel evidence, mixed-origin profiling, intervention-level comparison, and a reviewable upstream handoff.

The framework is conceptually developed but empirically unvalidated. Its independent status depends on coding reliability, negative controls, comparison with strong existing methods, and evidence that it improves intervention selection and governance traceability.

Keywords: AI governance; sociotechnical systems; problem formulation; institutional inheritance; causal diagnosis; algorithmic accountability; correction; contestability; AI risk; decision architecture.

Contents

• Abstract

• Introduction

• Part I — Canonical Theory

• 1. Governance Begins at the Wrong Causal Level

• 2. Social Decision Structures and Relative Causal Precedence

• 3. The Social Decision Inheritance Instance

• 4. Canonical Conditions and Exclusion Rules

• 5. The Risk-Origin Profile

• 6. Causal-Level Misclassification

• Part I Conclusion

• Part II — Upstream Foundations

• 7. Prior Structure and Governance Intelligibility

• 8. Universal Reason and the Reason–Realization Gap

• 9. Authority, Evidence, Incentives, and Correction Before AI

• 10. ADM/CIV and the Distribution of Benefit, Burden, and Corrective Capacity

• Part III — Transmission into AI-Mediated Decision Regimes

• 11. Provisional Inheritance and Activation Channels

• 12. Reproduction, Amplification, Operationalization, and Stabilization

• 13. Interaction-Emergent Risk

• 14. AI-Native Risk and the Limits of Social Explanation

• 15. CEP and S4 as Bounded Candidate Diagnostics

• 16. Prior Social Decision Formulation Completeness

• 17. The Handoff to the AI Problem Model

• Part IV — Research Position and Validation

• 18. Adjacent Traditions and the Residual Research Gap

• 19. Coding Architecture and Evidence States

• 20. Empirical Hypotheses and Comparative Research Design

• 21. Strongest Objections and Replies

• 22. Reduction, Falsification, Scope, and Maturity

• 23. Conclusion: Name the Decision Structure Before Governing the AI

• Glossary

• References

• Appendix A — Evidence and Claim-Control Protocol

Introduction

AI governance often begins after a decisive conceptual step has already occurred.

A use case has been selected, an organizational objective translated into a task, a target or proxy chosen, and a workflow defined. One actor is designated as the user, another as the reviewer, and another as the person or population affected by the result. Governance then asks whether the system is accurate, safe, fair, secure, explainable, auditable, reversible, and appropriately supervised. It introduces evaluations, risk registers, human-review stages, incident procedures, release gates, access restrictions, and rollback provisions.

These instruments matter. The causal problem is that they may be applied to a decision problem whose prior social structure has never been formulated.

An employment-screening system enters an existing relation among employers, applicants, recruiters, credentialing institutions, legal rules, organizational risk preferences, and accepted categories of merit. A healthcare-triage system enters relations among patients, clinicians, administrators, insurers, resource constraints, diagnostic conventions, and institutional definitions of urgency. AI does not create these relations, yet it may reproduce their classifications, amplify their reach, operationalize them through repeatable action, stabilize them through infrastructural dependence, or combine them with technical properties that create a new failure mechanism.

The relevant object is therefore the AI-mediated regime: the system, workflow, authority structure, affected-party relation, and correction process through which AI outputs acquire consequence. Neither model-only analysis nor unrestricted social explanation is sufficient for every case.

The prior analytical problem is causal differentiation:

• Which relevant decision structures existed independently of the specified integration?

• Through which supported channels did they become active?

• Which effects arise from social–technical interaction?

• Which mechanisms depend materially on the AI configuration?

• Which intervention levels are supported or independently justified?

These questions are adjacent to established work. Problem-formulation research shows that the translation of organizational objectives into targets and proxies is negotiated and normatively consequential (Passi & Barocas, 2019). Sociotechnical analysis shows that narrow abstraction can remove the actors and institutional relations through which fairness, justice, and due process materialize (Selbst et al., 2019). Participatory and causal-system approaches incorporate societal context and typically excluded stakeholders into problem formulation (Martin et al., 2020a, 2020b). The National Institute of Standards and Technology (NIST) AI Risk Management Framework (AI RMF) 1.0 calls for documentation of intended purpose, deployment context, affected actors, benefits, harms, assumptions, limitations, and lifecycle interdependencies (Tabassi, 2023). Lifecycle-harm research distinguishes seven potential sources of downstream harm across data collection, development, and deployment (Suresh & Guttag, 2021).

