AI 一夜之间能列出四十个可检验的下一步，但湿实验、长期随访、设备排队、伦理审批和复现，不会跟着一起变快。要问的其实是：这条具体的研究链上，哪一段被加速了，哪一段还卡在物理、制度或可信度的限制里？先把这些限制摊开，才谈得上在四十个候选里挑哪个先测，以及谁为”值得相信”这三个字签字。

AI can list forty checkable next steps overnight, but wet labs, longitudinal follow-up, instrument queues, ethics review, and replication do not speed up on the same clock. The question worth asking is: in this specific research chain, which segment got faster, and which one is still held back by physics, institutions, or credibility? Only once those limits are laid out can you ask which of the forty candidates to test first, and who signs their name to “worth believing.”

"种类之别，非程度之别"——地图隐喻把它说透

“A difference of kind, not degree”: the map metaphor makes it concrete

这一卷反复说"差别在种类，不在程度"，但这句话很容易听成口号。Asimov Press 的地图隐喻把它说到了骨头里：博尔赫斯写过一个寓言，帝国的制图师把地图做得越来越精确，最后做出一张与帝国等大、一比一的地图——细节拉满到极致，可它仍然只是同一种信息，没有变成新理解。把伦敦地铁画得再精确，标上每一段铁轨的真实曲率和地理坐标，它也还是一张越来越精确的地理地图。

直到 1933 年 Harry Beck 做了一件种类不同的事：他抛弃了地理精确性，把整张网络重画成一张电路图——线是直的、角是 45 度的、站距是均匀的，全都"不准"，却第一次让人能一眼看懂怎么换乘。这就是范式——一门学科当下公认的提问方式：一次"重新示意化（re-schematization）"，不是堆更多细节。AI 极擅长把地图做得更精确（填空白、加细节、提精度），但"该换一张什么样的图来示意"这个动作，是种类不同的，它不在"更精确"这条轴上，所以再多的算力也不会自动跨过去。这正是整卷"瓶颈搬家"最底层的道理。

This volume keeps saying “the difference is not degree but kind,” but that line is easily heard as a slogan. Asimov Press’s map metaphor drives it to the bone: Borges wrote a parable in which an empire’s cartographers made the map ever more accurate until they produced one the size of the empire, one-to-one. Detail was maxed to the limit, yet it remained the same kind of information, never becoming new understanding. Draw the London Underground ever more accurately, marking every rail’s true curvature and geographic coordinates, and it is still an ever-more-accurate geographic map.

Until 1933, when Harry Beck did something of a different kind: he threw away geographic accuracy and redrew the whole network as a circuit diagram: straight lines, 45-degree angles, even station spacing, all “inaccurate,” yet for the first time letting anyone see at a glance how to change trains. That is a paradigm, which at bottom is the way a discipline currently agrees to pose its questions: a re-schematization, not more detail. AI excels at making the map more accurate (filling blanks, adding detail, raising precision), but “what kind of map to redraw it as” is a different kind of act. It is not on the “more accurate” axis, so no amount of compute crosses to it automatically. This is the epistemological foundation of the whole volume’s “bottleneck moving.”

三个误读，把"嫁接"伪装成"原生"

Three misreadings that disguise “grafting” as “native”

把这一卷读窄，几乎都从同一个错位开始：把工具的更替当成稀缺的迁移。第一种误读是"更快即原生"：研究员用 AI 把六个月的文献综述压到六天，于是宣布自己 AI-Native 了。可它只把同一条流程的"执行"加速了，瓶颈仍卡在"这六天读完之后，谁来判断该信哪一条、该往哪个方向追"。第二种误读是"更多即更好"：把产出当成绩效，年产论文从 4 篇变 40 篇。但 RES 03 的 Nature 文献计量已经把这条路走到尽头：个人产出与影响力确实涨了，科学整体的主题覆盖却在收缩。第三种误读最隐蔽，"自动即自主"：以为接上一个 AI-Scientist 这样的 end-to-end 系统就等于把研究交了出去。它确实把"选题→调研→实验→结论"这条继承下来的旧流水线整段自动化了，这份验证价值、实验价值是真的——但流水线本身没被重新设计：它唯一能用来给自己想法打分的代理，仍是"与既有范式的距离"，于是它越自动，越把科学推向范式内的安全区。这是过渡态该有的样子，不是终态的证据。三种误读共用一个病灶：只搬了执行，没认出稀缺已经搬家。

Reading this volume too narrowly almost always begins from one dislocation: mistaking a change of tools for a migration of scarcity. The first misreading is “faster = native”: a researcher compresses a six-month literature review into six days with AI and declares themselves AI-Native. But that only sped the “execution” of the same pipeline; the bottleneck still sits at “after those six days of reading, who judges which thread to believe and which direction to chase.” The second is “more = better”: treating output as performance, going from 4 papers a year to 40. RES 03’s Nature bibliometrics already walked this road to its end: individual output and impact do rise, while science’s topical coverage contracts. The third is the most insidious, “automatic = autonomous”: believing that wiring up an end-to-end system like AI-Scientist equals handing research away. It genuinely does automate the whole inherited pipeline, from picking a question through survey, experiment, and conclusion, start to finish, and that verification value, that experimental value, is real. But the pipeline itself was never redesigned: the only proxy it has for scoring its own ideas is still “distance from the established paradigm,” so the more autonomous it gets, the harder it pushes science into the in-paradigm safe zone. This is what a transitional state looks like, not evidence of an end state. All three share one lesion: they moved execution without recognizing that scarcity had already moved.

科学是资源分配问题，不是智能问题

Science is a resource-allocation problem, not an intelligence problem

这卷之所以把研究当上游而非又一台 α 机器，背后有一个常被忽略的命题："科学根本是一个资源分配问题，不是智能问题"（chenhaot 的形式化）。意思是：限制科学进步的，从来不是"算得不够快、读得不够多"这类智能瓶颈，而是"有限的注意力、经费、人才该投向哪些问题"这个分配瓶颈。AI 把"算、读、跑"的成本压到近零，并没有解决分配问题，它只是把分配瓶颈暴露得更彻底了。产出更多不等于知识更多：如果一万篇论文全挤在同一个数据丰富的热门角落，知识的边界一寸没动。这就是为什么本卷反复说"产出量本身永远不是指标"，在一个执行充裕的世界里，唯一还稀缺的资源是注意力，而注意力该投向哪里，正是 RES 06 那个"哪个真相值得知道"的价值判断。把科学看成智能问题，你会去堆更多算力；把它看成分配问题，你才会去守那个决定"投向哪里"的判断节点。

The reason this volume treats research as upstream rather than one more alpha machine rests on an often-overlooked claim: “science is fundamentally a resource-allocation problem, not an intelligence problem” (chenhaot’s formalization). Meaning: what limits scientific progress was never an intelligence bottleneck like “not computing fast enough, not reading enough,” but an allocation bottleneck: “which problems should finite attention, funding, talent go to.” AI drives the cost of “compute, read, run” to near-zero and thereby does not solve allocation; it merely exposes the allocation bottleneck more starkly. Producing more does not equal knowing more: if ten thousand papers all cluster in the same data-rich hot corner, the edge of knowledge has not moved an inch. This is why the volume keeps saying “output volume itself is never the metric” — in a world of abundant execution, the one still-scarce resource is attention, and where attention should go is precisely RES 06’s value judgment of “which truth is worth knowing”. See science as an intelligence problem and you pile on more compute; see it as an allocation problem and you guard the judgment node that decides “where to point it.”

