This paper presents a hybrid classification strategy for sexism detection on social media. The proposed method, called “Thinking Twice”, combines a first-stage classifier with a confidence estimation mechanism based on attraction forces between vector representations of training and test instances. Instances predicted with high confidence are assigned a label directly, while low-confidence (untrustworthy) instances are forwarded to a more powerful second-stage model, which can be a large language model (LLM), enabling the use of sophisticated methods with lower computational cost. The method was evaluated on two Spanish-language datasets: AMI (misogyny detection), and EXIST (sexism detection). Results show that the approach is competitive, outperforming or matching strong baselines and prompting configurations.