Epistemic Limits of Local Interpretability in Self-Modulating Cognitive Architectures

Abdelaali MAHROUK^*

• Université Frères Mentouri Constantine 1, Constantine, Algeria

Local interpretability techniques—such as LIME (Ribeiro et al., 2016) and SHAP (Lundberg & Lee, 2017) —have become standard tools for probing the decision-making logic of complex machine learning models. These methods rely on the assumption of local continuity in the latent space: that small perturbations around an input yield semantically consistent and explainable model responses. However, this assumption often breaks down in recursive, self-modulating cognitive architectures, where internal states are dynamically restructured through feedback loops, cross-layer attention, and latent program rewriting. Under such conditions, local explanations may become unstable or misleading. In this paper, we present evidence—through formal analysis, simulation experiments, and epistemological reflection—that local proxy models are insufficient to capture the internal narrative dynamics of self-reflective systems. We propose a shift from post-hoc local approximations to causal-hierarchical traceability, integrating internal self-monitoring signals with meta-generative narratives. Our framework builds on recent developments in modular neuro-symbolic agents, structured world models (Ha & Schmidhuber, 2018 ; Guez et al., 2021), reflective prompting (Kojima et al., 2023), and interpretability research in large-scale language models (Ji et al., 2023 ; Chan et al., 2022 ; Huang et al., 2024). It argues for a holistic interpretability paradigm more suited to future architectures. Rather than claiming a final solution, we advance a new epistemological framing for Artificial General Intelligence (AGI : see Glossary in Section 2.6 for definition) : one that treats intelligent systems as narratively structured, self-explaining epistemic agents. This perspective is exploratory, but it highlights the need for interpretability frameworks that evolve with system complexity and narrative modulation.