Nish Tahir published an experiment restricting Qwen3-1.7B’s output to five choice tokens so it selects an answer in one forward pass. Accuracy on 1,221 CommonsenseQA examples was 59.38%, rising to 62.41% after fine-tuning, but the high-confidence group was overconfident: its average probability was 98.55%, compared with actual accuracy of 70.09%. Temperature scaling reduced the gap, and the public repository confirmed the experimental setup. The results show that preventing answers outside the choices and calibrating the probability of correctness are separate problems; generalization to other data needs additional evaluation.