Topos Causal Models: Categorical Causality

• Topos Causal Models are categorical frameworks integrating variables, mechanisms, and interventions within a topos, merging causal reasoning with sheaf theory and internal logic.
• They use morphisms and subobject classifiers to model interventions, enabling localized, compositional causal inference across varying contexts.
• TCMs provide a unifying structure that subsumes classical causal models while addressing issues like regime-locality, gluing, and contextuality in complex systems.

Topos Causal Models (TCMs) are categorical causal frameworks that place variables, mechanisms, interventions, and causal claims inside a topos or presheaf/sheaf environment, thereby combining causal modeling with subobject classifiers, exponentials, internal intuitionistic logic, and universal constructions such as limits and colimits. Across the current literature, TCMs appear in several closely related formulations: as a category of causal models with objects $\langle U,V,F\rangle$ and morphisms given by commutative squares (Mahadevan, 5 Aug 2025); as causal semantics internal to a sheaf topos $E=\mathrm{Sh}_J(C)$ with stochastic morphisms in $\mathrm{Kl}(\mathrm{Dist}_E)$ and interventions represented by subobjects and kernel replacement (Mahadevan, 20 Oct 2025); and as presheaf-theoretic organizations of local causal data over sites of contexts, histories, or research regions (Gogioso et al., 2023, Mahadevan, 13 May 2026). The common theme is that causal reasoning is localized, compositional, and internal to a categorical setting whose logic is generally Heyting rather than Boolean.

1. Foundational formulations of TCMs

A central formalization defines the category $C_{\mathrm{TCM}}$ with objects $M=\langle U,V,F\rangle$, where $U$ is a set of exogenous variables, $V=\{V_1,\dots,V_n\}$ a set of endogenous variables, and $F:U\to V$ a global function; morphisms $(h,g):M\to M'$ are commutative squares satisfying

$g\circ F = F'\circ h.$

In this formulation, TCM objects are “black box” causal models mapping exogenous to endogenous variables, while structural causal models (SCMs) appear as special cases in which local autonomous mechanisms $E=\mathrm{Sh}_J(C)$0 assemble into a unique global function $E=\mathrm{Sh}_J(C)$1 (Mahadevan, 5 Aug 2025).

A second formulation places TCMs in an ambient causal topos. Fix a small site $E=\mathrm{Sh}_J(C)$2, where objects $E=\mathrm{Sh}_J(C)$3 are stages or contexts, arrows $E=\mathrm{Sh}_J(C)$4 are refinements, and $E=\mathrm{Sh}_J(C)$5 is a Grothendieck topology of covering sieves. The causal topos is then

$E=\mathrm{Sh}_J(C)$6

the topos of $E=\mathrm{Sh}_J(C)$7-sheaves on $E=\mathrm{Sh}_J(C)$8. Random variables are internal objects $E=\mathrm{Sh}_J(C)$9, and stochastic mechanisms are morphisms in the co-Kleisli category $\mathrm{Kl}(\mathrm{Dist}_E)$0 of an internal finite-support distribution monad $\mathrm{Kl}(\mathrm{Dist}_E)$1; for example, a conditional $\mathrm{Kl}(\mathrm{Dist}_E)$2 is a stochastic morphism $\mathrm{Kl}(\mathrm{Dist}_E)$3 (Mahadevan, 20 Oct 2025).

The presheaf-oriented line of work derives TCMs by recasting causal structures over the locale of open lowersets of a space of input histories $\mathrm{Kl}(\mathrm{Dist}_E)$4. For any small poset or site, the category $\mathrm{Kl}(\mathrm{Dist}_E)$5 of presheaves is a topos, and causal data such as $\mathrm{Kl}(\mathrm{Dist}_E)$6, $\mathrm{Kl}(\mathrm{Dist}_E)$7, and $\mathrm{Kl}(\mathrm{Dist}_E)$8 are objects of that topos (Gogioso et al., 2023). This suggests that TCMs are not restricted to a single concrete category; rather, they form a family of causal semantics built in topos-theoretic settings.

This multiplicity of formulations is not merely terminological variation. It indicates that “TCM” names a program in which causal models are treated as objects in a category with enough internal structure to support interventions, equivalence classes of operations, gluing across contexts, and internal logical reasoning. A plausible implication is that the framework is intended to subsume classical SCM-style reasoning while also handling locality, regime dependence, and context-indexed causal information (Mahadevan, 5 Aug 2025, Mahadevan, 20 Oct 2025).