所以这卷的第一刀，是把“研究”这个词从“产出一篇可发表的东西”重新定义为“让一个值得相信、值得知道的真相，落进可追溯的结构里”。前者的稀缺是工时，后者的稀缺是判断。内核四步走完一遍，就是全卷命题：①执行先充裕——检索、跑实验、连批量出假设都近免费，“做出研究”不再稀缺；②判断随之成了瓶颈，而它沿“机器能不能检验”裂成两支，可机检的那半并回充裕，只剩“哪个真相值得知”留给人；③上下文沉成基础设施——一个可查询的证据库接住海量生成、当它的规格，让每条主张可追溯、可证伪；④人退到最后一格，不再当知识的生产者，改当“值得信、值得知”的担保人。下面这张图把这四步画成一个会自我纠偏的环。

So this volume’s first cut is to redefine the word “research” from “producing something publishable” to “landing a believable, worth-knowing truth into a traceable structure.” The former’s scarcity is hours; the latter’s is judgment. Walk the kernel’s four steps once and you have the whole thesis: ① execution becomes abundant first: search, running experiments, even batch-generating hypotheses go near-free, so “doing research” is no longer scarce; ② judgment then becomes the bottleneck, and it splits along “can a machine check it” into two branches: the machine-checkable half folds back into abundance, leaving only “which truth is worth knowing” with people; ③ context sinks into infrastructure: a queryable evidence base catches the mass generation and serves as its spec, keeping every claim traceable and falsifiable; ④ people retreat to the last cell, no longer producers of knowledge but guarantors of what is “worth believing, worth knowing.” The figure below draws those four steps as a loop that self-corrects.

内核第②步的分叉（全卷承重）：判断不是一整块"留给人"。它沿"能不能被机器检验"裂成两支——

The fork in kernel step ② (load-bearing for the whole volume): judgment is not one block “kept for humans.” It splits along “can a machine check it” into two branches:

一个具体的问题对，把这道分叉钉死："这款新药是否优于安慰剂？"——框架内，它能写成机器可检的验收标准（终点、样本量、显著性阈值都事先定死），AI 可以设计、跑、复核。而"我们该不该继续用肿瘤缩小、而不是病人多活了多久，来定义'有效'？"——框架级，它问的不是"哪个数字更高"，而是"该用哪个数字"，价值负载、没有机器能判的对错。前一个问题迟早并入①充裕；后一个问题，模型再强也退守在人这一侧。左支的稀缺只是暂时的能力门槛，右支的稀缺是"无对错可机检"这个结构性事实——这条界，也是后面一切操作判据的分水岭。

把两支混为一谈，会同时犯两个方向相反的错：要么把可机检的那支也当成"人的尊严最后防线"死守（结果没把执行充裕化、白白困在工时里），要么把只能人判的那支也当成"迟早会被自动化"提前交出去（结果把价值判断让给了生成层的默认偏置）。看清这道分叉，就是看清"哪些稀缺会被时间消解、哪些不会"，这是整卷一切操作判据的源头。

One concrete pair of questions nails the fork down: “Is this new drug better than placebo?”, in-frame, it can be written as machine-checkable acceptance criteria (endpoint, sample size, significance threshold all fixed in advance), and AI can design it, run it, re-check it. Whereas “Should we keep defining ‘effective’ by tumor shrinkage rather than by how much longer the patient lives?”, frame-level, it asks not “which number is higher” but “which number we ought to use,” value-laden, with no machine-checkable right answer. The first folds into ① abundance sooner or later; the second stays on the human side no matter how strong the model gets. The left branch’s scarcity is only a temporary capability threshold; the right branch’s is the structural fact of “no machine-checkable right answer,” and that line is the watershed for every operational test that follows.

Conflate the two and you commit two opposite errors at once: either defending the machine-checkable branch as “the last line of human dignity” (failing to make execution abundant, stuck in hours for nothing). Or you hand the human-judgment branch away early as “automatable sooner or later” (ceding value judgment to the generation layer’s default bias). Seeing this fork clearly is, in essence, seeing “which scarcities time dissolves and which it does not,” the source of every operational test in the volume.

于是研究卷的可证伪核心命题成形：执行充裕 → 提问稀缺 → 再退守"哪个真相值得知道"（价值判断）。它为假的条件很清楚：若能证明"哪个真相值得知道"可被无损地形式化、聚合、或交给系统自动判定——命题倒。写得出推翻它的条件，它才是命题。

So the volume’s falsifiable core thesis takes shape: execution becomes abundant → questioning becomes scarce → then a retreat to “which truth is worth knowing” (a value judgment). Its condition for being false is explicit: if one can show that “which truth is worth knowing” can be losslessly formalized, aggregated, or handed to a system to decide automatically, the thesis falls. Being able to write the condition that would refute it is what makes it a claim.

工程判"对不对"，研究判"值不值得"——同一个道理，却更深

Engineering judges “correct,” research judges “worth”: the same move, only deeper

工程面对代码，稀缺判断是"对不对"——判"这是不是真的"，有客观对错、有可机检的验收标准，终将被验证工具大幅自动化。研究面对真相，稀缺判断先是"提什么问题"，最终退守到"哪个答案值得信、哪个真相值得知"：前半截还在"算不算真"（可信度，有证据梯度），后半截跌进"值不值得知道"，无对错，只有"对谁、在哪个价值框架下"的归属。一道判断从"这是不是真的"变成"这值不值得知道"的那一刻，坐标系就换了，新坐标系里没有可机检的对错，这就是"更深"的确切含义。

Engineering faces code; its scarce judgment is “is it correct”, an epistemic judgment with objective right and wrong and machine-checkable acceptance criteria, so it will eventually be largely automated by verification tooling. Research faces truth; its scarce judgment is first “which question to ask,” retreating finally to “which answer deserves belief, which truth is worth knowing”: the first half still in epistemology (credibility, an evidence gradient), the second half fallen into axiology, no right answer, only belonging to “whom, under which value frame.” The moment a judgment turns from “is this true” into “is this worth knowing,” the coordinate system changes and the new one has no machine-checkable right answer; that is the exact meaning of “deeper.”

