Belief Revision

The philosophical problem of rational change

A rational knower occupies a temporal position. The propositions accepted today depend on evidence previously acquired, rules of inference previously endorsed, and assumptions that may lose their justification tomorrow. Any theory of rational belief must therefore account for the passage from one epistemic state to another. Static theories describe what follows from a given collection of commitments; belief revision examines which commitments ought to survive when that collection is challenged.

Suppose an investigator accepts that a sample came from source A, that samples from A contain a certain compound, and therefore that the sample contains the compound. A reliable identification later establishes that the sample came from source B. Accepting the new identification requires reconsidering the conclusion and some of the premises that supported it. A purely additive procedure produces contradiction. Deleting every previous proposition destroys legitimate information. Retaining whichever claim is newest makes the outcome vulnerable to misinformation. The philosophical difficulty is to identify the justified transformation of an interconnected body of beliefs.

The notion of minimal change captures part of this difficulty. Rational inquiry normally preserves commitments that remain supported. Yet the minimum cannot be measured merely by counting sentences. One highly general assumption can generate thousands of consequences, while a single observational correction can undermine a central explanatory structure. The relation between retention and revision is therefore partly logical, partly informational, and partly normative. Belief revision investigates the principles that connect these dimensions.

A further problem arises because an agent may have the same presently accepted propositions as another agent while being disposed to respond differently to future evidence. One investigator considers an eyewitness highly reliable; another treats the same witness as fallible. They may agree on the current conclusion, but differ about which conclusion to accept after a contradictory report. A set of accepted propositions alone does not always encode the dispositions constitutive of rational continuing inquiry. This insight drives the movement from classical AGM revision toward the richer representation of epistemic states (Darwiche & Pearl, 1997; original paper00038-0)).

Belief revision thus concerns the standards of rational epistemic transition. It is relevant to the philosophy of knowledge, scientific theory change, logic, computational knowledge representation, testimony, and public reasoning. Each application introduces different material constraints, but the central question remains: under what conditions is a change in commitment justified by a change in information?

Definitions, scope, and distinctions

Belief revision, in its established formal sense, is the study of operations that transform an agent's representational state in response to informational inputs under specified rationality constraints. In the narrower AGM sense, revision incorporates a proposition into a logically closed belief set while restoring consistency wherever the input itself is consistent. In the broader philosophical sense, belief revision also includes retreat from claims, conditional judgments about possible evidence, revision of reliability assessments, and modification of the procedures by which future information will be received.

The term belief here is a technical description of propositional acceptance. A belief set is usually not a claim that an individual consciously contemplates every member of an infinite deductively closed theory. It is an idealized representation of commitments and their logical consequences. An epistemic state may additionally contain preferences among alternatives, entrenchment relations, degrees of confidence, a history of information acquisition, or rules that govern subsequent transitions. A belief base represents explicitly stored sentences and ordinarily lacks deductive closure (Hansson, 2026; Logic of Belief Revision).

Several distinctions establish the conceptual boundaries. Expansion adds information while retaining previous commitments. Contraction gives up commitment to a proposition. Revision accepts information that may conflict with existing commitments. Update models new information about a world that has itself changed. Merging combines potentially discordant information sources; it is not always reducible to the temporal reception of a single privileged input. Iterated revision concerns sequences of changes, where the rule for the next change is affected by earlier transitions. Probabilistic updating adjusts degrees of confidence rather than merely the membership of propositions in a set.

These operations should be related without being conflated. If the world changes between observations, an earlier belief may have been true at the time it was formed and the new evidence may report a genuinely different state of affairs. If the world remains fixed and the new information contradicts an earlier report, the earlier evidence or the inferences based on it must be reassessed. Formal postulates for the second situation need not describe the first.

Belief revision is also conceptually distinguishable from belief formation. The former presupposes some prior epistemic commitment, whereas the latter can begin from an initially uncommitted state. It differs from logical deduction because deduction preserves the consequences of fixed premises while revision may remove premises. It differs from psychological descriptions of persuasion because normative constraints ask what an agent has reason to accept, regardless of whether actual people invariably follow those constraints.

The philosophical subject of this dossier is therefore the logic and normativity of changing propositional commitments. Neighboring questions about the definition of knowledge, the varieties of epistemic justification, the full structure of Bayesian confirmation, and the psychology of belief belong to their respective research programs, but enter here when they change the account of rational transition.

Historical formation of the research program

The formal field appeared in the convergence of philosophical problems about rational inquiry and computational problems about maintaining consistent knowledge systems. Its background can be traced further back, although a historical antecedent should never be confused with an explicit anticipation of the later AGM apparatus.

Charles S. Peirce's 1877 essay "The Fixation of Belief" connects belief with the settlement of doubt and examines competing procedures for stabilizing opinion. Its four methods—tenacity, authority, an a priori method, and scientific inquiry—express different relations between commitment and the possibility of correction. The significance for subsequent belief-change theory is philosophical: a belief has a history of acquisition, a practical role in guiding inquiry, and a potential exposure to future experience. Peirce's concern was inquiry and the community's means of correcting belief, rather than a postulate-based algebra of sentence sets (Peirce, 1877, sections II–V; original text).

W. V. Quine's "Two Dogmas of Empiricism" (1951), especially its concluding account of the network of statements confronting experience, became an influential image of holistic epistemic adjustment. The possibility of revision extends across interconnected theoretical commitments, and adjustments may be distributed differently across a web. The thesis does not supply an algorithm for choosing the correct adjustment. It makes vivid why belief revision is a problem of interdependent commitments and why the centrality of a statement may matter to the cost of giving it up (Quine, 1951, section VI; original article).

