The adoption of Explainable Artificial Intelligence (XAI) in Intrusion Detection Systems (IDSs) has grown substantially to improve the transparency of deep learning models used in cybersecurity. Recent literature reviews show that XAI methods are increasingly integrated into IDSs to improve decision interpretability, strengthen analysts’ trust, and facilitate the validation of security alerts [14][15]. Widely used methods, including SHAP, LIME, and Anchors, provide local or global explanations that help identify the features influencing model decisions.

However, several recent studies show that different XAI methods can produce substantially different explanations for the same prediction, thereby raising concerns about the reliability of the resulting interpretations [3][6][16][17]. In particular, Kumar and Thing [3] show that leading explainability methods used in deep learning-based IDSs frequently disagree and may produce conflicting explanations for the same network attack.

This phenomenon, commonly referred to as disagreement among XAI methods, arises primarily from fundamental differences between explanation paradigms. Attribution-based methods, such as SHAP, estimate the contribution of each feature to a model decision, whereas rule-based methods, such as Anchors, identify sufficient logical conditions that locally explain a prediction. Because these approaches rely on different explanatory principles, they may naturally produce divergent interpretations, particularly in intrusion detection systems characterized by highly nonlinear relationships among features [3][11][17].

Beyond being an interpretability issue, this divergence may reduce confidence in automated decisions, complicate the validation of security alerts, and directly affect the quality of decisions made by human operators [6][9][17].

In response to this problem, several families of approaches have been proposed to reduce or characterize conflicts among explainability methods.

A first family seeks to measure or establish consensus among multiple XAI methods. EXAGREE [5] selects models whose explanations exhibit the highest agreement with stakeholder expectations, whereas Consensus Importance [6] provides a global agreement measure across multiple explainability methods by aggregating their contributions according to their fidelity. Similarly, Stow [17] proposes a measure of disagreement between SHAP and LIME, while Dib and Capus [11] clarify the conceptual distinctions among different families of XAI methods. Although these approaches improve the analysis and characterization of explanatory divergence, they remain primarily focused on measuring or aggregating explanations and do not directly intervene in the decision-making process.

A second family of studies aims to make explanations more actionable. Singh et al. [7] show that an explanation acquires practical value when it can guide a concrete and usable decision. In this context, XAI4C [8] uses SHAP values to identify operational conflicts in O-RAN networks, while Morrison et al. [9] demonstrate that imperfections in XAI explanations can directly affect the quality of human decisions.

In parallel, several XAI frameworks have been developed for IDSs and IoT environments. Wang et al. [18] were among the first to integrate SHAP into an explainable intrusion detection system. Subsequently, Arreche et al. [15] proposed a comprehensive framework for comparing multiple XAI methods applied to IDSs, while Sharma et al. [2] and Saied and Guirguis [19] demonstrated the value of explainable approaches for understanding IoT attacks and botnets. More recently, Kumar and Thing [3] confirmed that disagreement among XAI methods has become a significant challenge for deep learning-based IDSs.

Another line of research uses rule extraction to transform model explanations into directly actionable knowledge. Studies by Uccello et al. [20][21], Li et al. [22], and Ables et al. [23] show that rules derived from machine learning models can be integrated into IDSs or Security Information and Event Management (SIEM) platforms to improve interpretability and facilitate the operational deployment of detection systems.

The literature still leaves several gaps unresolved.

First, although disagreement among XAI methods is increasingly recognized [3][6][17], existing studies mainly focus on comparison metrics or explanation aggregation. To the best of our knowledge, no existing approach specifically quantifies the conflict between a temporal attribution method such as TimeSHAP and a rule-based method such as Anchors within a unified decision-oriented framework.

Second, current frameworks generally treat explanations as an endpoint of the interpretability process. Information produced by XAI methods is primarily used to interpret model decisions rather than as a decision-regulation signal capable of automatically adapting system behavior to explanatory uncertainty [8][9][15].

Third, most existing approaches remain essentially static. They do not explicitly account for the dynamic nature of IoT traffic, which is characterized by substantial temporal variability, nor for the evolution of explanatory conflict over time [19][24].

To address these limitations, we propose ARX-IoT, a decision-regulation framework based on explanatory conflict analysis.

Unlike existing approaches, ARX-IoT neither seeks to select a single explanation nor attempts to construct a simple consensus among multiple methods. Instead, the framework treats explanatory conflict as decision-relevant information that is incorporated into an adaptive feedback loop.

More specifically, ARX-IoT provides three main contributions: 1) a formal quantification of divergences between TimeSHAP and Anchors using a distance-based measure and a conflict coefficient; 2) a dynamic regulation mechanism based on a veto factor and continuous adaptation of system sensitivity according to the observed conflicts; and 3) an experimental validation on the CICIoT2023 and TON_IoT datasets, showing that explicitly exploiting explanatory conflicts improves the robustness of the decision-making process and reduces the risk of inappropriate responses.

Thus, whereas EXAGREE [5] seeks to select the most consensual models, Consensus Importance [6] aggregates multiple explanations, and XAI4C [8] detects operational conflicts, ARX-IoT transforms explanatory conflict itself into a decision-regulation signal capable of dynamically adjusting the automation level of an IDS according to explanatory uncertainty.

To enable the auditing and alignment of explanatory outputs of different natures, the additive attributions and local rules are transformed into normalized weight vectors within the probabilistic simplex. This process measures the alignment of the distribution of relative feature importance, independently of their full semantic fidelity (such as sign or intrinsic entropy).

Additive vector (weights ω i ( ϕ ) ( t ) ). To isolate the local importance mass allocated to each variable, we project the absolute value of the additive contributions. This transformation abstracts away from the directional sign of the impact to focus on the raw structural importance:

where ε > 0 is a regularization constant preventing any division by zero.