研究是系列的耦合枢纽，而不是又一台 α 机器。把研究和工程、设计放在一起看，最容易犯的错是把它当成"再来一台把执行做便宜的机器"。它确实也把执行做便宜了，但它在系列里的位置不在产出端，在上游：它生产的不是代码、不是界面，是"哪个真相值得知道"这条最难外包的判断。下游的工程、设计、组织拿到这条判断，才知道该把昂贵的执行投在哪里。没有上游，下游是精密的空转：一支能在六天内交付任何东西的团队，如果没人回答"该交付什么真相"，它的高效只是把错的方向走得更快。这就是研究卷"耦合最深"的含义，它不只是被引用，它定义了整个系列的输入。

Research is the series’ coupling hub, not just one more alpha machine. Placing research beside engineering and design, the easiest mistake is to read it as “one more machine for making execution cheap.” It does make execution cheap, but its place in the series is not at the output end; it is upstream: what it produces is not code or interfaces but the hardest-to-outsource judgment of “which truth is worth knowing.” Only with that judgment in hand do downstream engineering, design, and the organization know where to spend their now-cheap execution. Without the upstream, the downstream is a precise idle: a team that can ship anything in six days, with no one to answer “which truth to ship,” merely walks the wrong direction faster. This is what “most deeply coupled” means: research is not just cited, it defines the series’ input.

"值得"从"算不算真"交到"值不值得知道"的那一刻，是全卷的枢轴

The pivot of the whole volume: the moment “worth” passes from epistemology to axiology

整卷如果只能留一句，是这一句：研究的稀缺判断沿一条路一直退，从"执行"退到"提对问题"，再退到"哪个答案值得信"，最后退到"哪个真相值得知道"——前几步都还在"算不算真"里，可机检、终将被部分自动化；只有最后一步换了坐标系，"值得"没有可机检的对错。但退到这一步只回答了"谁来接判断"，没回答更早的一问：把选题、调研、实验、结论这条继承来的流水线整段推掉重画，研究该长什么样？我们目前的方向感是——把"可复现"而非"可发表"当成唯一的记账单位：一个不为论文数、只为可复现因果知识的存量负责的机构会长成什么样，几乎没人真正试过。最强的反方是：复现出来的知识如果没有论文这类可传播、可评议的载体，就没法被共同体确认，"可复现"会退化成实验室笔记本里的私货，没有传播就没有被验证。能分辨这两种判断的观察很具体：一个绕开论文、只维护可复现证据库的机构，十年后积累的、被后来者独立验证过的知识增量，会不会稳定超过同等投入的论文体制——这是一个真押注，还没有一家机构跑到能看出答案的年头。这一步也是全卷的枢轴，也是研究卷向创新卷（价值发现）交棒的接口；它为什么最不可外包、具体怎么判，留到 第 7 节 展开。

If the whole volume could keep one sentence, it is this: research’s scarce judgment retreats along a path: from “execution” to “asking the right question,” then to “which answer deserves belief,” and finally to “which truth is worth knowing.” The first steps are still in the “right or wrong” coordinate system, machine-checkable, eventually partly automated; only the last step changes the coordinate system, where “worth” has no machine-checkable right answer. But landing on that step only answers who catches the judgment; it doesn’t answer an earlier question: strip the inherited pipeline (pick a question, survey, run the experiment, conclude) down to nothing and redesign it, and what does research look like? Our current lean is to make reproducibility, not publishability, the one unit of account: almost nobody has actually tried building an institution funded and judged purely on the stock of independently verified causal knowledge it produces, not on paper count. The strongest counter: reproduced knowledge with no paper-shaped, peer-debated vehicle to travel in can’t get confirmed by a community at all; “reproducible” degenerates into a private note in someone’s lab book, unverified because it never traveled. What would tell the two apart is concrete: whether an organization that skips papers and only maintains a reproducibility ledger accumulates, after ten years, a stock of independently verified knowledge that outpaces the paper-based system at equal spend; that is a real bet, and no institution has run long enough yet to show the answer. That step is also the volume’s pivot, and the interface where research hands off to Innovation (value discovery); why it is the least outsourceable, and how to actually judge it, is unfolded in Section 7.

这条耦合是双向的，每条接缝都落在一个具体章节：向上，研究在 第 7 节 把"哪个真相值得知道"交给创新（价值发现）；向下，在 第 8 节 把"谁有权定方向"交给组织（治理）；横向，与工程、设计、架构在 第 1 / 第 4 节 是同一招的对照。下面这张图把这六条接缝画在一起。

This coupling is bidirectional, and each seam lands on a concrete section. Upward, research hands “which truth is worth knowing” to Innovation (value discovery) at Section 7; downward, it hands “who owns the direction” to the Organization (governance) at Section 8; laterally, it mirrors Engineering, Design, and Architecture at Sections 1 and 4. The figure below draws all six seams together.

不切这一刀，核心命题会被自己证伪一半

Without this cut, the core thesis half-falsifies itself

为什么不能直接从"执行充裕"跳到"提问稀缺"？因为把"提问"当成铁板一块说成"永远属于人"，会被一个明摆的事实当场证伪一半：AI-Scientist 这类系统已能在既有框架内提出大量可检验的好问题，给它一张文献图，它列出的"下一步该测什么"常比初级研究者更全。

出路是诚实地把提问切成两层，不是嘴硬：框架内的提问（向数据丰富区找下一个可检验空白）确实在被充裕，划出去、并入①充裕；真正稀缺的是范式级重构（换框架、问旧框架问不出的题），它落在 AI 训练分布之外。切了这一刀，命题反而更稳，它不再押注"提问永远属于人"这个会被证伪的强主张，而是押注更耐打的一句：提问里"换框架"的那一半不会因模型变强而被充裕。好命题的标志，正是敢主动指出自己哪一半会塌，把承重移到不塌的那一半上。

Why not jump straight from “execution abundant” to “questioning scarce”? Because treating “questioning” as one monolithic block declared “forever human” would let an obvious fact half-falsify the thesis on the spot: systems like AI-Scientist can already pose plenty of checkable good questions inside an existing frame; give one a literature graph and its list of “what to test next” is often more comprehensive than a junior researcher’s.