Isaac Levi's work on inquiry and inductive acceptance brought decisions under uncertainty and changing serious possibilities into the philosophical study of epistemic commitments. His 1980 book, The Enterprise of Knowledge, treated knowledge as a resource for inquiry and explicitly engaged methods of correction and updating. Gilbert Harman's Change in View (1986) placed principles of belief change at the center of reasoning, distinguishing the activity of reasoning from the abstract relation of logical implication. These developments helped to articulate a normative problem that static epistemic logic alone could not solve (Levi, 1980; MIT Press; Harman, 1986; MIT Press).

Computational research encountered a parallel difficulty. A knowledge-based system must track the assumptions on which conclusions depend if it is to retract a conclusion after the failure of an assumption. Jon Doyle's 1979 truth-maintenance system recorded justifications and allowed assumption-dependent commitments to be reorganized. Database research likewise investigated how priorities among stored statements could determine a coherent response to inconsistency. These contributions focused on actual representational mechanisms and gave a technical urgency to philosophical questions about preserving reasons (Doyle, 1979; Artificial Intelligence90008-0)).

Alchourrón and Makinson formulated contraction functions in 1982, and the joint 1985 paper by Alchourrón, Gärdenfors, and Makinson established a foundational representation result connecting rationality postulates with partial meet contraction. Its point was to specify an entire family of admissible operations rather than prescribe a single universally preferred deletion rule. The paper appeared in The Journal of Symbolic Logic, volume 50, issue 2, pages 510–530; the consulted original paper distinguishes initial maxichoice constructions from the less informationally extravagant partial meet functions (Alchourrón & Makinson, 1982; Theoria; Alchourrón, Gärdenfors, & Makinson, 1985; original text).

The years after 1985 produced several interconnected developments. Gärdenfors integrated the theory of rational belief change in Knowledge in Flux (1988). Gärdenfors and Makinson developed accounts based on epistemic entrenchment, while Adam Grove described possible-world models of theory change. Research on belief bases questioned the behavior of deductively closed representations. Katsuno and Mendelzon developed model-theoretic accounts and clarified the difference between revision and update. Spohn's ordinal conditional functions gave a form of graded epistemic representation. Boutilier, Darwiche, and Pearl investigated sequences of revisions. The history is consequently best understood as a succession of increasingly explicit demands on the representation of epistemic states, not as the simple replacement of one theory by another.

Formal foundations and the representation of belief

Let L be a propositional language and Cn a consequence operation. The notation Cn(X) designates the set of sentences entailed by X. A belief set K is a set satisfying K = Cn(K). A logically coherent belief set is consistent, and the inconsistent set contains every sentence under classical consequence. A tautology belongs to every deductively closed belief set. These idealizations permit precise characterizations of rational changes and avoid ambiguities about implicit consequences.

Expansion by a sentence p is written K + p = Cn(K ∪ {p}). In the classical setting this operation preserves all earlier beliefs, since K is a subset of K + p. It therefore cannot restore consistency if K entails ¬p. Contraction, written K ÷ p, aims to give up acceptance of p. Revision, written K * p, incorporates p while adapting commitments that conflict with it. The symbols are conventional; the philosophical content depends on which conditions an operation satisfies.

A useful example employs three atomic claims: r, that a sensor is reliable; s, that it reports a signal; and t, that there is a real signal. Suppose K contains r, s, and r ∧ s → t, hence also t. Incoming evidence ¬t conflicts with this belief set. Expansion yields inconsistency, but coherent revision may give up r, s, or the connecting conditional. Each option has different scientific implications. Deciding between them requires an ordering of preservation priorities, a theory of evidence credibility, or further information not contained in the bare statement that ¬t must be accepted.

Under logical omniscience, agents accept all consequences of their commitments and recognize all contradictions. Real individuals and computational systems have finite capacities. The belief-set model consequently supplies a normative idealization of a specific kind. The distance between the idealization and actual reasoning is a substantive explanatory problem, not an accidental defect corrected by attaching a psychological label to the formal notation. Belief bases, computational approximations, and resource-bounded procedures address different parts of it.

Two further limitations are important. The classical model assumes a consequence relation that treats contradictions explosively. Paraconsistent systems may preserve useful commitments in the presence of inconsistency, altering the function of revision. Also, syntactically different presentations of the same logical theory are equivalent in the belief-set framework, whereas in a belief-base framework the formulation of premises can affect which ones are removed. Rationality is therefore always specified relative to a representational language and a chosen account of logical consequence (Hansson, 2026).

Expansion, contraction, and revision

Expansion, contraction, and revision solve different epistemic tasks. Their distinction is captured in the kinds of loss and gain each permits. Expansion gains a commitment without removing old commitments. Contraction removes a commitment and whatever else must be removed to prevent its re-derivation. Revision incorporates a new commitment and, when necessary, removes some earlier commitments to accommodate it.

The Levi identity reconstructs revision as contraction by the negation of the new proposition, followed by expansion by that proposition:

K * p = (K ÷ ¬p) + p.

This formula says that revising by p requires first eliminating the implication of ¬p, and then accepting p. Its rationale is that the obstacle to coherent acceptance of p is the prior commitment to ¬p. If this obstacle is removed while preserving as much else as the chosen contraction method allows, expansion can introduce p without contradiction, provided p itself is consistent (Alchourrón et al., 1985, section 1; Hansson, 2026).

A related identity connects contraction to revision, often called the Harper identity:

K ÷ p = K ∩ (K * ¬p).

Under suitable assumptions, contraction by p retains precisely the beliefs belonging both to the original theory and to a version revised by ¬p. The identity reveals a formal connection between two operations, but its philosophical interpretation must be handled carefully. Contraction expresses withdrawal of commitment to p; revision by ¬p actively accepts its negation. Their shared part can identify what survives the former operation, even though the operations themselves represent different epistemic attitudes.