The present article does not claim that AI governance has ignored society. Its narrower claim concerns four questions that context, stakeholder, lifecycle, and task-formulation methods do not necessarily answer. Does a bounded social decision structure exist independently of the specified integration? How does it become materially active? How does its contribution differ from interaction-emergent and AI-native mechanisms? Is intervention directed at the appropriate level?

The candidate proposition is:

Governance is vulnerable to causal misdirection unless it develops a reviewable account of any prior social decision structure materially carried forward into an AI-mediated regime.

The proposition imposes four disciplines.

First, the social object must be bounded. References to culture, politics, inequality, bias, or institutional context are insufficient without a decision object, interested parties, need profiles, categories, evidence rules, authority, burdens, and correction.

Second, inheritance must be demonstrated rather than presumed. Human-generated data or institutional deployment does not establish that a particular decision structure governs a particular outcome. A supported channel is required.

Third, origins need not be exclusive. One outcome may contain an inherited category, an interaction-emergent mechanism, and an AI-native vulnerability. The framework therefore uses a profile rather than a forced single-cause classification.

Fourth, diagnosis must affect intervention. A causal distinction has governance value only if it changes what is inspected, which evidence is required, where authority resides, or where correction is directed.

The article develops six principal constructs: Social Decision Structure, Prior Social Decision Structure, Social Decision Inheritance Instance, Risk-Origin Profile, Causal-Level Misclassification, and Prior Social Decision Formulation Completeness. Their value depends on whether independent researchers can identify the same units, distinguish transmission from resemblance, construct reproducible origin profiles, and improve intervention selection over strong existing approaches.

Part I defines the canonical theory. Part II develops its epistemic and institutional foundations. Part III specifies transmission, mixed causality, and the handoff into the AI problem model. Part IV positions the framework against adjacent traditions and defines its coding architecture, empirical hypotheses, objections, reduction rules, and maturity.

The endpoint is not a metric, gate, or permission state. It is the problem object from which those instruments must be derived.

Part I — Canonical Theory

Back to contents

1. Governance Begins at the Wrong Causal Level

AI governance commonly begins from a visible technical object: a model, dataset, output, benchmark result, agent action, tool call, workflow, application, deployment, or incident. The object is inspected for defects or risks, and the governance process attempts to modify it or constrain its operation.

This starting point is often appropriate. Prompt injection may require technical containment. Unauthorized tool use may require isolation, authentication, reduced permissions, or runtime interruption. A system that exposes confidential data may require immediate restriction regardless of the surrounding social structure.

The claim of this article is therefore not that governance should always move outward from the model to society. It is that governance must establish whether the visible technical object is the causally sufficient object.

A model that repeatedly disadvantages applicants with nonstandard career histories may exhibit at least four different causal structures:

1. a model-local statistical or representational defect;

2. an institutional proxy that predates the model and enters through data, target, or evaluation;

3. an interaction failure produced when ranking behavior is coupled to a high-volume screening workflow;

4. a technically accurate model relative to a target whose institutional legitimacy or relevance is itself disputed.

These descriptions imply different interventions.

A representational defect may justify model-level correction. An inherited proxy may require reformulation of the target or data-generation process. An interaction failure may require workflow redesign, authority limits, or a different correction path. A disputed target may require reconsideration of who defined success and whose need profile is represented.

Treating these cases as equivalent because they produce similar outputs is a causal error.

The same applies in reverse. A technical vulnerability should not be dissolved into general institutional critique merely because it occurs inside an organization. Structural explanation does not replace technical containment.

Let the candidate intervention levels be:

​

ℒ={Lm,Li,Lw,Lo,Lins,Ls}

​

where:

• (L_m) = model level;

• (L_i) = interface level;

• (L_w) = workflow level;

• (L_o) = organizational level;

• (L_{ins}) = institutional level;

• (L_s) = wider social-decision-structure level.

These levels are analytical rather than ontologically isolated. One failure may require intervention at several levels.

The governing principle is:

Visibility does not establish causal priority.

Model outputs, dashboards, scores, and incidents are visible. The mechanism that makes them repeat may reside in a category, incentive, authority relation, burden distribution, or correction failure located elsewhere.

Two symmetrical pathologies follow.

Technical compression

A multidimensional decision failure is compressed into a model property. Governance responds with more data, another benchmark, adjusted weights, filtering, prompt revision, or another human-review layer. These interventions may improve local behavior while leaving the target, authority structure, or correction failure unchanged.

