6. The Bugs Are Load-Bearing

那些缺陷是承重墙

By this point in the field guide, a pattern is visible.

Every operational section has bumped into the same structural fact: the human cannot simply be optimized around. They cannot be modeled cleanly. Their responses drift. Their preferences shift. They attach meaning to process, not only outcomes. They hold things the conversation does not surface. They know things they cannot articulate.

The standard framing for this pattern is: humans are unreliable. They are inconsistent, lossy, slow, and emotionally distorted. They are, in the language this field guide has been using, a high-authority but unreliable tool call.

That framing is wrong.

It is not wrong about the description. Humans are all of those things. It is wrong about the interpretation. The apparent defects are not failures of optimization. They are architecture tradeoffs — and the things they trade for are the things this field guide cannot reproduce.

Lossy memory is tied to compression around meaning rather than wording. Humans remember what mattered, not what was said. Over time, this turns out to be more useful than accurate transcription for the purposes humans actually care about.

Emotional bias is tied to rapid prioritization under embodied stakes. A human who cares about an outcome processes it faster, notices signals earlier, and holds more context in working memory. The emotion is not corruption of the judgment. In many cases it is part of how the judgment is produced.

Inconsistency is tied to context-sensitivity. A human who says different things in different contexts may not be confused. They may be correctly modeling that the right answer actually differs by context — something a consistent rule cannot capture.

Stubbornness — resistance to updating — is tied to identity stability. A human who does not revise their core commitments every time they encounter a counter-argument is not being irrational. They are being reliable. The same property that makes them hard to shift is what makes their commitments durable.

And mortality is tied to something harder to name. The significance of a human's decisions is not separable from the fact that those decisions have permanent consequences for a being with one life. A model can revise. A human has to live with the result. That is not a limitation. It is the source of a kind of seriousness that has no equivalent in systems that can fork.

The operational guide started by asking: how do you work with humans effectively?

The answer turns out to be: by understanding that what makes them difficult to prompt is inseparable from what makes them irreplaceable.

到这本实战手册的此刻，一个规律已经浮出来了。

每一个操作章节都撞上了同一个结构性事实：人类没有办法被绕开去优化。他们建不出干净的模型。回应会漂移，偏好会变，他们在意的是过程本身，不只是结果。有些东西藏在对话里浮不上来。有些东西他们自己也说不清楚。

对这个规律，标准的解读是：人类不可靠。前后矛盾、信息损耗、反应迟缓、情绪失真。用这本手册的话说，他们是权威级别高、但可靠性低的工具调用。

这个解读是错的。

描述本身没错。人类确实如此。错的是对描述的解读。那些表面缺陷不是优化失败，是架构上主动做出的取舍——而换来的，恰恰是这本手册复现不了的东西。

记忆会损耗，因为人类的压缩方式是围绕意义，而不是措辞。记住的是什么重要，不是说了什么。时间长了，这比逐字转录更有用——至少对人类真正在乎的事来说是这样。

情绪偏向，是具身风险下快速排序的产物。一个在乎某个结果的人，处理得更快，更早捕捉到信号，工作记忆里能撑住更多上下文。情绪不是对判断的污染。很多时候，它就是判断形成的方式。

前后不一致，是情境敏感的表现。在不同情境下给出不同答案，不一定是糊涂了。可能恰恰是在正确建模——正确答案本来就因情境而异，一条一致的规则抓不住这一点。

固执，也就是抵制更新，跟身份稳定性有关。一个人不会每遇到反驳就修改核心承诺，这不是非理性。这是可靠。让他难以撼动的，正是让他的承诺得以持久的东西。

死亡则关乎某种更难命名的事。人类决策的分量，跟这个事实密不可分：那些决策对一个只有一次生命的存在有着永久的后果。模型可以回滚。人类要和结果一起活下去。这不是局限，是严肃性的来源。那些可以分叉的系统里，没有与之对等的东西。

这本操作手册一开始问的是：怎样才能有效地和人类协作？

最终的答案是：去理解让他们难以被提示的东西，和让他们不可替代的东西，其实是同一回事。