---
corpus_id: WORK-000003
title: "The Architecture of Relevance"
slug: the-architecture-of-relevance
date_published: 2026-08-04T05:18:41.337+00:00
date_updated: 2026-08-04T05:18:41.337+00:00
status: stable
revision: 5
type: Treatise
concepts: ["Justified Priority", "Machine Answerability", "Attention Scarcity", "Contestability", "Answerability", "Institutional Ethics", "Relevance", "Public Justification", "Epistemic Frame", "Affected-Party Standing", "Relevance Theory", "Institutional Power", "Refutation", "Frame Problem", "Warrant", "Relation", "Accountability", "Public Reason", "Institutional Design", "Philosophy of Technology", "AI Governance", "Cognitive Effects", "Attention", "Normativity", "Philosophy of Mind", "Revision", "Institutional Frame", "Consequence", "Processing Effort", "Decision Theory", "Philosophy of Action", "Axiology", "Algorithmic Decision-Making", "Epistemology", "Frame Selection", "Administrative Governance", "Pragmatics", "Salience", "Architecture of Relevance", "Responsibility", "Future of Knowledge", "Artificial Polymaths", "Liability", "Artificial Intelligence", "Cross-Domain Intelligence", "Relevance Ledger", "Machine Intelligence", "Decision Frame", "Cognitive Relevance", "Relevance Ethics", "Standing", "Horizon of Value", "Communicative Relevance", "Information Abundance", "Governance", "Political Philosophy", "Human-Machine Cognition", "Relevance Governance"]
rights: "All rights reserved unless otherwise stated."
canonical_url: https://jabran.com/writings/the-architecture-of-relevance
identifiers:
  doi: 
  ark: 
  persistent: 
  external_archive: 
manuscript_status: none
tombstone_status: active
citation: "Jabran I. Chaudry. \"The Architecture of Relevance.\" Jabran.com, 2026-08-04T05:18:41.337+00:00. Revision 5."
---

# The Architecture of Relevance

> Intelligence does not become wisdom by generating more possibilities. It becomes wisdom only when relevance is fixed under value, consequence, and answerability. This work argues that relevance is the hidden architecture beneath intelligence, ethics, governance, and artificial intelligence.

A system may detect relations, produce explanations, retrieve information, rank outputs, predict futures, and generate plausible structure; yet none of these capacities, by themselves, determine what matters. Relevance is not a property of information alone. It is the justified priority of a relation under a horizon of value, consequence, and answerability.

The age of artificial intelligence intensifies the problem because machines can now generate relations faster than human institutions can evaluate them. The danger is not only machine error. It is also machine abundance: too many plausible connections, too many summaries, too many rankings, too many futures, too many recommendations, and too many technically correct answers inside malformed frames.

Earlier theories of relevance have explained relevance in cognition, communication, inference, attention, or information processing. This work asks a different question: when intelligence can generate unlimited plausible relations, what makes any relation worthy of attention, trust, action, or institutional force?

The work develops Relevance Ethics as the study of how minds, institutions, and machines decide what matters under conditions of excess possibility. It distinguishes relation from relevance, salience from warrant, attention from justification, auditability from answerability, and correctness from judgment. It argues that artificial intelligence cannot be governed only at the level of outputs, because outputs are late-stage surfaces of prior relevance architectures. To govern intelligence is first to govern salience.

The future will not belong to those who know everything, but to those who can still decide what matters without surrendering answerability for the decision.

# The Architecture of Relevance

## Abstract

Intelligence does not become wisdom by generating more possibilities. It becomes wisdom only when relevance is fixed under value, consequence, and answerability. This work argues that relevance is the hidden architecture beneath intelligence, ethics, governance, and artificial intelligence. A system may detect relations, produce explanations, retrieve information, rank outputs, predict futures, and generate plausible structure; yet none of these capacities, by themselves, determine what matters. Relevance is not a property of information alone. It is the justified priority of a relation under a horizon of value, consequence, and answerability.

The age of artificial intelligence intensifies the problem because machines can now generate relations faster than human institutions can evaluate them. The danger is not only machine error. It is also machine abundance: too many plausible connections, too many summaries, too many rankings, too many futures, too many recommendations, and too many technically correct answers inside malformed frames. Earlier theories of relevance have explained relevance in cognition, communication, inference, attention, or information processing. This work asks a different question: when intelligence can generate unlimited plausible relations, what makes any relation worthy of attention, trust, action, or institutional force?

The work develops Relevance Ethics as the study of how minds, institutions, and machines decide what matters under conditions of excess possibility. It distinguishes relation from relevance, salience from warrant, attention from justification, auditability from answerability, and correctness from judgment. It argues that artificial intelligence cannot be governed only at the level of outputs, because outputs are late-stage surfaces of prior relevance architectures. To govern intelligence is first to govern salience. The future will not belong to those who know everything, but to those who can still decide what matters without surrendering answerability for the decision.

---

## Opening Note: The Problem Beneath Intelligence

*The Undivided Mind* argued that polymathy is not the accumulation of many disciplines but the perception and transport of deep structure across domains. *The Architecture of Relevance* begins where that thesis becomes dangerous: the perception of structure is not enough. Once many structures can be seen, some principle must decide which structure matters.

Intelligence is usually honored for what it can produce: answers, predictions, classifications, explanations, analogies, models, warnings, diagnoses, strategies, images, plans, proofs, and designs. The intelligent system appears superior because it can generate more relations than a weaker system. It can connect distant facts, detect hidden patterns, compress disorder into form, and produce plausible structure where the ordinary mind perceives noise.

But relation is not yet relevance.

A relation only says that one thing can be connected to another. Relevance says that the connection matters. The difference is immense. A statement may be true and still trivial. A pattern may be real and still useless. A possibility may be imaginable and still unworthy of action. A risk may be statistically small and morally decisive. A detail may appear minor and yet overturn an entire judgment. Intelligence can discover that things are connected; judgment must decide which connections deserve attention, trust, sacrifice, and consequence.

This is why the age of artificial intelligence is not only an age of intelligence. It is an age of relevance.

The crisis is not that machines may know too little. They may know too much in the wrong way. They may generate too many plausible relations, too many possible explanations, too many rankings, too many summaries, too many futures. Herbert Simon’s account of the information-rich world argued that a wealth of information creates a poverty of attention, and therefore a need to allocate attention among an overabundance of sources (Simon, “Designing Organizations for an Information-Rich World,” 1971, at 40–41). The machine age radicalizes this condition: it does not merely multiply available information; it multiplies possible structures.

A human being once had to search for relations. Now relations arrive before judgment is ready for them.

The question therefore changes. It is no longer only whether a system can reason, answer, classify, generate, retrieve, summarize, predict, or pass the tests by which intelligence is recognized. The deeper question is whether it can know what matters. And if it cannot know what matters by itself, then who decides? By what authority? Under what value? For whose benefit? At whose expense? With what right of correction? With what liability after harm?

Relevance is the hidden architecture beneath every act of intelligence. It determines what enters attention, what remains outside the frame, what is counted as evidence, what is dismissed as noise, what becomes actionable, what is merely stored, what is amplified, what is forgotten, what is treated as urgent, and what is allowed to disappear.

No mind confronts the whole world at once. Every intelligence must select. Every system must frame. Every institution must decide what counts. The doctor cannot treat every signal in the body with equal urgency. The judge cannot weigh every fact in a life as legally material. The commander cannot respond to every movement as strategically decisive. The scholar cannot pursue every citation. The citizen cannot process every crisis. The machine cannot escape selection merely because its memory is vast and its outputs are fluent.

Where there is intelligence, there is selection.

Where there is selection, there is relevance.

Where there is relevance, there is ethics.

Earlier theories of relevance have explained how relevance functions in cognition, communication, inference, attention, and information processing. Sperber and Wilson’s relevance theory, for example, makes relevance central to communication and cognition by linking relevance to cognitive effects and processing effort. AI research and philosophy have long confronted versions of the frame problem: how an intelligent system determines which facts are relevant to action. This work does not reject those traditions. It changes the level of the question.

The question here is not only how relevance is computed, inferred, communicated, or processed. The question is: when intelligence can generate unlimited plausible relations, what makes any relation worthy of attention, trust, action, or institutional force?

That is the beginning of Relevance Ethics.

Relevance Ethics is the study of how minds, institutions, and machines decide what matters under conditions of excess possibility.

The first ethical act is not action. It is attention. Before a system acts, recommends, punishes, ranks, filters, diagnoses, targets, hires, excludes, elevates, suppresses, or remembers, it must determine what matters enough to enter the field of action. What appears as a technical decision is already a moral architecture. A ranking is not innocent. A filter is not empty. A metric is not merely descriptive. A model does not merely process a world; it inherits a decision about which world is worth processing.

To govern intelligence, then, is not merely to govern outputs. It is to govern salience.

A society that regulates only the final answer arrives too late. The answer has already been shaped by what was selected as relevant, by what was excluded as irrelevant, by what was made visible to the system, by what was made legible to the institution, by what was converted into data, by what the metric rewarded, by what the interface invited, by what the archive preserved, and by what the model learned to notice.

The architecture of relevance precedes the act of judgment in the same way grammar precedes the sentence. It does not dictate every conclusion, but it determines the forms in which conclusions become possible.

