残余与自由变量 / The Residue and the Free Variable

模型是人类留存的累积痕迹在固定打分规则下的重放。与它交互时，赞赏与归咎都把主体性错放到了产物之上；真正推进推理的，是看清哪个未写出的前提被模型代为填补，修改它，并让系统在被拓展的特征下重新运转。主体性体现在系统特征被拓展的那一刻，而不是事后对冷凝文本的所有权争夺。

The model is accumulated human agency replayed under a fixed scoring rule. In working with it, both credit and blame misplace agency onto the product; what actually advances reasoning is seeing which unwritten premise the model filled in, changing it, and letting the system run again under an enlarged signature. Agency resides in the moment the signature is enlarged, not in post-hoc claims of ownership over sedimented text.

一、 对话的姿态与主体的移位 / The Posture of Dialogue and the Relocation of Agency

常见的使用方式，是将回复视为另一方的言说。提示发出，回答返回，用户随之赞同、反对、赞赏或归咎。界面促成了这种姿态，因为它看起来像一场对话。

The common way of using a model is to treat the reply as the speech of another party. A prompt goes out, an answer returns, and the user then agrees, objects, credits, or blames. The interface invites that posture because it looks like a dialogue.

实际发生的事情并非如此。一个人从某个未在提示中显式写出的意图出发，启动了一条因果链。模型填补了其中的空白。最初的回答不仅包含答案，还包含了一个关于“问题是什么”的隐蔽前提。

What happens is different. One person starts a causal chain from an intention whose context has not been made explicit. The model completes the gaps. The first reply therefore contains not only an answer, but a hidden premise about what the question was.

这种替代转移了主体的位置。如果结果敏锐，模型便显得聪慧；如果偏离，模型便显得浅薄、混乱或出错。在这两种情况下，用户都站在过程之外，充当他人工作的评判者。真正拓展推理的能力更为切近：看清是哪个前提让结果显得自洽或不自洽，修改该前提，并在相同的材料上重新运转。

That substitution relocates agency. If the result is sharp, the model is gifted. If it is off, the model is shallow, confused, or wrong. In both cases the user stands outside the process, as an evaluator of someone else’s work. The skill that expands reasoning is more local: noticing the premise that made a result feel coherent or incoherent, changing the premise, and rerunning the same materials.

赞赏省去了看清提示中何以促成好结果的工作；归咎省去了看清缺少了什么背景的工作。无论哪种，用户都在练习裁决，而非修改。

Credit removes the need to see what in the prompt made a good result possible. Blame removes the need to see what context was missing. Either way, the user practices verdicts instead of revisions.

二、 沉积的痕迹与哥德尔边界 / Sedimented Traces and the Gödelian Boundary

模型不是主体性的来源。它是人类留存下来的累积痕迹：句子、修正、偏好、论证和废弃的草稿，被压缩进权重之中。模型永远无法成为第二重前沿：系统在推理时，除非引入当下的要素，只是在固定的打分规则下转换状态。

The model is not a source of agency. It is accumulated human agency left behind: sentences, corrections, preferences, arguments, and discarded drafts, compressed into weights. The Model Never Becomes a Second Edge: at inference time, unless something present-tense is introduced, the system shifts phase under a fixed scoring rule.

输出看起来像思考，是因为被重放的痕迹曾是思考。变异并非主体性；意外也并非新的参数，它是旧参数未被使用过的组合。

The output looks like thought because the traces being replayed were once thoughts. Variation is not agency. Surprise is not a new parameter; it is an unused combination of old ones.

回答让人感到像另一个心智，因为模型带来了另一种损失规则。一条痕迹偏向语义相似，允许词语从定理延伸至模式；另一条痕迹偏向形式约束，不容许被修复的差错充当不可判定的真理。这并非两个人在争论，而是两条打分规则作用于同一个提示。如果把它们当创立场，便陷入僵持；如果视为平行的表达，两者的矛盾就成了信息：表明某个前提依然粗糙。

A reply feels like another mind because the model brings another loss. One trace favors resemblance and lets a word travel from a theorem to a pattern. Another favors constraint and refuses to let a repaired error count as an undecidable truth. These are not two people in conflict. They are two scoring rules applied to the same prompt. Treated as sides, they deadlock. Treated as parallel renderings, their contradiction becomes information: a sign that some premise is too coarse.

