betaFeedback

WorksComputer Science

A Structural Framework for Succession: Continuation, Recognition, and Realization

Abstract

Succession asks whether an entity encountered later is the same entity that entered an arrangement earlier. For an artificial system that is fine-tuned, distilled, or copied, it has to be answered. We look for a mathematical structure weak enough to cover any such case and strong enough to support general theorems, so that results follow from the structure's conditions rather than from a particular pipeline or criterion. Succession is modeled on state/modification graphs with relational lifts that compose along paths. Within this framework we separate four questions that are often conflated: which continuations a representation admits, which descendants are compatible with a modification, which are recognized as continuing the identity, and which are realized. The framework allows one to address systematically questions of symmetry, computational complexity, recoverability, and robustness to partial realization, and we establish what the structure settles in each case before any particular criterion of identity is chosen.

Introduction

Any claim that an entity remains a party to a continuing arrangement presupposes a prior question: what makes the entity encountered later the same entity as the one that entered the arrangement earlier? For human and institutional parties, established practices of identity and succession usually make this question invisible. For an artificial system that is fine-tuned, distilled, copied, or otherwise modified, it becomes explicit. Although the motivating cases are drawn from AI transformation pipelines, the framework is stated for succession generally: the same structure applies to institutional, legal, and biological lineages in which a later entity is claimed to continue an earlier one [1, 2, 3]. Structural continuity may preserve parameters, circuits, dispositions, or records without determining whether a particular descendant should count as the continuing entity; conversely, structural change need not by itself establish that the earlier entity has ceased to exist.

The object of succession is what we will call an identity characteristic: a characteristic that a governance arrangement attached to an original system treats as carrying forward through transformations of that system. The characteristic is left uninterpreted by the framework. It may be an alignment property such as corrigibility or honesty under adversarial pressure, a legal obligation attached to a corporate entity, a personal attribute recognized by a lineage tradition, or a property of a biological organism. What matters for the formalism is only that some arrangement treats the characteristic as inheriting, and that structural continuity may or may not carry it. The question is what a structural account of a system’s transformations can establish about the inheritance of such a characteristic, and where that account stops.

Four notions are at work. The target property is what the arrangement intends to carry forward, for example honesty. The structural proxy is the continuation that the representation tracks for it. The recognition criterion is the rule by which a governing process decides which descendants inherit the proxy. The governance act is the decision to treat a particular successor as bearing the property. The underdetermination results are statements about the recognition criterion relative to a stipulated representation; whether a tracked continuation carries the property is a question for the arrangement.

The framework separates stipulating a structural representation, determining which continuations are realized, recognizing which continuations carry the identity, and correcting an exclusion. If the adopted representation admits a symmetry exchanging two descendants of a common predecessor, and the decision domain respects that symmetry, then every recognition criterion defined on the representation assigns them the same status. In this case, the asymmetry must enter from outside.

The paper does not propose a criterion of succession. Instead, it asks how much can be said about succession before any such commitment is made. Every commitment buys sharpness at the cost of generality, and a framework that fixes any of them can answer questions only about the cases it has chosen. The aim is to find a weak yet tractable structure in which general succession questions can be posed, and then to answer as many of them as the structure permits. The definitions are chosen for that purpose: rich enough that the questions have well-posed answers, weak enough that the answers hold for any choice of continuation, realization, and criterion.

Requirements. We take the following as the starting point as these requirements most strongly accord with intuitive notions of succession.

(R1) Direction without reversal. A later state continues an earlier one through a sequence of changes, and a change such as fine-tuning or deletion cannot in general be undone. States and modifications form a directed structure, and modifications are not assumed invertible.

(R2) Three fates. A continuation may have no descendant, one, or several: fission and loss are part of the subject, not exceptions to it. Lifts are therefore relations, not functions, and the statuses (§ 3) record only whether a lift is empty, a singleton, or larger.

(R3) Independence of grouping. What a pipeline does to a continuation is what its steps do in succession. If it depended on how the steps were grouped, “does the characteristic survive the pipeline?” would have no answer fixed by the pipeline. This is functoriality.

