Causal Stitching in Dynamic Systems

• Causal stitching is a compositional technique that integrates local causal units—such as state transitions or fault injections—into larger, executable causal models.
• The method relies on defined causal interfaces and compatibility conditions rather than simple concatenation, ensuring that only compatible causal fragments are stitched together.
• This approach underpins various applications, from modeling dynamic systems and cascading failures to transferring representations in diffusion models, enhancing both interpretability and performance.

Causal stitching denotes a family of constructions in which local causal fragments are linked into a larger causal object. In the cited arXiv literature, the stitched object may be a chain of state transitions in a dynamic system, a global causal function assembled from local causal sections, a cascading-failure cycle reconstructed from multiple fault-injection runs, a collapsed spatio-temporal causal model, a spatio-temporal sample formed by preserving causal regions while regenerating environmental regions, or a hybrid value model created by attaching a pretrained reward model to a diffusion backbone at an intermediate representation (Allen, 2023, Gogioso et al., 2023, Qian et al., 30 Sep 2025, Li et al., 25 Nov 2025, Duan et al., 2024, Go et al., 19 May 2026). The common operation is not mere concatenation: each local piece is joined only where a causal interface, causal constraint, compatibility condition, or representation match makes the composite object valid.

1. Core meaning and recurring structure

Across these papers, causal stitching is a structured way of moving from local causal information to a larger explanatory or inferential artifact. In “Using Causal Threads to Explain Changes in a Dynamic System,” the local units are state transitions; in “CSnake,” they are fault-propagation edges discovered in separate tests; in “The Topology of Causality,” they are local causal functions on open subspaces; in “Spatio-Temporal Hierarchical Causal Models,” they are subunit-level mechanisms and unit-level rollouts; in “CaPaint,” they are observed causal patches plus generated environmental completion; and in “StitchVM,” they are intermediate representations from two pretrained networks (Allen, 2023, Qian et al., 30 Sep 2025, Gogioso et al., 2023, Li et al., 25 Nov 2025, Duan et al., 2024, Go et al., 19 May 2026).

A recurring structure appears in all of these settings. First, a domain supplies partial causal objects: transitions, histories, propagations, subunit mechanisms, masked regions, or latent features. Second, the method defines an interface on which two pieces may be joined: a produced state that becomes the condition for the next transition, an overlap of open sets, a matched interference fault and injected fault, a shared macro-history, a binary mask, or a compatible feature space. Third, the stitched result is treated as more than a narrative summary: it is an executable, inspectable, queryable, or optimizable object. This suggests that “causal stitching” is best understood as a compositional pattern rather than a single formalism.

2. Dynamic-system explanation as chains of state transitions

In Allen’s framework, the target is a rich semantic model of a dynamic system in which causation is defined as “a change of state that leads to another change of state.” The system is represented as an executable, process-based dynamic knowledge graph built from objects, dimensions, and states; “objects are associated with Dimensions which are properties or processes,” and “most of those Dimensions are subdivided into States.” The key modeling unit is an object-transition pair, and causal relations are encoded by linking state transitions into threads that traverse the model across time. The resulting model is richer than a standard ontology because it includes not only classifications of things, but also how things interact and change (Allen, 2023).

The Snowball Earth example shows the stitching operation directly. A green causal path begins with continental drift, continues through increased albedo / reflected light, then to reduced photons absorbed at the surface, then to lower temperature, and finally to increased ice coverage. Each step is a state change that becomes the input condition for the next transition. The equilibrium process is shown separately, in red, and is “terminated” once the causal sequence changes the state of the system. The model also contains a positive feedback loop in which increased ice coverage reflects more photons, lowers temperature further, and increases ice coverage further, with eventual saturation because available liquid water is exhausted.