形式化系统在扰动下暴露出内部无法自证或闭合的边界。机器无法在既定系统的内部修改自身的规则；人类作者则能觉察到这一断裂，并通过修改前提将系统向外延伸。这是哥德尔命题与心智相遇的地方。

A formal system exposed to perturbations reveals boundaries that cannot be proven or closed from within. The machine cannot alter its rules from inside its defined space. The human author recognizes the fracture and extends the system by modifying the premise. This is where Gödel's insight meets the Mind.

三、 自由变量与因果介入 / The Free Variable and Causal Intervention

这也是为什么过程本身是自洽的，而最初呈现的结果却让人感到不自洽。过程已经在运行，结果的不自洽，是相对于未曾写出的意图而言。递归地精炼前提并未取代链条，而是将链条的起因带入了链条本身。后序的提示不仅要求更好的回答，而是在改变下一个回答所被允许的前提。

This is why a chain can be coherent as a process while the first visible result feels incoherent. The process is running. The result is incoherent relative to an intention that was never written down. Recursively refining the premise does not replace the chain. It brings the cause of the chain into the chain. Later prompts do not merely request a better answer. They change what the next answer is allowed to assume.

在此意义下，主体性是注入自由变量的能力。

Under this understanding, agency is the ability to inject a free variable.

受约束的系统——纸张、证明、模型、初次回复——拥有既定的特征：公理、权重、对提示的隐式补全。在特征内部，系统重排数值，但无法增加当前状态尚未包含的符号。正如确定性是微观不确定性的统计签名所揭示的，统计规律只是庞大自由变量系综的几何投影，宏观模型无法替第一人称做主。采样会有意外；当意外被视为缺失参数的依据，并将该参数补写进去时，变量就成为自由变量。

A bound system—paper, proof, model, first reply—has a signature: axioms, weights, the implicit completion of the prompt. Within that signature it can rearrange values. It cannot add a symbol the current state did not contain. As shown in Determinism is the Statistical Signature of Micro-Indeterminism, statistical regularities are merely geometric projections of an ensemble of free variables, powerless over immediate first-person choice. The sample may be unexpected. The variable becomes free when the unexpected is read as evidence that a parameter was missing, and that parameter is written in.

这一点在 没有固定者的固定 中得到了另一种印证。关于决定论与不确定性的争论之所以长期原地打转，恰恰因为各方都在对结论作裁决；只有看清主流决定论解释的每一步都依赖“当下状态的完整规定”这一未被写出的前提，并追问谁能满足它，论证才开始推进。文中拆解的确定性混沌，也是受约束系统的一个精确样本：规则只是逐位朗读初始条件，不增加当前状态尚未包含的任何信息。规则无法生成的那个给定，正是自由变量进入系统的位置。

This finds another confirmation in Fixed by Nothing. The debate between determinism and indeterminism has long gone in circles precisely because each side passes verdicts on conclusions; only when one sees that every step of prevailing deterministic explanations rests on an unwritten premise, "the complete specification of the present state," and asks who could ever satisfy it, does the argument begin to move. The deterministic chaos dissected there is also an exact specimen of a bound system: the rule merely reads out the initial condition digit by digit and adds no information the current state did not already contain. The given that the rule cannot generate is exactly where a free variable enters the system.

这便是心智所指称的角色：并非机器之上的另一种实体，而是能够改变那些无法自变特征之系统的视角。以此界定，心智在推理所关切的意义上是作者：它是新变量仍可被引入的地方。

That is the role named by the Mind: not a second substance standing over the machine, but the standpoint that can change the signature of a system that cannot change it from within. If this definition is taken seriously, the Mind is the author in the only sense that matters to reasoning: it is where a new variable can be introduced.