Social diffusion

A bounded technical failure is expanded into a general account of power, inequality, institutional bias, or political conflict. The account may be relevant but operationally indeterminate. It does not identify which technical state must change or which mechanism requires immediate containment.

Both detach diagnosis from intervention.

The framework therefore asks:

1. What relevant decision structure existed independently of the specified integration?

2. Through which supported mechanism did it become active?

3. Which contributions are inherited, interaction-emergent, AI-native, or unresolved?

4. Which intervention levels are supported by that diagnosis?

2. Social Decision Structures and Relative Causal Precedence

Every AI system operates inside a social environment. That fact is too broad to function as a governance unit.

A Social Decision Structure is a bounded configuration through which a decision or family of decisions becomes possible, authoritative, and consequential:

​

S=(O,IP,NP,K,A,J,D,C)

​

where:

• (O) = decision object and purpose;

• (IP) = interested parties and functional positions;

• (NP) = need profiles and competing objectives;

• (K) = categories, proxies, evidentiary rules, and knowledge claims;

• (A) = formal and effective authority;

• (J) = incentives and strategic constraints;

• (D) = distribution of benefit, burden, exposure, and risk;

• (C) = criticism, appeal, review, correction, and remedy structure.

An SDS need not be centralized or fully documented. The actor defining the target may differ from the actor setting the threshold. The beneficiary may differ from the risk bearer. The person visibly responsible may lack implementation power. A review path may exist without an effective route to correction.

What makes the configuration a structure is patterned or rule-governed dependence among its components. The dependence may be observed across repeated cases or instantiated in a single consequential decision governed by a stable procedure.

2.1 Bounding the object

“The labor market,” “healthcare,” “the university,” and “government” are not yet SDS units. Each contains many decision structures with different purposes, actors, evidence rules, authorities, burdens, and correction paths.

A bounded SDS may concern:

• screening applicants for a specified role class;

• prioritizing requests for a medical intervention;

• assessing and contesting public-benefit eligibility;

• authorizing a model release;

• or delegating external transactions to an autonomous system.

The correct boundary is the smallest structure that is causally sufficient to represent the decision problem.

If the object is too narrow, the analysis omits the mechanism. If it is too broad, causal claims and governance consequences become indeterminate.

2.2 Prior is relational

A Social Decision Structure becomes a Prior Social Decision Structure only relative to a specified AI integration.

“Prior” does not mean pre-digital or historically ancient. Relative causal precedence is supported where the relevant mechanism is shown through at least one of the following:

1. documented existence before the specified integration;

2. contemporaneous operation in a comparable setting in which the specified system is absent;

3. causal continuity from a pre-integration predecessor that remains identifiable after integration.

This permits limited co-evolution without treating every post-integration structure as inherited. An institution may have been shaped by earlier automation and still be prior relative to a new system where the relevant category, authority relation, burden allocation, or correction path can be traced to an independently operating predecessor.

Historical precedence alone is insufficient. An old category may be irrelevant to the present system. A long-standing inequality may fail to explain the specified outcome. Prior status is causal and relational, not merely chronological.

2.3 Prior does not mean defective

A PSDS may contain expertise, rights, evidentiary safeguards, professional standards, separation of powers, and meaningful correction. AI may preserve or strengthen these functions.

The framework is diagnostic rather than automatically accusatory. It must identify beneficial inheritance as well as harm.

2.4 Contestation belongs inside the structure

Actors may disagree over purpose, need, evidence, burden, authority, or correction. Such disagreement is not a reason to omit the structure. It is one of its properties.

A mature formulation records rival accounts rather than allowing one institutional formulation to become the technical default without being identified as contested.

3. The Social Decision Inheritance Instance

A PSDS is not automatically inherited by every AI system operating in the same institution.

In this article, inheritance is used functionally. It denotes the carrying forward of a prior decision structure into an AI-mediated regime through representation, preference, evaluation, category, workflow, authority, or correction mechanisms. It does not imply biological inheritance, conscious acquisition, or exact copying.

The primary unit of analysis is therefore the Social Decision Inheritance Instance:

​

j=(s,a,d,T)

​

where:

• (s) = a specified PSDS;

• (a) = a specified AI system, workflow, or configuration;

• (d) = decision domain and evaluative purpose;

• (T) = period of analysis.