The way out is not stubbornness but to honestly cut questioning into two layers: in-paradigm questioning (finding the next checkable gap toward data-rich regions) is indeed becoming abundant, so draw it out and fold it into ① abundance; the truly scarce thing is paradigm-level reframing (changing the frame, asking what the old frame cannot), which lies outside AI’s training distribution. Make this cut and the thesis is sturdier: it no longer bets on the falsifiable strong claim “questioning is forever human” but on a more durable one: the paradigm-level half of questioning is not made abundant by stronger models. The mark of a good claim is that it dares to name which half will collapse, then moves the load onto the half that will not.

〔探索清单〕此处的断裂点——"框架内被充裕、换框架仍稀缺"——目前是命题推演＋一篇讲"怎么算知道"的综述侧证（《The epistemic revolution of AI》论证 AI 正同时扰动经验论/证伪/库恩范式，但未给逐条实证），尚无单篇一手实证锚坐实"换框架的重构不可被 AI 充裕"；按两份清单的纪律标为"待坐实"，给出先行指标：AI 主导的研究里"换框架"型贡献占比是否长期低位。RES 03 的 Nature 文献计量给出了"AI 向数据丰富区聚集、收缩主题覆盖"的强侧证。

[exploratory ledger] This break — “in-paradigm made abundant, paradigm-level still scarce” — is for now thesis-derivation plus epistemology-review side-evidence (The epistemic revolution of AI argues AI simultaneously perturbs empiricism / falsification / Kuhnian paradigms, but offers no item-by-item empirics). It has no single first-hand empirical anchor nailing down “paradigm-level reframing cannot be made abundant by AI”; per the two-ledger discipline it is marked “to be grounded,” with a leading indicator: whether the share of “reframe” contributions in AI-led research stays durably low. The Nature bibliometrics in RES 03 give strong side-evidence that “AI clusters toward data-rich regions and contracts topical coverage.”

提问被充裕的机理：好问题＝知识边界上的最近邻搜索。为什么"提问"这个看起来最人性、最不可机械化的动作，会有一半被充裕化？机理藏在"好问题"的一个常见定义里："好问题＝在知识边界上识别最有价值的空白"。一旦把这个定义里的"价值"窄化成"可被现有数据检验、离已知最近的下一步"，它就变成了一个知识图谱上的最近邻搜索问题，而这恰恰是大规模图谱分析的强项。给 AI 一张足够全的文献图，让它找"哪些相邻领域之间还没有人架过桥""哪个被反复提及却从未被直接测量的变量""哪条假设链缺最后一环"，它能比大多数人更快、更全地列出这类框架内的好问题。这就是断裂点的左半截：框架内的提问，本质是一种检索，会被充裕。

右半截完全是另一回事："换一套框架、问一个旧框架里根本无法成立的问题"，没有图可搜——那张图本身就是要被换掉的东西。研究者最后的稀缺贡献，就落在"提问"里"换框架"的这一半。

The mechanism behind questioning turning abundant: a good question = nearest-neighbor search on the knowledge edge. Why would “asking a question,” seemingly the most human and least mechanizable act, be half made abundant? The mechanism hides in a common definition of “good question”: “a good question = spotting the most valuable gap at the edge of knowledge”. Once “valuable” in that definition is narrowed to “the next step checkable against existing data, nearest to the known,” it becomes a nearest-neighbor search problem on a knowledge graph: exactly the strength of large-scale graph analysis. Give AI a full-enough literature graph and ask it to find “which adjacent fields no one has yet bridged,” “which variable is repeatedly mentioned but never directly measured,” “which hypothesis chain is missing its last link”. It will list such in-paradigm good questions faster and more comprehensively than most humans. This is the left half of the break: in-paradigm questioning is in essence a retrieval, and becomes abundant.

The right half is entirely different: “switch the frame, ask a question that could not even hold in the old one”; there is no graph to search, because that graph is the very thing to be replaced. The researcher’s last scarce contribution lies in this paradigm-level half of “asking.”

peer review 的危机：投稿在涨，净知识可能在跌

The peer-review crisis: submissions rise, net knowledge may fall

"提问被充裕、判断变稀缺"不是抽象推演，它在 peer review 这个具体制度上已经显形：截至本版（2026-07），多家期刊与会议都在报告投稿量明显上涨、可观比例的评审带 AI 痕迹（作者估算·未入册·未独立核实）：生成端（写论文、写评审）都被加速，而判断端（决定哪篇值得发）没有等比例扩容。更尖锐的是一个 ODE 模型（arXiv:2604.05714）对评审系统动力学的预测：在生产加速、评审带宽恒定的参数下，系统的净知识可能损失约 40%。注意，这是模型预测，不是已证事实（Ⅲ 级，引用必须写"模型预测"）。但它把张力指明了：当人人都能让 AI 批量产"看似合格"的论文，评审作为唯一的判断闸口会被淹没，而被淹没的评审只能退回最廉价的代理（格式合规、与既有文献相似度、引用数），这恰恰又喂回 RES 03 的结构性偏置：奖励"框架内"、惩罚"换框架"。

“Questions made abundant, judgment scarce” is not abstract derivation; it has already surfaced in the concrete institution of peer review: as of this edition (2026-07), across journals and conferences, submission volumes are reported climbing markedly and a sizable share of reviews bear AI traces (author estimate · not in registry · not independently verified). The generation end (writing papers, writing reviews) is accelerated, while the judgment end (deciding which paper is worth publishing) has not scaled proportionally. Sharper still is an ODE model (arXiv:2604.05714) predicting review-system dynamics: under parameters of accelerating production and constant review bandwidth, the system’s net knowledge may lose about 40%; note, this is a model prediction, not a proven fact (grade Ⅲ; citations must read “the model predicts”). But it names the tension: when anyone can have AI mass-produce “seemingly qualified” papers, review as the sole judgment gate gets flooded. A flooded review can only fall back on the cheapest proxies (format compliance, similarity to existing literature, citation counts), which in turn feeds RES 03’s structural bias: rewarding in-paradigm, penalizing paradigm-level.