2. Topos structure, sheaves, and continuity-as-causality

One foundational claim is that $\mathrm{Kl}(\mathrm{Dist}_E)$9 has the structure of a topos: it has all limits and colimits, exponential objects, and a subobject classifier (Mahadevan, 5 Aug 2025). Limits and colimits are constructed by diagram-chasing in $C_{\mathrm{TCM}}$0 on the exogenous and endogenous sides, using the commutation constraints of TCM morphisms. Exponentials $C_{\mathrm{TCM}}$1 satisfy the adjunction

$C_{\mathrm{TCM}}$2

and the subobject classifier $C_{\mathrm{TCM}}$3 is explicitly described as a non-Boolean three-valued object $C_{\mathrm{TCM}}$4 with truth arrow $C_{\mathrm{TCM}}$5 (Mahadevan, 5 Aug 2025).

In the presheaf-topological development, the decisive structural insight is that causality is equivalent to continuity with respect to the lowerset topology $C_{\mathrm{TCM}}$6. Opens are down-sets, and a function between posets is continuous for $C_{\mathrm{TCM}}$7 iff it is order-preserving. For an extended function $C_{\mathrm{TCM}}$8, the main equivalence is that $C_{\mathrm{TCM}}$9 is causal, meaning it satisfies consistency or gluing, iff $M=\langle U,V,F\rangle$0 is continuous for $M=\langle U,V,F\rangle$1 (Gogioso et al., 2023). The lowerset specialization preorder induced by $M=\langle U,V,F\rangle$2 is exactly the original partial order.

The sheaf-theoretic status of causal functions is subtler. For each open lowerset $M=\langle U,V,F\rangle$3, one defines

$M=\langle U,V,F\rangle$4

with restriction maps given by ordinary restriction. These assemble into a separated presheaf on the locale of open lowersets. In tight spaces, $M=\langle U,V,F\rangle$5 is a sheaf; in general non-tight spaces, gluing can fail (Gogioso et al., 2023). The paper characterizes failure via a solipsistic contextuality witness and proves that $M=\langle U,V,F\rangle$6 is a sheaf iff $M=\langle U,V,F\rangle$7 admits no such witness.

The later sheaf-topos formulation of TCMs intensifies this locality principle. A Lawvere–Tierney topology $M=\langle U,V,F\rangle$8 turns global truth into local truth, and the Kripke–Joyal interpretation of $M=\langle U,V,F\rangle$9 states that $U$0 holds chartwise on a $U$1-cover (Mahadevan, 20 Oct 2025). PROMETHEUS adopts an operationally parallel perspective: a Topos World Model is a sheaf-like family of local causal predictive-state models over an explicit cover of a research substrate, with restriction maps and gluing diagnostics rather than a single universal DAG (Mahadevan, 13 May 2026).

Taken together, these results place TCMs within a general doctrine of localized causality. Causal content is represented on contexts, covers, and overlaps; exact global assembly may hold, may fail, or may hold only after passing through a topology or modality that records coverwise validity (Gogioso et al., 2023, Mahadevan, 20 Oct 2025, Mahadevan, 13 May 2026).

3. Causal mechanisms, interventions, and subobjects

In the deterministic TCM formalization, SCMs embed into TCMs by passing from local autonomous mechanisms to the induced global map. For a DAG-like system with local equations

$U$2

the induced global map is

$U$3

with $U$4 and $U$5 (Mahadevan, 5 Aug 2025). More generally, local compatible functions assemble into a unique global section by a sheaf-gluing principle.

Interventions are modeled categorically by subobjects. For an SCM $U$6, fixing $U$7 yields an intervened submodel $U$8, and the inclusion

$U$9

is a monomorphism (Mahadevan, 5 Aug 2025). The characteristic morphism $V=\{V_1,\dots,V_n\}$0 classifies the intervention as a predicate on generalized elements. In the explicit three-valued classifier, $V=\{V_1,\dots,V_n\}$1 records whether a generalized element factors through the submodel, is partially consistent, or not.

The sheaf-topos formulation internalizes interventions more finely. Observations are subobjects specified by internal predicates $V=\{V_1,\dots,V_n\}$2 and comprehension monos

$V=\{V_1,\dots,V_n\}$3

followed by normalization. Interventions perform graph surgery by cutting incoming edges to a variable $V=\{V_1,\dots,V_n\}$4 and replacing its parent-dependent kernel $V=\{V_1,\dots,V_n\}$5 with a chosen policy $V=\{V_1,\dots,V_n\}$6; propagation then occurs by Kleisli integration:

$V=\{V_1,\dots,V_n\}$7

where $V=\{V_1,\dots,V_n\}$8 is the structural kernel, $V=\{V_1,\dots,V_n\}$9 is the monad strength, and $F:U\to V$0 is multiplication (Mahadevan, 20 Oct 2025).

