CausalCompass: A Causal Navigation Framework

• CausalCompass is a design pattern that translates causal specifications into formal constraints for inference, diagnosis, and decision support.
• It integrates various methods such as DAG-driven symbolic bounds, geometric causaltopes, and robustness benchmarks for time-series causal discovery.
• The framework also extends to visual analytics and structured evaluation of LLM outputs, guiding exploratory analysis and causal reasoning.

In the cited literature, CausalCompass denotes several distinct but structurally related causal-analysis frameworks. Across these uses, the term refers to a system that starts from a causal specification, converts that specification into formal constraints or tests, and returns guidance for inference, diagnosis, or decision support. The label is used for a DAG-driven interface for exact symbolic bounds based on causaloptim (Jonzon et al., 2022), a device- and theory-independent geometric framework built from causaltopes (Gogioso et al., 2023), a benchmark suite for robustness evaluation in time-series causal discovery (Yi et al., 8 Feb 2026), and a causal-inference pipeline for assessing the effect of structured output formats on LLMs (Yuan et al., 26 Sep 2025). A closely related visual-analytics system, Causality Explorer, is explicitly described as a “causal compass” for exploratory causal analysis, validation, what-if reasoning, and action planning (Xie et al., 2020).

1. Shared orientation and uses

The recurring idea is to use causal structure as an organizing device for navigation under uncertainty. In one formulation, a “CausalCompass” should let a user start from a causal story given by a DAG, state a causal question in counterfactual notation, and then obtain either point identification or sharp bounds justified by the assumptions. In another, the compass is geometric: empirical behavior is mapped into a polytope sliced by linear causality equations, and linear programs quantify support for a target causal hypothesis or causal separability. In the time-series benchmark, the compass is evaluative: it systematically stress-tests methods under assumption violations. In the LLM setting, it is a causal inference framework over instruction, output format, and generation quality, used to decide which DAG is most consistent with the data. In the visual-analytics setting, it guides analysts from discovery to validation, uncertainty-aware reasoning, intervention simulation, and action planning (Jonzon et al., 2022, Gogioso et al., 2023, Yi et al., 8 Feb 2026, Yuan et al., 26 Sep 2025, Xie et al., 2020).

Instantiation	Domain	Core formal object
causaloptim-based CausalCompass	Partial identification from DAGs	Linear optimization over response-type distributions
Causaltopes CausalCompass	Device-independent causality	Polytope slice $A p = 0$
TSCD CausalCompass	Time-series causal discovery	Benchmark scenarios and evaluation metrics
Causality Explorer as “causal compass”	Visual analytics	DAG, F-GES, SCM-based simulation
LLM structured-output CausalCompass	Structured generation evaluation	Three-variable DAG selection over $I,F,Y$

A common misconception is that CausalCompass refers to a single package or formalism. The cited works instead use the name for different frameworks that share a navigational role in causal analysis. This suggests that the term functions more as a design pattern than as a single technical artifact.

2. DAG-to-bounds computation with causaloptim

In the symbolic-bounds setting, the stated vision is that a CausalCompass should accept a causal story as a DAG, accept a causal question in familiar counterfactual notation, and return the “best-possible answers justified by their assumptions”—point-identification when available, and otherwise sharp bounds holding for every data-generating process consistent with the DAG and optional side-conditions. The causaloptim package is presented as delivering this capability for a broad class of DAGs and queries through both a point-and-click GUI and a programmatic R API (Jonzon et al., 2022).

The workflow is explicitly DAG-first. Users draw DAGs in a Shiny web app whose canvas is split into two sides, $L$ and $R$, corresponding to the supported DAG class. Nodes and edges are added by shift-click and drag; nodes can be marked observed or latent and assigned cardinalities. The causal estimand is typed in counterfactual notation. Examples listed in the specification include

$\mathrm{ACE} = \mathbb{E}[Y_1 - Y_0],$

$P(Y(X = 1)=1),$

$\mathbb{E}[Y_x \mid Z=z],$

controlled direct effects $p\{Y(M=m, X=x)=y\}$, and natural direct effects

$p\{Y(M(X=x'), X=x)=y\} - p\{Y(M(X=x'), X=x')=y\}.$

Optional assumptions include monotonicity toggles on edges and text constraints such as $X(Z=1)\ge X(Z=0)$ or $I,F,Y$0. Exclusion and independence relations implied by the DAG are automatically enforced.