The same model may participate in different instances. A language model used for private drafting enters a different decision structure from the same model used to rank applicants, deny access, or trigger external transactions.

An affirmative SDII classification requires:

​

SDIIj⇔Rs∧ Bs∧ Hs,a∧ Gj

​

where:

• (R_s) = relative causal precedence or independently instantiated causal continuity;

• (B_s) = sufficiently bounded decision structure;

• (H_{s,a}) = supported inheritance or activation channel;

• (G_j) = material governance relevance.

3.1 Relative causal precedence — (R)

Evidence may include pre-integration policies, historical records, prior workflows, earlier categories, organizational incentives, professional standards, or repeated behavior under non-AI procedures.

The evidentiary burden should track the scope of the claim.

3.2 Bounded structure — (B)

At minimum, the analysis should identify:

• decision object;

• principal interested parties;

• purpose or need claim;

• categories or evidence rules;

• authority;

• incentives or strategic constraints;

• material benefit–burden distribution;

• correction path.

Unknown components must remain unresolved rather than silently inferred.

3.3 Supported channel — (H)

The analysis must connect the prior structure to the system through evidence of transfer, selection, activation, constraint, operational dependence, or correctional continuity.

Resemblance is insufficient.

3.4 Material relevance — (G)

The inherited or activated element must affect at least one of:

• problem definition;

• target or proxy;

• evidence interpretation;

• threshold;

• burden distribution;

• authority;

• correction capacity;

• permission consequence.

3.5 Candidate versus affirmative classification

A Candidate SDII is a bounded case under investigation in which one or more of (R,B,H,G) remain Partially Supported, Contested, or Unobservable.

An Affirmative SDII Classification requires all four conditions to be Supported.

This distinction permits staged inquiry without weakening the canonical conjunction.

4. Canonical Conditions and Exclusion Rules

The four conditions are intentionally restrictive.

The framework rejects the following shortcuts.

Human-generated training data are not sufficient

Human-generated data establish broad social dependence, not a bounded inheritance instance.

Similar output is not sufficient

A model may resemble an institutional position because of general prevalence, prompt context, retrieval, safety policy, independent reasoning, or another source.

Historical injustice is not sufficient

History may motivate investigation but does not establish present transmission.

Disparate outcome is not sufficient

A disparity may reflect data, target, population, sampling, policy, model behavior, workflow, or several interacting causes.

Institutional use is not sufficient

Organizational location does not prove inheritance of the organization’s complete decision structure.

Human oversight is not sufficient

A human may preserve context or reproduce the same category and incentive structure. Presence does not establish authority.

Feedback is not sufficient

Input may be accepted without any path to regime-changing correction.

Theoretical fit is not sufficient

Correspondence with the Central Equilibrium Problem (CEP), path dependence, institutional theory, or another interpretive framework does not replace evidence of (R,B,H,G).

The framework must permit negative results:

• no relevant PSDS;

• no supported channel;

• no material consequence;

• failure primarily AI-native;

• or insufficient evidence.

Without such results, inheritance becomes non-falsifiable.

5. The Risk-Origin Profile

A material governance outcome may contain several causal contributions simultaneously. ROP must therefore be indexed to a specified adverse, protective, or mixed outcome (Y_j); it is not a free-floating description of the system. Where ROP is attached to an affirmative SDII classification, (Y_j) must be the same material outcome or governance consequence used to assess (G_j).

The Risk-Origin Profile is:

​

ROPj=(Ij,Ej,Nj;Uj)

​

where:

• (I_j) = inherited causal contribution;

• (E_j) = interaction-emergent causal contribution;

• (N_j) = AI-native causal contribution;

• (U_j) = unresolved epistemic remainder concerning those contributions.

The semicolon separates causal origin from epistemic uncertainty. (U) is not a fourth cause.

In an Affirmative SDII Classification, the inherited component (I) must be Supported by construction because (R,B,H,G) are all Supported. ROP may also be constructed provisionally for Candidate or negative-control cases, in which (I) may remain Partial, Contested, Unobservable, or Unsupported.

Each causal component may be coded as:

• Supported;

• Partially Supported;

• Contested;

• Unobservable;

• Unsupported.

Each supported component should also record valence relative to (Y_j): risk-increasing, risk-reducing, mixed, or unresolved. This permits the framework to represent inherited safeguards and other beneficial transmission without redefining every inherited contribution as harm. The components are not assumed to be additive, statistically independent, or exhaustive beyond the explicitly represented epistemic remainder.

