Category-Theoretic AI Identity

• Category-Theoretic AI Identity is a mathematical framework that defines persistence and self-reference using fixed points, self-update endofunctors, and categorical invariants.
• It stratifies identity into structural, behavioral, and provenance-based criteria, clarifying the differences between strict equality and equivalence under change.
• Empirical and theoretical studies illustrate identity through Lawvere metric spaces, copresheaves, and transport-aware updates, offering actionable insights for AI system design.

Category-theoretic formalization of AI identity denotes a family of mathematical approaches that treat persistence, self-sameness, and transformation in AI systems as categorical structure rather than as a single primitive relation. In recent work, identity is formalized as a least fixed point of a self-update endofunctor, as non-geodesic structure in a Lawvere-enriched behavioral metric space, as the isomorphism class of a copresheaf state together with its realized provenance, as thin-category reachability among trustworthiness-preserving lifecycle states, and as layered structural, behavioral, and property-based equivalence of agent architectures (Alpay, 23 May 2025, Tanner, 20 Jun 2026, Wang et al., 31 May 2026, Ferrario, 30 Jun 2026, Riscos et al., 30 Mar 2026). The resulting literature does not yield one canonical criterion of AI identity; rather, it yields a hierarchy of criteria adapted to fixed-regime persistence, recursive self-reference, regime revision, behavioral drift, and governance transfer.

1. Arrow-first foundations and the hierarchy of sameness

A foundational strand begins from objectless and category-free formulations of category theory. In that setting, the primitive data are morphisms and their partially defined compositions, with identity morphisms replacing objects as the formal carriers of source and target. The paper "Category Free Category Theory and Its Philosophical Implications" states that objects can be eliminated by definition, but that the formal identity of objects remains essential through identity morphisms; it also develops the hierarchy of identities given by identity morphisms, isomorphism, natural isomorphism, and equivalence of categories (Heller, 2016). This establishes an arrow-first ontology in which identity is contextual and compositional rather than merely label-theoretic.

A related biological and enactive line reformulates autonomy as closure. In "Reformalizing the notion of autonomy as closure through category theory as an arrow-first mathematics," a monoid is a category with a single object, and operational closure is recast as a submonoid $O \subseteq \mathrm{End}(A)$ containing $1_A$ and closed under composition. The identity arrow $1_A$ functions as the categorical marker of persistence, while the coslice category $A/\Omega$ makes explicit the internal web of enabling relations and gives $1_A$ the role of an initial object in that coslice construction (Hirota et al., 2023). For AI identity, this furnishes a minimal formal vocabulary in which persistence can be expressed as closure of self-mediations under composition.

The comparative AGI literature generalizes this hierarchy beyond single agents. "Working Paper: Towards a Category-theoretic Comparative Framework for Artificial General Intelligence" models architectures as triples $A := (G_A, Know_A, \Phi_A)$, agents as strong monoidal functors into an implementation category, and properties through an institution and proof-carrying certificates. Within that scaffold, identity is stratified into structural identity, behavioral identity, and property-based identity: structural identity is isomorphism in the category of architectures or agents, behavioral identity is isomorphism in a behavior category under a behavior functor, and property-based identity is captured by equalizers of property valuations (Riscos et al., 30 Mar 2026).

Taken together, these foundations fix a recurring categorical pattern: identity may be strict, invertible, functorial, or only equivalential. A plausible implication is that disputes about whether an AI system remains “the same” often arise from mixing these levels.

2. Fixed-point emergence, observer coupling, and recursive self-reference

A second major formalization identifies identity with a fixed point of self-update. "Alpay Algebra II: Identity as Fixed-Point Emergence in Categorical Data" works in a small cartesian closed category $C$ and takes $\phi: C \to C$ to be an endofunctor encoding the self-referential update/operator of an agent’s self-model. Identity is a fixed-point object $I$ equipped with an isomorphism $\iota: I \cong \phi(I)$. The construction proceeds by a transfinite initial chain from an initial object $1_A$0: $1_A$1, $1_A$2, and $1_A$3 at limit ordinals. When the canonical map $1_A$4 is an isomorphism, the stabilized object $1_A$5 is $1_A$6, and under the stated hypotheses it is the initial $1_A$7-algebra $1_A$8, the universal least solution of $1_A$9 (Alpay, 23 May 2025).

