Dan Luu 写系统问题时，习惯先测量、再对照、最后才下结论。把「Why it's impossible to agree on what's allowed」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

On large platforms, it's impossible to have policies on things like moderation, spam, fraud, and sexual content that people agree on. David Turner made a simple game to illustrate how difficult this is even in a trivial case, No Vehicles in the Park . If you haven't played it yet, I recommend playing it now before continuing to read this document.

The idea behind the site is that it's very difficult to get people to agree on what moderation rules should apply to a platform. Even if you take a much simpler example, what vehicles should be allowed in a park given a rule and some instructions for how to interpret the rule, and then ask a small set of questions, people won't be able to agree. On doing the survey myself, one of the first reactions I had was that the questions aren't chosen to be particul

测量时容易踩的坑

I'm fascinated by the fact that my takeaway is the precise opposite of what the author intended.

To me, the answer to all of the questions was crystal-clear. Yes, you can academically wonder whether an orbiting space station is a vehicle and whether it's in the park, but the obvious intent of the sign couldn't be clearer. Cars/trucks/motorcycles aren't allowed, and obviously police and ambulances (and fire trucks) doing their jobs don't have to follow the sign.

对照之后能下的结论

So if this is supposed to be an example of how content moderation rules are unclear to follow, it's achieving precisely the opposite.

After going through the survey, you get a graph showing how many people answered yes and no to each question, which is where the "clear majority" comes from. First of all, 一个稳妥判断是 it's not correct to say that there is a clear majority. But even supposing that there were, there's no reason to think that there being a majority means that most people agree with you even if you take the majority position in each vote. In fact, given how "wiggly" the per-questi

落到生产里的动作

If you look at the data (which isn't available on the site, but Dave was happy to pass it along when I asked), as of when I pulled the data, there was no set of answers which the majority of users agreed on and it was not even close. I pulled this data shortly after I posted on the link to HN, when the vast majority of responses were HN readers, who are more homogeneous than the population at large. Despite these factors making it easier to find agreement,

Another way to look at this data is that 36902 people expressed an opinion on what constitutes a vehicle in the park and they came up with 9432 distinct opinions, for an average of ~3.9 people, per distinct expressed opinion. i.e., the average user agreement is ~0.01%. Although averages are, on average, overused, an average works as a summary for expressing the level of agreement because while we do have a small handful of opinions with much higher than th

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