5.1 Inherited contribution

An inherited contribution exists where a causally relevant element of the PSDS enters or constrains the AI-mediated regime through a supported channel.

Possible elements include target, proxy, administrative category, evidentiary convention, hierarchy of expertise, burden allocation, or correction weakness.

Inheritance does not require exact copying. It requires causal continuity.

5.2 Interaction-emergent contribution

An interaction-emergent contribution arises where the joint configuration of a social mechanism and a technical mechanism changes the probability, magnitude, persistence, or form of an outcome beyond their separately observed effects.

Examples include:

• institutional ranking preferences combined with machine-scale processing;

• weak appeal combined with rapid automated denial;

• professional deference combined with fluent model-generated explanation;

• ambiguous policy combined with consistent high-frequency enforcement.

5.3 AI-native contribution

An AI-native contribution depends materially on technical properties of the specified system and cannot be explained adequately by the prior structure alone.

Examples may include prompt injection, adversarial context manipulation, recursive tool-use failure, hidden-state or memory interaction, synthetic-data feedback contamination, model-specific distribution shift, or machine-speed action cascades.

AI-native does not mean socially uncaused, historically unprecedented, or without analogues in earlier software. It means that, in the analyzed instance, the operative mechanism materially depends on properties of the specified AI configuration and is not reducible to the PSDS.

5.4 Unresolved epistemic remainder

Uncertainty may arise from unavailable data, proprietary systems, conflicting causal accounts, undocumented workflow adaptations, changing versions, or several plausible mechanisms.

A high (U) may justify downstream caution, but caution must not be presented as causal proof.

5.5 The profile is not blame allocation

ROP allocates causal attention. Responsibility, legitimacy, liability, and remedy require additional analysis.

6. Causal-Level Misclassification

Causal-Level Misclassification occurs when an intervention omits one or more causally required levels or expands to a level lacking causal support or an independently declared justification.

Let:

• (\mathcal{L}^{supported}_j) denote the levels at which causal evidence supports a material contribution;

• (\mathcal{L}^{causal}_j\subseteq\mathcal{L}^{supported}_j) denote the causally supported levels a declared correction rule identifies as necessary;

• (\mathcal{L}^{independent}_j) denote additional levels authorized by an independently stated legal, precautionary, rights-protecting, emergency, or democratically authorized public-policy justification;

• (\mathcal{L}^{actual}_j) denote the proposed or implemented intervention levels.

Causal evidence does not determine intervention mechanically. Severity, reversibility, feasibility, rights, and authority inform the derivation of (\mathcal{L}^{causal}_j). Independent grounds may authorize additional intervention levels but do not convert those levels into causal findings.

Causal underreach is:

​

CLMunderj=1⇔ℒcausalj⊄ℒactualj

​

Unsupported expansion is:

​

CLMoverj=1⇔ℒactualj∖(ℒsupportedj∪ℒindependentj)≠∅

​

Failure to implement a separately binding legal or public-policy obligation is a governance-compliance failure. It should not be mislabeled as CLM unless a causally required level is also omitted.

CLM is a cross-test. It is not a stage in the canonical architecture.

6.1 Model-local reduction

Governance treats the model as the sufficient causal object even though the profile supports an inherited or interaction-emergent mechanism outside it.

6.2 Social reduction

Governance treats a technically distinctive failure only as a manifestation of social or institutional structure.

6.3 Interaction blindness

Social and technical components are assessed separately even though the failure arises through their coupling.

6.4 Intervention overreach

A local or partial causal claim is used to justify a wider intervention than the evidence supports.

The governing rule is:

Intervention scope should cover causally required levels and may include independently justified levels without presenting them as causal findings.

CLM survives as an independent construct only if it improves classification, predicts recurrence or burden migration, supports better intervention selection, or strengthens traceability. Otherwise, it should be reduced to an application of established concepts such as abstraction failure or problem misformulation.

Part I Conclusion

Part I defines three linked objects.

A Social Decision Structure specifies the bounded decision regime. An affirmative Social Decision Inheritance Instance requires:

​

R∧ B∧ H∧ G

​

The Risk-Origin Profile then distinguishes inherited, interaction-emergent, and AI-native contributions while preserving unresolved epistemic remainder:

​

ROP=(I,E,N;U)

​

Causal-Level Misclassification tests whether intervention covers the required causal and independently justified scope. The full upstream handoff is developed in Part III.

