This paper presents a dual-track system for conspiracy theory detection and psycholinguistic marker extraction. We evaluate multiple architectures, including DistilBERT, BERT-Base, DeBERTa-V3, RoBERTa, and instruction-tuned Qwen2.5 models. Qwen2.5-14B (full-shot) achieves the best performance with a Weighted F1-score of 0.80 in the detection task. Marker extraction remains challenging: while the fine-tuned LLM performs best on "Actors," its limited generalization in categories such as "Evidence" and "Effect" highlights persistent semantic ambiguity.