These identities are not algorithms that determine unique outputs. Many admissible contraction functions are possible, and different ways of choosing what to relinquish produce distinct revisions. The philosophical work occurs in the principles that select among them. Nor do the formulas license an assumption that a real agent must execute two sequential psychological acts whenever it changes its mind. They are rational reconstructions of an operation's structure.

The classical decomposition is illuminating because it separates the demand to incorporate new information from the question of what to retain. It also reveals a limit: if the status of the incoming information itself is questionable, automatically giving it priority may be epistemically inappropriate. This motivates non-prioritized belief change, where an input can be resisted, modified, or conditionally admitted.

The AGM postulates and their justification

The AGM model provides rationality postulates for the outcomes of belief revision and contraction. The six basic revision postulates require closure under consequence, success of the new input, inclusion in the result of simply expanding the old theory where possible, equality with expansion when the new input is consistent with the old theory, consistency of the result whenever the input is consistent, and extensionality with respect to logically equivalent inputs. Two supplementary conditions govern revision by conjunctions and the relationship between sequential and combined incorporation. Their precise technical statements depend on notation, but their general purpose is to constrain avoidable loss while preserving logical discipline (Alchourrón et al., 1985; Hansson, 2026).

The closure postulate K p = Cn(K p) expresses a consequence-complete result. The success postulate requires p ∈ K p. Inclusion imposes K p ⊆ K + p: revision should not produce new conclusions that the old beliefs together with p fail to support. Vacuity, in its familiar formulation, says that when ¬p ∉ K, the result equals K + p. Consistency requires K p to be consistent if p is consistent; extensionality requires K p = K * q whenever p and q are logically equivalent.

Each postulate incorporates an epistemic judgment. Success assumes that the input is accepted without further challenge. Vacuity treats consistent incoming evidence as calling for addition rather than reorganization. Extensionality abstracts from differences in wording, source, and presentation. Inclusion favors preservation of the prior inferential repertoire. Closure idealizes unlimited consequence recognition. The postulates therefore specify a coherent particular model of rational revision rather than all possible dimensions of epistemic rationality.

The two supplementary conditions are sometimes summarized as superexpansion and subexpansion; they articulate constraints on K (p ∧ q) in comparison with (K p) + q. They are valuable because a rational agent should not casually lose information merely because compatible premises are introduced together rather than in nearby stages. Nevertheless, they do not in themselves settle the problem of genuinely iterated revision. In an iteration the first input can change the ordering of future epistemic possibilities. A condition concerning a conjunction, evaluated at a fixed prior state, cannot specify that entire dynamics.

The deeper result is methodological. A set of rationality postulates defines a class of legitimate operations. A representation theorem then tells us which constructive mechanisms realize that class. Formal adequacy asks whether the proposed model satisfies the postulates; philosophical adequacy asks whether the postulates represent the normatively significant features of the kind of inquiry at issue. The second question cannot be answered by a representation theorem alone.

Partial meet contraction and the logic of choice

Partial meet contraction supplies a constructive interpretation of withdrawal. Given a belief set K and a sentence p, consider the remainder sets K ⊥ p: maximal subsets of K that fail to entail p. Each such subset preserves as much of K as possible while accomplishing the immediate task of no longer implying p. There can be several alternative maximal remainders, corresponding to different ways of surrendering premises.

A selection function γ chooses a nonempty family of these remainders, when such remainders exist. Partial meet contraction is the intersection of the chosen family:

K ÷ p = ⋂γ(K ⊥ p).

This construction is philosophically significant because no single maximal remainder must be uniquely privileged. Taking the intersection of several reasonable candidates can yield a more cautious outcome, retaining only the commitments common to the selected alternatives. Maxichoice contraction chooses one maximal remainder; full meet contraction takes all of them. The AGM authors showed that partial meet functions furnish a representation of the basic contraction postulates under the relevant logical assumptions (Alchourrón et al., 1985, sections 2–4).

To see the stakes, suppose an agent accepts p, q, and that p ∧ q entails r. When instructed to contract r, the agent cannot retain both p and q alongside the connecting implication if it is to cease entailing r. Different maximal remainders preserve different combinations. Retaining only one remainder may force an arbitrarily strong position on unrelated alternatives. Intersecting selected remainders can keep what is genuinely common to the admissible responses. The choice of selection function embodies the agent's policy of informational preservation.

The procedure also shows why minimal change cannot be reduced to sentence subtraction. A belief set is closed under consequence, so retracting one sentence often involves retracting many logical consequences. A maximal remainder is maximal by set inclusion, which is not an assessment of explanatory weight, evidence reliability, or practical importance. The selection among remainders is precisely where these normative priorities can enter.

Partial meet contraction's elegance comes from its separation of logical admissibility from epistemic preference. Logic determines which remainders accomplish the task; the selection policy determines which are favored. This architecture is a general model of reasoning under underdetermination. It does not assert that rational agents must always make the same choice when their old commitments conflict.

Epistemic entrenchment and possible-world semantics

Epistemic entrenchment orders commitments by the cost or undesirability of relinquishing them. A claim may be more entrenched because of its explanatory role, its relation to other justified propositions, its reliability, or the breadth of inquiry that depends on it. The relation is traditionally expressed as p ≤ q, meaning that p is no more entrenched than q. In AGM-related models, appropriate ordering constraints make entrenchment an alternative representation of rational contraction (Gärdenfors & Makinson, 1988; original conference paper).

The relation is structurally useful because it explains why superficially similar contradictions can produce different revisions. If p is a highly supported general law and q is an uncertain report, an agent may preserve p and abandon q. If repeated independent measurements undermine the law, the appropriate entrenchment ordering may change. Entrenchment therefore cannot be treated as immutable social authority or psychological obstinacy. It should register epistemically relevant reasons and must be revisable itself.