The representation is multi-level. Local transitions are organized into episodes, and episodes are organized into a discourse-level explanation. An episode consists of an Equilibrium phase followed by a Causal (Process) phase; the examples given are Initial Equilibrium followed by Freezing, Frozen Equilibrium followed by Thaw, and Initial Equilibrium followed by heavy Sedimentation. The prototype interface, implemented in Python with Tkinter, presents Earth, Atmosphere, Ocean, and Land as subregions, uses red for equilibrium and green for the causal process, and supports drill-down inspection of dimensions and richer descriptions. The methodological emphasis is direct representation using well-defined standardized vocabularies and resources such as SUMO, FrameNet, and VerbNet, rather than probabilistic text summarization. The paper also states several limitations: the models are laborious to build, rely on relatively well-defined scenarios, and face open issues in granularity, constraints, ambiguity or missing data, time, concurrency, and timelines.

3. Topological gluing of causal functions

“The Topology of Causality” reformulates causal stitching as a problem of gluing local causal data. For a space of input histories $\Theta$, a causal function assigns outputs to tip events of histories while respecting eventwise causal constraints. The paper’s general definition is

$$CausFun_{\Theta,\underline{O}} := \left\{ f \in \prod_{h \in \Theta}\prod_{\omega \in \Theta_h} O_\omega \;\middle|\; \#_\omega(h,h') \Rightarrow f(h)_\omega = f(h')_\omega \right\}.$$

The central structural claim is that explicit causal constraints on joint outputs are equivalent to the free assignment of outputs to tip events of histories once histories are used as the basic domain. In this formulation, causal order is internalized into the combinatorics and topology of histories rather than imposed externally as equations on a joint-input/joint-output map (Gogioso et al., 2023).

The paper then proves factorization results that make the compositional content explicit. For parallel composition, independent systems stitch by product. For sequential composition, the second stage becomes a family of causal functions indexed by the maximal extended histories of the first stage. For conditional sequential composition, the continuation is attached branchwise, with one causal function for each branch $k \in \maxExt(\Theta)$. This is a precise formalization of stitching across time and condition.

A further step is topological. The space of histories is given the lowerset topology, whose opens are downsets:

$$\mathcal{T}(X)=\{U\subseteq X \mid x\in U,\; y\le x \Rightarrow y\in U\}.$$

With this topology, causality is equivalent to continuity:

$$\hat{F}\text{ is causal} \quad\Longleftrightarrow\quad \hat{F}\text{ is continuous}.$$

Partial causal functions on open subspaces then form a presheaf under restriction. However, the presheaf of causal functions is not always a sheaf. In tight spaces, compatible local causal data glues uniquely to a global causal function, but in some non-tight spaces compatible local data fails to glue because the causal constraints themselves become context-dependent. The paper calls this phenomenon solipsistic contextuality. The result is a sharp limit on causal stitching: local validity and overlap compatibility do not always imply the existence of a global causal section.

4. Fault-propagation chains and self-sustaining cascading failure

In distributed systems, causal stitching is used to expose self-sustaining cascading failures before deployment. “CSnake” defines it as causally linking multiple single-fault injections across different workloads to simulate a complex propagation chain. The chain is not built by injecting multiple faults at once into one run. Instead, each edge comes from a separate fault-injection experiment, and the edges are stitched when the output fault of one edge matches the injected fault of another, provided the conditions are compatible. The causal criterion is explicitly counterfactual: $f_1$ is a counterfactual cause of $f_2$ iff $f_2$ would not occur unless $f_1$ occurs (Qian et al., 30 Sep 2025).