The supported DAG class is defined by a partition $I,F,Y$1 with no edges from $I,F,Y$2 to $I,F,Y$3 and no unmeasured confounding connecting the two sides; unmeasured confounding is allowed arbitrarily within each side, and all arrows crossing sides must originate in $I,F,Y$4. The method covers categorical observed variables, not only binary ones, although complexity grows with category counts. The supported estimands are any linear functionals of joint probabilities of nested counterfactuals whose outcomes lie on $I,F,Y$5 and satisfy the regularity conditions of Sachs et al. (2022a).

The mathematical formulation is linear. For each observed variable $I,F,Y$6 with parents $I,F,Y$7, a response function $I,F,Y$8 maps parent assignments to the value of $I,F,Y$9. A response-type distribution $L$0 over joint response functions on $L$1 parameterizes the latent heterogeneity relevant to the query, with coordinates $L$2. Consistency and exclusion yield a linear map

$L$3

where $L$4 denotes observed conditional probabilities and $L$5 is a $L$6-$L$7 matrix. The probabilistic constraints are $L$8 and $L$9. Any supported estimand becomes a linear functional $R$0, and the sharp lower bound is computed as

$R$1

The dual polytope is then built and its vertices enumerated using the double-description method through rcdd/cddlib with exact rational arithmetic.

The canonical worked example is the binary IV model $R$2 with unmeasured $R$3–$R$4 confounding. The target is

$R$5

under exclusion $R$6 and independence $R$7. With observed probabilities $R$8, the Balke–Pearl sharp bounds are returned as $R$9, where $\mathrm{ACE} = \mathbb{E}[Y_1 - Y_0],$0 is a maximum over lower-bound terms and $\mathrm{ACE} = \mathbb{E}[Y_1 - Y_0],$1 a minimum over upper-bound terms. The implementation prints, among other inequalities,

$\mathrm{ACE} = \mathbb{E}[Y_1 - Y_0],$2

and

$\mathrm{ACE} = \mathbb{E}[Y_1 - Y_0],$3

For this IV model, the full output includes eight max-terms for the lower bound and eight min-terms for the upper bound; adding monotonicity $\mathrm{ACE} = \mathbb{E}[Y_1 - Y_0],$4 removes defier types and yields the classical sharpened Balke–Pearl IV bounds.

Operationally, the GUI launches from specify_graph(), constructs an annotated igraph object, calls analyze_graph() to build the linear program, and calls optimize_effect_2() to compute symbolic bounds using exact vertex enumeration. Output includes MAX/MIN expressions, a legend mapping symbols to observed probabilities such as $\mathrm{ACE} = \mathbb{E}[Y_1 - Y_0],$5, and export to LaTeX or R functions for plug-in evaluation. The package is described as replacing legacy enumeration code with cddlib’s double-description method, yielding orders-of-magnitude speedups while preserving sharpness and numerical stability (Jonzon et al., 2022).

3. Geometric causality through causaltopes

A second use of the term presents a device- and theory-independent geometry of causality built from causaltopes, defined as slices of a product polytope of conditional distributions by linear causality equations (Gogioso et al., 2023). The basic objects are a space of input histories $\mathrm{ACE} = \mathbb{E}[Y_1 - Y_0],$6, a cover $\mathrm{ACE} = \mathbb{E}[Y_1 - Y_0],$7, and a family of per-context conditional distributions. For each context $\mathrm{ACE} = \mathbb{E}[Y_1 - Y_0],$8, the framework forms a simplex over joint outputs; vectorizing all contextwise distributions yields a single vector $\mathrm{ACE} = \mathbb{E}[Y_1 - Y_0],$9.