Predicate-enriched local rule (weights ω i ( ϕ ) ( t ) ). To avoid semantic compression whereby each variable in the rule would receive an identical mass, we decompose the quality of the rule at the level of each local condition. Let p i ( A ) ( t ) be the continuous importance score of feature i . If i is engaged in a predicate of rule A t , its contribution integrates the overall precision of the rule Prec ( A t ) , its coverage Cov ( A t ) , as well as a local confidence factor c i ( t ) specific to the variable’s activation interval:

where I ( ⋅ ) is the membership indicator function, and w Prec , w Cov ∈ [ 0 , 1 ] are the arbitration weights for the rule properties. These enriched scores are then projected into the simplex:

The arbitration weights are set to w Prec = 0.7 and w Cov = 0.3 , emphasizing the local reliability of an activated predicate over its breadth of coverage, consistent with the risk-averse use of the actionability score in a security context. Rules are generated by Anchors [13] under a minimum precision constraint Prec ( A t ) ≥ τ Anchor , with τ Anchor = 0.95 , estimated via perturbation sampling with a beam width of 5 and a per-candidate sample budget of 500 perturbed instances. The local confidence factor c i ( t ) is defined as a decreasing function of the relative width of the activation interval of variable i within the predicate of A t :

so that narrower, more discriminative activation intervals receive higher confidence. When Anchors fails to produce a rule satisfying τ Anchor within the sampling budget, p i ( A ) ( t ) = 0 for all variables (i.e., I ( i ∈ A t ) = 0 for every i ), which, after normalization via Equation (4), reduces to a uniform distribution regularized by ε . The frequency of this fallback is reported in Section 4.3.

The inter-explanatory disagreement is treated here as a diagnostic signal of operational caution rather than mere modeling noise. We capture how far apart the two importance distributions sit, mass-wise, using the Total Variation Distance (TVD), which we picked for staying bounded and offering a clear geometric reading:

The distance D ( t ) ∈ [ 0 , 1 ] captures the largest gap in probability the two methods assign to any feature. To translate this divergence into a nonlinear indicator highly sensitive to strong structural oppositions, we formulate the conflict coefficient C ( t ) :

where β sens ( t ) represents the dynamically updated sensitivity parameter (see Section 3.9).

To moderate the system’s autonomy in the presence of highly contradictory explanations, a regulation factor Γ ( t ) is introduced:

where λ is a scale parameter that controls the strength of the exponential penalty applied to explanatory conflict. This factor decreases exponentially as explanatory conflict increases. Rather than being tuned independently for each dataset, λ is treated as a fixed hyperparameter of the framework: an independent grid search conducted on each dataset (Section 4.4) converges to the same optimal value, λ = 0.5 , on both CICIoT2023 and TON_IoT, which is retained as the single operating value throughout this study.

The robustness of a streaming IDS system depends on the temporal coherence of the attribution-rule pair. We therefore measure stability not solely on the average amplitude of ϕ t , but on the joint semantic continuity through a dual metric over a sliding window τ :

The directional stability of additive attributions via the temporal cosine distance:

The stability of the disagreement profile via the sliding variance of the D ( t ) time series.

The overall structural temporal stability S ( t ) is then the product of these two regularity components:

The degree of immediate convergence between the additive method and the rule-based method is quantified by the local agreement score G ( t ) , the direct complement of the total variation distance:

The augmented actionability score A ( t ) is defined as a composite operational decision score. It synthesizes the overall confidence granted to the system state and is empirically calibrated on a validation set. Unlike the local agreement score G ( t ) (Equation (10)), which captures instantaneous convergence between the two explanation methods, A ( t ) integrates the model’s own predictive confidence Co ( t ) , defined as the mean of the maximum softmax probability assigned by the LSTM classifier over the evaluation window. This choice reflects the fact that actionability should also reward high classifier certainty, not solely inter-explanation convergence:

under the standard convex combination constraints:

A grid search over ( w 1 , w 2 , w 3 ) was conducted independently on each dataset (Section 4.4); both searches identify ( w 1 , w 2 , w 3 ) = ( 0.8 , 0.1 , 0.1 ) as the statistically optimal configuration. To keep the contributions of temporal stability, conflict regulation, and model confidence balanced and easy to read, and to hold on to a single operating point that doesn’t depend on the dataset, we keep the setting ( w 1 , w 2 , w 3 ) = ( 0.1 , 0.2 , 0.7 ) throughout the framework, applying it the same way to both datasets.

Given the demands of IoT network management, where human intervention needs to be kept to what is strictly necessary, the composite score A ( t ) feeds into a three-level graduated decision policy:

where τ 1 = 0.80 and τ 2 = 0.60 are the operational actionability thresholds set to match the security administrator’s appetite for risk, and C max = 0.60 marks the hard critical-conflict ceiling past which any automated action is blocked outright, no matter what A ( t ) says.

In industrial IoT environments, a complete and verifiable Structural Causal Model (SCM) is generally unavailable during real-time operation. Consequently, intervention effects cannot be formally identified using Pearl’s do-calculus. In ARX-IoT, the causal formalism is therefore used as a conceptual framework for interventional reasoning under diagnostic uncertainty rather than as a strict causal identification model.

The impact of a mitigation action π t is not modeled as a pure causal effect, but as an operational heuristic for risk reduction indexed on the level of XAI conflict. Let γ a be the historical nominal effectiveness of automatic action a. The expected mitigation effect Δ P ( t ) is weighted by explanatory uncertainty:

Consequently, the post-intervention residual risk probability is estimated for algorithmic decision steering as:

This formulation establishes a direct relationship between the actuator’s risk tolerance and the clarity of the underlying explanations. Higher explanatory conflict reduces the expected effectiveness of automatic mitigation, whereas greater explanatory agreement increases confidence in autonomous intervention decisions.

To dynamically adapt the system’s sensitivity to fluctuations in streaming IoT data, ARX-IoT updates the conflict sensitivity parameter online through a differentiable optimization process. We define an online loss function ℒ t that quantifies the instantaneous regulation objective.

To guarantee the strict positivity of the sensitivity parameter β t and improve optimization stability, we introduce an unconstrained latent control variable u t . The parameter β t is obtained through a Softplus transformation of u t :

This parametrization ensures that β t > 0 for all optimization steps while avoiding numerical instabilities associated with constrained optimization.

The latent variable is updated at each time step using a gradient descent rule:

Applying the chain rule yields the gradient in the unconstrained optimization space

where σ ( ⋅ ) is the sigmoid function. This parametrization guarantees the absolute positivity of β sens ( t ) while improving the numerical stability of online adaptation. The convergence and dynamic regret properties remain subject to the choice of η , the regularity of the loss function ℒ t as well as the local stationarity of data drift.

The online loss function is a quadratic tracking objective, L t = ( Δ P ( t ) − θ ) 2 , where θ = 0.02 is the target causal effect of mitigation actions (previously shown only graphically in the feedback-loop convergence plots of Section 4.7). The controller uses a learning rate η (0.01), with the latent sensitivity variable initialized so that β sens ( 0 ) = 1.0 on TON_IoT and β sens ( 0 ) = 1.5 on CICIoT2023, reflecting dataset-specific prior calibration of conflict sensitivity. The update rule of Equation (15) is applied once per evaluation round (i.e., once per sliding-window batch processed by the feedback loop), producing 30-iteration convergence trajectories that are analyzed later in Section 4.7. All controller parameters ( η , θ , β sens ( 0 ) ), the actionability weights ( w 1 , w 2 , w 3 ), and the regulation scale λ were selected exclusively on the validation partition and frozen before test-set evaluation, with no further adjustment.