The core analysis is fault causality analysis (FCA). FCA compares a fault injection run with a corresponding profile run, defined as the same workload without the injection, and identifies any additional faults triggered in the injection run. These additional faults are treated as causally related to the injected fault. Formally,

$I(f_i, t_j) = [f_{k_1}, f_{k_2}, \ldots, f_{k_m}],$

where $$CausFun_{\Theta,\underline{O}} := \left\{ f \in \prod_{h \in \Theta}\prod_{\omega \in \Theta_h} O_\omega \;\middle|\; \#_\omega(h,h') \Rightarrow f(h)_\omega = f(h')_\omega \right\}.$$0 is the list of faults triggered by injecting $$CausFun_{\Theta,\underline{O}} := \left\{ f \in \prod_{h \in \Theta}\prod_{\omega \in \Theta_h} O_\omega \;\middle|\; \#_\omega(h,h') \Rightarrow f(h)_\omega = f(h')_\omega \right\}.$$1 into test $$CausFun_{\Theta,\underline{O}} := \left\{ f \in \prod_{h \in \Theta}\prod_{\omega \in \Theta_h} O_\omega \;\middle|\; \#_\omega(h,h') \Rightarrow f(h)_\omega = f(h')_\omega \right\}.$$2. Once such edges are found, stitching searches for a sequence such as $$CausFun_{\Theta,\underline{O}} := \left\{ f \in \prod_{h \in \Theta}\prod_{\omega \in \Theta_h} O_\omega \;\middle|\; \#_\omega(h,h') \Rightarrow f(h)_\omega = f(h')_\omega \right\}.$$3 and $$CausFun_{\Theta,\underline{O}} := \left\{ f \in \prod_{h \in \Theta}\prod_{\omega \in \Theta_h} O_\omega \;\middle|\; \#_\omega(h,h') \Rightarrow f(h)_\omega = f(h')_\omega \right\}.$$4, yielding $$CausFun_{\Theta,\underline{O}} := \left\{ f \in \prod_{h \in \Theta}\prod_{\omega \in \Theta_h} O_\omega \;\middle|\; \#_\omega(h,h') \Rightarrow f(h)_\omega = f(h')_\omega \right\}.$$5. Beam-search cycle detection uses the condition

$$CausFun_{\Theta,\underline{O}} := \left\{ f \in \prod_{h \in \Theta}\prod_{\omega \in \Theta_h} O_\omega \;\middle|\; \#_\omega(h,h') \Rightarrow f(h)_\omega = f(h')_\omega \right\}.$$6

To manage the search space, the framework uses a three-phase allocation protocol of test budget. Phase 1 performs causally equivalent fault detection by injecting each fault into one workload, collecting the interference set, vectorizing it with inverse document frequency,

$$CausFun_{\Theta,\underline{O}} := \left\{ f \in \prod_{h \in \Theta}\prod_{\omega \in \Theta_h} O_\omega \;\middle|\; \#_\omega(h,h') \Rightarrow f(h)_\omega = f(h')_\omega \right\}.$$7

and clustering by cosine distance. Phase 2 explores causality by allocating budget round-robin across clusters. Phase 3 prioritizes clusters with lower intra-cluster similarity, using

$$CausFun_{\Theta,\underline{O}} := \left\{ f \in \prod_{h \in \Theta}\prod_{\omega \in \Theta_h} O_\omega \;\middle|\; \#_\omega(h,h') \Rightarrow f(h)_\omega = f(h')_\omega \right\}.$$8

A local compatibility check approximately tests whether connected propagations arose under compatible conditions by comparing local traces and the closest two call-stack levels, rather than solving full path constraints. On five distributed systems—HDFS 2.10.2, HDFS 3.4.1, HBase 2.6.0, OZone 1.4.0, and Flink 1.20.0—the method detected 15 new self-sustaining cascading failures, five were confirmed by developers, and two were fixed. The paper also reports that 11 of the 15 bugs could not be triggered by a naive single-fault strategy.