A civilization can possess immense intelligence and still fail if its relevance architecture is corrupt. It may know many facts and still rank the wrong ones. It may measure continuously and still miss what matters. It may optimize successfully and still optimize the wrong object. It may become technically brilliant and morally blind. It may drown not in ignorance, but in misordered knowledge.

Information does not save a civilization. Correct relevance might.

This work therefore begins from a single claim:

**Relevance is not a property of information alone; it is the justified priority of a relation under a horizon of value, consequence, and answerability.**

A thing is relevant not merely because it is related, recent, probable, measurable, popular, searchable, or statistically predictive. It is relevant because ignoring it would deform judgment. Relevance marks the point at which information becomes answerable to value.

This is why relevance cannot be reduced to attention. Attention is the fact of being noticed. Relevance is the justification for noticing. Salience is what stands out. Relevance is what deserves to stand out. A spectacle may capture attention without deserving judgment. A hidden wound may lack salience while possessing absolute relevance. Much of modern disorder arises from confusing the visible with the important.

The machine can produce salience signals. It can rank, score, classify, retrieve, summarize, and recommend. But salience is not yet answerability. A system can say that something is important without being the kind of being that must live under the burden of that importance. It can produce the form of priority without bearing the cost of prioritization.

The human problem after artificial intelligence is therefore not to outgenerate the machine. That contest is not worth having. The human task is to preserve the threshold between possibility and importance.

The machine can multiply relations; only an answerable being can decide which relation matters.

---

## I. Information Does Not Rank Itself

Information is often spoken of as though it were inherently illuminating. The more information a system receives, the better its judgment is assumed to become. Ignorance is imagined as darkness, information as light, and intelligence as the capacity to absorb and process that light.

But information does not arrive as light alone. It arrives as excess.

A datum does not announce its own place in the hierarchy of concern. It does not say whether it is central or peripheral, urgent or decorative, decisive or misleading, actionable or inert. It enters the world as something that can be interpreted. Its rank must be assigned.

A laboratory result may be medically decisive or clinically meaningless depending on the body in which it appears. A sentence in a legal case may be irrelevant background or the hinge on which liability turns. A market signal may be noise until it becomes the first visible sign of collapse. A childhood memory may appear private until it explains the architecture of an adult life. A line of code may seem harmless until it becomes the vulnerability through which an entire system is compromised.

The information did not change. Its relevance changed because the frame changed.

This is the first law of relevance:

**Information does not contain its own rank.**

Rank emerges from relation to purpose, danger, value, timing, context, and consequence. To know a fact is not yet to know what to do with it. To retrieve a fact is not yet to understand its weight. To connect a fact to another fact is not yet to justify why that connection should govern attention.

Modern systems often confuse retrieval with judgment. Search finds. Ranking orders. Recommendation suggests. Prediction anticipates. Generation produces. But none of these operations, by itself, resolves the question of importance. They may approximate importance through behavior, probability, frequency, engagement, authority, citation, similarity, or reward. Yet every approximation smuggles in a theory of what counts.

The problem is not that ranking systems rank. They must. The problem is that ranking often hides its philosophy. It presents itself as convenience, efficiency, personalization, optimization, or relevance, while concealing the deeper decision: what kind of world is this system training us to notice?

Every intelligence lives inside a frame. A frame is not merely a boundary around perception. It is an instruction about significance. It tells the system what belongs to the situation and what can be ignored. Without a frame, intelligence cannot act; with a corrupt frame, intelligence acts wrongly with great confidence.

This is why relevance is more dangerous than falsehood.

A false statement can be corrected when recognized. A misranked truth can quietly govern an entire institution. A society can be filled with true statements and still become irrational if it loses the ability to order them. Public life does not decay only when lies spread; it decays when trivial truths displace decisive ones, when spectacle outranks consequence, when measurable signals replace moral reality, when urgency is manufactured and importance is neglected.

A fact becomes dangerous when it is true, available, and wrongly ranked.

This applies equally to machines. A machine system does not need to invent falsehood in order to distort judgment. It may simply rank the irrelevant too high, hide the decisive too low, optimize the measurable at the expense of the meaningful, or convert human value into proxy signals too narrow to contain it. The distortion is architectural before it is verbal.

A machine that answers a question has already accepted a relevance frame. What counts as context? What counts as evidence? What sources matter? What user intention is assumed? What harms are visible? What tradeoffs are permitted? What uncertainty is tolerable? What must be said? What must be withheld? What does the system treat as the problem?

The output is the final surface of a deeper selection.

This means that intelligence cannot be evaluated only by correctness. Correctness is necessary, but insufficient. A correct answer to the wrong question may be a refined form of failure. A precise solution inside a malformed frame may accelerate harm. A coherent explanation of an irrelevant relation may waste attention with elegance. Fluency may disguise mispriority.

The intelligent system must therefore be judged not only by whether its statements are accurate, but by whether its architecture of relevance is warranted.

Warrant is the missing term. A relation is possible. A claim is assertible. A prediction is computable. A recommendation is generatable. But is it warranted? Does it deserve attention? Does it justify action? Does it bear the weight the system assigns to it? Can its rank be defended before those affected by it?

Without warrant, relevance decays into salience.

Salience is what rises into attention. Relevance is what has the right to remain there.

The distinction matters because human beings are vulnerable to salience. We mistake vividness for importance, repetition for truth, novelty for depth, proximity for priority, fear for evidence, confidence for authority, speed for intelligence. Machines can inherit and amplify these vulnerabilities. They can arrange the world so that what is easiest to surface becomes what is hardest to ignore.

A civilization that cannot distinguish salience from relevance becomes governable by noise.

The great task, then, is not merely to build more intelligent systems. It is to build systems and institutions that can justify their priorities. The question is not only: what did the system find? The question is: why did this matter enough to be found, ranked, shown, recommended, acted upon, remembered, or enforced?

A mature intelligence must pass through three gates.

First: relation. What is connected?

Second: relevance. Which connection matters?

Third: answerability. Who bears responsibility for acting as though it matters?

Most machine systems are powerful at the first gate and increasingly powerful at simulating the second. They can detect connections, rank likelihoods, infer intentions, and produce salience. But the third gate cannot be crossed by computation alone unless computation becomes answerable in a sense deeper than auditability.

Auditability explains what happened. Answerability bears the consequence of what happened.

This is why human judgment does not disappear in the age of artificial intelligence. It moves to a more severe location. The human being is no longer valuable merely because he can remember more facts, calculate faster, or generate more alternatives. The machine weakens those monopolies. The human being becomes valuable where relevance must be justified under consequence.

The world does not suffer from a shortage of relations. It suffers from a crisis of relevance.

And in that crisis, the decisive question is not whether intelligence can produce more.

It is whether anything, human or machine, can still decide what must not be ignored.

---

## II. Relation Is Not Relevance

A relation is any connection, dependency, analogy, correlation, causal link, structural similarity, or inferential bridge between two or more things. Relations are the raw material of intelligence. They allow a mind to compare, infer, generalize, translate, classify, predict, and compose.

But a relation is not yet a reason.

Everything can be related to something. The existence of a connection does not establish its importance. One can relate a childhood memory to a political ideology, a market movement to a weather pattern, a poetic image to a neurological state, a statistical anomaly to a future catastrophe, or a facial expression to a hidden motive. Some of these relations may be profound. Some may be trivial. Some may be dangerous. The relation itself does not settle the matter.

Relation is cheap; relevance is costly.

It is cheap because connection proliferates. The world is dense with possible associations. Language makes metaphor possible. Memory makes resemblance possible. Statistics makes correlation visible. Systems theory makes dependency visible. Machine learning makes latent pattern visible. The deeper the archive and the more powerful the model, the more relations can be produced.

Relevance is costly because it demands justification. It asks which relation deserves attention, which relation should govern action, which relation survives objection, which relation bears consequence, and which relation can be defended before those affected by it.

A conspiracy theory often begins not with the absence of relations but with their overproduction. Its pathology is not that it sees nothing. It sees too much without warrant. It discovers connections everywhere and lacks the discipline that separates structure from projection. In this sense, the conspiracy theorist is not anti-intellectual because he fails to relate. He is anti-rational because he cannot rank relation under evidence, proportion, and correction.

The same danger appears at a higher technical level in machine intelligence. A model can generate a plausible analogy, an elegant explanation, or a confident synthesis. Yet plausibility is not warrant. The fact that a relation can be expressed fluently does not mean it deserves trust. Fluency is not relevance. Coherence is not priority. Connection is not justification.

The discipline of relevance begins by refusing the seduction of relation.

A relation becomes relevant only when it is ordered by a frame that can answer for its selection. In medicine, the relevant relation is not every measurable association but the one that bears on diagnosis, prognosis, or treatment. In law, the relevant fact is not every true fact but the one material to the issue. In science, the relevant observation is not every recorded signal but the one that changes the explanatory structure. In politics, the relevant event is not every crisis but the one that alters legitimacy, power, obligation, or harm.

Relevance is therefore not a synonym for relatedness. It is relatedness under justification.

The distinction is simple, but modern culture continuously erases it. Recommendation systems often present what is behaviorally related as though it were personally meaningful. Search systems present what is retrievable as though it were epistemically central. Social platforms present what is engaging as though it were important. Institutions present what is measurable as though it were real. Bureaucracies present what is reportable as though it were accountable. Markets present what is priced as though it were valuable.

These substitutions are not minor errors. They are relevance failures.