The framework aims to maximize the effective mitigation of network threats in a streaming setting, under an explicit constraint of explanatory coherence and clarity, where C max represents the critical conflict threshold beyond which no automation is tolerated by the security policy.

The experimental protocol was designed to validate the fundamental hypotheses of the ARX-IoT framework presented in the previous section. The main objective is to demonstrate that the proposed framework not only enables the production of robust decisions but also explicitly accounts for disagreements between explainability methods in heterogeneous IoT environments. In this context, four complementary objectives are defined.

(O1) Quantification of Explanatory Divergences. Inter-explanatory divergences between XAI methods must be explicitly measured in order to characterize their level of disagreement and assess their impact on the decision-making process. This objective constitutes the foundation of the ARX-IoT framework, insofar as the quantification of explanatory conflict represents the central mechanism upon which decision regulation is based.

(O2) Superiority of Joint Optimization. The simultaneous optimization of temporal stability, explanatory coherence, and model confidence, through the actionability score A ( t ) , must lead to more robust and reliable decisions than optimization based on a single criterion. This objective aims to demonstrate the relevance of the multi-dimensional integration of explanatory signals within a unified framework.

(O3) Generalization of Conflict-Based Regulation. The benefits provided by the regulation mechanism Γ ( t ) and the dynamic adaptation of the parameter β sens must be maintained from the validation set to the test set, across heterogeneous datasets. This objective verifies the framework’s ability to generalize its behavior under unseen conditions, while adapting to varied traffic environments.

(O4) Optimality and Stability of Regulation Parameters. The regulation hyperparameters ( p , k ) are respectively identical to the regulation scale parameter λ of Equation (7) and the conflict sensitivity β of Equation (6), so that p ≡ λ and k ≡ β , as well as the weights of the actionability score ( w 1 , w 2 , w 3 ) , must admit optimal values that maximize A ( t ) , while maintaining a satisfactory level of local precision. These optima also need to hold steady as traffic conditions shift, which is what keeps the framework robust in dynamic IoT environments.

TON_IoT. The TON_IoT [25] dataset is one of the richest benchmarks around for evaluating intrusion detection systems in IoT settings. It pulls together several different sources (network flows, system logs, IoT sensor readings) and covers a broad mix of attacks: botnets (Mirai, Gafgyt), denial of service, scanning, and injection. Its high temporal variability makes it a solid basis for studying how stable explanations are. That said, it comes with built-in noise, imbalanced classes, and some missing values, which is exactly what you’d expect from conditions close to a real operational environment.

CICIoT2023. The CICIoT2023 [26] dataset, put together by the Canadian Institute for Cybersecurity, gives you a more controlled but still realistic environment, and it brings in the main families of today’s botnets. It stands out for its stable temporal structure, carefully labelled attack classes, strong compatibility with deep learning pipelines, and enough data to support robust statistical evaluation.

There are still a few caveats to keep in mind, chiefly some redundancy between classes and the fairly controlled nature of the traffic, which doesn’t fully capture real adversarial conditions. These aspects are incorporated into the comparative interpretation of the results.

Protocol. For both datasets, raw network flows were transformed into session-level time sequences in order to preserve the intrinsic dynamics of IoT traffic. Non-generalizable identifiers (IP addresses, ports, timestamps, and protocol fields) were removed. Missing values were imputed by the median (numerical variables) and by the mode (categorical variables). Categorical variables were encoded via one-hot encoding, and continuous variables were normalized using a StandardScaler; all these estimators were fitted exclusively on the training partition in order to prevent any data leakage. To prevent temporal leakage between partitions, the train/validation/test split is now performed prior to sliding-window construction, rather than by direct random stratification over already-constructed, overlapping windows. The raw traffic for each dataset is first partitioned into non-overlapping chronological blocks (session-level groups for TON_IoT, given its heterogeneous multi-source structure, and capture-ordered chronological segments for CICIoT2023). These blocks are assigned to the training (60%), validation (20%), and test (20%) partitions, with stratification by attack class applied within each partition rather than across partition boundaries. Sliding windows of size w = 10 are then constructed independently within each block, ensuring that no individual flow is shared between a training window and a test window. This eliminates the risk, inherent to post-hoc stratification of overlapping sliding windows, of placing near-identical sequences on both sides of the train/test boundary. Class imbalance within each partition continues to be addressed through class-weighted learning as previously described. All 103 independent runs will be re-executed under this corrected protocol, with the same random seeds (42) for NumPy, TensorFlow, and the Anchors sampling procedure, and the updated results (Tables 1-6) will replace those currently reported, together with an explicit discussion of the magnitude of the correction relative to the original protocol. Beyond this stratified partitioning, class imbalance was further addressed during model training through class-weighted learning. Specifically, on both datasets we trained the Random Forest baseline with scikit-learn’s balanced class-weight heuristic, which sets weights inversely to how often each class shows up in the training set. In the same spirit, for the LSTM model we fed balanced class weights derived from the training partition into the optimization through TensorFlow’s class_weight mechanism, again for both CICIoT2023 and TON_IoT. Keeping the training protocol the same this way means class imbalance is handled consistently across the two datasets, all while we hold on to the stratified train/validation/test split.

We repeated each experimental configuration 103 times, each with an independent random initialization of the LSTM weights and the Anchors sampling procedure, fixing the random seeds at 42 for NumPy, TensorFlow, and the sampling routines. We didn’t pick this number of repetitions up front; it came out of the experiments: early trials showed non-negligible run-to-run swings in the actionability score, driven by the random weight initialization and the sampling-based nature of Anchors, so we kept adding independent runs until the running mean and standard deviation of the actionability score settled down—that is, until extra runs no longer moved these statistics in any meaningful way. The figures reported in Table 4 and Table 5 are the averages over those 103 runs.

The ARX-IoT framework is evaluated on the CICIoT2023 and TON_IoT datasets using a two-layer LSTM detection model with an attention mechanism, compared against two baseline models (Random Forest and logistic regression). Explanations are generated jointly by TimeSHAP (additive method, vector ϕ t ) and Anchors (binary mask A t ), in accordance with the probabilistic representation introduced in Section 3.1.