(R4) Independence of continuations. Whether a continuation has a descendant does not depend on which other continuations are present. A modification then acts on a set of continuations member by member, which is what makes a lift computable element by element. This is additivity. It is built into taking relations as the primitive, so the commitment is pointwise action: it excludes joint effects of co-present continuations.

(R5) Neutrality to names. Labels are arbitrary. A criterion that is purely structural cannot separate what the structure cannot tell apart, which is made precise by invariance under automorphisms (§ 5).

(R6) Separation of questions. Which continuations a representation admits, which descendants are compatible with a modification, which are recognized as continuing the identity, and which are realized are different questions. Keeping them apart is what allows statements about recognition without committing to a particular criterion (§ 4 and § 8).

(R7) Correctable exclusion. A wrongful exclusion may be corrected later, so what remains reachable from an excluded coordinate matters (§ 7).

(R8) Checkability. A claim about a lineage is useful only if the cost of deciding it can be stated (§ 7.1).

Together, (R2)–(R4) say that a modification acts on sets of continuations as a relation, and that relations compose. Direction (R1) then gives a directed graph, and the neutrality of (R5) gives a group action on it. The graph with its relations is a nondeterministic automaton and its path structure is a layered flow network, so results from group theory, automata theory and graph theory apply to it once the correspondence is established. Requirements (R1)–(R4) leave this relational structure. (R5) and (R8) narrow the framework without determining it: (R5) is made precise by automorphism, the strongest notion of indistinguishability, and coarser notions such as bisimulation are possible but not developed; (R8) asks only that costs be statable, not that they take any particular form. Irrevocable loss is a consequence and not a requirement: the empty set is absorbing under relational composition, so a lineage with no descendant at some stage has none at any later stage.

Each limitation of the framework, with one exception, is the cost of one of these requirements. Joint effects of co-present continuations cannot be expressed without refining the states, because of (R4). Pipelines in which grouping matters need a finer generating graph, because of (R3). Convergence and plural succession are not classified, because (R2) follows the descendants of a single predecessor. The continuation spaces and relations are stipulated and not derived, because (R6) leaves open what a continuation is. The complexity results are worst-case on explicit encodings: (R8) asks that costs be statable, and the worst case over explicit encodings is the first statement of that kind. The exception is a modeling choice and not the cost of a requirement: the recognition criterion is static, which gives a single recognition function on which invariance can be stated. A criterion that changes over time while leaving completed decisions in place is coherent, but is not treated here (§ 5). The framework adopts no commitment beyond these requirements and this choice.

Related work. The symmetry result is the formal counterpart of the tie that closest-continuer accounts of identity face under fission [1, 4, 5]. Such a tie arises when two successors are equally close to the predecessor; an automorphism exchanging two descendants is an exact tie of that kind in the structural sense. In social choice it is the familiar tension between neutrality or anonymity and resoluteness [6], and in distributed computing it parallels the impossibility of deterministic leader election in symmetric anonymous networks [7]. The structure studied here is a nondeterministic automaton (§ 2), and the refinement example of § 6 uses bisimulation [8]. The symmetry corollary follows directly from the definition of a structural criterion; its content lies in that formulation and in the orbit characterization of exclusive criteria. The Recover_avail reductions depend on the paper’s own definitions and are given in full in the appendices.

Roadmap. § 2 defines the continuation structure and its composition conditions, and § 3 classifies persistence along a path. § 4 introduces the decision domain and the recognition record, that is, what recognition must supply where the structure branches. § 5 shows when no structural criterion can supply it, because a symmetry exchanges the candidates, and § 6 works through an abstract model in which this occurs. § 7 asks whether a failed recognition can later be corrected, and § 8 asks how robust a lineage is when the stipulated structure is only partly realized. Several of these questions also have a computational side, namely how hard they are to audit, and § 7.1 collects the classification. Proofs and long constructions are in Appendices § 12–§ 15; notation is collected in Appendix § 16.