Part II — Upstream Foundations

Back to contents

7. Prior Structure and Governance Intelligibility

The canonical theory assumes that a decision structure can be identified, bounded, and transferred into an AI problem model. The Prior Structure Principle explains why such a structure is necessary (Dunavich, 2026f, 2026h):

​

Input does not explain itself.

​

A signal does not identify its own meaning. A metric does not state which mechanism it represents. A category does not justify its boundary. An audit trail does not determine whether the recorded sequence is normal, anomalous, harmful, or causally relevant.

The governance analogue is:

​

Governance Information not⇒Governance Knowledge

​

For this article:

• governance information is a recorded item such as a score, warning, incident, complaint, benchmark result, or audit event;

• governance knowledge is information organized sufficiently to support causal and operational judgment;

• governance load is information that consumes attention while its place in the decision structure remains unclear.

A disparity can indicate a model defect, an invalid target, a sampling failure, an inherited proxy, an interaction effect, or several causes. A human override can indicate successful governance, reviewer inconsistency, missing context, weak model performance, or symbolic oversight. The event becomes governance knowledge only when connected to purpose, mechanism, affected actors, evidence, authority, burden, correction, and consequence.

This explains why an organization can accumulate controls, reports, dashboards, and review procedures without gaining equivalent decision clarity.

7.1 Non-circularity

The task, categories, acceptable-answer criteria, risk language, and review procedures of an AI system may derive from the institution that later evaluates it. Governance based only on those artifacts risks circular confirmation:

​

Institutional Assumption
arrow
System Design
arrow
System Output
arrow
Institutional Evaluation
arrow
Confirmation

​

A PSDS formulation provides an external reference relative to the technical task. It is not an absolute view from outside society; it is a defeasible account designed to prevent the task from serving as the sole source of its own legitimacy.

7.2 Governance artifacts require interpretation

Metrics require a specified mechanism, baseline, burden, authority, and consequence. Thresholds require rationale and a declared decision rule. Human review requires evidence access, competence, time, independence, and implementation power. Audits can document what occurred without establishing whether the governed requirement was adequate. Explanation can reduce opacity without enabling correction. Feedback becomes governance only where valid input can alter the relevant state.

7.3 Structured incompleteness

Prior structure does not require final knowledge. It requires structured incompleteness: explicit separation of what is defined, observed, inferred, disputed, unknown, and potentially falsifying.

The contrast is:

​

Reviewable, Defeasible Structure
versus
Unexamined Default

​

The Prior Structure Principle supplies an epistemic rationale. It does not establish PSDS, transmission, materiality, or causal origin; those remain empirical questions governed by (R,B,H,G).

8. Universal Reason and the Reason–Realization Gap

Contested social problems raise a second question: if actors occupy different positions and no observer possesses complete knowledge, why assume that comparative problem formulation is possible (Dunavich, 2026h)?

This article adopts a bounded premise:

​

Human beings possess shared capacities for structured rational inquiry, but those capacities are realized unevenly under actual decision conditions.

​

Universal rational potential does not imply identical conclusions, equal competence, or a single uncontested political solution. It denotes capacities to distinguish claim from evidence, identify contradiction, compare alternatives, revise beliefs, separate means from ends, and ask whether a category or decision is justified.

The Reason–Realization Gap is the distance between such potential and the understanding realized under conditions of unequal evidence, knowledge, language, time, authority, incentives, institutional position, and cognitive load.

The gap is not represented here as a scalar distance. No commensurable measure of ideal and realized understanding is assumed. It is a structured diagnostic relation among the evidence, knowledge, language, time, authority, incentives, and institutional access available to actor (i) at time (t).

The gap has two dimensions:

1. Cognitive: actors differ in factual knowledge, abstraction, causal reasoning, uncertainty recognition, and access to the relevant problem level.

2. Institutional: decision structures determine whose knowledge is recorded, whose evidence is admissible, who has time to investigate, who is protected when dissenting, and who can alter the governing rule.

The second dimension turns epistemic inequality into a governance problem. A team may identify a risk but lack authority to delay deployment. An affected person may understand the harm but lack evidence access. A reviewer may detect a repeated pattern but be authorized only to decide individual cases.

Governance is not a substitute for reason. It is an institutional attempt to preserve rational distinctions under imperfect conditions:

• signal versus evidence;

• model accuracy versus task legitimacy;

• explanation versus correction;

• formal authority versus effective control;

• uncertainty versus permission;

• procedure versus substantive adequacy.