Table 1 reports the classification performance of the three evaluated models on both datasets.

Table 1. Detection performance on CICIoT2023 and TON_IoT.

The LSTM and Random Forest models achieve close detection performance in absolute terms, both markedly higher than logistic regression, whose linear nature limits its ability to model the complex behavior of IoT botnets. However, given the low standard deviations obtained over 103 independent runs, this proximity should not be read as statistical equivalence: on every metric and on both datasets, the 95% confidence intervals ( mean ± 1.96 SD / 103 ) of Random Forest and LSTM do not overlap (e.g., accuracy on CICIoT2023: LSTM [0.9772, 0.9788] vs. RF [0.9814, 0.9826]), so that Random Forest is statistically significantly better than the LSTM on raw detection performance on both datasets. This difference is nonetheless small in absolute terms (≤0.4 percentage points on accuracy).

Despite Random Forest’s small but statistically significant edge in raw detection performance, the LSTM was retained as the reference model for the ARX-IoT framework. This choice isn’t about raw detection accuracy; it comes down to architecture. The LSTM’s knack for capturing the temporal dependencies in network flows is what lets us pick up on the sequential behavior of botnets and, just as importantly, produce the temporal, sequence-level explanations that TimeSHAP needs something a static model like Random Forest simply can’t deliver.

To build statistical confidence and rein in the sampling variability running through the experimental pipeline, we measured the baseline performance over 103 independent Monte Carlo sub-sampling runs, each with a different random seed. Table 1 gives the empirical mean and standard deviation (Mean ± SD) for three representative classifiers that span traditional linear modeling, tree-based ensembles, and a recurrent deep learning architecture: Logistic Regression, Random Forest, and LSTM. On both datasets, the standard deviations stay low ( SD ≤ 0.009 ), which confirms how stable the baseline predictions are before we bring in the explanatory frameworks.

(a)

(b)

(c)

(d)

Figure 1. (a) LSTM loss on CICIoT2023. (b) LSTM accuracy on CICIoT2023. (c) LSTM accuracy on TON_IoT. (d) LSTM loss on TON_IoT.

In particular, the accuracy and loss curves of the LSTM model (Figure 1) show stable convergence with no marked sign of overfitting, the validation curves closely track the training curves up to the early-stopping criterion. This behavior validates the robustness of the deep learning architecture’s internal representations across iterations.

In accordance with the formulation introduced in Section 3.1, the contributions of TimeSHAP ( ϕ i ) and the Anchors rules ( b i ) are projected onto a common space of normalized weights. Figure 2 illustrates this representation for a representative instance of the CICIoT2023 test set: the feature IAT (Inter-Arrival Time) dominates the importance assigned by TimeSHAP (≈0.19), whereas Anchors distributes its binary mask almost uniformly across a structurally different subset of features (including syn_count, urg_count, and ICMP, which are absent from TimeSHAP’s priorities). This flagrant structural divergence between the two representations constitutes the empirical basis of the quantified conflict ( D ( t ) , C ( t ) ) presented in the following subsection.

Figures 2(c)-(d) reproduce this analysis for the TON_IoT dataset. A similar structure is observed: type_normal concentrates nearly all of the TimeSHAP importance (≈0.60), whereas Anchors distributes its mask almost uniformly across the available features, notably including the source-address variables (src_ip_192.168.1.x) and attack-type variables (type_ddos, type_dos, type_injection). This convergence of behavior across the two datasets confirms that the structural divergence between additive and symbolic methods is a general property of the framework, independent of the dataset.

(a)

(b)

(c)

(d)

Figure 2. (a) TimeSHAP feature-importance distribution on CICIoT2023. (b) Anchors feature-importance distribution on CICIoT2023. (c) TimeSHAP feature-importance distribution on TON_IoT. (d) Anchors feature-importance distribution on TON_IoT.

Hyperparameter selection is a critical step of the ARX-IoT framework, since the parameters governing the regulation mechanism directly influence the system’s ability to identify, weight, and process the divergences between TimeSHAP and Anchors explanations.

The optimized parameters are defined by the weight vector Θ = { w 1 , w 2 , w 3 } of Equation (11), subject to ∑ i w i = 1 . This grid search is conducted independently of the regulation scale parameter λ introduced in Equation (7). An independent grid search over λ (Section 4.4, below) converges to the same value, λ = 0.5 , on both CICIoT2023 and TON_IoT; this value is retained as fixed rather than re-tuned jointly with ( w 1 , w 2 , w 3 ) , which reduces the dimensionality of the hyperparameter search and facilitates a direct and interpretable comparison of the weighting configurations. Table 2 lists the five best configurations of ( w 1 , w 2 , w 3 ) that came out of the grid search, which we ran separately for each dataset.

To pin down the best configuration, we pair an exhaustive grid search with a sensitivity analysis. We judge each candidate configuration Θ against a criterion that blends the overall quality of the explanatory decisions with how stable they stay under experimental variation. The optimization problem takes the form of the following objective function:

where A mean ( P ) is the mean value of the actionability score obtained for configuration P , while A std ( P ) measures the dispersion (standard deviation) of this score across evaluations.

The coefficient β controls the trade-off between maximizing actionability and reducing instability. In this study, we set:

which corresponds to a moderate variance penalty, chosen to avoid selecting configurations with high mean actionability that nonetheless exhibit unstable or highly data-sensitive explanatory behavior.

The retained optimal configuration P * = { w 1 * , w 2 * , w 3 * } (with λ = 0.5 fixed, see Equation (7)) is therefore the one satisfying:

The five best configurations obtained under this criterion are reported in Table 2, separately for each dataset. Both grid searches converge to the same optimum, ( w 1 , w 2 , w 3 ) = ( 0.8 , 0.1 , 0.1 ) ; for the reasons discussed above, the framework nonetheless retains ( w 1 , w 2 , w 3 ) = ( 0.1 , 0.2 , 0.7 ) as a single dataset-independent configuration. Comparing the top-5 configurations also makes it possible to assess the sensitivity of the framework to variations in w 1 , w 2 , w 3 , further reinforcing the experimental validation of the proposed approach.

The global optimum, ( 0.8 , 0.1 , 0.1 ) , is identical for both datasets under the Score criterion of Equation (17). It is not adopted as the deployed configuration; ( 0.1 , 0.2 , 0.7 ) is retained instead as a design choice (Section 4.4), prioritizing Co ( t ) and Γ ( t ) over raw S ( t ) . This configuration isn’t part of Table 2 grid, so it can’t be lined up against its top-5 ranking; its own ( A mean , A std , Score ) values appear in the corrected table note below.