它是一道梯度，而不是一条界线。"范式内 / 范式级"读起来像一刀两段，真相更像一道连续的可验证性梯度：从"有现成 benchmark、对错可机检"那一端，平滑滑到"无对错、只有归属、对谁在哪个价值框架下值得"那一端。中间是大片灰区——可机检但判据本身要价值权衡（"多少证据算够"因域而异）。把它画成梯度而不是界线很重要，因为充裕的前线在持续右移：今天还要人判的"提一个范式内好问题"，明年可能被一个够强的知识图谱 agent 吃掉。命题没有押注"某个具体任务永远属于人"，它押注的是梯度最右端那一段——"哪个真相值得知道"这类由人定义的判断——不会因为模型更强而左移，因为它的稀缺不来自能力门槛，来自"无对错可机检"这个结构性事实。下面这张谱系图把若干真实研究动作钉在这道梯度上。

It is a gradient, not a line. “In-paradigm / paradigm-level” reads like a clean cut, but the truth is more a continuous verifiability gradient: from the “has a ready benchmark, correctness machine-checkable” end, sliding smoothly to the “no right answer, only belonging, worth to whom under which value frame” end. In between is a wide grey band: machine-checkable yet the criterion itself needs a value trade-off (“how much evidence is enough” varies by field). Drawing it as a gradient rather than a line matters, because the frontier of abundance keeps moving right: “posing a good in-paradigm question,” which needs a human today, may be eaten next year by a strong-enough knowledge-graph agent. The thesis does not bet that “some specific task stays human forever”; it bets that the rightmost segment — constitutive value judgment — does not move left as models get stronger, because its scarcity comes not from a capability threshold but from the structural fact of “no machine-checkable right answer.” The spectrum below pins several real research actions onto that gradient.

Q-RES-01 · AI 更会选题之后，研究议程该归谁？一个实验室装了个定议程的 agent：喂它全领域文献，它排出下一季度最该做的问题。跑满一年，它选的题在引用和复现上都压过资深 PI，而 PI 偏爱的那个不时髦方向次次垫底。经费续约到了——下一步做什么，谁说了算？

Q-RES-01 · Once AI picks problems better, whose is the research agenda? A lab installs an agenda-setting agent: feed it the field’s literature and it ranks next quarter’s problems. A year in, its picks beat the senior PI’s on citation and replication, while the PI’s unfashionable pet direction lands last every time. Grant renewal comes due: who decides what gets done next?

梯度回答的是"值得"能不能被机器算，不是它归谁。就算得由人定义"值得"，也还有一问：哪个人、哪个机构？AI 一旦稳定选出更高产的问题，冲突就从"人对 AI"变成"谁的价值框架拿到经费和算力"——不是打分能裁的，三方理由都强、且指向相反。

The gradient answered whether “worth” can be machine-computed, not whose it is. Even granting a human must define “worth,” which human, which institution? Once AI reliably selects the higher-yield problems, the conflict shifts from “human vs AI” to “whose value frame gets the funding and compute”: no score adjudicates it, and three strong reasons point opposite ways.

更深的一问。这场听证默认了该问"议程归谁"。更利的一刀在下面：若同一笔资源不养一个被拥有的议程，而是同时养许多互相竞争的小议程，"归谁"之争就自行消解——可谁来护住那些还没兑现就被撤资的输家（创新卷守的"散木"）？再往里一层：当主张生成近乎无限、复现依旧昂贵，谁获得被验证的那次机会本身就是一次分配——而"这个真相值得验"到底是科学判断，还是已经是政治判断？

The sharper question. This hearing assumed the thing to ask is “whose agenda.” The sharper cut is underneath: if the same resources funded many small competing agendas instead of one owned agenda, the “whose” fight dissolves on its own, but who then protects the losers, defunded before they can pay off (the “useless tree” Innovation guards)? One layer deeper: when claim-generation is near-infinite and replication stays expensive, who gets the one chance to be validated is itself an allocation, and is “this truth is worth verifying” a scientific judgment, or already a political one?

一手信号（瓶颈正搬向"判断可信"）：Sakana 的 AI Scientist-v2等级 Ⅳ 一手陈述 + Ⅲ 独立评估出处：Sakana AI 官方陈述 + 第三方独立复核；属探索性侧证，非硬锚——hover 可见其证据等级与出处。 自动化了"提想法 → 设计实验 → 跑 → 写论文 → 评审"的全生命周期，其生成的论文已通过 ICLR 2025 workshop 同行评审（一篇均分越过人类录用阈值）。但独立评估指出：它的文献综述靠简单关键词检索、新颖性判断薄弱（把已确立概念误判为新）。执行被充裕化后，最弱、最稀缺的恰是"判断可信 / 判断新颖 / 判断值得"，这正是命题预言的瓶颈搬家。〔等级 Ⅳ 一手陈述 + Ⅲ 独立评估〕

First-hand signal (the bottleneck is moving toward “judging credibility”): Sakana’s AI Scientist-v2grade Ⅳ first-hand + Ⅲ independent assessmentProvenance: Sakana AI’s own account + third-party independent review; exploratory side-evidence and not a hard anchor; hover to see its evidence grade and source. automated the full lifecycle “ideate → design experiments → run → write the paper → review,” and a paper it generated passed peer review at an ICLR 2025 workshop (one mean score above the human acceptance threshold). But independent assessment notes its literature review relies on simple keyword search and its novelty judgment is weak (mistaking established concepts for new). Once execution is made abundant, the weakest and scarcest things are exactly “judging credibility / novelty / worth”: precisely the bottleneck-move the thesis predicts. [grade Ⅳ first-hand account + Ⅲ independent assessment]

peer review 的结构性冲突：若让 AI 自评"新颖性"，它唯一能用的代理就是"与既有文献分布的距离"：而这恰恰把真正新颖的工作压低分（越偏离现行框架，越像"离群/不可信"）。于是 AI 评审天然奖励框架内、惩罚换框架，把 RES 02 的保守偏置又放大一层。这就是为什么"评 AI 研究者的可信度"是与"评质量"根本不同的工作：前者要的是人去抵抗这条结构性偏置。

The structural conflict in peer review: if AI self-assesses “novelty,” the only proxy available to it is “distance from the existing-literature distribution”. This is exactly what scores genuinely novel work low (the more it departs from the established paradigm, the more it reads as “outlier / not credible”). So AI review intrinsically rewards in-paradigm and penalizes paradigm-level, amplifying RES 02’s conservative bias one more turn. This is why “judging the credibility of an AI researcher” is fundamentally different work from “judging quality”: the former needs a human to resist this structural bias.