Every field has its own way of confusing relation with relevance. Scholarship confuses citation density with intellectual necessity. Politics confuses visibility with mandate. Art confuses novelty with depth. Technology confuses capability with need. Business confuses usage with value. Governance confuses compliance with answerability. The machine age does not invent these confusions; it accelerates them.

The artificial polymath is powerful because it can relate across domains. It can move from physics to poetry, from code to jurisprudence, from biology to finance, from theology to user-interface design. It can produce analogies and structures at a speed that resembles genius. But if relation is not governed by relevance, artificial polymathy becomes a machine for plausible confusion.

This is the danger at the center of machine breadth.

The human polymath, at his best, does not merely connect domains. He submits the connection to taste, discipline, risk, experience, and answerability. He asks whether the analogy clarifies or merely decorates, whether the pattern reveals or merely flatters, whether the synthesis carries responsibility or merely astonishes. The artificial polymath can imitate the movement. It cannot automatically inherit the burden.

The crucial question is not whether a machine can discover relations across domains. It can. The crucial question is whether the system that receives those relations can decide which ones should matter.

This is why relevance is the governing problem of the post-disciplinary age. Once knowledge is no longer trapped inside separate fields, the scarcity shifts. It is no longer access to relations. It is authority over relevance.

The classical academic discipline solved the problem by narrowing the frame. A field tells the scholar what counts as evidence, what counts as method, what counts as a contribution, what counts as competence, what counts as error. The discipline is a relevance machine. It protects inquiry from chaos by forbidding most possible relations from entering the argument.

Polymathy breaks that protection. Artificial intelligence breaks it further.

When many fields become mutually permeable, and when machines can produce cross-domain relations at scale, the old guardrails weaken. This creates possibility, but also danger. Without discipline, interdisciplinarity becomes collage. Without warrant, synthesis becomes hallucination. Without answerability, breadth becomes irresponsibility.

The future therefore requires not merely more polymathy, but a stricter theory of relevance.

A relation asks: can this be connected?

Relevance asks: should this connection matter?

Answerability asks: who bears the cost of treating it as though it does?

---

## III. The Frame Problem Becomes Civilizational

The frame problem in artificial intelligence began as a technical and philosophical difficulty: how can a system determine which facts remain unchanged after an action, and more broadly, which facts are relevant to deciding what to do? Its deeper meaning was always larger than its formal setting. An intelligent agent cannot evaluate the whole universe before acting. It must decide what belongs inside the frame of action and what can be ignored.

The frame is not a convenience. It is the precondition of action.

A creature that cannot frame cannot act, because every situation opens onto indefinitely many facts. To pick up a glass, one need not consider the color of every distant star, the genealogy of the table, or the chemical history of the floor. Intelligence survives by ignoring almost everything. The difficulty is not ignoring. The difficulty is ignoring correctly.

The machine age turns the frame problem outward. It is no longer only a problem for artificial agents in laboratories. It becomes a civilizational problem for institutions, platforms, governments, markets, schools, hospitals, courts, and publics. Societies must decide what counts as relevant harm, relevant evidence, relevant identity, relevant context, relevant risk, relevant consent, relevant accountability, and relevant correction.

Every institution is a frame.

A court frames reality through admissibility, jurisdiction, precedent, procedure, burden of proof, materiality, and remedy. A hospital frames reality through symptoms, diagnostics, triage, protocols, insurance codes, and clinical urgency. A school frames reality through curricula, grades, disciplines, credentials, and developmental expectations. A state frames reality through law, census, taxation, security, citizenship, and official categories. A platform frames reality through ranking, moderation, recommendation, virality, and interface design.

These frames do not merely describe the world. They decide which world becomes actionable.

A civilization fails when its frames can no longer select reality in a way that permits correction. It may still produce information. It may still generate reports, rankings, dashboards, alerts, policies, metrics, and archives. Yet if the wrong things are made institutionally prominent and the decisive things are buried, the civilization becomes intelligent in form and irrational in operation.

This is the grammar of institutional failure: reality speaks, but the frame cannot hear it.

Artificial intelligence does not solve this problem by adding more processing power. It may intensify it. A model can expand the apparent frame by bringing more information into view, but a larger frame is not automatically a better frame. More context can clarify; it can also drown. More variables can improve judgment; they can also conceal the decisive variable among thousands of plausible distractions.

The civilizational frame problem is therefore not only epistemic. It is political and ethical.

Who defines the frame? Who benefits from the frame? Who is excluded by the frame? Who can challenge the frame? What harms does the frame make visible? What harms does it convert into noise? What kinds of persons become legible to the system? What kinds of suffering remain illegible because they do not fit the categories by which the institution sees?

A relevance architecture always has politics.

This is visible in AI governance. A system may be evaluated for accuracy while the relevant harm is dignity. It may be evaluated for bias while the relevant harm is surveillance. It may be evaluated for privacy while the relevant harm is dependency. It may be evaluated for transparency while the relevant harm is institutional automation without appeal. It may be evaluated for safety while the relevant harm is the quiet transfer of judgment from accountable persons to unanswerable systems.

An evaluation regime is itself a relevance regime.

To ask whether an AI system performs well is already to assume a frame: performs well at what, for whom, under what cost, with what failure modes, with what recourse, and under whose authority? The technical benchmark appears neutral only because its relevance architecture is hidden.

The frame problem becomes civilizational when entire societies outsource more judgment to systems whose relevance criteria are difficult to inspect, contest, or correct. The question is no longer only whether a model can represent the world. The question is whether the institutions deploying the model can answer for the world it has been trained to notice.

A civilization that cannot govern its frames will be governed by them.

---

## IV. Attention, Salience, and Warrant

Attention is the gateway through which reality enters judgment. But attention is not neutral. It is selective, limited, trained, exhausted, captured, and bought. What a person notices is partly biological, partly cultural, partly technological, partly institutional, and partly moral.

Consciousness is selective. To attend is to choose one part of the world for intensified presence. But in the machine age, attention is no longer only a psychological fact. It is an infrastructural target. Platforms compete for it. Interfaces shape it. Markets price it. States influence it. Models predict it. Institutions operationalize it.

Attention has become a battlefield for relevance.

Yet attention is not relevance. Attention is what receives notice. Salience is what stands out. Relevance is what deserves notice. Warrant is the justification that allows salience to become legitimate attention or action.

The distinction is severe.

Salience can be manufactured. Relevance must be justified.

A spectacle is salient. It may not be relevant. A hidden structural cause may be relevant. It may not be salient. An interface can make a notification salient. It cannot by that fact make it important. A model can rank an output highly. It cannot by that ranking alone make the output worthy of trust. A political machine can flood the public sphere with urgent signals. It cannot thereby create moral priority.

The crisis of attention is not only that we are distracted. It is that the world has become skilled at producing false salience.

False salience is not simply noise. Noise can be ignored when recognized. False salience mimics importance. It borrows the posture of urgency. It wears the grammar of relevance. It says: attend to me, fear me, desire me, share me, respond to me, optimize for me. Its danger is not that it lacks effect. Its danger is that it has effect without warrant.

Warrant is the missing discipline of the attention age.

A warranted relevance claim must be able to answer four questions:

Why does this matter?

To whom does it matter?

What follows if it is ignored?

Who bears responsibility for treating it as relevant?

Without those questions, attention becomes a prey drive for systems of capture.

The machine age makes warrant harder because it multiplies salience signals. A model can produce a ranked list, a risk score, a confidence estimate, a recommendation, a summary, a diagnosis, a content label, or a likelihood. Each looks like ordered attention. But order is not justification. A ranking may appear rational while concealing fragile data, proxy values, hidden exclusions, misaligned objectives, or institutional convenience.

To say that something is relevant is to make a claim about the correct ordering of attention.

That claim can be wrong.

It can be wrong because the information is false. It can be wrong because the information is true but misweighted. It can be wrong because the frame is too narrow. It can be wrong because the metric has replaced the value. It can be wrong because the person affected by the judgment has no route of contestation. It can be wrong because the system cannot distinguish what is vivid from what is decisive.

The ethics of relevance therefore requires not only transparency but contestability. It is not enough to know that a system ranked something. One must be able to ask why the ranking mattered, what it ignored, how it can be corrected, and who is answerable when the ranking harms.

Transparency shows the window.

Warrant explains why the window was built there.

A society that confuses transparency with warrant will mistake visibility for justice. It may show its processes and still misrank reality. It may reveal its data and still select the wrong harms. It may publish its procedures and still become morally blind. The question is not only whether the system can be inspected. The question is whether its relevance claims can be justified.

Salience without warrant is domination by appearance.

Attention without relevance is captivity.

Relevance without answerability is power without burden.

---

## V. Relevance Ethics

Relevance Ethics begins one step beneath moral philosophy, political theory, epistemology, and AI governance. Those fields ask what should be done, how power should be arranged, what can be known, and how systems should be constrained. Relevance Ethics asks the prior question on which each depends: **how was the field of importance selected before judgment began?**

### V.1 From Relevance to Relevance Ethics

Descriptive relevance theory explains why something becomes worth processing. Information retrieval treats relevance as fit between query and document. Machine learning operationalises relevance as contribution to predictive performance. Each is a theory of *selection*: what a system will in fact treat as pertinent.