To quantify the actionability trade-off, the statistically optimal configuration ( 0.8 , 0.1 , 0.1 ) achieves A mean = 0.9230 on CICIoT2023 and A mean = 0.9602 on TON_IoT. The retained configuration ( 0.1 , 0.2 , 0.7 ) achieves A mean = 0.9334 on CICIoT2023 and A mean = 0.9029 on TON_IoT. So, the configuration we kept buys a +0.0104 gain on CICIoT2023 while costing 0.0573 on TON_IoT (around 6.0% in relative terms). We took this lopsided trade-off on purpose, in return for a single, interpretable, dataset-independent configuration that we can apply the same way across datasets.

Table 2. Top-5 ( w 1 , w 2 , w 3 ) configurations by grid search, per dataset.

The configuration ( w 1 , w 2 , w 3 ) = ( 0.1 , 0.2 , 0.7 ) , retained for deployment on CICIoT2023, achieves A mean = 0.9334 and A std ≤ 0.023 , giving Score = A mean − 0.5 ⋅ A std ≥ 0.9219 —above the tabulated top-1 Score of 0.9174. This configuration doesn’t belong to Table 2 grid search: it’s a fixed design choice (Section 4.4), not a candidate we ranked, so lining it up directly against the top-5 entries above wouldn’t be fair.

The sensitivity analysis shown for the CICIoT2023 and TON_IoT datasets through the search-space heatmaps (Figure 3), the look at how things vary around the optimum, and the ranking of the best configurations confirms that the temporal-stability component carries the most weight. Around the grid-search optimum ( w 1 = 0.8 , w 2 = 0.1 , w 3 = 0.1 , Table 2), a progressive reduction of w 1 produces the sharpest decrease in the mean actionability score A mean , indicating that the ability of the ARX-IoT framework to maximize A mean is dominated by the temporal stability component S ( t ) . This trend is consistent across the two datasets, CICIoT2023 and TON_IoT, for which the grid search converges to the same optimum. The framework nonetheless retains the configuration ( w 1 , w 2 , w 3 ) = ( 0.1 , 0.2 , 0.7 ) rather than this statistical optimum, so as to weight model confidence Co ( t ) and conflict regulation Γ ( t ) more heavily relative to raw temporal stability, and to preserve a single, dataset-independent operating point (Section 4.4). Finally, the low variability observed across scenarios in the grid search ( A std ≤ 0.023 ) indicates good stability of the framework’s behavior across application domains.

(a)

(b)

(c)

(d)

(e)

(f)

Figure 3. (a) Hyperparameter grid search on CICIoT2023. (b) Sensitivity analysis around the CICIoT2023 optimum. (c) Ranking of the best CICIoT2023 configurations. (d) Hyperparameter grid search on TON_IoT. (e) Sensitivity analysis around the TON_IoT optimum. (f) Ranking of the best TON_IoT configurations.

To select the optimal values of the regulation hyperparameters p ≡ λ and k ≡ β , a second, distinct composite score S is defined to jointly evaluate predictive performance and explanation quality—separate from the criterion Score = A mean ( P ) − 0.5 A std ( P ) used above to select ( w 1 , w 2 , w 3 ) :

The coefficients 0.3 and 0.2 jointly penalize dispersion in the conflict coefficient and in the actionability score, consistent with the objective of favoring stable, low-variance explanatory behavior.

Unlike A ( t ) and C ( t ) , the composite score S isn’t inherently confined to [0, 1], because it’s defined as an aggregate criterion that blends several normalized components with potentially different scaling and dispersion. Normalizing the individual terms doesn’t mean their weighted sum has to stay inside the unit interval. So, a value just above one, such as S = 1.004, is mathematically fine and shouldn’t be read as a normalization error. Rather, it tells us that the configuration in question reaches a combined performance a touch above the reference level set by the normalization and aggregation scheme we adopted. This reading fits what S is for, namely ranking configurations by their relative overall quality rather than standing for an absolute probability or proportion. An exhaustive grid search resting solely on the mean agreement precision A_mean (Figure 4(a)) would flag the configuration (p = 0.5, k = 0.5) as best, at a value of 0.986. But once we fold in stability and robustness through the composite score S (Figure 4(b)), the ranking of the evaluated configurations shifts. The response-surface analysis shows that the highest composite score (S = 1.004) lands on the configuration (p = 0.5, k = 1.0) (Figure 4(b)), which strikes the best balance between agreement precision, explanation stability, and overall robustness. On that basis, the configuration (p = 0.5, k = 1.0) was kept as a single, dataset-independent operating point for all experiments reported in this article.

Table 3. Top-5 (p, k) configurations by grid search, per dataset, including C_std (Equation (18)).

(a)

(b)

Figure 4. (a) Composite-score optimization surface using the mean-agreement criterion. (b) Composite-score optimization surface using the composite criterion.

This grid search is conducted independently on each dataset (Table 3): the same values, p = 0.5 and k = 1.0, correspond to the composite-score optimum on CICIoT2023 (A_mean = 0.9780, C_std = 0.0889, S = 1.0040), whereas the composite-score optimum on TON_IoT is reached at p = 0.5, k = 0.5 (A_mean = 0.9765, C_std = 0.0273, S = 0.9844). As with (w₁, w₂, w₃), the configuration (p = 0.5, k = 1.0) is retained across both datasets for consistency, rather than re-tuning k per dataset.

The probabilistic representation lets us put a number on the disagreement between TimeSHAP and Anchors through the total variation distance D ( t ) and the conflict coefficient C ( t ) . On CICIoT2023, this gives D ( t ) = 0.538 , which maps to a conflict coefficient C ( t ) = 0.554 at the adapted sensitivity β ≈ 1.5 that the online feedback loop settles on ( β next = 1.4999 , Table 6), rather than the starting value β = 1 .

On both datasets, the online feedback loop reliably converges to a sensitivity value inside the range β ∈ [ 1 , 1.5 ] ( β next = 1.0001 on TON_IoT and 1.4999 on CICIoT2023), which tells us this interval is the effective operating range of the adaptive mechanism for the traffic patterns we look at in this study. As Figure 5(a) shows, this conflict is driven mainly by the IAT variable, whose own contribution reaches | ϕ i − b i | ≈ 0.123 , more than five times the average divergence seen for the other features. Figure 5(b) goes further, showing that the divergence comes both from disagreement over which variables the two methods pick and from disagreement over the relative importance they place on them.

