通过一次评测就可以上线Passing One Evaluation Is Enough to Ship

上线前把评测跑通一次，之后就可以不用管了。Run the evaluation once before launch and you never have to look again.

那一次通过只覆盖当时的版本组合：模型版本、提示、检索到的材料、用户问法与数据分布，任何一项变动都可能让结论失效。发布后要留抽检、监控与回滚路径，改动时复跑同一套样本。（虚构示例）输入是一次「只更换检索源、模型与提示都没变」的改动 → 处理是复跑同一套样本 → 结果判断看的是「引用是否命中现行文档」这一项口径 → 若该项通过率下降，下一步是重新做上线判断并给检索源加一条变更检查，而不是沿用旧一次的「已通过」。That single pass covers only the version combination at the time: model version, prompt, retrieved material, user phrasing and data distribution - a change in any of them can void the result. After release keep sampling, monitoring and a rollback path, and rerun the same sample set on changes. (Fictional example) Input is a change that swaps only the retrieval source, with model and prompt untouched -> processing reruns the previous sample set -> the check looks at the single criterion "does the citation hit the current document" -> if that pass rate drops, the next step is to redo the go-live decision and add a retrieval-source change check, not to keep relying on last round's "passed".