评测分数高就等于可靠A High Evaluation Score Means It Is Reliable

评测分数高，系统用起来就可靠。A high evaluation score means the system is reliable in use.

可靠性关心的是波动与分布：同一输入跑多次是否一致、没见过的输入表现如何、边界场景的失败长什么样。分数是某个版本在某个任务集、某个设置下的一次读数，两个口径不同。要判断可靠，就用自己的样本复跑并记下波动。（虚构示例）输入是同一条退款期限咨询 → 处理是用同一提示、同一模型版本各跑三遍 → 结果判断是三遍里有一遍把期限说错，说明这条链路的波动不为零 → 下一步不是去引用它的高分，而是把该问法加进自己的评测样本，看修复后是否还会波动。Reliability is about spread and distribution: whether the same input gives consistent results over repeats, how unseen inputs behave, what failures look like at the boundaries. A score is one reading of one version on one task set under one setting - a different question. Judge reliability by rerunning your own samples and recording the variance. (Fictional example) Input is one query about a refund deadline -> processing runs it three times with the same prompt and the same model version -> the check finds one of the three stating the deadline wrongly, so variance on this path is not zero -> the next step is to add that phrasing to your own sample set and recheck after a fix, not to cite the high score.