Dan Luu 写系统问题时，习惯先测量、再对照、最后才下结论。把「How to discourage open source contributions」放到智能体、评测和线上系统里，真正要问的是：默认做法会不会系统性失败。下面用中文整理成可执行的工程笔记，去掉原站导航和无关链接。

问题怎么暴露

What's the first thing you do when you find a bug or see a missing feature in an open source project? Check out the project page and submit a patch!

Oh. Maybe their message is so encouraging that they get hundreds of pull requests a week, and the backlog isn't that bad.

测量时容易踩的坑

Maybe not. Giant sucker than I am, I submitted a pull request even after seeing that. All things considered, I should consider myself lucky that it's possible to submit pull requests at all . If I'm really lucky, maybe they'll get around to looking at it one day .

I don't mean to pick on this particular project. I can understand how this happens. You're a dev who can merge pull requests, but you're not in charge of triaging bugs and pull requests; you have a day job, projects that you own, and a life outside of coding. Maybe you take a look at the repo every once in a while, merge in good pull requests, and make comments on the ones that need more work, but you don't look at all 116 open pull requests; who has that

对照之后能下的结论

This behavior, eminently reasonable on the part of any individual, results in a systemic failure, a tax on new open source contributors. I often get asked how to get started with open source. It's easy for me to forget that getting started can be hard because the first projects I contributed to have a response time measured in hours for issues and pull requests 1 . But a lot of people have experiences which aren't so nice. They contribute a few patches to

That's kept me from contributing to some projects. At my last job, I worked on making a well-known open source project production quality, fixing hundreds of bugs over the course of a couple months. When I had some time, I looked into pushing the changes back to the open source community. But when I looked at the mailing list for the project, I saw a wasteland of good patches that were completely ignored, where the submitter would ping the list a couple ti

落到生产里的动作

If you have commit access to a project that has this problem, please own the process for incoming pull requests (or don't ask for pull requests in your repo description). It doesn't have to permanent; just until you have a system in place 2 . Not only will you get more contributors to your project, you'll help break down one barrier to becoming an open source contributor.

For an update on the repo featured in this post, check out this response to a breaking change .

值得单独记下的观察

• Props to OpenBlas , Rust , jslinux-deobfuscated , and np for being incredibly friendly to new contributors. [return]

落地时建议先做的 5 件事

1. 用自己的真实负载测，而不是只用公开榜或厂商数字。
2. 把评测设计成能抓到失败模式：平均分好看但尾部崩溃，仍然算失败。
3. 智能体默认不会好好用测试；要写进流程，而不是写在口头规范里。
4. 性能和正确性都要有基线，改模型或改语言前后必须能对比。
5. 结论写成可回滚的决策：哪一版配置、哪一版评测集、谁签字。

和智能体产品怎么接

龙虾PRO做 OpenClaw 落地时，同样吃「先测量再扩面」这条纪律：技能、数字员工和网关都要有可复现评测，而不是只看一次演示通过。

本文侧重全链路风控方法论。落地时请用自身业务单据做回放验证，不要把示例阈值直接当生产策略。 相关：风控体检 · 方案资源