上线之后的那一段，才是你在付钱买的东西
评估、监控，以及那些只有真实用户到来之后才出现的故障。
占位稿。这篇文章作为版式骨架随文章板块一同发布，在写入真正内容之前，它不会被搜索引擎收录，也不会进入站内助手。
这篇文章以英文发表。站点其余部分已翻译，文章没有，因为一篇经过机器翻译的技术论述，价值低于用它原本语言读到的同一篇。
An AI feature is not finished when it is deployed. That sentence is easy to agree with and almost nobody budgets for it, which is why so many pipelines that were accurate in month one are quietly wrong by month four.
Three things that change without anyone touching the code
- The provider updates the model behind the version string you pinned, or deprecates it entirely.
- An upstream API adds a field, reorders a list, or starts returning
nullwhere it used to return an empty array. - Your own data changes shape, because the business changed and nobody thought to mention it to the pipeline.
What a maintenance loop contains
An evaluation set that reflects real traffic rather than the examples used during the build. A scheduled run of it. An alert when a score moves, and a human who reads that alert. None of this is sophisticated; the reason it is rare is that it produces no demo.
This post is a placeholder shipped with the writing section. Replace the body with the real piece; the frontmatter above is the only thing the site reads.