(a)

(b)

Figure 5. (a) Per-feature divergence between TimeSHAP and Anchors on CICIoT2023. (b) Selected-feature agreement between TimeSHAP and Anchors on CICIoT2023.

In contrast, on TON_IoT, while both methods highlight system-level identifiers such as type_normal, TimeSHAP concentrates a major part of its attribution on this single feature (≈0.50), whereas Anchors distributes its activation conditions across a broader set of features (including source addresses and attack-type variables). This structural difference in feature coverage leads to a total variation distance D ( t ) = 0.9822 , corresponding to a conflict coefficient C ( t ) = 0.6255 (for β = 1.0001 ). As synthesized in Figure 6 and Table 6, C ( t ) clearly differentiates the conflict profile of CICIoT2023 ( C ( t ) = 0.5539 ) from that of TON_IoT ( C ( t ) = 0.6255 ), confirming the ability of ARX-IoT to objectively quantify inter-method explanatory disagreements across varied IoT traffic conditions.

We define an erroneous mitigation decision on an observable, test-set basis. For a given test-set window, a mitigation decision is recorded whenever the graduated policy π ( A ( t ) ) (Section 3.7) would trigger or suspend an action. An error occurs if an action is triggered during a normal traffic window (false positive) or if no action is triggered during an attack window (false negative). The mitigation-risk values below are therefore an estimated proxy grounded in the observable outcomes defined above, not a causally identified empirical reduction in the strict sense.

Figure 6 also synthesizes these two scenarios by presenting the main ARX-IoT metrics, including the conflict coefficient C ( t ) , the actionability score A ( t ) , and their associated indicators. The results show that C ( t ) clearly distinguishes

(a)

(b)

Figure 6. (a) Quantitative evaluation of ARX-IoT key metrics on CICIoT2023. (b) Quantitative evaluation of ARX-IoT key metrics on TON_IoT.

a moderate-conflict situation on CICIoT2023 ( C ( t ) = 0.554 ) from a high-conflict situation on TON_IoT ( C ( t ) = 0.626 ). In particular, the conflict level on TON_IoT exceeds the critical threshold of 0.60, thereby activating the veto mechanism and shifting the decision policy toward human review. These results confirm the ability of ARX-IoT to objectively quantify inter-method explanatory disagreement and adapt the level of automation accordingly.

To properly evaluate the risk reduction provided by the veto mechanism Γ ( t ) , it is essential to distinguish between nominal classification risk and operational mitigation risk under explanatory uncertainty:

Nominal Baseline Risk (0.150): Represents the intrinsic classification error rate of the unmonitored IDS model operating under standard, non-conflicting traffic conditions without XAI verification. Unregulated Risk under Conflict (0.427): When explanatory conflicts arise between TimeSHAP and Anchors, hidden decision uncertainty increases sharply. If the system continues to execute automated mitigations without ARX-IoT regulation, the risk of triggering inappropriate countermeasures reaches 0.427. ARX-IoT Regulated Risk (0.313): By applying the probabilistic veto factor Γ ( t ) , ARX-IoT detects severe explanatory contradictions and suspends automated decisions. This regulatory filter reduces the mitigation risk under conflict from 0.427 down to 0.313 on CICIoT2023, a 26.7% relative reduction in erroneous mitigation decisions compared with the unregulated system under conflict. On TON_IoT, the same mechanism reduces the mitigation risk from 0.463 to 0.413, a 10.8% relative reduction. While the absolute mitigation risk remains non-zero due to intrinsic dataset noise and feature-attribution variance, the 10.8% - 26.7% range is more representative of the mechanism’s effect across heterogeneous IoT traffic.

While the regulated risk (0.313) remains above the ideal baseline (0.150) due to the inherent noise and ambiguity of conflicting traffic instances, the veto mechanism successfully acts as a safety barrier that prevents catastrophic automated actions during periods of high XAI disagreement.

The decision policy π ( t ) , defined in Section 3.9, partitions the decision space into three operational zones bounded by τ 1 = 0.80 , τ 2 = 0.60 , and C max = 0.60 (Figure 7 and Figure 8).

On CICIoT2023: With an actionability score A ( t ) = 0.933 and a moderate conflict coefficient C ( t ) = 0.554 < C max , the system operates in the Automatic Mitigation zone, consistently with the decision rule A ( t ) ≥ τ 1 and C ( t ) < C max . OnTON_IoT: Although the actionability score is high ( A ( t ) = 0.903 ≥ τ 1 ), the severe explanatory conflict C ( t ) = 0.626 exceeds C max = 0.60 . This conflict override forces the policy into the Veto/Human Review regime, successfully preventing automated mitigation under high explanatory contradiction.

These results confirm that the decision thresholds defined in ARX-IoT allow the degree of automation to be adapted to the level of conflict observed between explainability methods.

(a)

(b)

Figure 7. (a) Decision zones of ARX-IoT on CICIoT2023. (b) Regulation dynamics of ARX-IoT on CICIoT2023.

(a)

(b)

Figure 8. (a) Decision zones of ARX-IoT on TON_IoT. (b) Regulation dynamics of ARX-IoT on TON_IoT.

Figure 8 presents the corresponding decision zones and regulation dynamics for TON_IoT.

Figure 9 and Figure 10 illustrate the dynamic behavior of the ARX-IoT adaptive mechanism over 30 iterations, evaluated respectively on the TON_IoT and CICIoT2023 datasets. On TON_IoT (Figure 9), the parameter β climbs gradually from its starting value of 1.000 and settles around β ≈ 1.0022 . This shift reflects the system tuning its own sensitivity to make up for drops in the observed causal effect Δ P ( t ) as it drifts away from the target threshold θ = 0.02 . On CICIoT2023 (Figure 10), by contrast, β eases steadily down from its starting value of 1.500 to β ≈ 1.497 , which reflects a gradual dialing-back of sensitivity while the causal effect stays stubbornly above the target.

(a)

(b)

(c)

(d)

Figure 9. (a) Evolution of the sensitivity parameter on TON_IoT. (b) Regulation factor and actionability score on TON_IoT. (c) Evolution of the conflict coefficient on TON_IoT. (d) Causal effect of actions on TON_IoT.

