Neuro-Symbolic Class-Contrast Evidence Audit for Reliable Cross-Subject Wearable Activity Recognition

Reliable wearable activity recognition requires not only a class label but also an auditable indication of whether that label is supported by historical sensor evidence. We present CC-NSIEA, a label-preserving neural-plus-rule-based class-contrast evidence audit for cross-subject wearable activity recognition. A Temporal Residual Perception Network supplies the sole activity label, posterior probabilities, and a normalized temporal embedding. A read-only Training-Subject Evidence Memory retrieves global, predicted-class, and competing-class records. A rule-based Evidence Consistency Audit combines data validity, dynamic/static motion coherence, retrieval support, and class separation. When first-round evidence is insufficient, Class-Contrast Evidence Refinement performs one deterministic contrast between the predicted class and the strongest posterior competitor; the audit cannot change the neural label. The term neuro-symbolic is used only in this restricted architectural sense: a neural predictor is coupled to explicitly represent deterministic predicates and a finite rule-based controller; the method does not perform symbolic inference, theorem proving, or knowledge-graph reasoning. On five subject-disjoint outer folds of the UCI HAR official training partition, the shared perception model achieved 90.13% accuracy and 90.55% macro-F1 across 7352 out-of-fold windows from 21 subjects. Relative to a matched dynamic deterministic controller, CC-NSIEA increased Error AUPRC from 0.423802 to 0.433057 and reduced AURC from 0.035941 to 0.035913. The 10,000-resample subject-cluster bootstrap interval for the AUPRC difference was [0.001595, 0.019547]. CC-NSIEA provides an evidence-centered complement to confidence-based reliability estimation.