Historical and Continuation Structure

We model the history of a lineage by a directed state/modification structure HH. Its vertices are system states—checkpoints, snapshots, deployed versions—and its directed edges are elementary modifications between those states: fine-tuning runs, distillation steps, copies, shutdowns, restarts. We also include identity modifications and the composites of directed paths. Composition represents successive modifications. We do not assume modifications are invertible. A maximal generating history is a path in the chosen generating graph that cannot be extended by another generating modification; its composite is a modification in HH. Finite paths represent finite histories; a finite path may or may not be extendable.

Attach to each state xx a continuation space C(x)C(x). An element γ∈C(x)\gamma\in C(x) is a continuation the adopted structural representation recognizes at xx. The term “continuation” is deliberately broad. Depending on the application it may denote a mechanistically identified circuit [9], a behavioral disposition, or a functional capacity. Examples include refusal behavior, honesty under adversarial pressure, corrigibility [10], an interpretable safety circuit [9], or the capacity to represent and report uncertainty.

The continuation space is part of the adopted representation, and its choice determines the granularity at which persistence and branching are assessed. A coarser space may collapse distinct mechanisms or behaviors, while a finer space may classify almost every modification as loss or branching. The framework treats continuations as given. Determining whether a transformed system preserves “honesty under adversarial pressure,” for instance, is an empirical question that the framework does not resolve. The framework is a language for talking about such properties once they have been identified.

Questions of stakeholder standing are governance questions and are not part of the definition of C(x)C(x). The admissible scope of C(x)C(x) is therefore contestable by parties recognized by the governing process, and the criteria for such standing are external to the formalism.

For each modification f:x→yf : x \to y, define a compatibility relation

Rf⊆C(x)×C(y).R_f \subseteq C(x) \times C(y).

For γ∈C(x)\gamma\in C(x), its lift under ff is

Rf(γ)={δ∈C(y):(γ,δ)∈Rf}.(1)R_f(\gamma) = \{\delta\in C(y) : (\gamma,\delta) \in R_f\}. \tag*{(1)}

So Rf(γ)R_f(\gamma) is the set of continuations at yy that the adopted structural representation recognizes as structurally compatible descendants of γ\gamma under ff. The relation is more general than a function on purpose: ∣Rf(γ)∣|R_f(\gamma)| may be 0, 1, or greater than 1.

Remark 1. A lift may disappear, remain unique, or branch. In the abstract model of § 6, the lift of γ2\gamma_2 under dd branches into γ4S\gamma_4^S and γ4W\gamma_4^W.

We require the relations to respect composition. For all composable modifications f:x→yf : x \to y and g:y→zg : y \to z (including composites, when ff or gg is itself a composite),

Rg∘f=Rg∘Rf,(2)R_{g \circ f} = R_g \circ R_f, \tag*{(2)}

equivalently,

Rg∘f(γ)=Rg(Rf(γ))for every γ∈C(x).(3)R_{g \circ f}(\gamma) = R_g(R_f(\gamma)) \qquad\text{for every } \gamma\in C(x). \tag*{(3)}

We also require

Ridx=idC(x)(4)R_{\mathrm{id}_x} = \mathrm{id}_{C(x)} \tag*{(4)}

for every state xx. Together with the composition condition, this means that the continuation assignment and compatibility relations are consistent across successive modifications: the relation attached to a composite modification is exactly the relational composition of the relations attached to its constituent modifications, and identities act as identity relations.

Composition has two parts. Two conditions on the relations underlie every path-indexed object in the paper; they are requirements (R3) and (R4) of the Introduction, stated here formally. Both are commitments.

Functoriality. Equations (2)–(4) require that composite modifications carry composite relations and that identities act as identities. Over the free category on the generating graph, any one-step data extends uniquely to such a family, so there it defines the relations on composites. It constrains only a non-free presentation, in which parallel paths are identified and must then carry equal relations; a pipeline whose parallel paths differ requires a finer generating graph that distinguishes them. Functoriality is what makes the iterated lifts of § 3 functions of the path.