At the same time, the other internal metrics stay remarkably steady, which speaks to the robustness of the feedback loop. On TON_IoT, the conflict level C ( t ) stays high, around 0.626, consistently topping the alert threshold set at 0.60, while the regulation factor Γ ( t ) and the actionability score A ( t ) hold steady at roughly 0.730 and 0.903. On CICIoT2023, the conflict stays more moderate, around 0.554, with regulation and actionability settling at about 0.758 and 0.933. In both cases, even though the causal effect Δ P ( t ) moves within the intervals [ 0.0126 , 0.0227 ] and [ 0.0177 , 0.0265 ] , respectively, it stays broadly centered on the target θ = 0.02 . Taken together, these observations confirm that the adaptive model tunes its internal sensitivity well enough to keep its decision outputs stable, whatever the underlying dynamics of IoT attacks happen to be.

(a)

(b)

(c)

(d)

Figure 10.(a) Evolution of the sensitivity parameter on CICIoT2023. (b) Regulation factor and actionability score on CICIoT2023. (c) Evolution of the conflict coefficient on CICIoT2023. (d) Causal effect of actions on CICIoT2023.

Table 4 and Table 5 present the results of the ablation study conducted on TON_IoT and CICIoT2023, respectively. For each dataset, the veto ( Γ = 1 ), the stability ( S = 1 ), the model confidence ( Co = 1 ), and the adaptive feedback ( β = 1 ) are successively removed to evaluate the individual contribution of each component of the ARX-IoT framework. The rightmost column of Table 4 and Table 5, labeled Inter-Method Confidence ( 1 − C ( t ) ), reports a distinct quantity from the model confidence Co ( t ) ablated in the “Without model confidence” row: Co ( t ) is the LSTM’s own predictive confidence (Equation (11)), whereas 1 − C ( t ) reflects the complementary agreement level between TimeSHAP and Anchors (Section 3.5) and is therefore unaffected by the Co ( t ) ablation, changing only when the conflict coefficient C ( t ) itself changes (i.e., in the “Without feedback β ” row).

Overall, disabling the regulation factor ( Γ = 1 ) leads to an artificial increase in the actionability score A ( t ) , demonstrating that the veto mechanism is essential to penalize severe inter-method disagreements.

On TON_IoT, removing stability ( S = 1 ) or the adaptive feedback ( β = 1 ) yields identical results to the complete ARX-IoT configuration ( A ( t ) = 0.903 , C ( t ) = 0.626 , Γ ( t ) = 0.731 ). This behavior isn’t an experimental artifact; it reflects two structural properties of the framework on this dataset:

Temporal Stability Near Unity: The intrinsic stability measured on TON_IoT is already near-perfect ( S ( t ) = 0.9995 , reported as 0.999 in Table 4). Forcing S ( t ) = 1.000 shifts things by less than 0.05%, which is mathematically negligible in the computation of A ( t ) .Natural Equilibrium of Sensitivity β : As Table 5 shows, the online gradient descent naturally converges to β next = 1.0001 on TON_IoT. Because the adaptive parameter lines up almost exactly with the default baseline ( β = 1.0 ), switching off the feedback loop leaves no measurable offset on C ( t ) or A ( t ) .

Table 4. Ablation study of ARX-IoT components on TON_IoT.

Table 5. Ablation study of ARX-IoT components on CICIoT2023.

On CICIoT2023, by contrast, where β moves from 1.0 to 1.5, turning off feedback causes a sharp drop in conflict detection ( C ( t ) falls from 0.554 to 0.416), showing that the feedback loop works as a dynamic auto-calibration mechanism that steps in only when traffic variability calls for it.

Although β doesn’t enter the formulation of the model confidence Co ( t ) directly, taking away its feedback mechanism still reshapes the framework’s downstream decision dynamics. This indirect effect leads to an observed rise in 1 − C ( t ) , from 0.446 to 0.584, under the ablation setting considered here.

Removing the veto produces the biggest swings, which confirms it as the main decision-regulation mechanism. On CICIoT2023, actionability A ( t ) climbs from 0.933 to 0.982, a 5.2% rise. That jump comes down to the moderate conflict level on this dataset ( C ( t ) = 0.554 ): by imposing a regulatory filter, the veto reins in actionability to keep things safe. On TON_IoT, where the conflict level runs higher, the veto’s direct effect on the actionability score stays in the same ballpark, with A ( t ) moving from 0.903 to 0.957 (+6.0%). This tells us the veto works not just through the size of the actionability penalty but also through the decision policy, where the conflict threshold flatly overrides automated mitigation.

Removing the adaptive feedback β reveals a complementary picture. On CICIoT2023, it only barely touches actionability (0.933 → 0.944) but sharply cuts the system’s ability to detect conflict (0.554 → 0.416), a 24.9% drop. On TON_IoT, by contrast, removing β brings no notable change in either A ( t ) or C ( t ) , which stay at 0.903 and 0.626. What these observations tell us is that the feedback mechanism isn’t meant to improve the decision directly, but to tune the system’s sensitivity to explanatory divergence. It therefore plays a fine-tuning role in conflict measurement, especially handy in the trickiest scenarios.

Finally, removing stability S ( t ) and model confidence Co ( t ) moves things only slightly across all metrics. Their contribution thus looks confined to a fine-calibration role rather than serving as a central regulation lever. All in all, these results back up the architecture of the framework, in which the veto and the adaptive feedback handle the critical work, while stability and confidence chip in on the fine-grained adjustment of decisions.

Table 6 synthesizes the final ARX-IoT metrics for both datasets, allowing a direct comparison of the decision regimes.

Table 6. Synthesis of key ARX-IoT metrics.

Contrast of decision regimes π ( t ) . The conflict levels stay fairly close ( C ( t ) = 0.554 versus 0.626), yet that small gap is what decides the regime: only TON_IoT’s C ( t ) tops the veto threshold C max = 0.60 . Actionability runs high on both datasets ( A ( t ) = 0.933 on CICIoT2023, 0.903 on TON_IoT), which confirms that, in this framework, it’s the conflict coefficient C ( t ) rather than the actionability score A ( t ) that mainly settles the choice between the Automatic Mitigation and Veto regimes. On CICIoT2023, with the corrected A ( t ) = 0.9334 , the system stays in the Automatic Mitigation zone, since A ( t ) ≥ τ 1 = 0.80 and C ( t ) = 0.5539 < C max = 0.60 . On TON_IoT, the conflict crossing C max forces a veto policy with human review, no matter how comparably high the actionability score is. This contrast holds together: the framework brings in human oversight whenever explanatory ambiguity crosses the critical threshold, however confident and exploitable the underlying model may otherwise look.