"评质量"与"评 AI 研究者的可信度"是两份不同的工作。peer review 的性质改变，是工作种类的质变，不是"评审变多了"这种量变。旧的 peer review 评的是这一篇研究做得好不好：它默认背后有一个会为自己声誉负责、会被同行追责的人类作者。当论文由 AI 研究者批量生成时，评审的对象悄悄从"这一篇"变成了"这个 AI 研究者的产出整体上有多可信"。这是两份工作：前者是逐篇的质量判断，后者是对一个生成源的可信度担保。

为什么这个区别承重？因为它决定了人该把判断投在哪里：如果还按"评质量"的老办法逐篇精读，带宽会被瞬间淹没（RES 09 的天平正是为此而设）；只有认识到该评的是"生成源的可信度"，才会去建可信度评估的机制——抽样复核、追踪某个源的历史命中率、对它的系统性偏置打补丁。把新工作当旧工作做，是 peer review 在 AI 时代失效的第一步。

“Judging quality” and “judging an AI researcher’s credibility” are two different jobs. Peer review’s change of kind is not the quantitative “more reviews” but a qualitative change of job. Old peer review judged how well this one study was done: defaulting to a human author behind it who answers for their reputation and is held accountable by peers. When papers are batch-generated by an AI researcher, the object of review quietly shifts from “this one” to “how credible, overall, is this AI researcher’s output.” These are two jobs: the former is per-paper quality judgment, the latter is vouching for the credibility of a generation source.

Why does this distinction bear weight? Because it dictates where the human should spend judgment: keep reading each paper closely the old “judge quality” way and bandwidth is instantly flooded (RES 09’s ledger exists precisely for this); only by recognizing that what to judge is “the source’s credibility” do you build credibility-assessment mechanisms: sample re-checking, tracking a source’s historical hit-rate, patching its systematic bias. Doing the new job as the old job is peer review’s first step toward failing in the AI era.

AI 科学案例账：把"已发生 / 正在发生 / 推演"分开记

An AI-science case ledger: book “happened / happening / projected” separately

命题最容易被两种姿态毁掉：用一个炫目的成功案例当成"自主科研已成"的证据，或用一个失败案例当成"AI 做不了科学"的反证。两者都把证据等级压平了。诚实的做法是建一张案例账，每一条都标三件事：它是什么时态（已发生 / 正在发生 / 推演）、它的证据等级（Ⅰ–Ⅴ）、以及它支持还是挑战命题。下面这张表把研究卷用到的主要 AI 科学案例摆在一起——读它的方式是看瓶颈搬到哪去了，不是"看 AI 多强"：几乎每一条成功都在执行端，几乎每一条短板都在判断端（新颖性、可信度、方向选择）。这正是命题预言的形状。

The thesis is most easily wrecked by two postures: using one dazzling success as proof that “autonomous science has arrived,” or using one failure as a counter-proof that “AI cannot do science.” Both flatten the evidence grades. The honest move is to build a case ledger where every row is tagged with three things: its tense (happened / happening / projected), its evidence grade (Ⅰ–Ⅴ), and whether it supports or challenges the thesis. The table below sets the main AI-science cases this volume draws on side by side; the way to read it is not “how strong AI is” but where the bottleneck moved: almost every success sits on the execution side, almost every shortfall on the judgment side (novelty, credibility, direction selection). That is exactly the shape the thesis predicts.

读这张表的纪律：把"已发生"的硬锚（Hao 4129.8 万篇是唯一的 Ⅱ 级文献计量）[R9]与"推演"的前沿命题（模式生物是 Ⅴ）分开记账，是这卷不被一次性证伪的关键。Sakana 过评审是真信号，但它是侧证（Ⅳ 一手 + Ⅲ 独立）[R7]，不是因果实验；GNoME 的 220 万材料是一手成果，[R16]但"绝大多数是已知结构内替换"这条限定，才是它对命题的真正贡献。最锋利的一条是太阳系模型：它把"预测准"和"理解对"彻底分开了——AI 能把轨道预测到任意精度，却从未在内部表示里长出"引力"，因为它的目标函数只奖励预测误差，从不奖励"这个变量是不是对的描述层级"。

The discipline for reading this table: booking the “happened” hard anchor (Hao’s 41.3M is the only grade-Ⅱ bibliometrics) [R9]separately from the “projected” frontier claim (the model organism is Ⅴ) is what keeps this volume from being falsified in one shot. Sakana passing review is a real signal, but it is side-evidence (Ⅳ first-hand + Ⅲ independent) [R7], not a causal experiment; GNoME’s 2.2M materials are a first-hand result, [R16] but the qualifier “the vast majority are substitutions inside known structures” is its true contribution to the thesis. The sharpest row is the solar-system model: it cleanly separates “predicting accurately” from “understanding correctly”: AI can predict orbits to arbitrary precision yet never grows a “gravity” representation internally, because its objective rewards only prediction error and never “is this variable even the right level of description.”

赢的是底层四属性，不是某个图谱工具。凡满足这四条的证据载体都被放大，凡是锁在 PDF 截图、私有数据库、不可追溯综述里的都被边缘化——和工程那五条贯穿原理同源：

It is not a particular graph tool that wins, but four underlying properties. Any evidence carrier that meets these four gets amplified; anything locked in PDF screenshots, proprietary databases, or untraceable reviews gets marginalized; the same source as engineering’s five through-lines:

证据库是规格，不是事后归档。最常见的把知识图谱用错的方式，是把它当成研究做完之后放结果的仓库，先生成、再归档。这恰恰颠倒了承重的次序。在 AI-Native 研究里，可追溯证据库是生成的规格，它必须先于生成存在并约束生成：你先在库里写下"什么算可信证据、什么算冲突、什么必须可追溯到原始数据"，生成层才有一个明确的靶子去对齐。这和工程"上下文即基设、先于实现"是同一句话——上下文不是给模型的补充材料，它定义了什么算"做对了"。

次序一旦颠倒，问题立刻显形：先让 agent 批量产出再想办法归档，你会得到一堆格式各异、来源残缺、彼此矛盾却无人察觉的主张——一座无法整合的垃圾山，清理它的成本远超当初省下的生成成本。先立库、后生成，库就成了一道实时护栏：无来源的当场被拦、与既有证据冲突的当场被标记、不可追溯的根本进不来。这就是为什么 RES 13 的研究环把"①框定（先立证据库）"放在"②生成"之前——这是命题本身，不是流程洁癖。

The evidence base is the spec, not after-the-fact archiving. The most common way to misuse a knowledge graph is to treat it as a warehouse for results after the research is done: generate first, archive later. That inverts the load-bearing order. In AI-Native research, the traceable evidence base is the spec for generation; it must exist before generation and constrain it: you write into the base first what counts as credible evidence, what counts as a conflict, what must be traceable to raw data, and only then does the generation layer have a clear target to align to. This is the same line as engineering’s “context is infrastructure, prior to implementation”: context is not supplementary material for the model; it defines what counts as “done right.”

