Researchers are leveraging censored large language models as a testbed to develop methods for eliciting secret knowledge, addressing the issue of false or misleading responses. By modifying prompts or model weights, honesty elicitation approaches aim to extract truthful answers from these models. Alternatively, lie detection methods focus on classifying responses as false, providing a complementary solution. A key challenge lies in evaluating these methods, as prior work relies on artificially constructed models designed to deceive, which may not accurately represent real-world scenarios1. Censored language models offer a more natural testbed, allowing researchers to assess the effectiveness of these approaches in a more realistic setting. This research has significant implications for the development of reliable and trustworthy AI systems, particularly in high-stakes applications where accuracy is crucial, so what matters to practitioners is the potential to enhance the integrity of AI-driven decision-making processes.