In that framework, the least fixed point is not merely a static identifier. The paper explicitly interprets the transfinite chain as an evolving trace of a self-updating agent, and interprets stabilization as the point at which applying $1_A$0 yields no new structure. The colimit stages internalize prior approximants, yielding the paper’s three characteristic interpretations of identity: symbolic memory, recursive coherence, and semantic invariance. The universal property

$1_A$1

expresses the compatibility of stabilized identity with any further recursive definition on $1_A$2 (Alpay, 23 May 2025).

"Alpay Algebra III: Observer-Coupled Collapse and the Temporal Drift of Identity" extends this fixed-point program by introducing observer and verification structure. It adds an observer functor $1_A$3, a verification functor $1_A$4, a natural automorphism $1_A$5 as phase structure, and an observer-indexed collapse family $1_A$6. The resulting ordinal-indexed flow is $1_A$7, with colimit or limit constructions at limit ordinals according to cumulative or convergent interpretation. The paper defines curvature by the commutator

$1_A$8

so that nonzero curvature measures noncommuting observer coupling and hence temporal drift; it also introduces entropy-based bounds and a Lyapunov functional $1_A$9 to control drift (Alpay, 26 May 2025).

This recursive literature therefore gives a strong ontological thesis: identity is the stabilized outcome of self-reference. In Alpay II it is the least fixed point $A/\Omega$0; in Alpay III it is a fixed-point or coalgebraic structure modulated by verification, phase, and observation. The common claim is that identity emerges from within the logic of categorical change.

3. Behavioral geometry, non-geodesicity, and measurable drift

A third line formalizes identity behaviorally rather than recursively. "Measuring What Persists: Conditioning Mechanisms and a Geometric Framework for AI Agent Identity" models an AI agent’s behavior space as a Lawvere metric space enriched over $A/\Omega$1, with probe contexts $A/\Omega$2 as objects and behavioral distances defined by information geometry. At each probe $A/\Omega$3, an empirical response distribution $A/\Omega$4 is collected, and the paper defines

$A/\Omega$5

using the fact that $A/\Omega$6 is a proper metric bounded by $A/\Omega$7. The paper’s conceptual claim is that identity is non-geodesic structure, while drift is relaxation toward the geodesic (Tanner, 20 Jun 2026).

Magnitude and magnitude homology supply the invariants. For a finite metric space $A/\Omega$8, the similarity matrix is $A/\Omega$9, and magnitude is

$1_A$0

Magnitude homology is then built from chains $1_A$1 of fixed total length $1_A$2, with a boundary operator that removes internal points only when the geodesic equality

$1_A$3

holds. Chains protected from this cancellation by strict triangle inequality survive in homology, so nontrivial $1_A$4 records non-geodesic organization (Tanner, 20 Jun 2026).

The paper also places this geometry into enriched semantics of language. Bradley–Terilla–Vlassopoulos are cited as showing that a LLM defines a $1_A$5-enriched category of texts with $1_A$6, and the paper extends that framing by treating the identity specification as a copresheaf assigning characteristic response distributions to contexts. In that setting, a geodesic factorization corresponds to $1_A$7, while identity-specific responses are precisely those that do not factorize (Tanner, 20 Jun 2026).

The strongest empirical finding reported in that work is a two-mechanism conditioning structure revealed by cross-condition distances in a seven-probe battery.

Cluster	Probes	Cross-condition separation
Identity-vacuum	Q3, B3, B1	0.8326, 0.8326, 0.7361
Safety-basin/intermediate	Q1, B2, B4, Q2	0.2762, 0.4169, 0.5000, 0.6103

Within-condition geometry for the three sentinel probes $1_A$8, $1_A$9, and $A := (G_A, Know_A, \Phi_A)$0 was equilateral at maximum separation $A := (G_A, Know_A, \Phi_A)$1 with zero bootstrap variance in both Card and base conditions, and scalar magnitude $A := (G_A, Know_A, \Phi_A)$2. The paper reports that Card-conditioned $A := (G_A, Know_A, \Phi_A)$3 produced 55 unique 10-token prefixes at $A := (G_A, Know_A, \Phi_A)$4, versus 1 for the base model, and reports a cross-condition scalar magnitude of $A := (G_A, Know_A, \Phi_A)$5 on the seven-probe battery (Tanner, 20 Jun 2026).

