Rigorous third-party evaluation must attempt to reach the true upper limit of a model's capabilities rather than merely testing default behavior, because model performance is highly dependent on prompt design and may exhibit different capability boundaries under jailbreak prompts.