The same source expresses conditional independence and interventional equalities internally. For context $F:U\to V$1 and objects $F:U\to V$2, the conditional independence $F:U\to V$3 in the cut model $F:U\to V$4 is the equality

$F:U\to V$5

meaning that $F:U\to V$6’s kernel ignores $F:U\to V$7 (Mahadevan, 20 Oct 2025). Interventional claims are internal sequents such as

$F:U\to V$8

interpreted as equality of morphisms $F:U\to V$9.

DEMOCRITUS adopts a related but systems-oriented presentation. Within a topos $(h,g):M\to M'$0, causal entities are objects, and causal relations may be represented either as typed morphisms or as subobjects

$(h,g):M\to M'$1

with characteristic map $(h,g):M\to M'$2; domain-local structure is organized in slice categories $(h,g):M\to M'$3 (Mahadevan, 8 Dec 2025). Interventions are described as endomorphisms of slices that replace structural arrows by constants and cut incoming influences into the intervened variable, but this paper treats such reasoning as a future “to-be-plugged-in” reasoner rather than as an implemented inference engine (Mahadevan, 8 Dec 2025).

4. Internal logic, Kripke–Joyal semantics, and the $(h,g):M\to M'$4-do-calculus

Because TCMs live in topoi, their truth values are generally intuitionistic. In the elementary-topos formulation, the internal language is a Mitchell–Bénabou language in which objects are types, equality is classified by diagonals, and membership is interpreted through evaluation into $(h,g):M\to M'$5 (Mahadevan, 5 Aug 2025). Kripke–Joyal semantics evaluates a formula $(h,g):M\to M'$6 at a generalized element $(h,g):M\to M'$7 by the forcing clause

$(h,g):M\to M'$8

and it satisfies monotonicity and local character (Mahadevan, 5 Aug 2025).

The sheaf-topos elaboration replaces global truth by coverwise truth using a Lawvere–Tierney topology. A Lawvere–Tierney topology is a natural operator $(h,g):M\to M'$9 satisfying

$g\circ F = F'\circ h.$0

and it is monotone and inflationary (Mahadevan, 20 Oct 2025). Semantically,

$g\circ F = F'\circ h.$1

Thus $g\circ F = F'\circ h.$2 acts as a modality of local truth.

On this basis, the paper “Intuitionistic $g\circ F = F'\circ h.$3-Do-Calculus in Topos Causal Models” defines $g\circ F = F'\circ h.$4-stability for conditional independences and interventional claims. For disjoint $g\circ F = F'\circ h.$5,

$g\circ F = F'\circ h.$6

Interventional equalities are $g\circ F = F'\circ h.$7-stable if their internal equality holds on a $g\circ F = F'\circ h.$8-cover; equivalently, their characteristic subobject is $g\circ F = F'\circ h.$9-closed (Mahadevan, 20 Oct 2025).

The framework introduces three inference rules mirroring Pearl’s insertion/deletion and action/observation exchange rules:

Rule	Premise	Conclusion
J1	$E=\mathrm{Sh}_J(C)$00	$E=\mathrm{Sh}_J(C)$01
J2	$E=\mathrm{Sh}_J(C)$02	$E=\mathrm{Sh}_J(C)$03
J3	$E=\mathrm{Sh}_J(C)$04	$E=\mathrm{Sh}_J(C)$05

These rules are sound in Kripke–Joyal semantics (Mahadevan, 20 Oct 2025). The soundness argument uses the factorization form of conditional independence, the stagewise interpretation of equality, and the fact that interventions compute $E=\mathrm{Sh}_J(C)$06; when $E=\mathrm{Sh}_J(C)$07, the dependence on $E=\mathrm{Sh}_J(C)$08 disappears chartwise, and the equality then glues through the $E=\mathrm{Sh}_J(C)$09-modality (Mahadevan, 20 Oct 2025).

The paper also proves conservativity with respect to Pearl’s classical calculus: under the trivial topology $E=\mathrm{Sh}_J(C)$10, $E=\mathrm{Sh}_J(C)$11 reduces to the identity, the $E=\mathrm{Sh}_J(C)$12-premises collapse to standard d-separation in mutilated graphs, and the conclusions become Pearl’s equalities (Mahadevan, 20 Oct 2025). This suggests that the intuitionistic generalization preserves the classical theory while enlarging the class of premises from global graph-separation statements to coverwise-valid ones.