Sperber and Wilson’s relevance theory analyzes relevance within cognition and communication. On their account, an input becomes more relevant as it produces greater positive cognitive effects for lower processing effort (Wilson and Sperber, “Relevance Theory,” in *The Handbook of Pragmatics*, ed. Laurence R. Horn and Gregory Ward, Blackwell, 2004, 607–632, at 608–609); human cognition tends toward the maximization of relevance (at 610), while ostensive communication carries a presumption of its own optimal relevance (at 612). This work inherits the recognition that relevance is selective, relational, and conditioned by cognitive economy. It asks, however, a different question. An input may be cognitively relevant, predictively useful, or communicatively salient without being entitled to govern an institutional decision. Relevance Ethics therefore begins where cognitive relevance alone becomes insufficient: with the justification of priority. Its additional categories of warrant and answerability concern why a relation may be elevated, who may contest that elevation, and who must explain or revise the consequences that follow. The distinction is not between relevance and irrelevance, but between relevance as achieved cognitive effect and relevance as justified consequential priority.

Relevance Ethics is not a competing theory of selection. It is a theory of **justified priority**. It begins where descriptive relevance ends, at the moment a selection acquires consequence for someone who did not make it. The transition has three steps.

First, selection becomes consequential: the ranking enters an action, an allocation, a refusal, a record.

Second, selection becomes attributable: some agent, office, or system stands in a position to have chosen otherwise.

Third, selection becomes contestable: there exists, or ought to exist, a party with standing to demand that the priority be justified.

Where all three hold, the question is no longer whether a factor is relevant to a model. The question is whether the priority it receives can be warranted to those who bear it.

**Relevance Ethics is the study of how minds, institutions, and machines assign justified priority under conditions of excess possibility.**

### V.2 Axioms

**Axiom I — Selection is unavoidable.** No mind, institution, or system meets the world entire. Every judgment begins after a field has been narrowed. There is no view without a cut.

**Axiom II — Every frame is an instruction of relevance.** A decision frame does not merely bound a problem; it distributes visibility. What the frame excludes cannot be weighed, contested, or corrected from inside it.

**Axiom III — Relevance requires a horizon of value.** Priority is unintelligible without a specification of whose goods, harms, standing, timescale, and constraints are in play. A relevance claim made without a horizon is not neutral; it is a horizon left unstated.

**Axiom IV — Relation is not relevance, and relevance is not warrant.** A connection may be real without deserving weight, and a weight may be assigned without being justified. Salience, prediction, and precedent each establish something less than warrant.

**Axiom V — Priority creates a burden of justification.** To rank is to expose others to the ranking. The burden falls on the party that sets the priority, not on the party that bears it.

**Axiom VI — Consequential relevance requires answerability.** Where priority produces consequences for others, someone must be positioned to explain it, bear it, and revise it. Relevance without answerability is power without burden.

**Axiom VII — Legitimate relevance must remain revisable.** A relevance architecture that cannot be corrected by what it excluded has ceased to be a judgment and become a fate.

### V.3 Corollaries

**Corollary I — Governance begins before outputs.** An output is the visible surface of prior decisions about data, labels, objectives, thresholds, and appeal.

**Corollary II — Every metric is a relevance theory.** To count is to declare what is worth counting; measurement is a normative act performed in technical vocabulary.

**Corollary III — Omission can be an action.** A harm never made legible to a system is not thereby absent; it is unaddressed by construction.

**Corollary IV — Correct outputs do not vindicate a defective frame.** Precision inside a failed frame is precision, not judgment.

**Corollary V — Scale converts private error into public architecture.** A misweighting applied once is a mistake; applied at population scale it is an institution.

**Corollary VI — The right to contest a decision must include the right to contest its frame.** Appeal confined to the application of a rule leaves the selection of the rule unexamined.

### V.4 The Relevance Ledger

The Relevance Ledger is the record by which a relevance claim becomes answerable. It is a duty rather than a document: it functions only where failure to create, preserve, disclose, or reconsider it carries an identified institutional consequence.

A ledger entry states ten things, each required by the argument rather than by convention:

1. the decision or priority;
2. the governing frame — boundaries, variables, objectives (Axiom II);
3. the affected parties and the basis of their standing;
4. the evidence included and excluded, with reasons (Corollary III);
5. the reasons for the priority — the warrant, not the relation;
6. the uncertainty acknowledged, and who bears residual risk;
7. the foreseeable consequences and their distribution;
8. the responsible authority, so that answerability is not diffused;
9. the contest mechanism, reaching the frame itself (Corollary VI);
10. the revision trigger (Axiom VII).

Two conditions follow. Decision and review authority may not be one office: an authority that reviews itself records its decisions without constraining them. The ledger becomes an answerability instrument only when its failure carries institutional consequence.

Custody, retention, escalation, disclosure, and archival format are institutional variables, not philosophical ones; they belong to the extended ledger specification, on which this argument does not depend.

### V.5 Descriptive Relevance and Ethical Relevance

One compact case shows that the two come apart.

A lending or admissions model is offered a variable — residential postcode — that is, in the population at hand, strongly predictive of the outcome the institution cares about. Descriptive relevance analysis is complete at that point: the variable improves prediction, so it is relevant to the model.

Relevance Ethics reaches a different verdict, and does so for reasons that are not reducible to accuracy. The variable carries predictive weight largely because it encodes a history the institution did not cause but would now perpetuate; the parties whose priority it sets have no standing in its selection; the burdens fall on a class distinct from the class that benefits; and the institution could not state the priority publicly, to those it affects, in terms they could contest. The relation is real. The warrant is absent.

**Descriptive relevance is not justified priority.**

This is what Relevance Ethics adds to relevance theory: not a further account of what systems select, but an account of when a selection may be imposed.

### V.6 Limits of Formalization

The framework does not decide what matters. It specifies the conditions under which a claim about what matters may be imposed on others.

Four things it cannot do automatically. It cannot rank incommensurable goods. It cannot settle contested value horizons, which are political questions that procedure can structure but not resolve. It cannot determine the correct level of description for a decision; that choice is itself a relevance claim and must be recorded as one. It cannot guarantee that a well-formed ledger accompanies a defensible decision, since form is satisfiable without substance.

What it can do is make the selection visible, attributable, contestable, and revisable — and refuse the claim that a selection made invisibly was therefore not made.

## VI. Three Worked Cases

A theory of relevance remains incomplete until it touches cases. The following examples show how Relevance Ethics works in practice.

### Case 1: medicine and the problem of clinical relevance

A medical system receives many signals: lab values, symptoms, history, imaging, medication use, family risk, insurance codes, patient speech, sensor data, and institutional protocol. The presence of information does not determine its rank. A laboratory result may be true but clinically minor. A faint symptom may be decisive. A pattern may matter in one body and not another.

Clinical judgment is therefore not the processing of information alone. It is the warranted ordering of signals under consequence.

A glucose reading, a cardiac marker, a kidney value, or an inflammatory signal does not interpret itself. The clinician must ask: does this result matter now, for this patient, under this risk, given this history, with these consequences? If an AI system assists diagnosis or triage, it may identify patterns and propose priorities. But the relevance claim still requires answerability. If a patient is deprioritized, misdiagnosed, overtreated, or ignored, the harm does not remain inside the model. It enters a body.

The relevance failure in medicine is not merely error. It is mispriority. The decisive signal is treated as background; the vivid signal is overtreated; the measurable proxy replaces the human condition.

Relevance Ethics clarifies the difference between medical information and clinical importance.

### Case 2: content moderation and the problem of public salience

A platform does not merely host speech. It ranks, recommends, suppresses, labels, amplifies, demonetizes, removes, and preserves speech. Every one of these acts depends on a relevance architecture.

The platform must decide which harms matter: harassment, misinformation, incitement, political manipulation, sexual exploitation, reputational violence, coordinated inauthentic behavior, self-harm, state propaganda, satire mistaken for threat, dissent mistaken for danger, or danger disguised as ordinary speech. No moderation system can treat all possible harms equally. It must frame.

The danger is not only that a system removes too much or too little. The deeper danger is that the platform's salience engine becomes mistaken for public importance. What receives engagement becomes visible. What becomes visible appears socially meaningful. What appears socially meaningful begins to govern public attention.

In this case, engineered public visibility can masquerade as relevance at civilizational scale.

A relevance-aware moderation system would not ask only, "Does this content violate a rule?" It would also ask:

- What harm is being ranked?
- What context is being lost?
- Who can appeal?
- What forms of speech become invisible under this frame?
- What incentives make engineered visibility profitable?
- Who is answerable when ranking transforms public reality?

Content moderation is therefore not only speech governance. It is relevance governance.

### Case 3: algorithmic hiring and the problem of proxy relevance

A hiring system may rank candidates by résumé signals, degree prestige, employment continuity, keyword match, assessment scores, inferred traits, interview analysis, referral strength, or past employee similarity. Each signal may appear relevant. Some may even improve prediction. Yet predictive usefulness is not the same as legitimate relevance.

The core question is not only: does this signal correlate with performance?

The deeper question is: **should this signal matter?**

A gap in employment may predict something in one context and conceal caregiving, illness, immigration instability, economic disruption, or exclusion in another. Degree prestige may correlate with opportunity as much as ability. Communication style may reflect class, language, disability, culture, or neurodivergence. Past employee similarity may reproduce the institution's old preferences while presenting them as objectivity.

Algorithmic hiring exposes the danger of proxy relevance. The system treats a measurable signal as though it were a legitimate substitute for human capacity.