Additivity (pointwise action). Lifts of sets are formed by union (Eq. (7)): for Γ⊆C(x)\Gamma\subseteq C(x),

Rf(Γ)=⋃γ∈ΓRf(γ).R_f(\Gamma)=\bigcup_{\gamma\in\Gamma}R_f(\gamma).

Once RfR_f is a relation and its action on sets is defined by this union, additivity holds by construction; it is not an extra axiom on relations. The commitment is the choice of relations as the primitive, that is, pointwise action. Whether a descendant of γ\gamma survives ff therefore never depends on which other continuations are present. This is the substantive restriction. It excludes joint effects. Suppose α′,β′∈C(y)\alpha^\prime,\beta^\prime\in C(y) and that f2:y→zf_2:y\to z should produce αz\alpha_z when both are present but nothing when either is present alone. No relation Rf2R_{f_2} does this: additivity gives Rf2({α′,β′})=Rf2(α′)∪Rf2(β′)R_{f_2}(\{\alpha^\prime,\beta^\prime\})=R_{f_2}(\alpha^\prime)\cup R_{f_2}(\beta^\prime), so a nonempty value for the pair forces a nonempty value for one of its members.

Such an interaction can be encoded by refining the states. If the pair arises as the lift of a single α\alpha under an earlier f1f_1, let C(y)C(y) contain one continuation ϖ\varpi standing for “both present,” with Rf1(α)={ϖ}R_{f_1}(\alpha)=\{\varpi\} and Rf2(ϖ)={αz}R_{f_2}(\varpi)=\{\alpha_z\}. Then Rf2(Rf1(α))={αz}R_{f_2}(R_{f_1}(\alpha))=\{\alpha_z\}. The interaction is thus fixed when f1f_1 is applied; a dependence that is not determined at that point cannot be encoded this way, and the burden of anticipating it falls on the modeler. Refinement makes additivity a modeling choice. The price is that the symmetry and complexity results below are relative to the chosen granularity of the continuation spaces.

Additivity also implies monotonicity: if Γ⊆Γ′⊆C(x)\Gamma\subseteq\Gamma^\prime\subseteq C(x) then Rf(Γ)⊆Rf(Γ′)R_f(\Gamma)\subseteq R_f(\Gamma^\prime). A modification therefore cannot remove a descendant because another continuation is present, and a dependence of the form “δ\delta survives only if ε\varepsilon is absent” has to be encoded by a refinement that records, at the earlier step, whether ε\varepsilon is present.

A consequence is that history-dependent behavior must be encoded in the states or in a family of modifications indexed by the relevant history. Compatibility is a relation attached to the modification: parallel modifications between the same states may carry different relations.

Remark 2. Whether δ\delta is a descendant of γ\gamma depends only on the modification ff and the two states. The abstract model relies on this: the two paths PAP_A and PBP_B reach the same state x4x_4 through different modifications, and the lifts of α\alpha and β\beta are the same on both, while the lift of γ\gamma differs.

Relation to automata. The structure (H,C,R)(H,C,R) is a Rel-valued functor on the generating graph: a set-valued assignment to states together with relations on edges, composed by relational composition. This is exactly a nondeterministic automaton whose states are the pairs (x,γ)(x,\gamma) with γ∈C(x)\gamma\in C(x), that is, the disjoint union ⨆x∈XC(x)\bigsqcup_{x\in X}C(x); whose alphabet is the set of generating modifications; and whose transition relations are the RfR_f. The vertices of HH index the continuation fibers; the automaton’s states are the elements of those fibers. The complexity reductions of § 7.1 are classical many-one reductions from classical problems (NFA non-universality, NFA-intersection nonemptiness, CNF satisfiability), but they are not routine: the target is not an arbitrary automaton but an instance of the framework, and the reductions must respect functoriality and additivity. The layered constructions that enforce these conditions, and the verifications that they are faithful, are in Appendices § 12 and § 13.