5. Spatio-temporal and hierarchical forms of causal stitching

In spatio-temporal causal inference, stitching refers to recovering a unit-level or system-level causal object from finer-grained local structure. “Spatio-Temporal Hierarchical Causal Models” studies units $$CausFun_{\Theta,\underline{O}} := \left\{ f \in \prod_{h \in \Theta}\prod_{\omega \in \Theta_h} O_\omega \;\middle|\; \#_\omega(h,h') \Rightarrow f(h)_\omega = f(h')_\omega \right\}.$$9, subunits $k \in \maxExt(\Theta)$0, and times $k \in \maxExt(\Theta)$1 in the presence of unobserved, time-invariant unit-level confounders $k \in \maxExt(\Theta)$2. The framework introduces macro-states $k \in \maxExt(\Theta)$3 and $k \in \maxExt(\Theta)$4-variables that represent collapsed subunit-level causal mechanisms, and extends hierarchical causal modeling with temporal dependence and spatial spillovers. Its central theorem states that, under mechanism convergence, regularity of state spaces, and spatial assumptions, the marginalized macro-state distribution converges in Kullback-Leibler divergence to that of a corresponding Spatio-Temporal Dynamic Collapsed Model:

$k \in \maxExt(\Theta)$5

The proof constructs a Super-Unit model by concatenating all spatial units into one temporal hierarchical system. This establishes a mathematically explicit form of stitching: many subunit observations approximate the latent local mechanism increasingly well, so the hierarchical structure collapses into a tractable causal representation (Li et al., 25 Nov 2025).

The same paper turns the theorem into identification and estimation. The target effect is $k \in \maxExt(\Theta)$6, and identification is obtained in the collapsed graph by adjustment on observable spatio-temporal history or, when needed, by instrumental variables. The practical estimator is a two-stage hierarchical g-computation procedure: learn unit-specific conditional outcome models $k \in \maxExt(\Theta)$7 using LMM, GBM, or GP, then simulate forward under an intervention and average final outcomes across units. In synthetic data with true $k \in \maxExt(\Theta)$8, the mean estimate is about $k \in \maxExt(\Theta)$9, and the estimation error decreases as the number of subunits $$\mathcal{T}(X)=\{U\subseteq X \mid x\in U,\; y\le x \Rightarrow y\in U\}.$$0 increases. In the Chicago traffic study, the data comprise 29 traffic regions, 1,025 road segments, about 1.2 million observations, and 1,943 crash events.

A different spatio-temporal usage appears in “CaPaint.” The method first performs causal deciphering by splitting each frame into non-overlapping patches, embedding them into a Vision Transformer, and computing an attention-derived importance score

$$\mathcal{T}(X)=\{U\subseteq X \mid x\in U,\; y\le x \Rightarrow y\in U\}.$$1

Top-$$\mathcal{T}(X)=\{U\subseteq X \mid x\in U,\; y\le x \Rightarrow y\in U\}.$$2 high-scoring patches are retained as causal patches $$\mathcal{T}(X)=\{U\subseteq X \mid x\in U,\; y\le x \Rightarrow y\in U\}.$$3 and the remainder are treated as environmental patches $$\mathcal{T}(X)=\{U\subseteq X \mid x\in U,\; y\le x \Rightarrow y\in U\}.$$4. The second stage keeps the causal region intact and inpaints the environmental region with a fine-tuned unconditional DDPM. At each reverse step, a binary mask merges the preserved causal region with the generated environmental region:

$$\mathcal{T}(X)=\{U\subseteq X \mid x\in U,\; y\le x \Rightarrow y\in U\}.$$5

The paper describes this as preserving the regions that the model deems causally important and regenerating the less important regions from a learned data distribution. It also states that the method reduces data generation complexity from exponential to quasi-linear levels by replacing approximately $$\mathcal{T}(X)=\{U\subseteq X \mid x\in U,\; y\le x \Rightarrow y\in U\}.$$6 with roughly $$\mathcal{T}(X)=\{U\subseteq X \mid x\in U,\; y\le x \Rightarrow y\in U\}.$$7, and reports improvements on five real-world ST benchmarks (Duan et al., 2024).