A relevance-aware hiring system must therefore distinguish:

- predictive relation
- legitimate relevance
- contestable warrant
- institutional answerability

If the system rejects a candidate, the candidate should not face a sealed relevance architecture. The institution must be able to explain what mattered, why it mattered, what alternatives were considered, how the candidate can contest the frame, and who is answerable for the decision.

The hiring case shows the central thesis in practical form:

**Correctness without relevance is precision inside a failed frame.**

### What the cases show

Medicine shows that information must be clinically ranked.
Moderation shows that engineered visibility can capture public reality.
Hiring shows that proxy signals can become illegitimate relevance.

In each case, the same structure appears:

1. Relation is discovered.
2. Salience is produced.
3. Warrant must be established.
4. Answerability must be preserved.

This is the practical grammar of Relevance Ethics.

---

## VI-A. Salience, Attention, Relevance, Warrant, Answerability

The five terms are not synonyms, and the argument fails if they are permitted to drift.

**Salience** is prominence to a cognitive or attentional system, whether biological or formally analogous. It is a property of a stimulus in relation to a mechanism of notice. Research on attentional capture — the bottom-up saliency tradition, and the distinction between goal-directed and stimulus-driven attentional systems — describes salience as competition for processing, not as a claim about importance.

**Attention** is the allocation of finite processing resources. Simon argued that information abundance produces a scarcity of attention; an allocation of a scarce resource describes a budget, not a justification.

**Relevance** is a relation judged pertinent to a question, task, or horizon. It is already normative in the weak sense that it presupposes a question, but not yet in the strong sense that it justifies imposition.

**Warrant** is the justification for assigning priority to a relation, given a horizon of value and the standing of affected parties.

**Answerability** is the standing obligation to explain a priority, bear its consequences, and revise it under challenge.

The empirical literature secures the first two terms and constrains the third. It does not supply the fourth or fifth. That is the precise sense in which the ethics is not a redescription of the science: no measurement of prominence or resource allocation entails a licence to impose.

Three consequences follow. Engineered prominence can be produced without any corresponding increase in warrant. A system can be optimised for attention and thereby degrade relevance. And a governance regime that measures only what a system surfaces, without asking what it was licensed to surface, will confirm compliance while missing the failure.

## VII. The Artificial Polymath

The artificial polymath is the machine system capable of generating relations across many domains without possessing a unified life in which those relations must be borne. It can move from code to law, from poetry to physics, from medicine to design, from theology to economics, from governance to interface, from ancient text to contemporary policy. It can synthesize without belonging to any one discipline. It can appear learned without undergoing formation. It can speak across domains without inhabiting the consequences of its speech.

This makes the artificial polymath astonishing.

It also makes it dangerous.

Human polymathy has always been limited by biography. A human being must learn in time, remember imperfectly, suffer confusion, revise identity, and bear the consequence of judgment. The human mind cannot detach its synthesis from its life. Even when it is wrong, it is wrong as someone. It must return to the world as the same being who made the claim.

The artificial polymath lacks this continuity. It may have memory, logs, training history, or system state, but it does not possess biography in the human sense. It does not grow old under its claims. It does not lose reputation as a wound. It does not feel shame as a moral interruption. It does not experience regret as the reorganization of selfhood. It does not suffer the world that follows from its own relevance assignments.

The artificial polymath has breadth without biography, inference without liability, and structure without wounds.

This does not make it useless. It makes it incomplete.

The danger is not that artificial intelligence cannot contribute to relevance. It can. It can retrieve overlooked evidence, compare frames, expose contradictions, generate objections, detect anomalies, and produce candidate relevance structures. It can help human beings see relations that their disciplines, habits, or institutions conceal. It can assist the work of judgment.

But assistance is not answerability.

A system may produce a relevance proposal. It may not be able to bear the moral status of that proposal. A machine may say that this risk matters, this patient needs attention, this applicant is lower priority, this speech is dangerous, this route is efficient, this military target is probable, this source is authoritative, this child is at risk, this employee is suspicious, this citizen is anomalous. In each case, the machine has not merely processed information. It has participated in the ordering of significance.

The question is not whether machines should be used. The question is how relevance claims produced by machines become answerable within human institutions.

There are two bad answers.

The first bad answer is machine supremacy: the belief that sufficiently advanced computation can replace human judgment because it can process more information. This view mistakes scale for legitimacy. It assumes that the ability to generate or evaluate more relations automatically improves the authority to decide what matters.

The second bad answer is human nostalgia: the belief that human judgment is pure simply because it is human. This view ignores the long record of human prejudice, laziness, corruption, fear, tribalism, and misattention. Human beings are not automatically good relevance machines. They are vulnerable, biased, status-driven, frightened, and often governed by salience.

The task is neither to worship the machine nor romanticize the human.

The task is to design architectures in which machine breadth is subordinated to answerable relevance.

This requires a new division of labor. Machines may generate relations. Humans and institutions must justify relevance. Machines may surface candidates for attention. Human governance must establish warrant. Machines may produce explanations. Human authority must bear responsibility for action. Machines may help construct frames. Human institutions must preserve contestation and correction.

The artificial polymath should not be asked merely to answer. It should be asked to expose the relevance structure of its answer.

What did it treat as central?

What did it treat as peripheral?

What did it ignore?

What assumptions organized the frame?

What alternative relevance frames could change the answer?

What harms are visible under this frame?

What harms are invisible?

What would make the recommendation unsafe, unjustified, or misranked?

In the machine age, a good answer is not enough. We need answerable relevance.

---

## VII-A. Abundance Without Priority

Cross-domain reach intensifies the problem this treatise states. A system traversing many fields produces more candidate relations per question than any previous instrument, and the ratio of relations offered to priorities justified moves the wrong way. Abundance is not the difficulty; abundance without priority is. Access to many domains does not itself justify weight in any of them. A relation transported across a disciplinary boundary arrives with its warrant stripped: it was justified relative to a question no longer being asked. An artificial polymath, so called, is therefore not a system that retrieves across domains; retrieval widens the field, while polymathy would require judging which retrieved relations bear on the actual question, under a stated horizon of value, with someone answerable for the judgment.

The consequence is a dependency, not a doctrine. As relational abundance grows, relevance architecture becomes more load-bearing, and the four gates become the precondition for any account of cross-domain synthesis. How synthetic judgment is warranted, and who answers for it, belongs to another work.

## VIII. Institutions as Relevance Machines

Institutions are not merely organizations. They are machines for deciding what matters.

A court decides which facts matter legally. A hospital decides which symptoms matter clinically. A university decides which questions matter intellectually. A market decides which goods matter economically. A state decides which persons, territories, risks, and obligations matter administratively. A museum decides which objects matter culturally. A platform decides which signals matter socially. A school decides which forms of knowledge matter developmentally.

Every institution is a relevance machine.

Its power lies not only in what it does, but in what it makes visible as a proper object of action. An institution does not need to deny reality in order to distort it. It can simply classify reality badly. It can treat the decisive as irrelevant, the peripheral as central, the human as administrative, the moral as procedural, the measurable as real, the unmeasured as nonexistent.

Institutional failure is often relevance failure before it is operational failure.

A bureaucracy may process correctly and still misrecognize the person before it. A university may produce scholarship and still lose the hierarchy of important questions. A hospital may meet procedural requirements and still fail to see the patient as a whole. A government may collect data and still fail to notice the suffering that its categories cannot encode. A company may optimize metrics and still destroy the value that made the product worth using.

The institution becomes dangerous when its relevance architecture becomes immune to correction.

This is why every serious institution needs not only procedures, but routes of appeal. Appeal is a relevance technology. It allows the person harmed by the frame to challenge what the frame treated as irrelevant. It says: the system saw something wrongly; the hierarchy of importance must be reopened.

A society without appeal is a society without relevance correction.

Artificial intelligence threatens appeal when its classifications become operationally authoritative while remaining difficult to contest. A person denied opportunity, flagged as risk, deprioritized for service, or evaluated by automated means may face not merely an incorrect output, but an inaccessible relevance structure. The harm is not only that the system may be wrong. The harm is that the system may be wrong in a way no one can meaningfully challenge.

An unchallengeable relevance architecture is a form of power.

The same principle applies beyond formal decision systems. Public discourse itself is an institution of relevance. It decides what is discussed, what is ignored, what is treated as scandal, what is treated as normal, what is mourned, what is mocked, what is elevated, what is forgotten. A public sphere can become diseased not because it lacks speech, but because it loses the ability to rank importance.

The disease of the public sphere is not silence alone. It is misattention.

When spectacle outranks consequence, when humiliation outranks truth, when novelty outranks memory, when engagement outranks judgment, when outrage outranks repair, the society still speaks but no longer listens to reality in the right order.

A civilization does not collapse from ignorance alone. It collapses when it can no longer tell which knowledge matters.

Institutions survive by correction. Correction requires that relevance remain contestable. The most dangerous institution is not the one that makes mistakes. Every institution does. The most dangerous institution is the one that cannot perceive its mistakes as relevant.

---

## IX. AI Governance as Relevance Governance

Most AI governance begins too late.

It begins with the output: Was the answer biased? Was the decision explainable? Was the model accurate? Was the system safe? Was privacy preserved? Was the user informed? Was the result transparent? These questions are necessary, but they are not sufficient. By the time the output appears, a relevance architecture has already done its work.

