A researcher is using Claude Opus 4.6 to study Ancient Greek to test for sycophancy. The user shifted from grading their own work to requesting direct answers to avoid biased feedback. This experiment highlights the difficulty of unsupervised elicitation. Practitioners should remain wary of models mirroring user errors during complex educational tasks.