5. Composition, contextuality, and regime-local reasoning

A major motivation for TCMs is compositionality. In the category-theoretic TCM formalization, every diagram $E=\mathrm{Sh}_J(C)$13 has a limit and a colimit, and these are interpreted as canonical solutions or approximations to the causal specification given by the diagram (Mahadevan, 5 Aug 2025). A cone $E=\mathrm{Sh}_J(C)$14 is a causal approximation along incoming morphisms, and the universal approximation is the limit cone $E=\mathrm{Sh}_J(C)$15; dually, outgoing approximations are organized by colimits (Mahadevan, 5 Aug 2025).

The presheaf-topological framework proves explicit factorization theorems. For disjoint event sets, parallel composition satisfies

$E=\mathrm{Sh}_J(C)$16

while sequential composition satisfies

$E=\mathrm{Sh}_J(C)$17

A conditional sequential composition theorem is also given for causally complete spaces (Gogioso et al., 2023). These results underpin constructive composition of TCMs.

The same paper establishes that causal functions may fail to form a sheaf, and that this failure manifests as causally-induced contextuality. A solipsistic contextuality witness is a quadruple $E=\mathrm{Sh}_J(C)$18 satisfying a specific non-gluing condition; the paper proves that $E=\mathrm{Sh}_J(C)$19 is a sheaf iff no such witness exists (Gogioso et al., 2023). It further constructs deterministic empirical models over solipsistic covers that cannot arise by restriction from any standard model on the standard cover. By contrast, for causal switch spaces, including total orders, any standard empirical model is local and arises as restriction of a classical empirical model (Gogioso et al., 2023).

The sheaf-topos TCM work reframes these ideas in terms of regimes and chartwise certification. Its Earthquake–Alarm example uses variables $E=\mathrm{Sh}_J(C)$20 with edges $E=\mathrm{Sh}_J(C)$21 and $E=\mathrm{Sh}_J(C)$22, and a $E=\mathrm{Sh}_J(C)$23-cover $E=\mathrm{Sh}_J(C)$24 consisting of a purely observational chart and an interventional chart cutting incoming edges to $E=\mathrm{Sh}_J(C)$25 (Mahadevan, 20 Oct 2025). Chartwise conditional independences satisfy

• $E=\mathrm{Sh}_J(C)$26 on both charts,
• $E=\mathrm{Sh}_J(C)$27 on both charts,
• $E=\mathrm{Sh}_J(C)$28 fails on $E=\mathrm{Sh}_J(C)$29.

Therefore,

$E=\mathrm{Sh}_J(C)$30

and an application of $E=\mathrm{Sh}_J(C)$31-Rule 1 yields

$E=\mathrm{Sh}_J(C)$32

in $E=\mathrm{Sh}_J(C)$33 (Mahadevan, 20 Oct 2025). The paper’s “chartwise/sewing” construction defines the sieve

$E=\mathrm{Sh}_J(C)$34

so that if chosen charts generate $E=\mathrm{Sh}_J(C)$35 and each validates $E=\mathrm{Sh}_J(C)$36, then $E=\mathrm{Sh}_J(C)$37.

These results clarify a recurring misconception: TCMs are not simply DAGs translated into category theory. They are frameworks in which DAG-based separation, presheaf gluing, local charts, and coverwise truth can all be expressed, and in which failure of global assembly is itself a mathematically meaningful feature rather than an anomaly (Gogioso et al., 2023, Mahadevan, 20 Oct 2025).

6. Operational systems, large-scale causal atlases, and open problems

Two later systems papers adapt TCM ideas to large heterogeneous corpora. DEMOCRITUS constructs, organizes, and visualizes large causal models extracted from LLM-generated text, storing them as domain-local slices $E=\mathrm{Sh}_J(C)$38 and integrating slices through pushouts, pullbacks, and coequalizers (Mahadevan, 8 Dec 2025). Its six-module pipeline comprises topic graph construction, causal question generation, causal statement generation, triple extraction, a Geometric Transformer plus UMAP manifold stage, and topos slice storage with cross-slice integration (Mahadevan, 8 Dec 2025). The system aggregated 90,016 synthetic relational causal statements across 9 domains; triple extraction yielded 54,514 unique concepts and 57,390 typed relations, and the Geometric Transformer with diagrammatic backpropagation constructed multi-relational simplicial complexes with between 553 and 1,336 regime triangles per domain (Mahadevan, 8 Dec 2025). The paper explicitly distinguishes LCMs from TCMs: an LCM is the aggregate node-edge-manifold structure built from textual triples, while a TCM is that structure embedded as slices of a topos with internal logic and categorical operations available.