The drift analysis in that same paper is explicitly qualified. A magnitude decrease observed under repetitive padding was later found to be an artifact; with diverse padding, the paper reports no measurable deformation through 150K tokens, with within-condition distances remaining at $A := (G_A, Know_A, \Phi_A)$6 to eight decimal places and magnitude and magnitude homology unchanged. The same study develops first-order perturbation theory for equilateral configurations and argues that $A := (G_A, Know_A, \Phi_A)$7 symmetry cancels first-order shape perturbations, leaving only perimeter-sensitive “breathing” effects in magnitude. This sharply distinguishes uniform contraction from structural collapse (Tanner, 20 Jun 2026).

4. Provenance, copresheaves, and identity across regime change

A fourth formalization treats identity as transport-aware persistence of typed knowledge states under self-revision. "Self-Revising Discovery Systems for Science: A Categorical Framework for Agentic Artificial Intelligence" defines a regime $A := (G_A, Know_A, \Phi_A)$8, where $A := (G_A, Know_A, \Phi_A)$9 is a small schema category of artifact types and admissible operations. The AI system state at time $C$0 is a copresheaf $C$1, and the category of elements $C$2 realizes typed provenance: objects are pairs $C$3, and morphisms $C$4 are schema arrows $C$5 such that $C$6 (Wang et al., 31 May 2026).

Within a fixed regime, identity is defined as the isomorphism class $C$7 in $C$8 together with the isomorphism class of realized provenance $C$9. This makes identity explicitly type-sensitive and lineage-sensitive. Fixed-regime updates are endofunctorial only under provenance-preserving refinements, taken in the paper’s strict presentation as componentwise injective natural transformations. Injectivity is not incidental: it models the audit rule that accepted distinct artifacts are never silently identified (Wang et al., 31 May 2026).

Across regimes, identity becomes a transport problem. A verified regime transition consists of a schema functor $\phi: C \to C$0 and a preservation map $\phi: C \to C$1, subject to faithfulness, injectivity, and gate conditions. Old artifacts are transported by left Kan extension,

$\phi: C \to C$2

and compared with the post-transition state by the adjoint transpose $\phi: C \to C$3. The image $\phi: C \to C$4 is the transported-evidence subobject, while residual content beyond transport is formalized as novelty $\phi: C \to C$5, either as a complement-like subobject or as a quotient when admissible (Wang et al., 31 May 2026).

The paper therefore defines identity across regime transitions as preservation “up to verified residuals.” Search, retrieval, and discovery are separated categorically: retrieval is refinement within the same schema, search is a fixed-regime endofunctorial update, and discovery is a verified regime transition with Kan transport and residual identification. In the CategoryScienceClaw setting, this is further embedded in a proof-carrying knowledge–computation graph in which typed artifacts, skills, open needs, gates, stress tests, and public discourse are all typed and provenance-bearing (Wang et al., 31 May 2026).

This transport-aware view changes the meaning of persistence. Identity is not merely what is preserved; it is what is preserved under explicit recoding, verification, and declaration of residual novelty.

5. Reachability categories, realized histories, and architecture-level comparison

A fifth strand makes temporal organization explicit. "A Category Theory Account of AI Identity" begins with an AI system type datum $\phi: C \to C$6, where $\phi: C \to C$7 is a techno-function, $\phi: C \to C$8 is a trustworthiness profile, and $\phi: C \to C$9 maps quantified assessments $I$0 to a finite set of trustworthiness levels $I$1. A $I$2-relative state is $I$3, with $I$4. Primitive lifecycle transformations generate a free path category; restricting to trustworthiness-level-preserving paths and quotienting by the paper’s trustworthiness equivalence yields the AI system state category

$I$5

The key structural result is that $I$6 is thin: $I$7 iff $I$8, and whenever nonempty the hom-set is a singleton (Ferrario, 30 Jun 2026).

That thinness generates the paper’s distinction between weak and strong identity. Weak state identity is equality of trustworthiness level, $I$9. Strong state identity is categorical isomorphism $\iota: I \cong \phi(I)$0, which in a thin category is mutual trustworthiness-preserving reachability. Histories are functors $\iota: I \cong \phi(I)$1 with time-ordered representatives, and realized histories form a category $\iota: I \cong \phi(I)$2 whose morphisms are time-synchronous natural transformations. Strong synchronic identity is then natural isomorphism of realized histories, while weak synchronic identity is pointwise level equality induced by the existence of a history morphism (Ferrario, 30 Jun 2026).

This temporal account recovers earlier propositional trustworthiness-based identity criteria but enriches them categorically. It distinguishes equality of level from directed reachability, and directed reachability from mutual reachability. It also makes explicit that level drops interrupt identity-preserving history functors, because no single history in $\iota: I \cong \phi(I)$3 can contain both sides of a level change (Ferrario, 30 Jun 2026).