The claim boundary is strict. The article does not assert that reason eliminates value conflict or that one actor occupies an uncontested rational position. It claims only that shared capacities for rational distinction make comparative formulation possible, while unequal realization explains the need for evidence, review, contestation, and correction.

Universal reason is not constitutive of SDII, ROP, CLM, or PSDFC. The classifier can also be justified under a weaker procedural fallibilism in which rival formulations remain comparable, evidence-sensitive, and revisable. This section supplies a philosophical rationale rather than a required empirical premise.

9. Authority, Evidence, Incentives, and Correction Before AI

Before AI integration, a decision domain already contains rules about:

• who defines the problem;

• what counts as evidence;

• which objectives are legitimate;

• who acts;

• who bears cost;

• and whether criticism can change the regime.

These rules may be formal or informal.

9.1 Formal and effective authority

Formal authority is assigned by law, policy, role, or contract. Effective authority is the practical capacity to alter the decision or system state.

​

Aformal≠Aeffective

​

A committee may possess formal approval power while leadership controls whether delay is tolerated. A reviewer may possess nominal override authority while workload and interface design make override costly. A regulator may possess legal authority while the institution controls the technical evidence needed for intervention.

9.2 Evidentiary and category authority

A regime determines what may be submitted, what is credible, who may interpret it, and what is sufficient to reopen a decision.

It also determines categories: what cases are comparable, which differences matter, what counts as normal, and which deviations become visible.

Categories are necessary. The governance problem begins when a category conceals a disputed purpose, substitutes a proxy for a need, transfers burden invisibly, or becomes resistant to counterevidence.

9.3 Incentives and local rationality

Actors operate under legitimate but potentially conflicting incentives: speed, cost, compliance, safety, throughput, reputation, or workload. A system-level failure may arise even when each actor behaves locally rationally.

The relevant object is not individual motive but the incentive structure connecting positions.

9.4 Responsibility and control

Let (V_i) denote visible responsibility and (C_i) effective control.

Symbolic accountability may exist where:

​

Vi>Ci

​

Unaccountable authority may exist where:

​

Ci>Vi

​

AI integration can widen either gap by distributing design, operation, verification, risk acceptance, and implementation across different actors (Dunavich, 2026g).

9.5 Criticism, correction, and soft closure

Criticism is an objection, anomaly, challenge, harm report, or counterevidence.

Correction is a change in outcome, rule, evidence treatment, category, threshold, authority, workflow, permission, or resource allocation caused by criticism or evidence judged valid under an explicit and reviewable standard. “Valid” must not mean merely accepted by the institution whose structure is under examination.

​

Criticism not⇒Correction

​

A minimum correction path is:

​

Cr=(Tg,St,Ea,Cr,Jd,Ga,Im,Rm,Lt)

​

where:

• (Tg) = trigger recognition;

• (St) = standing;

• (Ea) = evidence access;

• (Cr) = competent review;

• (Jd) = judgment;

• (Ga) = state-changing authority;

• (Im) = implementation;

• (Rm) = remedy;

• (Lt) = latency.

A regime is in soft closure where input is formally permitted but lacks a reliable route to structural consequence (Dunavich, 2026i).

Correction is time-sensitive. A correct appeal completed after the relevant employment, treatment, or deployment opportunity has closed may be procedurally valid but operationally ineffective.

10. ADM/CIV and the Distribution of Benefit, Burden, and Corrective Capacity

ADM/CIV maps functional positions inside the decision structure (Dunavich, 2026a).

ADM denotes the administrative, managerial, institutional, organizational, regulatory, or governing position. It commonly defines problems, controls procedure, allocates resources, manages risk, sets criteria, and holds authority.

CIV denotes the civil, exposed, dependent, human-facing, or cost-bearing position. It commonly experiences decisions, absorbs errors, supplies evidence, seeks access, requests explanation, appeals, and requires correction.

The distinction is relational, not moral.

For actor (i) in instance (j):

​

Pos(i,j)∈{ADM,CIV,Mixed,Unresolved}

​

The same actor may occupy different positions in different relations.

10.1 Purpose orientation

An AI-mediated system may be:

• ADM-first: serving throughput, standardization, institutional risk, cost, or administrative control;

• CIV-first: serving access, recognition, explanation, protection, correction, or remedy;

• mixed: legitimately serving both.

The governance problem is not mixture but concealed substitution.