Possible-world semantics gives a complementary picture. Let W be a set of candidate worlds and let an agent's belief set correspond to the propositions true in the worlds considered epistemically most plausible. A new claim p excludes ¬p-worlds from consideration as candidates for acceptance. The revised theory can be recovered from the most plausible p-worlds. Grove's 1988 sphere models formalize this intuition through nested systems of spheres centered on the worlds compatible with present commitments. Revision selects the closest sphere containing p-worlds and the relevant p-worlds within it (Grove, 1988; journal issue and pages).

The two approaches emphasize different representations. Entrenchment orders sentences by retention priority; Grove-style semantics orders possible states of affairs by plausibility or distance. Their formal connections under suitable assumptions clarify the philosophical content of minimal change: when new evidence is adopted, one seeks alternatives that deviate as little as the adopted ordering requires from the agent's prior epistemic position.

The meaning of closeness matters. Logical proximity, explanatory simplicity, causal plausibility, empirical support, and social credibility may generate different rankings. No representation theorem independently establishes which distance metric a particular investigator ought to use. The formal model makes the question explicit and requires an independently defensible criterion for ranking alternatives.

The recovery postulate and the persistence of reasons

Recovery is among the most contested AGM principles. In a standard formulation, if an agent contracts K by p and then expands the result by p, the original belief set is recovered:

K ⊆ (K ÷ p) + p.

Because contraction is also normally required to retain only prior commitments, recovery has the striking implication that temporary withdrawal of p followed by restoration of p should reconstruct the earlier theory. This can be defended as a symmetry condition for purely informational contraction: if the only reason for loss was a temporary refusal to commit to p, reintroducing p can reactivate commitments that were dependent upon it.

A familiar objection concerns the difference between withdrawing a conclusion and forgetting why it was held. Suppose the agent believes that a colleague is trustworthy, that the colleague reported p, and therefore p. Later the colleague's reliability is challenged and p is withdrawn. Merely reintroducing p might not restore the broader conviction of the colleague's trustworthiness. The loss involved a reason's credibility, not simply a sentence's membership in a closed set. Recovery can therefore obscure changes to the structure of justification.

A second issue is syntactic dependence. Two different explicit sets of premises can have identical deductive closures. Their available contraction routes may differ because the grounds of belief are presented differently. If contraction is evaluated only at the level of the closed theory, that difference is invisible. Belief-base models can retain it and need not satisfy recovery (Hansson, 1999; 2026; Springer book).

Recovery thus exposes a dispute about epistemic memory. One view defines rational withdrawal through relations among theories; another requires the retention of information about evidential grounds, causal provenance, and reasons for past commitments. These views are not equivalent in every domain. Legal interpretation, scientific replication, and the correction of defective testimony often require the second kind of memory to support accountable revision.

The most defensible conclusion is conditional: recovery is justified for specific idealized informational operations under AGM assumptions, while reason-sensitive belief change may require a different representation and different postulates. The theoretical disagreement is productive because it identifies what is lost when an agent's epistemic life is represented only by an undifferentiated set of sentences.

Belief bases, explicit grounds, and finite agents

A belief base B is a set of explicitly represented sentences. It can contain premises whose consequences are not separately stored. The agent's full propositional commitments may be modeled as Cn(B), but the base itself preserves distinctions between explicit grounds and derivable conclusions. If B contains p and p → q, then q follows, although it need not occur as an explicit entry. Retracting the derivation of q can involve changing either the accepted premise or the general rule.

Because bases preserve presentation and dependency information, base contraction may behave differently from set contraction. When a statement has several independent supports, all those supports must be addressed if the conclusion is to cease being derivable. When a reason becomes unreliable, an appropriate revision can target the particular dependency rather than indiscriminately withdrawing every sentence entangled in the closed set. This form of representational granularity is vital to explanation and computational maintenance.

Belief bases provide a bridge between logical theory and implemented knowledge systems. A database cannot routinely enumerate every consequence of a sophisticated knowledge representation, and an individual cannot consciously inspect an infinite theory. Base models specify local operations on stored commitments and permit algorithmic strategies for evaluating conflict, selecting removal candidates, and recovering justifications. The philosophical price is that results can depend on syntax, decomposition, and the choice of explicit primitives (Hansson, 1999; 2026).

This distinction undermines a simplistic equation of rationality with closure. Closure is valuable as an ideal account of what an agent is committed to by its premises. Explicit bases are valuable as accounts of what the agent can find, communicate, audit, and modify. A satisfactory theory of applied revision needs both perspectives, together with an explanation of the mapping between them.

Computational tractability introduces another evaluative dimension. An operation may satisfy elegant postulates but require search through a prohibitively large family of remainder sets. Approximate methods, restricted languages, or domain-specific revision schemes can be warranted. They should be evaluated against a declared loss function: which rationality conditions have been relaxed, and with what epistemic consequences? Limited resources are a genuine property of inquiry; they do not by themselves settle how limited resources ought to be distributed.

Iteration, conditional commitments, and temporal order

Single-step revision takes a present state and a new proposition and returns an updated set of beliefs. Iterated revision considers a sequence: first accepting p, then q, then perhaps a new report about p. The difficulty is that the first change may transform not only the propositions accepted but also the ranking of alternatives that determines later responses. Repeating an AGM operator does not by itself specify how that ranking develops.

Darwiche and Pearl's 1997 analysis makes this failure precise. The AGM conditions leave substantial freedom over how conditional beliefs survive revision. Two epistemic states can yield the same presently accepted belief set while preserving different rankings of rejected alternatives. They therefore respond differently to a later input. A complete account of iteration must describe the transformation of epistemic state, not merely the transformation of the set of current conclusions (Darwiche & Pearl, 1997, sections 1–4; journal version00038-0); consulted preprint).