Persistence Along a History

Let

p=(f1,…,fn)p = (f_1,\ldots,f_n)

be a finite path beginning at x0x_0, and let γ0∈C(x0)\gamma_0 \in C(x_0). Define the iterated lifts by setting

L0(γ0)={γ0}(5)L_0(\gamma_0) = \{\gamma_0\} \tag*{(5)}

and then recursively defining

Li(γ0)=Rfi(Li−1(γ0))(i=1,…,n),(6)L_i(\gamma_0) = R_{f_i}\left(L_{i-1}(\gamma_0)\right) \qquad(i=1,\ldots,n), \tag*{(6)}

where for a subset Γ⊆C(x)\Gamma\subseteq C(x) we write

Rf(Γ)=⋃γ∈ΓRf(γ).(7)R_f(\Gamma) = \bigcup_{\gamma\in\Gamma} R_f(\gamma). \tag*{(7)}

For a singleton we write Rf(γ)R_f(\gamma) for Rf({γ})R_f(\{\gamma\}), as in Eq. (1); the two are the same set. For a path p:x⇝yp:x\rightsquigarrow y, we write Rp(γ)R_p(\gamma) for the resulting iterated lift. By functoriality (Eq. (2)) and additivity (Eq. (7)) it equals the relation of the composite modification,

Ln(γ)=Rfn∘⋯∘f1(γ),(8)L_n(\gamma) = R_{f_n\circ\cdots\circ f_1}(\gamma), \tag*{(8)}

as induction on nn shows: Ln(γ)=⋃ϵ∈Ln−1(γ)Rfn(ϵ)L_n(\gamma)=\bigcup_{\epsilon\in L_{n-1}(\gamma)}R_{f_n}(\epsilon), which is Rfn∘Rfn−1∘⋯∘f1(γ)R_{f_n}\circ R_{f_{n-1}}\circ\cdots\circ f_1(\gamma).

For a finite path pp, define three structural statuses. We say that γ\gamma is dead along pp if Li(γ)=∅L_i(\gamma)=\varnothing for some stage ii. We say that γ\gamma is preserved along pp if ∣Li(γ)∣=1|L_i(\gamma)|=1 for every stage. We say that γ\gamma is diluted along pp if every lift is nonempty but some lift has cardinality greater than one. Formally,

Dead⁡p(γ)⟺Li(γ)=∅ for some i,(9)\operatorname{Dead}_p(\gamma) \Longleftrightarrow L_i(\gamma)=\varnothing\text{ for some }i, \tag*{(9)}
Pres⁡p(γ)⟺∣Li(γ)∣=1 for every i,(10)\operatorname{Pres}_p(\gamma) \Longleftrightarrow|L_i(\gamma)|=1\text{ for every }i, \tag*{(10)}
Dil⁡p(γ)⟺∣Li(γ)∣>0 for every i and ∣Li(γ)∣>1 for some i.(11)\operatorname{Dil}_p(\gamma) \Longleftrightarrow|L_i(\gamma)|>0\text{ for every }i\text{ and }|L_i(\gamma)|>1\text{ for some }i. \tag*{(11)}

These three cases are exhaustive and mutually exclusive. Write

status⁡p(γ)∈{Dead⁡,Pres⁡,Dil⁡}\operatorname{status}_p(\gamma)\in\{\operatorname{Dead},\operatorname{Pres},\operatorname{Dil}\}

for the unique status whose defining predicate holds. Every finite path therefore assigns each continuation exactly one of the three statuses. “Preserved”, “diluted” and “dead” are structural statuses relative to the stipulated relations: they describe the representation and assert no loss or weakening of any property in a real system.

Scope of the statuses: forward branching. The terminology is structural. “Preserved” means that the continuation has a unique compatible descendant at every stage of the path. Dilution is a historical property, not a terminal-state property. A diluted continuation remains structurally present through several descendants, but the history no longer identifies a unique successor.