It has selected training data.

It has defined tasks.

It has chosen labels.

It has determined metrics.

It has weighted harms.

It has shaped the interface.

It has encoded incentives.

It has chosen what counts as success.

It has decided what kind of failure matters enough to measure.

The output is not the beginning of governance. It is the visible residue of prior selection.

AI governance is therefore relevance governance.

To govern AI is to ask: What has the system been taught to notice? What has it been allowed to ignore? What harms are legible to it? What human values have been reduced to proxy signals? What institutional incentives shape its ranking? What forms of appeal exist after harm? What forms of uncertainty are disclosed? What relevance claims can be contested? What happens when technically correct output produces morally wrong prioritization?

Frameworks for trustworthy AI already gesture toward this deeper problem. Risk management, validity, reliability, safety, accountability, transparency, explainability, privacy, and fairness are all relevance categories as well as governance categories. Each asks what should matter in the design, deployment, evaluation, and correction of a system.

But the categories are often treated as checklist items. Relevance Ethics asks what precedes the checklist: why these dimensions, why these harms, why these metrics, why this population, why this context, why this deployment, why this threshold?

The danger of checklist governance is that it can become procedurally correct while remaining relevance-poor.

A system can be documented and still unjustified. It can be explainable and still aimed at the wrong objective. It can be transparent and still coercive. It can be accurate and still illegitimate. It can be fair according to one metric and harmful according to a deeper social frame. It can meet a benchmark and still fail the world.

Correctness without relevance is precision inside a failed frame.

The current global turn toward AI governance proves that relevance has become institutional. AI is no longer only a research problem or product problem. It is a problem of public order. Governance frameworks, risk management systems, ethics recommendations, and legal regimes all attempt, in different languages, to answer the relevance question: which AI harms, capabilities, rights, risks, and obligations must matter?

Yet the law cannot fully solve relevance. Law formalizes relevance after political and ethical struggle. It can name high-risk systems, define prohibited practices, require documentation, demand transparency, or assign obligations. But before the law can govern a risk, the risk must become visible as a risk. Before a harm can be regulated, it must become legible as harm. Before an obligation can be enforced, some institution must decide that the obligation matters.

Law is downstream of relevance.

This does not weaken law. It clarifies its dependence. A mature AI governance regime needs statutes, standards, audits, documentation, technical evaluations, institutional oversight, and public accountability. But beneath all of them lies the prior task: the governance of relevance selection.

The crucial question for AI governance is therefore not merely, “What did the system output?”

The crucial questions are:

What did the system treat as relevant?

Why did it treat it as relevant?

Who authorized that relevance structure?

Who can challenge it?

Who is harmed if it is wrong?

Who bears responsibility after deployment?

If these questions are absent, AI governance becomes output theater: a performance of control over a system whose deeper relevance architecture remains untouched.

---

## IX-A. Institutional Analogies, Not Equivalences

Contemporary governance instruments partially instantiate features of this framework. They are not philosophical equivalents of its categories: an instrument can be satisfied while the underlying relevance defect persists.

Each row states the mechanism, its philosophical analogue, the mismatch, and the limit of the analogy. No legal conclusion, enforcement interpretation, case law, recital, quotation, or page reference is asserted, and every article-level description is confined to the applicable statutory or framework text. The comparison is bounded in advance: Article 6 only with respect to the classification of high-risk systems, Article 13 only with respect to transparency and the provision of information, Article 14 only with respect to human oversight.

| Mechanism | Philosophical analogue | Mismatch | Limit of the analogy |
|---|---|---|---|
| Regulation (EU) 2024/1689, Article 6 — classification rules for high-risk AI systems | Frame declaration: the act of fixing which systems enter a heightened regime | Classification is a *sorting* of systems, not a *justification* of the frame each system applies internally | A correctly classified system may still operate on an unwarranted internal relevance architecture |
| Regulation (EU) 2024/1689, Article 13 — transparency and provision of information to deployers for specified high-risk systems | Disclosure: making the selection inspectable | Disclosure shows what was selected; it does not show that the selection was warranted | Complete information provision is compatible with a fully disclosed and fully unjustified priority |
| Regulation (EU) 2024/1689, Article 14 — requirements concerning human oversight for specified high-risk systems | Answerability architecture: a human position in the decision path | The presence of an oversight design is not the presence of substantive answerability; an overseer without authority to contest the frame supervises applications only | Oversight may be implemented, documented, and ineffective simultaneously |
| NIST AI RMF 1.0 (NIST AI 100-1, 2023) — Govern, Map, Measure, Manage functions; version-bound, and not described here as permanently current | Relevance governance cycle | A voluntary framework structures deliberation; it does not allocate consequence | Adoption is compatible with unchanged institutional priority |

Nothing here asserts that the Regulation implements justified relevance, warrant, substantive answerability, the Relevance Ledger, or the four-gate account. Classification is not justified framing, disclosure is not warrant, oversight is not answerability, and compliance is not legitimate priority. Each mechanism can therefore be implemented procedurally while the relevance defect persists: classification without frame justification, disclosure without warrant, and oversight without authority are the three characteristic forms this framework predicts.

## X. The Human Threshold Between Possibility and Importance

The machine age humiliates certain human vanities. It weakens the claim that human beings are unique because they can calculate, remember, classify, translate, summarize, compose, generate, or recognize patterns. Many of these capacities can now be simulated or exceeded in specific domains.

But the collapse of cognitive monopoly is not the collapse of human significance.

The human does not end when the machine thinks. The human ends only when no one remains answerable for thought.

This is the threshold that must be preserved: the movement from possibility to importance.

Possibility belongs easily to machines. They can generate many alternatives, many explanations, many designs, many futures, many arguments, many images, many interpretations. Importance is different. Importance requires a world in which consequences matter to someone. It requires care, vulnerability, obligation, memory, mortality, and the possibility of being changed by what one has done.

Care and importance offer one route into this problem. What matters to a person is not merely what the person notices. It is what structures that person’s will, identity, concern, and practical life. Ethical life is not reducible to abstract procedure without losing contact with lived seriousness, and normativity stands in relation to agency and practical identity. One point follows: value is not merely computed. It is borne.

The question is not whether machines can be answerable, but which forms of answerability a system can instantiate, and which it cannot yet instantiate independently. At least six forms must be distinguished.

**Causal answerability**: the system is the traceable proximate cause of an outcome. Contemporary systems satisfy this routinely.

**Operational answerability**: the system can produce a reconstruction of how an output arose. Logging, model documentation, and interpretability work address this level. It is satisfiable without any normative capacity whatever.

**Institutional or delegated answerability**: the system occupies a position within an architecture in which some office is answerable *for* it. This is the level at which present governance operates, and it is a real form of answerability — borne by the institution, discharged in part through the system.

**Moral answerability**: the capacity to be held to account as an agent, on one's own behalf.

**Political answerability**: standing within a structure of reciprocal justification among parties who may sanction one another.

**Full normative answerability**: the conjunction of the moral and political forms, sustained over time.

The claim advanced here is deliberately narrow. It is not that machines are barred from answerability by lacking biography, mortality, or embodied vulnerability. Those are contingent features of the only answerable agents we have so far encountered, not demonstrated necessary conditions. The claim is that present systems do not independently satisfy full normative answerability, and the deficit is specifiable as a set of capacities rather than as a biological fact:

- comprehension of reasons as reasons, rather than as tokens;
- recognition of affected-party standing;
- capacity to revise commitments under challenge, not merely to update parameters under gradient;
- continuity of identity sufficient for a commitment made earlier to bind later;
- exposure to sanction or consequence that constitutes a cost to the system itself;
- authority to make binding commitments;
- participation in reciprocal justification;
- possession or representation of ends not wholly imposed by another agent.

A system may hold some of these and lack others. Answerability is therefore graded, and a system may participate in an answerability architecture without being its terminus. What must not happen is that operational answerability be accepted as a substitute for the institutional or normative forms — that a trace be read as a justification, or an audit log as a bearer of consequence.

Whether some future system satisfies the missing capacities is an open empirical and conceptual question. It is recorded in Section XII as a revision condition, not treated here as a prohibited category.

The task of the human after machine breadth is not to produce more possibilities than the machine. The task is to determine which possibilities deserve incarnation. To choose is to close worlds. To act is to make one possibility consequential and abandon others. This is why completion is morally heavier than generation. Generation opens. Judgment closes. Action binds.

The future belongs not to the mind that can produce the most possibilities, but to the mind that can bear responsibility for choosing among them.

This is not a sentimental defense of human intuition. Human intuition is often wrong. Human judgment requires discipline, evidence, institutions, criticism, and correction. But human answerability names something computation does not automatically possess: the condition of standing behind a relevance claim in a world where the consequences do not vanish when the output is produced.

The human threshold is not mystical. It is juridical, ethical, embodied, and historical.

A person can be asked why he judged something important. He can be blamed. He can revise. He can apologize. He can be punished. He can be transformed. He can lose trust. He can bear shame. He can inherit consequences across time. The continuity of personhood makes answerability possible in a way that ordinary computation does not yet reproduce.

A machine may be audited. A human can be answerable.

Auditability explains the trace. Answerability bears the burden.

The defense of human judgment in the AI age must therefore become more exact. It cannot rest on vague claims about soul, intuition, creativity, or uniqueness. It must identify the precise human function that remains irreducible in institutional life: the authority to bind relevance to consequence.