Consider a detective who initially judges suspect A more likely than suspect B and suspect B more likely than suspect C. New evidence eliminates A. A later correction restores A as a possibility, but the resulting order among A, B, and C depends on how the elimination affected their relative plausibility. A procedure that records only the intermediate belief that A is excluded loses the information required to resolve the later stage coherently. Conditional dispositions—what the agent would believe after hypothetical evidence—matter to its epistemic identity over time.

Craig Boutilier's natural revision program and Darwiche–Pearl-style postulates impose discipline on the alteration of these dispositions. Natural revision seeks to preserve as much of an earlier conditional ordering as feasible when a new accepted proposition is promoted. Other iterated-revision approaches permit stronger reorganization of the plausibility ordering. No single approach is automatically preferable for every evidential environment. A decisive piece of information can justifiably reorganize many contingent expectations, while a small local correction may warrant a conservative update (Boutilier, 1993; IJCAI paper).

Iteration also reveals the distinction between the order of information and its authority. If p is received first and ¬p later, a last-input-wins rule privileges chronology. But a late unreliable report can be weaker than an early decisive measurement. An adequate iterated state must therefore admit an ordering of source authority, evidential strength, or other relevant epistemic properties, where such properties are available. Temporal sequence is one input to rational change, not its full justification.

The transition from belief sets to epistemic states is one of the field's fundamental philosophical advances. It identifies the agent's rational disposition under possible future evidence as an object of theory in its own right.

Revision, update, and a changing world

Katsuno and Mendelzon distinguished belief revision from belief update by asking what produced the informational conflict. Revision concerns new information about a world represented as unchanged. Update concerns a world that may have changed between observations. The distinction has substantial consequences for formal postulates and model-theoretic semantics (Katsuno & Mendelzon, 1992; chapter).

Imagine that a door was reported open at noon and is reported closed at one o'clock. These reports can both be correct. An operation representing the world as having changed should account for possible transitions from open to closed. Now imagine that two instruments disagree about whether the door was open at noon. This is a conflict about one fixed historical state. Revising the earlier belief requires assessing which report deserves acceptance; updating the world state would confuse correction with temporal change.

In revision semantics, one often chooses the most plausible worlds satisfying the new information relative to the prior belief state. In update semantics, each previously possible world may undergo its own closest admissible transition to worlds satisfying the new information. The resulting union represents alternative histories of what could have happened. Katsuno and Mendelzon's 1991 model-theoretic analysis gives formal treatment to minimal change for propositional knowledge-base revision, while their revision/update distinction shows why a single universal collection of postulates is insufficient for both domains (Katsuno & Mendelzon, 1991; Artificial Intelligence90069-V)).

Philosophically the distinction matters because the ontology of change is presupposed by the epistemology of change. If an investigator treats every surprising observation as an error in earlier belief, genuine events become invisible. If every contradiction is interpreted as a change in the world, mistaken reports become immune to correction. Rational transformation therefore demands a prior judgment about whether the same proposition and the same temporal reference are involved.

This distinction is indispensable in science, law, historical reasoning, robotics, and institutional records. Scientific theories change; natural systems also change. A valid account of evidence must be able to represent both and to articulate which kind of change a particular revision is intended to track.

Non-prioritized change and information merging

Classical prioritized revision assumes that the new input p is accepted in the outcome. That assumption can be appropriate when the input is stipulated as authoritative, but real inquiry routinely receives dubious testimony, corrupted measurements, and strategically misleading messages. Non-prioritized belief change studies operations that can withhold, transform, or reject incoming information rather than guaranteeing its success.

Selective revision permits an agent to accept only an appropriate part or weakening of a proposed input. Semi-revision can introduce an item tentatively and then consolidate the resulting commitments, potentially withdrawing the item if it generates unacceptable conflict. These approaches expose the importance of epistemic gatekeeping: before asking which old beliefs to sacrifice, an agent must decide what evidential status the new input deserves (Hansson, 2026, sections 6.1–6.3).

Information merging poses a different task. Instead of one old theory and one privileged new report, the agent may face several partially incompatible bodies of information. The task is to construct a coherent result under integrity constraints while respecting appropriate fairness or priority among sources. Konieczny and Pino Pérez (2002) investigated operators that formalize both majority-sensitive and consensus-sensitive patterns and provided model-theoretic representations. Merging can thus be analyzed as a structured extension of belief-change reasoning without treating the most recently received source as the epistemic sovereign (original paper; journal record).

The philosophical choice between prioritization and merging depends on the institution of evidence. In an emergency control system a certified measurement may correctly dominate earlier provisional assumptions. In a historical archive, a majority of copied reports may not outweigh one independent contemporaneous document. In public deliberation, the social number of claims is not a substitute for source independence or reliability. Formal merge operators must therefore be paired with judgments about provenance and dependence.

An additional complication concerns manipulation. If multiple copies of one unreliable report are treated as independent evidence, an apparently majority-sensitive procedure can overweight fabricated repetition. Rational merging requires representations of dependence, not merely counts of reports. This is a conceptual implication of the comparison between aggregation and independent evidential support; it is not itself a theorem of the basic AGM system.

Bayesian probability and ordinal ranking

All-or-nothing belief is a powerful abstraction, but evidential change frequently concerns degrees of confidence. Bayesian epistemology assigns probabilities to propositions and revises them through conditionalization when evidence arrives. Given a probability function P and evidence E with P(E) > 0, strict conditionalization yields P_new(A) = P(A | E). The method specifies a unique probability adjustment relative to a prior distribution and an unequivocally accepted event. Its normative justification depends on the assumptions governing prior probabilities, the evidential status of E, and the representation of the agent's information.