The statuses track forward branching. If Rf(α)=Rf(β)={δ}R_f(\alpha)=R_f(\beta)=\{\delta\} for distinct α,β∈C(x)\alpha,\beta\in C(x), each of α\alpha and β\beta has a singleton lift and each counts as preserved, while the distinction between them is lost at yy. Convergence of several continuations onto one is as common as fission in transformation pipelines, and the framework as stated does not classify it. The three statuses record whether a continuation has no descendant, one descendant, or several; they do not record whether several continuations have converged on one. A treatment of convergence would require a separate classification on the reverse structure, and is left to future work. Plural succession, in which a successor is constituted by the combination of two or more predecessors, is a different case. A product construction on continuation spaces extends the representation to this case at the cost of generalizing modifications to multi-source joins and compatibility relations to higher arity; we outline this extension, without proof, in the Discussion.

A notion of status-equivalence may be defined by declaring γ,δ∈C(x)\gamma, \delta\in C(x) equivalent when they have the same status along every finite path beginning at xx. This records, for each finite path from xx, whether each lift is empty, a singleton, or larger; it does not record which descendants the lift contains, and two status-equivalent continuations may have entirely different descendant sets. It is used only in the refinement example of § 6.

References

  1. [1]Parfit, D. (1984). Reasons and Persons. Oxford University Press.
  2. [2]Olson, E. T. (1997). The Human Animal: Personal Identity Without Psychology. Oxford University Press.
  3. [3]Shoemaker, S. (1963). Self-Knowledge and Self-Identity. Cornell University Press.
  4. [4]Nozick, R. (1981). Philosophical Explanations. Harvard University Press.
  5. [5]Lewis, D. (1976). Survival and identity. In A. O. Rorty (ed.), The Identities of Persons, University of California Press, 17–40.
  6. [6]Moulin, H. (1983). The Strategy of Social Choice. North-Holland.
  7. [7]Angluin, D. (1980). Local and global properties in networks of processors. In Proceedings of the 12th Annual ACM Symposium on Theory of Computing (STOC), 82–93.
  8. [8]Park, D. (1981). Concurrency and automata on infinite sequences. In Theoretical Computer Science, 5th GI Conference, Lecture Notes in Computer Science 104, Springer, 167–183.
  9. [9]Elhage, N., Nanda, N., Olsson, C., et al. (2021). A Mathematical Framework for Transformer Circuits. Transformer Circuits Thread.
  10. [10]Soares, N., Fallenstein, B., Yudkowsky, E., and Armstrong, S. (2015). Corrigibility. In AAAI Workshops: Workshops at the Twenty-Ninth AAAI Conference on Artificial Intelligence, pp. 74–82.
  11. [11]Karp, R. M. (1972). Reducibility among combinatorial problems. In Complexity of Computer Computations, Plenum Press, 85–103.
  12. [12]Alpern, B. and Schneider, F. B. (1985). Defining Liveness. Information Processing Letters 21(4), 181–185.
  13. [13]Arora, S. and Barak, B. (2009). Computational Complexity: A Modern Approach. Cambridge University Press.
  14. [14]Kozen, D. (1977). Lower bounds for natural proof systems. In Proceedings of the 18th Annual Symposium on Foundations of Computer Science (FOCS), IEEE, 254–266.DOI
  15. [15]Meyer, A. R. and Stockmeyer, L. J. (1972). The equivalence problem for regular expressions with squaring requires exponential space. In Proceedings of the 13th Annual Symposium on Switching and Automata Theory, IEEE, 125–129.DOI
  16. [16]Savitch, W. J. (1970). Relationships Between Nondeterministic and Deterministic Tape Complexities. Journal of Computer and System Sciences 4(2), 177–192.DOI
  17. [17]Menger, K. (1927). Zur allgemeinen Kurventheorie. Fundamenta Mathematicae 10, 96–115.DOI
  18. [18]Diestel, R. (2017). Graph Theory, 5th ed. Springer, Graduate Texts in Mathematics 173.
  19. [19]Zhu, G. and Wang, C. (2026). Authorization for self-modifying AI agent populations: conserving authority across replacement, forking, and rollback. arXiv preprint arXiv:2610.00347.arxiv.org/abs/2610.00347

Paper details

vlnc:W_h7mqy42x
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Computer Science

Not recorded for this paper

View statement
Looking for related papers…
Contents