The ambient polytope is

$P(Y(X = 1)=1),$0

with per-context normalization and non-negativity constraints. Causality is then imposed by equality of marginals across overlaps. For any lowerset $P(Y(X = 1)=1),$1, the restriction maps $P(Y(X = 1)=1),$2 and $P(Y(X = 1)=1),$3 induce the causality equations

$P(Y(X = 1)=1),$4

Collecting these equalities yields a linear system

$P(Y(X = 1)=1),$5

The causaltope is therefore

$P(Y(X = 1)=1),$6

In the standard cover, this specializes to “no backwards influence,” and in discrete spaces it reduces to standard no-signalling constraints.

The same paper makes the CausalCompass operational by attaching linear programs to sub-causaltopes. For a sub-normalized component $P(Y(X = 1)=1),$7, the maximal supported fraction in a target sub-causaltope $P(Y(X = 1)=1),$8 is obtained by maximizing its mass subject to quasi-normalization, linear subspace constraints, and $P(Y(X = 1)=1),$9. The reported form is

$\mathbb{E}[Y_x \mid Z=z],$0

For causal separability over a family $\mathbb{E}[Y_x \mid Z=z],$1, the LP maximizes the total mass of components $\mathbb{E}[Y_x \mid Z=z],$2 supported in the respective sub-causaltopes: $\mathbb{E}[Y_x \mid Z=z],$3 subject to quasi-normalization and linear constraints for each $\mathbb{E}[Y_x \mid Z=z],$4, non-negativity, and $\mathbb{E}[Y_x \mid Z=z],$5. The resulting optimum is the causally separable fraction.

The framework is explicitly broader than standard definite-order formalisms. It admits arbitrary spaces of input histories, arbitrary covers, and causal separability relative to arbitrary causal constraints. Proposition 4.4 states that the causally separable fraction is bounded below by the separable local, or non-contextual, fraction. The paper also introduces “causal contextuality,” defined as a phenomenon in which causal inseparability is correlated with, or directly implied by, non-locality and contextuality.

The reported examples are quantum-switch and contextual-control models. For a single switch with Bell entangled control, at $\mathbb{E}[Y_x \mid Z=z],$6 and $\mathbb{E}[Y_x \mid Z=z],$7, the model is 100% supported by the relevant indefinite-order causaltope and has a causal separability of approximately $\mathbb{E}[Y_x \mid Z=z],$8, split approximately $\mathbb{E}[Y_x \mid Z=z],$9 on each completion. For the GHZ-control single switch, the corresponding causal separability is approximately $p\{Y(M=m, X=x)=y\}$0. For two switches with Bell entangled controls, the causally separable fraction is approximately $p\{Y(M=m, X=x)=y\}$1, approximately $p\{Y(M=m, X=x)=y\}$2 on each definite completion. In the contextually controlled classical-switch example, causal separability over the 16 completions is $p\{Y(M=m, X=x)=y\}$3, realized by exactly six completions with $p\{Y(M=m, X=x)=y\}$4 each, and the paper states that the causal-separable fraction equals both the model’s local fraction and the Bell local fraction, which it interprets as evidence of “causal contextuality.” Computationally, the LPs are polynomial-time, the matrices are extremely sparse and block diagonal by context, and one large landscape evaluation reportedly required 15 hours on 30 cores with approximately 260GB RAM (Gogioso et al., 2023).

4. Robustness benchmarking for time-series causal discovery

A third instantiation, explicitly titled “CausalCompass: Evaluating the Robustness of Time-Series Causal Discovery in Misspecified Scenarios”, is a benchmark suite rather than an inference engine (Yi et al., 8 Feb 2026). Its stated purpose is to evaluate the robustness of time-series causal discovery methods when key modeling assumptions are violated. The motivation is that TSCD methods often rely on untestable assumptions such as causal sufficiency, measurement-error-free observations, stationarity, linearity, or specific noise forms, whereas existing benchmarks largely evaluate methods on “vanilla” data satisfying those assumptions.