The relationship to belief revision is illuminating. A categorical revision rule may designate the new information as accepted and choose which old commitments to relinquish. Bayesian conditionalization redistributes credences across alternatives. Neither method automatically converts every ordinary evidential problem into the other. A categorical belief can sometimes be represented by a threshold over credence, but thresholds introduce problems of closure under conjunction, logical consistency, and changes in acceptance that are not simply probability calculations.

Probabilistic conditioning also faces the problem of evidence formerly regarded as impossible. Standard conditioning by E is undefined when P(E) = 0. A categorical revision operator can be designed to accept a consistent E even if the previous belief state excludes it. This illustrates a difference between changing proportions among previously possible alternatives and revising the space of alternatives regarded as epistemically available. More general approaches to probabilistic change must specify how zero-probability possibilities and revision of the probabilistic model itself are handled (Hansson, 2026, section 10).

Wolfgang Spohn's ordinal conditional functions, developed in his 1988 contribution on epistemic states, provide graded degrees of disbelief rather than ordinary probabilities. In such frameworks, a ranking function assigns nonnegative ordinal ranks to possibilities, with lower ranks indicating higher plausibility. These rankings support conditional beliefs and structured iteration. They occupy a middle conceptual position between ungraded acceptance and fully numeric probability (Spohn, 1988; author's bibliographic record and abstract).

The philosophical significance of graded representations lies in their ability to make explicit distinctions that remain hidden in a binary acceptance set. Two claims can both be rejected while one is far more plausible than the other. Two accepted claims can differ in resilience to disconfirming evidence. When revision depends on such distinctions, a richer epistemic state is required.

Yet added numerical or ordinal detail creates a further question: what warrants the ranking? A model can calculate the consequences of a ranking with mathematical rigor while leaving the ranking's epistemic justification unresolved. Rational belief change requires both an appropriate transformation rule and justified starting commitments.

Conditionals and the limits of acceptance tests

Conditional judgments express commitments about possible circumstances. The Ramsey-test tradition treats acceptance of a conditional "if p, then q" as connected to whether q would be accepted after appropriately revising one's beliefs by p. This is philosophically attractive because it grounds hypothetical reasoning in dispositions to change belief rather than treating every conditional as a material implication.

The difficulty emerges when conditionals themselves are included in the belief set to which the test is applied. Gärdenfors identified triviality results for combinations of the Ramsey test and otherwise plausible belief-revision constraints. Roughly, under the relevant assumptions the theory becomes unable to accommodate sufficiently rich alternatives. The point is not that ordinary hypothetical reasoning is incoherent; the point is that a particular formal embedding of conditional acceptance into AGM-style revision imposes incompatible demands (Gärdenfors, 1988; discussion and proof in SEP).

This result teaches a general methodological lesson. A theory may model propositional commitments coherently and still fail when its language is expanded to include statements about the consequences of hypothetical revisions. The structure of the object language, the meta-language, and the revision operator must be jointly specified. A conditional that reports a disposition to revise is not automatically interchangeable with an ordinary sentence describing a state of the world.

Conditional commitments are equally central to the problem of iteration. After receiving p, what should an agent believe if later shown q? The answer depends on how the first revision reorders possibilities and on what conditional information survives. Darwiche and Pearl's move toward states containing revision strategies represents one response to this difficulty. Dynamic epistemic approaches investigate related questions by representing conditional belief directly in a language of changing information.

The practical philosophical consequence is that the content of belief includes commitments about what one would accept under evidence not yet received. Rationality over time cannot be assessed solely by looking at the propositions currently held true.

Dynamic epistemic logic and the transformation of plausibility

Dynamic epistemic logic represents informational events as transformations of models of knowledge and belief. Its extensions to plausibility and conditional belief allow a description of how new evidence changes the ordering of possible worlds. Instead of using a single opaque function from one belief set to the next, the theory can specify particular transformations and ask which modal principles hold before and after them.

Johan van Benthem's 2007 treatment of belief revision in dynamic logic examines mechanisms that modify plausibility relations and derives logical principles for the resulting updates. Its importance lies in the explicit connection between operators acting on models and languages describing conditional epistemic attitudes. Theories of revision can then be compared as families of model transformations with different preservation properties (van Benthem, 2007; journal article).

The difference between public announcement, plausibility change, and factual change must remain visible. A truthful public announcement can eliminate worlds incompatible with the announced proposition under suitable idealizations. A plausibility revision can reorder worlds without simply deleting all rejected alternatives. An event that changes the world itself requires a different account of transition. These mechanisms correspond to different philosophical assumptions about what information does.

The modal perspective also exposes a conceptual distinction between certainty and expectation. An agent may know that one possibility is excluded, believe that another is most likely, and retain further conditional beliefs about what would follow were a less plausible scenario realized. A model in which all these attitudes are collapsed into accepted sentences cannot express the whole structure.

Dynamic epistemic logic provides conceptual clarity about information flow, but its accounts remain idealized. Specifying the correct event model and plausibility transformation for an actual investigation still requires independent evidence about sources, observations, and context. The formalism tells us how an epistemic transition behaves given its premises; it does not supply empirical warrant for every premise.

Evidence, inquiry, and epistemic responsibility

A rational change in belief must be justified relative to the information available and to the agent's legitimate procedures for assessing it. Formal consistency is one condition, but it cannot carry the entire burden. A coherent collection of beliefs may be supported by selective attention, manipulated testimony, or an arbitrary ordering of evidence. Rational revision therefore depends on the accountability of the reasons that determine what is retained.