Regulation and trade-off Γ ( t ) . With a single fixed value of λ = 0.5 across both datasets (Equation (7)), the regulation coefficient Γ ( t ) comes out similar in size on TON_IoT (0.731) and CICIoT2023 (0.758), tracking the similarly moderate conflict levels C ( t ) on both datasets (0.626 and 0.554, respectively). This consistency confirms that a single, dataset-independent regulation strength is enough to produce an appropriately calibrated penalty: the veto mechanism doesn’t need re-tuning per dataset to stay effective, and the different decision outcomes (Automatic Migration on CICIoT2023, Veto on TON_IoT) come down to the conflict coefficient C ( t ) relative to the threshold C max , not to any difference in regulation strength.

Adaptive feedback β next . How β next evolves also differs across datasets. On TON_IoT, it stays practically flat (1.0001), a sign that the initial calibration already fits well. On CICIoT2023, it climbs to 1.4999, reflecting a deliberate tweak of conflict sensitivity to better anticipate future shifts in this dataset.

Causal impact Δ P . The causal variations stay small (0.0223 and 0.0187), confirming that the regulation ARX-IoT performs doesn’t meaningfully alter the distribution of predictions. The framework instead acts as a fine-grained safety filter, nudging the decision at the margin without distorting the behavior of the underlying model.

1) XAI divergence as an uncertainty signal (O1). The experiments confirm the existence of a structural explanatory conflict between additive and symbolic approaches. Where TimeSHAP concentrates its probability mass on a single temporal variable (IAT, ≈0.19 for CICIoT2023), Anchors distributes its binary mask across morphological traffic criteria (syn_count, urg_count), ignoring TimeSHAP’s priorities. Measured by a total variation distance D ( t ) = 0.5382 on CICIoT2023, this divergence demonstrates that a single, isolated XAI method offers a biased view. Unlike the state of the art, which treats such divergence as a failure, ARX-IoT exploits this disagreement as an uncertainty indicator to modulate the system’s conservatism.

2) Validation of the optimization and impact of the veto (O2, O4). The sensitivity analysis validates the multi-criteria configuration of the augmented actionability score A ( t ) retained for the framework ( w 1 = 0.1 , w 2 = 0.2 , w 3 = 0.7 ), giving priority to model confidence Co ( t ) ( w 3 ), while the independent grid search identifies temporal stability S ( t ) as the dominant driver of the statistically optimal configuration ( w 1 = 0.8 , Table 2). The ablation study demonstrates the safeguard role of the veto mechanism Γ ( t ) : its removal raises actionability from 0.933 to 0.982 (+5.2%) on CICIoT2023. The veto therefore effectively curbs automation under strong XAI contradiction, protecting infrastructures against erroneous countermeasures.

3) Dynamic regulation and causal robustness (O3). The feedback loop adapts properly to fluctuations: the regulation parameter β ( t ) rises on TON_IoT to correct the drift from the target ( θ = 0.02 ) and falls on CICIoT2023 once the effect steadies. With the measured causal impact Δ P ( t ) staying marginal (<0.023), ARX-IoT shows its topological neutrality: it leaves the detection performance of the underlying LSTM model untouched and works as a transparent peripheral filter.

4) Synthesis of decision regimes. The comparative analysis brings out two distinct decision regimes depending on the traffic environment. On CICIoT2023, the framework meets moderate conflict C ( t ) and regulation Γ ( t ) (0.55 and 0.76). With an actionability score A ( t ) of 0.9334, the system takes an automatic mitigation regime, since A ( t ) ≥ τ 1 = 0.80 and C ( t ) < C max ; this business logic steers clear of the kind of over-cautiousness that would block the network for no good reason. On TON_IoT, conflict is comparable (0.62) but regulation is lower (0.73). Even with strong actionability at 0.9029, the sharp XAI contradiction triggers a veto regime that calls for human arbitration, putting maximum safety first in the face of explanatory ambiguity.

5) Limitations and perspectives. For all the results we obtained, this study has its limits. First, the evaluation of ARX-IoT leans on just two explainability methods, TimeSHAP and Anchors. While these two do represent complementary paradigms (additive and rule-based), trying it with other XAI methods, such as Integrated Gradients, DeepSHAP, or Layer-wise Relevance Propagation (LRP), would help gauge how general the framework really is. Second, mixing several explanatory methods carries extra computational cost that could hamper deployment in tightly constrained IoT environments; future work will look at trimming this overhead to keep real-time performance intact. Finally, the experiments ran on two benchmark datasets; validating on real IoT infrastructures and operational scenarios would help confirm how robust and generalizable ARX-IoT is. Beyond these technical extensions, wider adoption of ARX-IoT in operational settings would also gain from lining up its decision-regulation outputs with established practices for XAI-supported security decision-making, as Sahito et al. [27] discussed. Fourth, the experimental evaluation pits ARX-IoT against raw detection baselines (Random Forest, logistic regression) and against internal ablations of its own components, but stops short of a direct empirical comparison with the conflict-resolution frameworks covered in Section 2 (EXAGREE [5], Consensus Importance [6], XAI4C [8]).

These methods pursue related but distinct objectives. EXAGREE focuses on a priori model selection, Consensus Importance performs post-hoc explanation aggregation, and XAI4C addresses operational conflict detection in O-RAN networks. None of these methods produces the continuous, closed-loop decision-regulation signal targeted by ARX-IoT. Consequently, a like-for-like empirical benchmark would require re-implementing each method within the same IoT decision-regulation setting.

Fifth, the mitigation-risk reduction reported in Section 4.6 relies on an operational definition of “erroneous mitigation decision” grounded in observable test-set outcomes (action condition, ground-truth label mapping, and conflict-conditioned denominator, as specified in the revised Section 4.6). This quantity should be interpreted as an estimated proxy derived from the heuristic interventional model of Section 3.8 (Equations (12) - (13)), rather than as a formally identified causal reduction in the strict do-calculus sense.

The contribution of ARX-IoT relative to these methods is therefore primarily conceptual and methodological. It is supported by the component ablation study presented in Section 4.8 rather than by a direct empirical benchmark. Establishing such a benchmark, for example, against a Consensus-Importance-style weighted aggregation or a static ensemble threshold using the same datasets and evaluation metrics is left for future work.

More fundamentally, this study shows that conflict between explainability methods should no longer be regarded as a weakness of XAI techniques, but rather as an exploitable source of information for decision-making. This perspective constitutes a central scientific contribution of ARX-IoT. Whereas existing approaches primarily seek to produce a single explanation or to maximize the fidelity of an isolated method, ARX-IoT transforms disagreement between explanations into an adaptive regulation mechanism that automatically adjusts confidence, the level of automation, and the mitigation strategy.