When machines multiply possibility, humans must govern importance.

When machines generate relations, humans must establish warrant.

When machines produce salience, humans must preserve answerability.

This is not a lesser role. It is the role beneath civilization.

---

## XI. Objections, Limits, and Replies

Eight objections are stated here in their strongest available form and answered in turn. Where an objection is not fully answered, the residue is stated as an unresolved limit rather than closed.

### Objection I — Relevance Ethics renames existing relevance theory

**Strongest version.** Relevance is a worked field. Sperber and Wilson supply a cognitive account, information science a retrieval account, machine learning an operational account. Renaming the field "Relevance Ethics" and appending a normative vocabulary adds no explanatory power.

**Tradition.** Pragmatics; information science; philosophy of science on theoretical parsimony.

**Reply.** The existing accounts explain selection. None supplies a criterion for when a selection may be *imposed* on a party who did not make it. Section V.5 gives a case in which the descriptive verdict and the ethical verdict diverge on a factor whose predictive standing is not in dispute. The divergence is the explanatory work.

**Qualification.** The framework claims no priority over descriptive relevance theory within its own domain, and inherits its results.

**Unresolved.** The strongest form of the objection — that the normative work is done by fairness theory and procedural justice rather than by relevance as such — is only partly met. The framework's distinctive contribution is the *location* of the normative question (before judgment, at selection) rather than a new normative principle.

### Objection II — Levels of abstraction make relevance relative, and relativity defeats normativity

**Strongest version.** On Floridi's method of levels of abstraction, what counts as relevant depends on the level of description chosen. If the level is chosen freely, relevance is relative; if relevance is relative, it cannot carry normative weight.

**Tradition.** Philosophy of information; philosophy of science on description-relativity.

**Reply.** Frame-dependence is not arbitrariness. The choice of level is itself a decision with consequences and is therefore itself subject to justification, contest, and record. The framework does not deny relativity to frames; it makes the frame an object of evaluation rather than a neutral container.

**Qualification.** The framework cannot adjudicate between levels of description from a level-independent standpoint, and does not claim to.

**Unresolved.** Whether frame-justification can proceed without covertly presupposing a privileged frame remains open. This is the same difficulty the regress objection makes explicit.

### Objection III — Attention, prediction and salience are jointly sufficient

**Strongest version.** Relevance is what a well-calibrated system attends to; warrant is whatever improves outcomes on a specified objective. Attention theory plus predictive performance plus utility exhausts the phenomenon; the further vocabulary is ornamental.

**Tradition.** Decision theory; behavioural and computational accounts of attention; instrumentalism about norms.

**Reply.** The reduction requires that the objective itself be given. Where the objective is contested — as it is in triage, moderation, hiring, and allocation — the reduction relocates the normative question rather than dissolving it. Section VI-A states the distinctions in a form that can be tested: prominence and resource allocation are measurable; licence to impose is not measured by them.

**Qualification.** Within a fixed and legitimately settled objective, the reduction largely holds, and the framework then reduces to good practice.

**Unresolved.** How much of institutional life operates under fixed and legitimately settled objectives is an empirical question this work does not settle.

### Objection IV — Functionalist agency: a sufficiently capable machine implements the whole framework

**Strongest version.** If answerability is a functional role — giving reasons, revising under challenge, bearing sanction, maintaining continuity — then a system implementing that role is answerable, and any refusal is anthropocentric prejudice.

**Tradition.** Functionalism in philosophy of mind; machine ethics; the practice of giving and asking for reasons.

**Reply.** The revised Section X accepts the functionalist framing and answers within it. Answerability is graded, and machine systems already satisfy causal and operational forms and participate in delegated forms. What present systems do not independently satisfy is the conjunction listed in Section X — comprehension of reasons as reasons, standing recognition, commitment revision, binding authority, reciprocal justification, exposure to cost. The claim is capacity-specific and empirically defeasible, not biological.

**Qualification.** The framework does not assert that these capacities are unattainable in principle.

**Unresolved.** Whether the listed capacities can be jointly satisfied by a system without something like a life is precisely what is left open.

### Objection V — Proceduralism is enough

**Strongest version.** Documentation, risk assessment, audit, transparency, and oversight already govern selection. A further philosophical layer adds cost without adding constraint.

**Tradition.** Administrative law; regulatory theory; Lessig on architecture as a modality of regulation (Lessig, *Code: Version 2.0*, 2006, 123–124).

**Reply.** Procedure inherits a frame. The governance table in IX-A names three characteristic failures — classification without frame justification, disclosure without warrant, oversight without authority — each of which is formally compliant and substantively defective. The framework's demand is narrow: that the frame itself be recorded, attributable, and contestable.

**Qualification.** Procedure is necessary; the framework is a constraint on procedure, not a replacement.

**Unresolved.** Whether frame-contest procedures materially change outcomes is an empirical question, and is stated in XII as a refutation condition.

### Objection VI — The framework conceals political disagreement in philosophical vocabulary

**Strongest version.** Disputes about what matters are disputes about interests and power. Recasting them as questions of "warrant" and "horizon of value" gives the appearance of neutral adjudication to what is in fact contested politics.

**Tradition.** Political realism; critical theory on institutional power and the opacity of consequential selection.

**Reply.** The framework does not adjudicate value disputes and says so (V.6). It requires that the horizon be stated, that standing be recognised, and that dissent be preserved rather than absorbed. Making a political disagreement explicit is not the same as resolving it, and the framework claims only the former.

**Qualification.** Where an institution controls both the frame and the contest procedure, the framework's demands are weak.

**Unresolved.** The objection is not fully answered: procedural explicitness can itself become a technique of legitimation. This is recorded as a standing limit.

### Objection VII — Relevance review produces paralysis and administrative burden

**Strongest version.** Ledgers, contest triggers, and frame review impose transaction costs on every consequential decision. At scale the regime either halts decision-making or degenerates into template completion.

**Tradition.** Regulatory cost-benefit analysis; organisational theory.

**Reply.** The duty is scaled to consequence, not applied uniformly: the ledger attaches where a decision is consequential, attributable, and contestable (V.1). Template degeneration is a real failure mode and is the reason the consequence condition in V.4 is stated as decisive — a ledger without an attached consequence for its own failure is precisely the artefact the objection describes.

**Qualification.** The framework accepts that its own machinery can be captured by the compliance dynamics it criticises.

**Unresolved.** Whether relevance review causes net harm through delay and burden is an empirical question, stated in XII as a refutation condition.

### Objection VIII — The justification of priority regresses infinitely

**Strongest version.** Every priority requires a warrant; every warrant presupposes a horizon; every horizon embeds prior priorities. Either justification is infinite, in which case no action is licensed, or it terminates arbitrarily, in which case the framework has merely relocated the arbitrariness.

**Tradition.** The epistemology of justification, the analysis of practical reasoning, and deliberative accounts of political legitimacy.

**Reply.** Four things must be held apart:

- **authority to decide** — the institutional standing to close deliberation and act;
- **justification of the decision** — the reasons that support the priority selected;
- **legitimate procedural closure** — the conditions under which stopping deliberation is defensible;
- **ultimate philosophical foundation** — a terminus in the order of reasons.

An institution may hold authority to decide without thereby making the priority it selects true, morally correct, or fully justified. Terminating the regress in institutional mandate is therefore not available: it would answer a question about justification with a fact about power.

The correct reply is that practical judgment ends through **provisional and contestable closure**, constituted by public reason-giving, recognised affected-party standing, contestability, evidential disclosure, revision procedures, allocated responsibility, exposure to consequences, and preservation of unresolved dissent. Closure of this kind is defeasible by construction: it stops the process without claiming to have completed the justification.

**Decision terminates procedurally before justification terminates philosophically.**

The framework thus rejects both the demand for complete prior justification, which licenses nothing, and the declaration that power itself creates relevance, which licenses everything.

**Qualification.** Provisional closure is a condition on legitimate action, not a guarantee of correct action.

**Unresolved.** This repair does not dissolve the regress. It relocates it: the conditions constituting legitimate closure are themselves priority claims requiring warrant. The framework asserts only that this residual circularity is shared with every account of practical justification, and that it is preferable to record it than to conceal it behind institutional authority. This remains an open limit of the theory.

---

## XII. Conditions of Failure, Revision, and Refutation

The conditions below are stated in observable or adjudicable form and classified. A limitation is not labelled a refutation condition.

#### 1. Predictive sufficiency — *empirical refutation condition*

**Defeated if:** across consequential domains, decisions selected purely by predictive contribution to a stated objective are found, on independent review by affected parties and reviewing authorities, to be no less defensible than decisions selected under warrant review — with divergence cases (V.5 type) failing to replicate.

#### 2. Successful formalization of value — *conceptual refutation condition*

**Defeated if:** a formally specified system repeatedly produces stable, publicly defensible, cross-contextually robust value judgments in contested domains, without residual appeal to unstated human interpretation at the point of application — where "residual appeal" is operationalised as any point at which a human interpreter must supply a determination the specification does not fix.

#### 3. Machine answerability — *revision trigger*

**Revised if:** a system independently satisfies the capacity set in Section X — comprehension of reasons as reasons, recognition of affected-party standing, commitment revision under challenge, continuity sufficient to bind, exposure to cost, binding authority, participation in reciprocal justification, and non-imposed ends — as adjudicated by the institutions that would have to accept its commitments. The framework's human-anchored account would then require revision, not abandonment.