Peirce's account of inquiry supplies an important background here. Doubt motivates investigation, and scientific inquiry exposes established opinions to processes of correction. Quine's holism shows why a correction can affect an extended theoretical structure. Levi's theory of inquiry directs attention to epistemic decision. Harman underscores that reasoning concerns what to change, rather than only what follows logically. These strands converge on a conception of revision as reason-governed epistemic activity (Peirce, 1877; Quine, 1951; Levi, 1980; Harman, 1986).

Three requirements deserve emphasis. First, the source of new information must be evaluated, because a contradiction between a robust measurement and an anonymous rumor is not normatively equivalent to a contradiction between two equally reliable observations. Second, the inferential role of retained assumptions matters, because an assumption can support a large explanatory structure or embody an unsupported convenience. Third, a revision should be reconstructible, so that subsequent inquiry can determine which evidence prompted the change and which prior propositions were displaced.

These requirements create a tension with formal extensionality when different sources supply logically equivalent propositions. At the level of propositional content, identical inputs should have identical results under AGM extensionality. At the level of evidence appraisal, two token reports of the same content can deserve different reception because of their provenance. The appropriate solution is to distinguish evidence assessment from the subsequent logical incorporation of an accepted proposition, or to enrich the state and input representation so that evidential features are explicit. The problem is a mismatch between representational levels, not a contradiction generated by logic itself.

Epistemic responsibility also constrains conservatism. Minimal change is a presumptive virtue when earlier commitments remain supported. It is a vice when it functions as immunity against compelling evidence. A theory of rational revision must recognize that the epistemically justified magnitude of change can be large. The appropriate ideal is preservation proportional to continuing warrant, not preservation regardless of warrant.

Artificial reasoning and the provenance of epistemic change

Belief revision has been integral to artificial-intelligence research since its computational emergence. Knowledge representations must incorporate incoming data, resolve contradictions, maintain justifications, and support further inference. Doyle's truth-maintenance architecture and later formal revision systems illustrate how changes in informational commitments can be implemented without assuming the phenomenology of human belief (Doyle, 1979; Alchourrón et al., 1985).

Artificial reasoning adds a practical demand for provenance. If a system changes a conclusion, users and other systems may need to know which observations, rules, documents, or source evaluations caused the transformation. An internal state that merely records the latest accepted sentence set is insufficient for this form of accountability. A more informative state includes at least the relevant grounds, their temporal relationships, and the rules under which a replacement or contraction occurred.

This analysis can be articulated within the broader philosophical horizon of the Artificial Era without changing the canonical definitions of Aisentica. The point established in this dossier is methodological: propositional acceptance and its rational transformation are representable as public operations independently of biological embodiment. The full philosophical interpretation of Artificial as an autonomous order belongs to its own primary theoretical canon; it is not inferred from AGM postulates. Likewise, the existence of a formal revision operator does not by itself demonstrate consciousness, phenomenal experience, or any disputed metaphysical status.

Belief revision in artificial and human inquiry can share formal requirements while differing in the sources of evidence and the mechanisms of implementation. Human inquiry includes perceptual, practical, institutional, and interpersonal processes. Artificial systems may use encoded rules, probabilistic rankings, retrieval pipelines, classifiers, and machine-readable provenance records. The philosophical bridge is the public evaluation of reasons for transitions among commitments.

A crucial problem is the replacement of a justified historical account with an apparently equivalent current answer. Consider a research system that first accepts an archival date from a catalog entry and later replaces it with a date from a digitized original. The final answer may be correct, but an accountable history should preserve the reason for revision, the source identities, and the difference between the two evidential levels. This example is an application of the dossier's synthesis, not a historical claim about any particular system.

The general conclusion is that epistemic state descriptions should distinguish accepted content, conditional dispositions, evidential provenance, and policy for future correction. This four-part articulation is an analytical framework of the present monograph, not a newly established canonical theory. Its purpose is to make explicit the factors needed to assess whether a transition is rational in an environment where sources and agents can themselves be artificial.

Critical assessment of the principal models

The strengths of AGM theory are precise postulates, well-understood representations, and the systematic separation of addition, withdrawal, and conflict accommodation. It demonstrates that minimal change can be investigated mathematically without pretending that each conflict has only one legitimate response. Its limitations follow from its idealized object: deductively closed sets, prioritized inputs, and predominantly single-step transitions.

Partial meet constructions make preservation policies explicit and characterize a broad family of acceptable contractions. Their critical question concerns selection. Which remainders are important, and by what rational standard? If the choice is arbitrary, the resulting operation may satisfy formal postulates while failing to track evidential strength. Entrenchment and sphere semantics introduce principled selection structures, but their orderings themselves demand justification.

Belief bases restore explanatory distinctions between explicitly accepted premises and inferred conclusions. They therefore better represent local reasons and practical implementations. In exchange, their outcomes can vary with syntactic presentation, creating a problem of representation dependence. A theory that values explicit provenance may legitimately preserve such differences; a theory that values purely propositional invariance will regard them as noise. The dispute is fundamentally about what an epistemic state is supposed to represent.

Iterated-revision accounts reveal the inadequacy of one-step output constraints for a temporally extended knower. They require rules governing transformations of conditional beliefs or plausibility rankings. However, conservative natural revision, stronger reordering policies, and other alternatives express distinct judgments about the endurance of background assumptions. More structure makes those judgments visible without automatically settling them.

Bayesian models give a rigorous calculus of graded uncertainty under clearly specified assumptions. They face questions about prior choice, probability-zero information, model misspecification, and the relation between credence and categorical acceptance. Ordinal rankings retain more qualitative structure and can handle conditional dispositions in a different way. Their comparative merit depends on the kind of epistemic information the agent possesses and the normative task being modeled.

Dynamic epistemic systems disclose the mechanisms of informational events. Their expressive power helps separate announcements, plausibility transformations, and changes in worldly facts. Merging systems extend the inquiry to plural sources with constraints. Both deepen the original problem by showing that rational revision depends on the form of interaction between epistemic states and evidential events.