Invert the order and trouble surfaces at once: batch-generate first and archive later, and you get a heap of claims in varied formats, with broken provenance, mutually contradictory yet unnoticed: an un-integratable garbage mountain whose cleanup cost far exceeds the generation cost you saved. Stand up the base first and the base becomes a real-time guardrail: the sourceless is blocked on the spot, conflicts with existing evidence are flagged on the spot, the untraceable never enters. This is why RES 13’s loop puts “① FRAME (stand up the base)” before “② GENERATE”: not process fastidiousness, but the thesis itself.

• 对 agent 可读：主张、证据、来源是结构化、可被模型直接读写的节点，不是只能人读的散文。
• Legible to agents: claims, evidence, and sources are structured nodes a model reads and writes directly, not prose only a human can parse.
• 可追溯：每条主张挂着它的证据边，能回到原始数据/论文，能被独立复现追踪。
• Traceable: every claim carries its evidence edges: back to raw data/papers, trackable for independent replication.
• 可证伪：相互矛盾的主张在图里显形为冲突边，而不是被淹没在两篇互不引用的论文里。
• Falsifiable: contradictory claims surface as conflict edges in the graph rather than drowning in two papers that never cite each other.
• 可整合：跨领域的主张能被缝合、比对、综合，为下一张"整合而非检索"做基设。
• Integrable: claims across fields can be stitched, compared, synthesized: the infrastructure for the next section, “integration, not retrieval.”

结构成因 · 旧 / 新：旧世界里证据散在 PDF 截图、私有数据库、互不引用的综述里，整合靠人脑里偶然的连接；新世界里满足"对 agent 可读 / 可追溯 / 可证伪 / 可整合"四属性的证据载体被放大，其余被边缘化——不是因为某个图谱工具赢了，是因为这四条属性本身就是"让海量生成不塌成噪声"的最小充分条件。它和工程那五条贯穿原理同源：能被模型直接读写、能回到原始出处、矛盾能显形、能跨域缝合。任何缺一条的载体，在生成充裕的世界里都会被自然选择淘汰。

Structural cause · old / new: in the old world evidence scatters across PDF screenshots, proprietary databases, and reviews that never cite each other, and integration relies on accidental connections in a human head; in the new world any evidence carrier meeting the four properties — “legible to agents / traceable / falsifiable / integrable” — gets amplified and the rest marginalized, not because some graph tool won, but because those four properties are themselves the minimal sufficient condition for “mass generation not collapsing into noise.” They share a source with engineering’s five through-lines: readable/writable by models directly, traceable to origin, conflicts made visible, stitchable across fields. Any carrier missing one is selected against in a world of abundant generation.

四属性里，"可证伪"是最容易被偷工的一条

Of the four properties, “falsifiable” is the one most easily skimped

四条属性里，"对 agent 可读"和"可追溯"是工具厂商最爱讲的，因为它们好演示、好卖；真正承重却最容易被偷工的，是可证伪。一个只追求"可读+可追溯"的证据库，会变成一个高效的赞同机器：它能把一万条彼此一致的主张整整齐齐地存好、查好，却从不让相互矛盾的两条主张正面相撞。可证伪的具体形态，是让冲突显形为图里的一条冲突边，而不是任由它淹没在两篇互不引用的论文里——一篇说 X、一篇说非 X，各自有引用、各自"可追溯"，系统却从不报告它们矛盾。当生成把主张推到每小时上千条，这种沉默的矛盾会指数级累积，最后你拥有的是一个自洽的幻觉，而不是知识库。所以建证据库时，"冲突检测"不是锦上添花的高级功能，它是把"库"和"堆"区分开的那条线。

Of the four properties, “legible to agents” and “traceable” are the ones tool vendors love to talk about, because they demo well and sell well; the truly load-bearing yet most easily skimped is falsifiable. An evidence base that chases only “legible + traceable” becomes an efficient agreement machine: it can neatly store and query ten thousand mutually consistent claims yet never let two contradictory claims collide head-on. The concrete form of falsifiable is making a conflict visible as a conflict edge in the graph, rather than letting it drown in two papers that never cite each other. One says X, one says not-X, each cited, each “traceable,” yet the system never reports that they contradict. When generation pushes claims to thousands an hour, such silent contradictions accrue exponentially, and what you end up with is not a knowledge base but a self-consistent illusion. So when building the base, “conflict detection” is not a nice-to-have advanced feature; it is the line that separates a “base” from a “pile.”

生产速度与可消化速度的剪刀差，是整合升值的根

The scissors-gap between production and digestion is the root of integration’s rising value

为什么整合会从"研究的一个环节"升值成"人最稀缺的贡献"？因为两条曲线的剪刀差正在张开。科学文献的生产侧本就在指数增长——约 250 万篇/年、每 9 年翻倍；AI 在其上又叠了一层质变加速，把"产出一篇"的成本压到近零。可消化侧呢？人类的认知带宽近乎恒定：一个研究者一天能真正读懂、判断、并入自己理解结构里的论文数，几十年没变多少。生产曲线陡升、消化曲线平躺，两者之间的剪刀差就是"已生成但无人整合"的知识堆积量，它在以生产曲线的速率累积。这道剪刀差不是某个工具能补的，因为补它需要的恰恰是带宽，而带宽是被锁住的那一侧。所以整合的升值是结构性的：在一个生产近无限、消化近恒定的世界里，唯一还在涨价的，就是那个能把碎片缝成理解的认知主体的注意力。〔此为命题推演＋数量级侧证，"堆积速率"缺一手实证，标待坐实〕[thesis-derivation + order-of-magnitude side-evidence; the “accrual rate” lacks first-hand empirics, flagged to be grounded]

Why does integration appreciate from “one step of research” into “the human’s scarcest contribution”? Because the scissors-gap between two curves is opening. The production side of the scientific literature was already growing exponentially: ~2.5 million papers a year, doubling every 9 years; AI layered on a qualitative acceleration, driving the cost of “producing one” to near-zero. And the digestion side? Human cognitive bandwidth is nearly constant: the number of papers a researcher can truly understand, judge, and fold into their structure of understanding in a day has barely grown in decades. Production curve steep, digestion curve flat: the scissors-gap between them is the stock of “generated but un-integrated” knowledge, accruing at the production curve’s rate. No tool fills this gap, because filling it requires precisely bandwidth, and bandwidth is the locked side. So integration’s appreciation is structural: in a world of near-infinite production and near-constant digestion, the one thing still rising in price is the attention of a cognitive subject that can stitch fragments into understanding. 〔此为命题推演＋数量级侧证〕

推至极限：检索是"找到已存在的那条"——可被向量库 + agent 充裕化到近免费。整合是"把从未被并置的几条缝成一个新理解"，它要求一个能同时持有多个框架、判断哪些该缝、缝出的东西是否成立的认知主体。当 AI 每小时新增成千上万条可检索主张，"已生成但无人整合"的知识会堆积成山。研究者的杠杆，是站在这座山顶做综合，而不是在山脚多搬几块砖。

Pushed to the limit: retrieval is “find the one that already exists”: can be made abundant to near-free by a vector store plus an agent. Integration is “stitch several never-juxtaposed claims into a new understanding”: it demands a cognitive subject that can hold several frames at once, judge which to stitch, and judge whether the stitch holds. As AI adds thousands of retrievable claims an hour, “generated but un-integrated” knowledge piles into a mountain. The researcher’s real leverage is to synthesize from the summit, not to haul a few more bricks at the base.