PROMETHEUS develops a different operationalization. It organizes retrieved literature, data, code, simulations, and reports into a causal atlas, described as a sheaf-like family of local causal predictive-state models over an explicit cover of a research substrate (Mahadevan, 13 May 2026). In this framework,

$E=\mathrm{Sh}_J(C)$39

where $E=\mathrm{Sh}_J(C)$40 is the context category, $E=\mathrm{Sh}_J(C)$41 is a Grothendieck-style coverage, $E=\mathrm{Sh}_J(C)$42 assigns local predictive-state structures

$E=\mathrm{Sh}_J(C)$43

$E=\mathrm{Sh}_J(C)$44 are restriction maps, and $E=\mathrm{Sh}_J(C)$45 contains diagnostics and metadata including gluing tension, support, and provenance (Mahadevan, 13 May 2026).

PROMETHEUS formalizes overlap discrepancy by

$E=\mathrm{Sh}_J(C)$46

and gluing tension by

$E=\mathrm{Sh}_J(C)$47

When compatibility is sufficient, local sections are glued by support-weighted aggregation,

$E=\mathrm{Sh}_J(C)$48

while incompatible cells become obstruction records (Mahadevan, 13 May 2026). Localized interventions are aggregated over covers via

$E=\mathrm{Sh}_J(C)$49

which the paper describes as an intervention-conditioned probe rather than an automatically identified Pearl effect (Mahadevan, 13 May 2026).

PROMETHEUS reports three literature-atlas case studies and four grounded-counterfactual case studies. For ocean-temperature impacts on marine populations, the run acquired 11 studies, extracted 3,065 events, built 199 local PSRs, checked 198 restrictions, found 160 compatible, 194 compatible gluing overlaps, 4 tense gluing overlaps, and mean gluing loss 0.0179 (Mahadevan, 13 May 2026). For GLP-1 weight-loss evidence, the system processed 11 studies, 3,376 events, 191 local PSRs, 149 compatible restrictions, 41 divergent restrictions, 186 compatible gluing overlaps, 4 tense gluing overlaps, and mean gluing loss 0.0189 (Mahadevan, 13 May 2026). For resveratrol/red-wine health-benefit claims, it processed 13 studies, 4,057 events, 227 local PSRs, 177 compatible restrictions, 49 divergent restrictions, 221 compatible gluing overlaps, 5 tense gluing overlaps, and mean gluing loss 0.0178 (Mahadevan, 13 May 2026). Its grounded counterfactuals include a microplastics forcing intervention with area-weighted mean forcing changing from $E=\mathrm{Sh}_J(C)$50 to $E=\mathrm{Sh}_J(C)$51, a drop of $E=\mathrm{Sh}_J(C)$52 or approximately $E=\mathrm{Sh}_J(C)$53 (Mahadevan, 13 May 2026).

Despite these extensions, the theoretical literature identifies substantial open problems. The sheaf-topos $E=\mathrm{Sh}_J(C)$54-do work lists the choice of $E=\mathrm{Sh}_J(C)$55 as methodological, notes that completeness of $E=\mathrm{Sh}_J(C)$56-do-calculus relative to an internal separation criterion remains to be established, and identifies robustness under misspecification, finite-sample uncertainty, computational issues in constructing and verifying covers, and extensions to cyclic or latent-variable models and richer modalities as open questions (Mahadevan, 20 Oct 2025). The category-theoretic TCM paper notes that cyclic cases may require additional structure such as domain-theoretic fixpoints, that identifiability remains limited up to equivalence, and that causal homotopy theory and higher algebraic $E=\mathrm{Sh}_J(C)$57-theory are future directions (Mahadevan, 5 Aug 2025). DEMOCRITUS emphasizes that its current system is a builder rather than a reasoner, that it lacks identifiability guarantees, and that ontology alignment and full topos reasoning remain incomplete (Mahadevan, 8 Dec 2025). PROMETHEUS likewise states that local tests are not identified causal effects unless paired with valid designs or models, and that coverage and overlap may be sparse (Mahadevan, 13 May 2026).

Across these works, TCMs define a research program in which causal models are local-to-global structures in a topos-theoretic environment. Variables and mechanisms become objects and morphisms, interventions become subobjects or slice endomorphisms, truth becomes internal and often local, and incompatibility across contexts becomes diagnostically explicit rather than suppressed. This suggests that the distinctive contribution of TCMs is not a replacement of existing causal formalisms, but a reorganization of them around categorical locality, gluing, and internal reasoning (Mahadevan, 5 Aug 2025, Mahadevan, 20 Oct 2025, Gogioso et al., 2023, Mahadevan, 8 Dec 2025, Mahadevan, 13 May 2026).