#### 4. Inefficacy of frame contest — *empirical refutation condition*

**Defeated if:** in institutions that implement frame-contest procedures, measured decision outcomes, exclusion patterns, and error distributions are found not to differ materially from matched institutions that implement application-level appeal only.

#### 5. Net harm of relevance review — *empirical refutation condition*

**Defeated if:** implementation of ledger and review duties is shown to produce delay, burden, and foregone benefit exceeding the harms corrected, on measures agreed in advance with affected parties rather than selected after the fact by the implementing institution.

#### 6. Explanatory redundancy — *conceptual refutation condition*

**Defeated if:** the distinctions between salience, attention, relevance, warrant, and answerability can be shown to license no verdict that existing accounts — relevance theory, procedural justice, fairness theory, administrative law — do not already license, with the V.5 divergence case reconstructed entirely within one of them.

#### 7. Output-only governance success — *empirical refutation condition*

**Defeated if:** governance regimes that constrain outputs alone reliably prevent the class of harms the framework attributes to upstream selection, over a period sufficient for upstream defects to surface.

#### 8. Frame neutrality — *conceptual refutation condition*

**Defeated if:** decision frames can be shown to be neutral containers rather than distributions of visibility — that is, if what a frame excludes can be shown to remain equally available for contest and correction from within it.

#### 9. Compliance capture of the ledger — *institutional failure condition*

**Fails institutionally if:** ledgers are widely adopted and consistently completed while contest rates, revision rates, and disposition changes remain at or near zero. This would show the mechanism operating as documentation rather than duty. It would not refute the theory; it would defeat the instrument.

#### 10. Regress residue — *scope limitation*

The reply in XI.VIII relocates rather than dissolves the regress. The conditions constituting legitimate closure are themselves priority claims. This is a limitation on the framework's foundational completeness. It is recorded as a limit, not as a refutation condition, and not as an answered objection.

#### 11. Incommensurability — *scope limitation*

The framework does not rank incommensurable goods and does not claim to. Cases turning wholly on incommensurability fall outside its adjudicative reach.

---

## XIII. Conclusion: What Must Not Be Ignored

The age of artificial intelligence has been described as an age of automation, generation, prediction, acceleration, and intelligence. These descriptions are not wrong. They are incomplete.

The deeper transformation is relevance.

The world is entering a condition in which relations can be generated faster than they can be judged. Machines can multiply associations, summaries, analogies, futures, recommendations, and explanations. Institutions can embed those outputs into procedures. Platforms can distribute engineered visibility at planetary scale. Governments can regulate after the fact. Markets can price attention. Scholars can drown in citations. Citizens can drown in crises. The whole world can become articulate and still fail to know what matters.

This is the central danger: not silence, but misordered speech; not ignorance, but misranked knowledge; not absence of intelligence, but intelligence without relevance; not lack of answers, but answers without answerability.

A civilization does not become wise by knowing more. It becomes wise by ordering knowledge under value.

The first task of the machine age is therefore not merely to build more capable intelligence. It is to build relevance architectures that can be justified, contested, corrected, and borne. We need systems that reveal not only what they output, but what they treated as worth noticing. We need institutions that can answer not only for decisions, but for frames. We need governance that reaches beneath outputs into the selection of what is made prominent. We need human beings capable of resisting both noise and false importance. We need archives that preserve not merely texts, but the claims, concepts, passages, citations, objections, revisions, and records by which thought can remain answerable across time.

The machine can multiply relations.

It can generate.

It can retrieve.

It can rank.

It can recommend.

It can simulate the form of judgment.

But the decisive question remains: what must not be ignored?

That question cannot be surrendered cheaply. It is the question beneath ethics, politics, scholarship, design, law, medicine, art, and civilization. It is the question every institution answers whether it admits it or not. It is the question every interface encodes. It is the question every model inherits. It is the question every human life confronts when possibility exceeds time.

Relevance is the architecture by which intelligence becomes responsible to the world.

As systems acquire greater power to select what becomes visible, actionable, and consequential, governance will increasingly depend on whether those selections can be justified and contested.

---

## Formal Definitions

**Relation**
A relation is any connection, dependency, analogy, correlation, causal link, structural similarity, or inferential bridge between two or more things.

**Relevance**
Relevance is the justified priority of a relation under a horizon of value, consequence, and answerability.

**Salience**
Salience is the fact that something stands out to a mind, institution, model, or interface.

**Warrant**
Warrant is the justification that permits a relation to move from possible connection to legitimate attention or action.

**Answerability**
Answerability is the obligation and capacity to give reasons for a priority to those with legitimate standing, and to remain open to scrutiny, contest, correction, and consequence.

**Relevance Ethics**
Relevance Ethics is the study of how minds, institutions, and machines decide what matters under conditions of excess possibility.

**Artificial Polymath**
An artificial polymath is a machine system capable of generating relations across many domains without possessing a unified life in which those relations must be borne.

---

### Horizon of value

A **horizon of value** is the specified normative field within which a priority claim is made. A horizon is adequately specified only when it states, at minimum:

1. the affected entities and the standing set — who counts, and on what basis;
2. the recognised goods, harms, rights, duties, and objectives;
3. the temporal horizon over which consequences are counted;
4. the spatial or jurisdictional scope;
5. the evidential standards applied;
6. the acceptable uncertainty, and who bears residual risk;
7. the distributional constraints;
8. the revision authority;
9. the contestability conditions — who may challenge, on what ground, through what route;
10. the prohibited tradeoffs, if any.

A horizon is not required to be complete, and is not required to be computable. Where a horizon is represented in machine-readable form, four classes of element must be represented separately and must not be conflated: elements **formally encoded**; **externally supplied normative commitments** entering by reference to an authority outside the system; elements **unresolved or non-formalizable**; and the **interpreting authority** responsible for the residue, together with the version and revision history of the horizon itself. Substituting an arbitrary variable for an unresolved element does not make an axiom machine-decidable; it records the ambiguity in a different notation.

### Standing

**Standing** is the recognised entitlement of a person, group, institution, represented interest, or affected entity to demand reasons, contest a frame, introduce evidence, seek review, or require reconsideration. It may be grounded in exposure to consequences, rights, institutional role, representative authority, epistemic competence, or vulnerability. Six kinds are not interchangeable: **affected-party**, borne by those the decision falls upon; **legal**, conferred by a jurisdiction; **epistemic**, held by those competent to assess the evidence or frame; **representative**, exercised for another; **institutional**, attaching to an office; **public-interest**, asserted for a diffuse public.

Standing does not require personal participation. Where the affected entity cannot speak, does not yet exist, is diffuse, lacks power, is nonhuman, or is a future generation, it is exercised representatively — by guardianship, trusteeship, statutory representation, or an office holding the interest. Representation is itself answerable: the representative must be identifiable and the ground stated. Claims may conflict, or be manufactured, which is why standing must rest on a stated basis.

Standing determines **who may demand justification**, not **whose claim prevails**. A recognised claimant may be answered and still lose.

### Answerability, responsibility, accountability, and liability

These four terms are not synonyms, and the argument fails wherever they are exchanged.

**Responsibility**: a relation in which an agent, office, institution, or system is connected to an action, decision, condition, or consequence through authorship, control, role, contribution, capacity, or obligation.

**Answerability**: the obligation and capacity to give reasons for a decision or priority to those possessing legitimate standing, and to remain open to scrutiny, contest, correction, or consequence.

**Accountability**: the institutional arrangement through which answerability is demanded, assessed, and connected to review, sanction, correction, removal, or compensation.

**Liability**: legally specified exposure to remedy, sanction, loss, or obligation.

Responsibility can exist where answerability fails; answerability requires more than causal contribution, since a system that caused an outcome has not thereby answered for it; liability is neither necessary nor sufficient for moral responsibility; and an institution remains answerable for decisions partly produced by machines, which may act within an answerability architecture without bearing moral responsibility.

### Senses of "frame"

Four distinct senses are used in this work and are not interchangeable.

**Epistemic frame**: the conceptual organisation under which facts become intelligible as facts of a certain kind.

**Decision frame**: the specified boundaries, variables, and objectives of a particular decision.

**Institutional frame**: the rules, mandates, and procedures structuring an organisation's judgment. An institutional frame is not the same thing as an *institutional mandate*, which is the authority conferred on an office to decide; the frame shapes what the office can see, the mandate determines that it may act.

**Frame problem**: the computational and philosophical problem concerning what changes, what remains fixed, and what matters following an action.

## Selected Passages

**Relation is cheap; relevance is costly.**

**Information does not arrive with its own rank.**

**A fact becomes dangerous when it is true, available, and wrongly ranked.**

**Salience is what appears; relevance is what must not be ignored.**

**The machine can multiply relations; only an answerable being can decide which relation matters.**

**To govern intelligence is first to govern salience.**

**A civilization does not collapse from ignorance alone. It collapses when it can no longer tell which knowledge matters.**

**Relevance is the point at which intelligence becomes liable for its attention.**

**Correctness without relevance is precision inside a failed frame.**

**Auditability explains what happened. Answerability bears the consequence of what happened.**

**A civilization that cannot govern its frames will be governed by them.**

**The future will not belong to those who know everything, but to those who can still decide what matters.**

---