The comparative AGI framework broadens this further to architectural comparison. In that literature, the total category $\iota: I \cong \phi(I)$4 is a Grothendieck construction over architectures, and the projection $\iota: I \cong \phi(I)$5 is a fibration. Structural identity is isomorphism in $\iota: I \cong \phi(I)$6 or in $\iota: I \cong \phi(I)$7; behavioral identity is isomorphism in a behavior category $\iota: I \cong \phi(I)$8 under a behavior functor $\iota: I \cong \phi(I)$9; property-based identity is membership in the equalizer of property valuation maps. The paper also writes machine isomorphism in the standard commuting form

$1_A$00

and specializes behavioral identity to RL, CRL, and SBL settings (Riscos et al., 30 Mar 2026).

Across these frameworks, category theory replaces one undifferentiated identity relation with a structured hierarchy of equality, reachability, isomorphism, and natural isomorphism. This suggests that whether two AI systems are “the same” depends on which categorical layer is operationally relevant: state, history, architecture, behavior, or certified property profile.

6. Assumptions, failure modes, and open directions

The literature is explicit that each formalization depends on strong structural assumptions. In the fixed-point setting of Alpay II, existence of $1_A$01 requires an initial object, colimits of ordinal-indexed chains, and preservation of those colimits by $1_A$02; if $1_A$03 does not preserve the relevant colimits, or if the ambient category lacks them, the initial chain may fail to stabilize. The same paper stresses that multiple fixed points can exist, and that the relevant uniqueness claim is uniqueness of the least fixed point up to unique isomorphism in $1_A$04-Alg, not uniqueness among all fixed points in $1_A$05 (Alpay, 23 May 2025).

The arrow-first philosophical literature makes a related restriction at the ontological level. Even when objects are eliminated and categories are reformulated solely in terms of morphisms, formal identity does not disappear: identity morphisms and identity functors remain indispensable. The formalism therefore does not support a simple thesis that AI identity can be reduced away; it supports a shift from object-first identity to compositional identity (Heller, 2016).

The behavioral-geometry approach also records substantial empirical and methodological limits. The observed contraction under context pressure was induced by repetitive padding rather than by context length per se; magnitude homology stability under metric perturbation lacks a general theorem; Card-conditioned magnitude homology configurations were fragile at $1_A$06; thresholds for the Heartbeat were calibrated on artifact-affected data and require recalibration; and the study reports results for a single agent/model, with multi-model replication still needed. The paper also emphasizes that equilateral within-condition geometry characterizes probe design, not identity architecture, and that identity in its experiments manifests as intra-probe richness and cross-condition separation rather than as within-condition distance dispersion (Tanner, 20 Jun 2026).

The governance-oriented lifecycle account imposes further caution. Even strong categorical identity—whether state isomorphism in $1_A$07 or natural isomorphism in $1_A$08—is not sufficient by itself for transferring responsible-AI claims, evidence, or governance procedures across versions or deployments. The paper lists additional conditions including evidential adequacy, operational comparability, robust design of $1_A$09, and lifecycle documentation (Ferrario, 30 Jun 2026). The self-revising discovery framework imposes analogous conditions on verified transitions: faithfulness and injectivity of transport, acceptance of transported evidence, and explicit declaration of residual novelty (Wang et al., 31 May 2026).

Several extensions recur across the corpus. Alpay II mentions enriched settings, probabilistic or higher-categorical lifts, and the coalgebraic dual via terminal coalgebras for persistent, non-well-founded identities (Alpay, 23 May 2025). Alpay III adds learning observer couplings, robustness to adversarial observation, and higher structures such as topos-based semantics and cohomological invariants for identity drift (Alpay, 26 May 2025). The trustworthiness-based history framework points toward an epistemology of AI identity in MLOps and governance practice (Ferrario, 30 Jun 2026). The comparative AGI paper states that it does not explicitly define AI identity, but proposes definitions consistent with its fibration-based framework (Riscos et al., 30 Mar 2026).

A plausible synthesis is that category-theoretic AI identity is best understood as a family resemblance among persistence structures: fixed points for self-reference, non-geodesic invariants for behavior, isomorphism classes for typed provenance, reachability and natural isomorphism for histories, and transport-preserving equivalence for self-revision. The common categorical achievement is not a single criterion, but a disciplined way to state exactly which invariants must survive when an AI system changes.