Multimodal Diagnostic Overlays for Tabular Deep Learning: Fractal, Spectral, and Reflex-Aware Interpretability

Deep learning models for tabular data often behave unpredictably during training, and these internal dynamics directly influence the reliability of post hoc interpretability methods. Static techniques such as SHapley Additive exPlanations (SHAP) and Integrated Gradients offer useful but incomplete snapshots of model behaviour, while tabular models frequently undergo noise-driven fluctuations, structural reorganisations, and collapse phases that these methods cannot detect. This study introduces a temporally aligned, multimodal diagnostic framework that observes and interprets model behaviour as it evolves. The framework integrates fractal complexity, spectral energy, persistent homology, gradient-noise statistics, reflex responsiveness, attribution instability, and exploratory sonification into a single phase-aligned system. The framework is evaluated on Multilayer Perceptron (MLP) and Gated Recurrent Unit (GRU) models across benchmark and real-world datasets using a reproducible protocol with consistent configurations and autocorrelation-robust analysis. An ablation study confirms the distinct diagnostic value of each modality, with reflex-aware metrics validated against optimiser behaviour and sonification assessed through event detection. The multimodal diagnostics reveal instability regimes, fragmentation events, attribution drift, and optimiser–noise coupling failures that static Explainable Artificial Intelligence (XAI) methods cannot detect. GRU models show repeated structural transitions, whereas MLPs exhibit simpler but noisier optimisation dynamics. These behaviour-aware, temporally aligned diagnostics provide a more reliable foundation for interpreting tabular deep learning models.