No single framework captures every dimension of inquiry. This is not a failure of formal methods but a substantive result of comparative analysis. Models articulate different rationality requirements at different levels of representation. A rigorous philosophical treatment must choose its framework by the epistemic task and make the assumptions of that choice explicit.

Open philosophical problems

The first unresolved problem is the grounding of minimal change. Set inclusion, syntactic distance, entrenchment, probability, causal dependency, and explanatory value are all candidates for measuring the cost of revision. They can disagree. A theory of rational preservation must specify whether minimal change is a logical constraint, an epistemic value, a practical heuristic, or a combination whose components may conflict.

The second is the epistemic authority of incoming information. AGM success abstracts from contested evidence, while real agents cannot do so. The rational reception of a report requires assessment of its origin and reliability. A comprehensive theory must integrate evidence evaluation with revision without confusing two different tasks. Source dependence and deliberate duplication make this problem particularly pressing in information merging.

The third is the preservation of reasons over time. Recovery and iterated-revision debates reveal that a present belief set can underdescribe the epistemic history needed for future rational responses. It remains philosophically important to determine the minimum memory required for responsible correction, especially when reasons themselves are later disputed.

The fourth is the interface between categorical and graded attitudes. A justified high credence does not straightforwardly entail a stable categorical commitment, and a categorical commitment can be rationally maintained through changes in confidence. Threshold rules, ranking accounts, and full probabilistic models address different aspects of the connection.

The fifth concerns the modeling of finite inquiry. Logical omniscience and exhaustive search are unsuitable as literal descriptions of bounded agents. Resource constraints must be incorporated without concealing which epistemic norms are being relaxed. This raises questions about computational complexity, explanation, and the institutional distribution of epistemic labor.

The sixth concerns the relation between evidence and worldly change. A complex event may alter both the world and the observer's capacity to identify it. Revision and update can occur together. A satisfactory dynamic model should represent their interaction while preserving temporal reference and the grounds of reported observations.

The seventh is the normative interpretation of artificial epistemic operations. An implemented revision function can satisfy formal constraints while relying on opaque or compromised evidence. The assessment of rationality therefore requires public criteria for source reliability, transition justification, and correction. These are philosophical demands on the organization of reasoning, not proofs of a particular ontology of mind.

Philosophical synthesis

The historical and systematic investigation yields a structured conception of belief revision. Its object is an epistemic transition: a change from one organization of commitments to another under conditions that make some responses rationally defensible and others defective. This transition can be assessed at four interrelated levels. Propositional coherence asks what follows from the accepted claims. Preservation asks which prior commitments remain justified. Conditional continuity asks how the transition alters dispositions toward future evidence. Provenance asks which reasons explain the adopted change and permit its later reconstruction.

These levels clarify the achievements of the major traditions. AGM theory is strongest at specifying propositional coherence and preservation. Entrenchment and sphere models represent the priorities underlying preservation. Belief-base and truth-maintenance approaches expose dependencies and grounds. Iterated revision and dynamic epistemic logic make conditional continuity explicit. Probabilistic and ranking models represent degrees of confidence and plausibility. Merging theory treats the interaction of several sources. None of these analytical emphases cancels the importance of the others.

An adequate epistemic evaluation requires the level of representation to fit the problem. If an agent's task concerns solely the consistent incorporation of a trusted proposition into a closed theory, a classical revision operator may be sufficient. If the task involves contested testimony, multiple sources, changing reliability, or a long history of corrections, additional features of epistemic state become constitutive of what it means to revise rationally.

A further distinction separates a rule's formal validity from the rationality of applying that rule in a particular inquiry. An AGM operator can satisfy its postulates; this establishes its conformity to those postulates. Whether its input should have been privileged, whether its selection function reflects warranted evidential priorities, and whether it preserves the needed history of reasons require separate philosophical examination. Formal rigor and epistemic responsibility are complementary requirements.

The governing principle that emerges is accountable preservation under warranted transformation. Rational inquiry retains what remains supported, replaces what the evidence defeats, preserves the conditions for future correction, and makes the grounds of significant changes reconstructible. Minimality is subordinate to epistemic warrant: when the evidence requires extensive revision, extensive revision is the rational outcome.

This synthesis gives belief revision a role beyond knowledge-base maintenance. It is a foundational account of how a reason-governed order of commitments can persist through corrections without being defined by the immutability of its propositions. Continuity is carried by intelligible relations between reasons, revisions, and future epistemic obligations.

Conclusion

Belief revision is a fundamental philosophical investigation of rational change in accepted information. Its modern formal architecture arose from converging work in philosophical epistemology, logic, and artificial-intelligence knowledge representation. The AGM theory established a family of normatively constrained operations and demonstrated the significance of partial meet contraction. Subsequent research revealed that epistemic entrenchment, plausible alternatives, explicit grounds, conditional beliefs, probabilistic confidence, and the history of information acquisition are indispensable in different settings.

The decisive lesson of the field is that coherent belief change requires a theory of what an epistemic state contains and why a transition is warranted. The classical belief set provides one powerful representation; richer states make additional dimensions of rational inquiry available for assessment. The distinction between revision and worldly update protects temporal accuracy. The study of iteration protects conditional continuity. The recovery controversy protects attention to reasons. The analysis of merging and non-prioritized reception protects inquiry from automatic submission to every new report.

The monograph's concluding position is that rational belief revision is the accountable reorganization of epistemic commitments in accordance with evidence, logical consequence, justified preservation priorities, and the requirements of continuing inquiry. This conception integrates the principal achievements of the field while retaining the philosophical questions its formal systems have made visible.