The benchmark uses two graph types. The summary causal graph $p\{Y(M=m, X=x)=y\}$5 has one node per process $p\{Y(M=m, X=x)=y\}$6, with an edge $p\{Y(M=m, X=x)=y\}$7 if some lagged or contemporaneous influence occurs over time. The window causal graph $p\{Y(M=m, X=x)=y\}$8 has nodes indexed by variables within a lag window $p\{Y(M=m, X=x)=y\}$9, with edges $p\{Y(M(X=x'), X=x)=y\} - p\{Y(M(X=x'), X=x')=y\}.$0, $p\{Y(M(X=x'), X=x)=y\} - p\{Y(M(X=x'), X=x')=y\}.$1, and $p\{Y(M(X=x'), X=x)=y\} - p\{Y(M(X=x'), X=x')=y\}.$2 for self-links. Two vanilla generators are used: a sparse linear VAR,

$p\{Y(M(X=x'), X=x)=y\} - p\{Y(M(X=x'), X=x')=y\}.$3

with $p\{Y(M(X=x'), X=x)=y\} - p\{Y(M(X=x'), X=x')=y\}.$4, and a nonlinear Lorenz–96 system,

$p\{Y(M(X=x'), X=x)=y\} - p\{Y(M(X=x'), X=x')=y\}.$5

The benchmark defines eight assumption-violation scenarios, all applicable to both vanilla generators: measurement error, nonstationarity, latent confounders, z-score standardization, mixed data, min–max normalization, missing data under MCAR, and trend plus seasonality. The formulas are explicit. Measurement error is

$p\{Y(M(X=x'), X=x)=y\} - p\{Y(M(X=x'), X=x')=y\}.$6

with $p\{Y(M(X=x'), X=x)=y\} - p\{Y(M(X=x'), X=x')=y\}.$7 in the main setting and up to $p\{Y(M(X=x'), X=x)=y\} - p\{Y(M(X=x'), X=x')=y\}.$8 in the appendix. Z-score standardization is

$p\{Y(M(X=x'), X=x)=y\} - p\{Y(M(X=x'), X=x')=y\}.$9

and min–max normalization is

$X(Z=1)\ge X(Z=0)$0

Mixed data discretize a fraction $X(Z=1)\ge X(Z=0)$1, $X(Z=1)\ge X(Z=0)$2, by thresholding $X(Z=1)\ge X(Z=0)$3 at $X(Z=1)\ge X(Z=0)$4. Nonstationarity is implemented through Gaussian-process modulation of the noise scale: $X(Z=1)\ge X(Z=0)$5

The suite benchmarks 11 representative methods across six families: Granger-based, constraint-based, noise-based, score-based continuous optimization, topology-based, and deep learning-based. The named methods are VAR, LGC, PCMCI, VARLiNGAM, DYNOTEARS, NTS-NOTEARS, TSCI, cMLP, cLSTM, CUTS, and CUTS+. Metrics are AUROC and AUPRC on off-diagonal adjacency entries, self-links excluded. The benchmark uses $X(Z=1)\ge X(Z=0)$6, $X(Z=1)\ge X(Z=0)$7, Lorenz–96 forcing $X(Z=1)\ge X(Z=0)$8, and 5 random seeds per configuration, reporting mean $X(Z=1)\ge X(Z=0)$9 standard deviation.