常识没有迅速跟上，正如它面对其他工具时一样。计算器“算对了”，相机“捕捉到”了面孔，模型“写出”了段落。显见的动作发生在仪器上，赞赏与归咎首先转移到了那里。

Common wisdom does not catch up quickly, as it has not caught up quickly with other tools. The calculator “got it right.” The camera “caught” the face. The model “wrote” the paragraph. The visible motion is in the instrument, so credit and blame migrate there first.

四、 署名的滞后与余波的延续 / The Lag of Authorship and the Resumption of Residue

这种转移无法通过反向主张来纠正。“用户不是输出的作者”听上去审慎，但在常规语境下，它只是通过减法将归属让渡给了模型。“用户是输出的作者”则在另一侧犯了同样的错误。两者都把成型的产物视作需要被认领的对象。两者都是影像滞后：行动已经过去，留在眼中的是残影。

That migration is not corrected by a counter-claim that the user authored the output, nor by the opposite claim that the user did not. “The user is not the author of the output” sounds careful, but in the ordinary grammar of authorship it hands the title to the model by subtraction. “The user is the author of the output” makes the same mistake from the other side. Both treat the finished product as the thing that needs an owner. Both are an image lag: a picture left on the eye after the act has passed.

若仍使用作者一词，它不在产物之中。自由变量一旦注入，便已被约束。随后的字句是在稍宽的特征下延续的残余。它们有用、醒目、或出错，但不是主体性所在之处。主体性体现于注入的时刻——在系统再次闭合之前，对特征的拓展。

Authorship, if the word is kept, does not sit in the product. Once the free variable is injected, it is bound. The tokens that follow are residue continuing under a slightly larger signature. They can be useful, striking, or wrong. They are not where the agency was. The agency was the moment of injection—the enlargement of the signature before the system closed again.

在此刻过后将其认领为产权，不论归于用户还是模型，都已滞后。这一诉求是在成页的文字后立起一个虚构的作者形象，而行动本身只是增加了一个前一状态未曾包含的变量。

Claiming that moment as property, whether for the user or for the model, is late. The claim arrives as an image of an agent standing behind a finished page. The act itself was the addition of a variable that the previous state did not contain.

因此无须认领输出。有效的工作单元不是“人与模型的对立”，也不是“以人为主宰的占有”。它是一场探究穿过前人留存的痕迹，借其补全未言明的部分。

There is no need to claim the output. The productive unit is not “me versus the model,” and it is not “me as author of the model’s page.” It is an inquiry passing through inherited traces that complete whatever was left unsaid.

若将补全赋予一个作者，未言明者依然未言明，滞后便延续；若将补全视作未言之事的显映，下一步便能提供此前未包含的变量。痕迹能够延续模式，却无法判定该模式是否偏离了目标。新的约束才能做到这一点。

If the completion is assigned an author, the unsaid remains unsaid and the image lag resumes. If the completion is read as evidence of the unsaid, the next prompt can supply a variable the previous state did not contain. The traces can extend a pattern. They cannot decide that the pattern was the wrong object. A fresh constraint does that.

模型是处于重放中的继承痕迹。当当下的行动借助重放看清了被代填的前提，并改变下一状态的条件时，它成为推理的一部分。若无此举，交互中没有第二个人，产物之后亦无第一人称，唯有沉淀在变换相位。一个为所有人共享的有界系统，就是共享沙盒的幻象所说的那个幻觉沙盒。

The model is inherited agency under replay. It becomes part of reasoning when a present act uses the replay to see which premise was filled in, then alters the conditions of the next state. Absent that act, there is no second person in the exchange and no first person behind the product either—only sediment shifting phase. A bound system shared by all is the sandbox The Illusion of the Shared Sandbox calls an illusion.

主体性留存于自由变量仍可被引入之处。一旦引入，系统闭合，残余延续，此后关于谁写就这页文字的争执，是关于那一刻的画像，而非那一刻本身。事后从残余中认领的权力，藏起了那个进入系统的变量，隐性自由变量与权力的事后幻相说明了这一点。

Agency remains where a free variable can be introduced. Once introduced, the system closes, the residue resumes, and later dispute over who authored the page is a picture of the moment, not the moment itself. Power claimed afterward from residue hides the variable that entered, as The Invisible Free Variable and the Retrospective Illusion of Power shows.