6. Representation transfer, terminological variation, and limits

“Stitched Value Model for Diffusion Alignment” uses model stitching to construct a value function on noisy latents by attaching a frozen diffusion backbone to a truncated pixel-space reward model. The hybrid is

$$\mathcal{T}(X)=\{U\subseteq X \mid x\in U,\; y\le x \Rightarrow y\in U\}.$$8

where $$\mathcal{T}(X)=\{U\subseteq X \mid x\in U,\; y\le x \Rightarrow y\in U\}.$$9 is the diffusion backbone truncated at layer $$\hat{F}\text{ is causal} \quad\Longleftrightarrow\quad \hat{F}\text{ is continuous}.$$0, $$\hat{F}\text{ is causal} \quad\Longleftrightarrow\quad \hat{F}\text{ is continuous}.$$1 is the reward-model suffix starting from layer $$\hat{F}\text{ is causal} \quad\Longleftrightarrow\quad \hat{F}\text{ is continuous}.$$2, and $$\hat{F}\text{ is causal} \quad\Longleftrightarrow\quad \hat{F}\text{ is continuous}.$$3 is a lightweight stitching layer. Candidate stitch points are scored by a closed-form least-squares fit,

$$\hat{F}\text{ is causal} \quad\Longleftrightarrow\quad \hat{F}\text{ is continuous}.$$4

after which only the stitch layer and reward-model suffix are finetuned. The paper explicitly connects this setup to the notion of “causal stitching” or representation transfer: once a compatible intermediate representation exists, the downstream function can be transplanted without relearning the entire model. The reported stitching and fine-tuning of CLIP ViT-L and SD 3.5 Medium takes about 10 GPU-hours; DPS becomes up to $$\hat{F}\text{ is causal} \quad\Longleftrightarrow\quad \hat{F}\text{ is continuous}.$$5 faster while using about 50% lower peak GPU memory, and DiffusionNFT becomes $$\hat{F}\text{ is causal} \quad\Longleftrightarrow\quad \hat{F}\text{ is continuous}.$$6 faster (Go et al., 19 May 2026).

This usage is related but not identical to the explanatory and causal-inference usages above. Here the stitched object is a learned value model for noisy diffusion latents, not a causal graph or causal section. The paper’s own connection is analogical and representational: the diffusion backbone provides the right state representation for noisy latents, the reward model provides task semantics for the clean-image objective, and the stitch layer transfers information between them with minimal retraining. This suggests that, in current arXiv usage, “causal stitching” sometimes names a broader transplantation pattern in which an intermediate interface supports the reuse of downstream function.

A terminological distinction is also necessary. “Stitching and dimensionality effects on large artificially generated volume datasets” studies patch assembly in large-image and large-volume generation. There, stitching refers to assembling output patches and mitigating border artifacts through tile-and-stitch, padded convolutions with averaged overlap, or valid convolutions with no overlap. The paper shows that FID scores fail to detect subtle stitching artifacts that significantly impact downstream segmentation performance, that 3D models with artifact-free stitching marginally outperform 2D models on downstream tasks, and that 2D models train more stably due to larger batch sizes (Chamier et al., 18 Jun 2026). This is not causal stitching in the sense of gluing causal descriptions; it is a separate use of the word “stitching” in generative imaging.

Several limits and misconceptions are clarified by the cited literature. Causal stitching is not mere correlation or happens-before: CSnake uses counterfactual causality rather than correlation, and Allen distinguishes genuine causal relations from mere necessary conditions or sequences of events in equilibrium. It is not unrestricted natural-language explanation: Allen’s framework relies on direct representation with well-defined vocabularies. It does not always succeed globally: the topological framework proves that compatible local causal functions can fail to glue in non-tight spaces because the causal constraints themselves are context-dependent. It is also often laborious or assumption-heavy: Allen notes laborious construction and open problems in time and concurrency; ST-HCMs rely on spatial Markov, spatial causal ordering, and spatial homogeneity; and CaPaint depends on a causal/environmental partition derived from attention scores.