The central finding is that there is no universal winner, but deep learning-based methods dominate overall. The paper states that the best average performance across vanilla and all eight violation scenarios “typically comes from deep neural TSCD” and that CUTS+ leads overall, followed by cLSTM and cMLP. Several quantitative examples are reported. On linear 10-node $I,F,Y$00 vanilla data, NTS-NOTEARS performs poorly in the vanilla setting, with AUROC $I,F,Y$01 and AUPRC $I,F,Y$02, but after standardization jumps to AUROC $I,F,Y$03 and AUPRC $I,F,Y$04. In a moderate nonstationary linear 10-node setting, cLSTM achieves AUPRC $I,F,Y$05, CUTS+ $I,F,Y$06, and DYNOTEARS $I,F,Y$07. Under severe measurement error in nonlinear data with $I,F,Y$08 and $I,F,Y$09, CUTS+ still achieves AUPRC $I,F,Y$10, whereas classic methods drop near chance on AUPRC. Under latent confounding with $I,F,Y$11 in nonlinear $I,F,Y$12 data, reported AUPRCs include CUTS+ $I,F,Y$13, cLSTM $I,F,Y$14, cMLP $I,F,Y$15, and DYNOTEARS $I,F,Y$16.

The benchmark also provides tuning guidance. It states that continuous-optimization methods are highly sensitive to thresholding and regularization, and that NTS-NOTEARS is “extremely sensitive” to input scaling; one recommendation is therefore always to test with and without z-score standardization. CUTS and CUTS+ are recommended under missingness or irregular sampling, and fast baselines such as VAR, LGC, or PCMCI are suggested for coarse screening before heavier models are applied. Limitations listed for the current release include the absence of sampling-rate mismatch, explicit contemporaneous cycles with feedback loops, interventions or regime switches beyond noise-scale modulation, high-dimensional settings above 100 variables, and metrics such as SHD or SID (Yi et al., 8 Feb 2026).

5. Visual analytics as a causal compass

The paper “A Visual Analytics Approach for Exploratory Causal Analysis: Exploration, Validation, and Applications” presents a system named Causality Explorer, and the specification explicitly states that it can be understood as a “causal compass” for practitioners (Xie et al., 2020). Its purpose is to support the full workflow from exploratory causal discovery in high-dimensional observational data to validation, uncertainty-aware reasoning, what-if analysis, and action planning.

The causal representation is a DAG grounded in Structural Causal Models, $I,F,Y$17, with the causal Markov factorization

$I,F,Y$18

The discovery backend is Fast Greedy Equivalence Search (F-GES) under Bayesian Information Criterion scoring,

$I,F,Y$19

with a forward phase that adds the single edge producing maximal score improvement and a backward phase that deletes the single edge producing maximal score improvement until no improvement remains. Edge confidence is then quantified as

$I,F,Y$20

where $I,F,Y$21 denotes the graph with edge $I,F,Y$22 removed. Thicker links in the interface represent higher confidence.

The visualization emphasizes scalable inspection of large DAGs. Nodes are pie charts encoding categorical value distributions. The graph is laid out in layers by topological order, with chain aggregation to reduce clutter, cross-layer glyphs to suppress crossings, and leaf nodes positioned to make sinks explicit. Subgraph projection can be invoked by double-clicking a node. The system coordinates this graph view with a dimension view, a table view, and dedicated controls for intervention and attribution.

Intervention reasoning is implemented through SCM-style sampling rather than closed-form adjustment. Given $I,F,Y$23, the system samples variables in topological order, using $I,F,Y$24 for roots unless intervened and $I,F,Y$25 for downstream variables, producing post-intervention empirical distributions $I,F,Y$26 that are compared to baseline distributions $I,F,Y$27. Multiple interventions $I,F,Y$28 are supported. Attribution is defined through interventional probabilities: $I,F,Y$29 and the influence of $I,F,Y$30 is summarized by $I,F,Y$31, which is then mapped to node size.

The paper reports two case studies. In the education study, based on 3,500 university students, analysts validated expected links such as Region $I,F,Y$32 Graduated Highschool and Gender/Region $I,F,Y$33 Major, identified Major $I,F,Y$34 Student Status, and used attribution for Student Status = dropout to highlight Fail as the largest influence. In the marketing study, based on 10,000 visits and 32 dimensions, analysts validated known links such as Country $I,F,Y$35 City and Referral Channel $I,F,Y$36 Landing Page, discovered thinner-edge links including Referral Channel $I,F,Y$37 Number of Searches and Browser Type $I,F,Y$38 Operating System, and used what-if interventions on Landing Page to compare purchase rates. The largest graph explored had approximately 100 nodes and 186 edges. Limitations noted in the paper include hidden confounding, selection bias, data-quality pathologies such as Berkson’s paradox, and temporal complications such as lagged effects and evolving causality (Xie et al., 2020).