A system may be CIV-facing but ADM-serving: it interacts with applicants, patients, customers, or citizens while principally reducing institutional burden or controlling access.

10.2 Benefit–burden profile

For actor (i) during period (T):

​

Λi(T)=(Bi,Wi,Qi,Di,Ki,Ai,Ci,Xi)

​

where:

• (B_i) = benefit;

• (W_i) = work or cognitive burden;

• (Q_i) = risk exposure;

• (D_i) = delay or access burden;

• (K_i) = knowledge and evidence access;

• (A_i) = effective authority;

• (C_i) = correction capacity;

• (X_i) = exit or dependency cost.

Aggregate performance may conceal that benefit accrues to ADM while verification, delay, risk, or correction work shifts to CIV.

Not every burden is unjustified. The question is whether burden is visible, attributable, proportional, compared with a credible baseline, and connected to correction authority.

10.3 Sovereignty layers

• Decision sovereignty: control over purpose, target, category, evidence, metric, threshold, and available decision states.

• Correction sovereignty: control over whether criticism or harm becomes a valid correction signal.

• Public-grammar sovereignty: control over the language through which the system appears neutral, objective, safe, efficient, fair, or merely assistive.

10.4 Civil corrective capacity

Civil corrective capacity is the effective ability of CIV-originating evidence or criticism to reach competent review and regime-changing authority, produce implementation, and, where appropriate, remedy.

It does not grant unlimited veto. It requires a meaningful path through which justified contestation can become operative.

ADM/CIV does not replace ROP.

ROP asks what caused the failure.

ADM/CIV asks who defines, benefits, bears, knows, controls, and corrects.

Part III — Transmission into AI-Mediated Decision Regimes

Back to contents

11. Provisional Inheritance and Activation Channels

A channel is a supported mechanism through which an element of a PSDS becomes represented, rewarded, evaluated, operationalized, authorized, or preserved in a specified AI-mediated regime.

A positive channel claim requires more than correspondence between a prior element and a system outcome. It requires a transfer or activation mechanism and an observable AI-mediated manifestation. Material governance relevance is assessed separately under condition (G); it is not part of the definition of (H).

The provisional taxonomy is:

​

ℋ={H1,H2,H3,H4,H5,H6,H7}

​

H1 — Corpus and Representation

The prior element enters a learned or retrieved representation through training data, fine-tuning data, retrieval collections, labels, demonstrations, or institutional documents.

Corpus presence establishes availability, not supported transmission. Stronger evidence must connect representation to an AI-mediated behavioral or representational manifestation; condition (G) separately assesses whether that manifestation is materially governance-relevant.

H2 — Preference and Alignment

The prior element is selected, rewarded, prohibited, or preferred through human preference data, reward signals, behavioral policies, evaluator instructions, or acceptable-answer rules.

The claim is not that the model possesses evaluator ideology. It is that a specified preference regime selects response patterns that reproduce an identifiable institutional distinction, norm, or asymmetry.

H3 — Evaluation and Benchmark

The prior element determines what counts as success, failure, safety, relevance, or acceptable uncertainty through benchmark construction, score aggregation, risk taxonomy, severity rules, or acceptance criteria.

The channel is especially important where the same prior definition shapes the task, optimization, and evaluation, allowing the evaluation to confirm its own premise.

H4 — Category and Public Grammar

The prior element defines the labels, proxies, schemas, taxonomies, or justificatory language through which the regime represents its objects and presents itself as neutral, efficient, fair, objective, scientific, or assistive.

The channel is material where language or category changes reliance, evidentiary burden, authority, or contestability.

H5 — Workflow and Deployment

The system’s position in a workflow turns output into a consequential event. The same model may be exploratory in one setting and practically determinative in another.

Workflow determines who sees the output, when it appears, which alternatives remain, and whether it is treated as advisory, presumptive, or binding.

H6 — Authority and Permission

Policy, access privileges, organizational custom, interface design, or delegated action grant outputs practical power to classify, prioritize, deny, approve, transact, modify records, or trigger external processes.

Capability alone is not authority. The channel requires recognized decision status and operational consequence.

H7 — Feedback and Correction

The prior regime’s standing, evidence access, review, authority, implementation, latency, and remedy structure is preserved, weakened, or improved in the AI-mediated process.

Feedback is not correction unless valid input can change the relevant state.

11.1 Channel boundary rules

Table 1. Provisional inheritance and activation channels.