Dan Luu 写系统问题时，习惯先测量、再对照、最后才下结论。把「Culture matters」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。

Appendix: Bay area culture

Three major tools that companies have to influence behavior are incentives, process, and culture. People often mean different things when talking about these, so I'll provide an example of each so we're on the same page (if you think that I should be using a different word for the concept, feel free to mentally substitute that word).

If you read "old school" thought leaders, many of them advocate for a culture-only approach, e.g., Ken Thompson saying, to reduce bug rate, that tools (which, for the purposes of this post, we'll call process) aren't the answer, having people care to and therefore decide to avoid writing bugs is the answer or Bob Martin saying " The solution to the software apocalypse is not more tools. The solution is better programming discipline ."

Appendix: Centaur's hiring process

The emotional reaction those kinds of over-the-top statements evoke combined with the ease of rebutting them has led to a backlash against cultural solutions, leading people to say things like "you should never say that people need more discipline and you should instead look at the incentives of the underlying system", in the same way that the 10x programmer meme and the associated comments have caused a backlash that's led to people to say things like vel

But if we look at quantifiable output, we can see that, even if processes and incentives are the first-line tools a company should reach for, culture also has a large impact. For example, if we look at manufacturing defect rate, some countries persistently have lower defect rates than others on a timescale of decades 1 , generally robust across companies, even when companies are operating factories in multiple countries and importing the same process and i

Appendix: importing culture

Coming back to programming, Jamie's post on "moving faster" notes:

The main thing that helped is actually wanting to be faster.

值得单独记下的观察

• Getting people to show up to meetings on time Incentive: dock pay for people who are late
• Process: don't allow anyone who's late into the meeting
• Culture: people feel strongly about showing up on time
• Getting people to build complex systems Incentive: require complexity in promo criteria
• Process: make process for creating or executing on a work item so heavyweight that people stop doing simple work
• Culture: people enjoy building complex systems and/or building complex systems results in respect from peers and/or prestige
• Avoiding manufacturing defects Incentive: pay people per good item created and/or dock pay for bad items
• Process: have QA check items before shipment and discard bad items

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