6. Structured-output evaluation for LLMs

The paper “Navigating the Impact of Structured Output Format on LLMs through the Compass of Causal Inference” uses CausalCompass as a causal inference framework for deciding whether structured output formats such as JSON, XML, or YAML affect LLM generation quality (Yuan et al., 26 Sep 2025). The framework is defined over three observed variables: instruction $I,F,Y$39, output format $I,F,Y$40, and generation quality $I,F,Y$41, with answer correctness as the proxy for quality. It combines matched re-prompting, stratified observational comparisons, and formal statistical tests to choose among five candidate DAGs.

The guaranteed constraints are central. The paper states marginal independence $I,F,Y$42, maintained either by reusing the same template across formats or, in function-calling experiments, by specifying format outside the prompt. It also states temporal precedence: $I,F,Y$43 precede $I,F,Y$44. Under these conditions, the framework analyzes five structures: collider without m-bias, collider with m-bias, single cause from instruction, single cause from output format, and independence. The formalizations include

$I,F,Y$45

for collider without m-bias, and

$I,F,Y$46

for collider with m-bias. The format effect is written as

$I,F,Y$47

The empirical design covers seven public tasks and one developed task: GSM8K, LLC, ELLC, SOT, GCF, GCC, OpsEval, and XCodeEval. Outcomes are binary and evaluated by exact match or accuracy depending on the task. Two implementation modes are used: format-restricting prompts and function calling. The inferential toolkit includes McNemar’s test for paired binary format contrasts, Cochran’s $I,F,Y$48 test for instruction variants, Stouffer’s method for combining p-values across correlated strata, mixed-effects logistic regression for m-bias detection, and Bonferroni correction for DAG decisions.

The main result is that coarse metrics alone can suggest positive, negative, or neutral effects, whereas the causal analysis concludes that for GPT-4o, 43 of 48 scenarios show no causal impact of structured output on $I,F,Y$49. The remaining 5 scenarios include 3 collider structures without m-bias and isolated single-cause cases. The paper gives examples: LLC–YAML at $I,F,Y$50 is classified as a collider without m-bias, with m-bias test $I,F,Y$51; XCodeEval–YAML at $I,F,Y$52 yields a format-only case with format $I,F,Y$53 and instruction $I,F,Y$54; GCC–XML yields an instruction-only case with instruction $I,F,Y$55 and format $I,F,Y$56; OpsEval–JSON is an independence case with instruction $I,F,Y$57 and format $I,F,Y$58. The function-calling experiments preserve the predominance of independence, and the paper notes that structured outputs via function calling outperform format-restricting instructions in 5 of 8 datasets.

The stated interpretation is that many apparent gains or drops are non-causal artifacts. In collider structures, both $I,F,Y$59 and $I,F,Y$60 affect $I,F,Y$61, and conditioning on correctness can create spurious association between instruction and format. The paper therefore advises not conditioning on $I,F,Y$62 when estimating effects, using function calling when available, ensuring template parity across formats, maintaining the independence constraint $I,F,Y$63, applying Bonferroni correction, and repeating API calls to assess non-determinism. A broader limitation is that the outcome variable $I,F,Y$64 measures answer correctness rather than reasoning quality, with the latent reasoning variable $I,F,Y$65 reserved for future work (Yuan et al., 26 Sep 2025).

Across these uses, CausalCompass consistently denotes a causal-navigation framework rather than a single method. Its concrete instantiations differ substantially—symbolic bounds, polyhedral geometry, benchmark stress-testing, visual analytics, and causal diagnosis of LLM prompting—but each treats causal structure as the primary object through which uncertainty is formalized, alternative assumptions are made explicit, and downstream conclusions are disciplined.