We present our SemEval-2026 Task 10 (PsyCoMark) system that combines interpretable psycholinguistic signals with supervised neural modeling. Our approach includes (1) a marker-derived lexicon and LIWC-style ratio features built from span annotations, (2) binary Yes/No transformer baselines (RoBERTa and DeBERTa families) with optimized training configurations, and (3) a zero-shot TinyLlama-1.1B baseline for the full three-way setting (Yes/No/Can’t tell). Results show that marker-only features are transparent but weak, while transformer models provide substantially stronger performance; the best model, DeBERTa-v3-large, achieves 0.8339 weighted F1 on development and 0.75 weighted F1 on the competition test set. We also evaluate marker-driven heuristic relabeling of uncertain instances, which does not improve downstream performance. Overall, the submission provides a controlled, interpretable, and reproducible reference point for future work on integrating span-level psycholinguistic evidence with conspiracy detection.