〔探索清单·待坐实〕"整合鸿沟急剧扩大"目前是命题推演＋一篇讲"怎么算知道"的综述侧证（《The epistemic revolution of AI》直指"知识生产速度超出单一人类认知"），尚无单篇一手实证锚量化"已生成未整合"的堆积速率；标先行指标：整合产物（综述/理论缝合）相对原始产出的比率，以及"已生成但无人整合"的知识堆积量。证伪条件：若自动综述系统能在专家盲评下产出被判"真正缝合而非拼贴"的整合，则整合的人类专属性松动。

[exploratory · to be grounded] “The integration gap explodes” is for now thesis-derivation plus epistemology-review side-evidence (The epistemic revolution of AI points directly at “the rate of knowledge production outpacing single-human cognition”). It has no single first-hand empirical anchor quantifying the accrual rate of “generated-but-un-integrated”; leading indicators: the ratio of integration artifacts (reviews / theory-stitching) to raw output, and the stock of “generated but un-integrated” knowledge. Falsifier: if automated review systems produce, under blinded expert judging, integrations rated “genuinely stitched, not collaged,” the human-exclusivity of integration weakens.

缝合是一种换框架的动作，不是更长的检索

Stitching is not longer retrieval; it is a paradigm-level act

最容易把整合贬低成"高级检索"的误解是：以为只要把检索窗口拉得足够长、把上下文塞得足够满，模型就能"读完一切"然后自动综合。这混淆了两个种类不同的动作。检索是在同一个框架内找到相关条目，它有标准答案、有最近邻、可被向量库充裕化。缝合是跨框架地判断"这几条本来不在一起的主张，并置之后是否生出一个新理解"，它没有最近邻可循，因为"该把哪几条放在一起"这个判断本身就在框架之外。这正是 RES 02 那道可验证性梯度的右段：缝合往往要求一次小型的换框架动作——决定用哪个新框架来组织这些碎片。所以整合不是检索的延长线，它和"提一个换框架的问题"是同一种稀缺判断的两种表现。

The easiest way to demote integration to “advanced retrieval” is to assume that a long-enough retrieval window and a full-enough context will let the model “read everything” and synthesize automatically. This conflates two different kinds of act. Retrieval finds relevant items within one frame: it has a right answer, a nearest neighbor, and can be made abundant by a vector store. Stitching judges across frames whether “these claims that were never together generate a new understanding once juxtaposed”: it has no neighbor to follow, because “which ones to put together” is itself a judgment outside any frame. This is precisely the right end of RES 02’s verifiability gradient: stitching often demands a small paradigm-level act, deciding which new frame organizes the fragments. So integration is not retrieval extended; it and “posing a paradigm-level question” are two expressions of the same scarce judgment.

带宽问题，不是存量问题，这一区分有操作后果。如果整合是存量问题（"读得不够多"），解法就是让 AI 多读、多产摘要，把存量补齐。但它是带宽问题（"同时持有多个框架并判断如何缝合的认知容量有限"），那么多产摘要只会让堆积更快，让带宽更紧。操作上的差别巨大：把工时投回"让 AI 多读多摘"，是在加速病因；把工时投回"让人专注做跨框架综合，AI 只负责把候选材料喂到手边"，才是对症。所以 RES 13 的研究环把④整合设为承重瓶颈阀门——不是因为整合"重要"这种空话，而是因为它是唯一不能靠多产来缓解、反而会被多产加重的环节。盯住一个具体指标：整合产物（真正缝出新理解的综述/理论）相对原始产出的比率，是升还是降。

A bandwidth problem, not a stock problem, and this distinction has operational consequences. If integration were a stock problem (“haven’t read enough”), the fix would be to have AI read more and summarize more, filling the stock. But it is a bandwidth problem (“limited cognitive capacity to hold several frames at once and judge how to stitch”), so producing more summaries only piles faster and tightens bandwidth further. The operational difference is large: reinvesting hours into “let AI read and summarize more” accelerates the cause; reinvesting into “let humans focus on cross-frame synthesis while AI only feeds candidate material to hand” treats it. This is why RES 13’s loop sets ④ integration as the load-bearing bottleneck valve. This is not from the empty phrase that integration “matters,” but because it is the one step that cannot be relieved by more output and is in fact worsened by it. Watch one concrete metric: whether the ratio of integration artifacts (reviews/theory that genuinely stitch a new understanding) to raw output is rising or falling.

可复现是那道墙：错误不该被删，该被回流。整合的反面，是把无法整合的东西悄悄丢掉。当生成每小时新增成千上万条主张，最省事的处理是把"复现不出来的""与既有库冲突的""看起来离群的"统统标记为噪声删掉，这恰恰是最危险的动作。可复现性在这卷里不是一个质检环节，它是那道墙：它把"会自我纠偏的研究环"和"高速生成器"分开。一条主张复现不出来，有两种可能，它是错的（该证伪），或它揭示了现有评估方法的盲区（该建新 eval）。把它当噪声删掉，你两种信息都丢了；把它回流成证据库里的一条新规则、一个新节点类型、或一道新的复现检查，错误就变成了护栏，下一轮少犯。下面这张图把这条"错误回流"画成闭环。

Reproducibility is the wall: errors should not be deleted, they should feed back. The opposite of integration is quietly discarding what cannot be integrated. When generation adds thousands of claims an hour, the laziest handling is to tag everything “non-replicable,” “conflicting with the base,” “looking like an outlier” as noise and delete it, which is precisely the most dangerous move. Reproducibility in this volume is not a QC step; it is the wall that separates “a self-correcting research loop” from “a fast generator.” A non-replicable claim has two possibilities: it is wrong (falsify it) or it reveals a blind spot in current evaluation methods (build a new eval). Delete it as noise and you lose both kinds of information; feed it back as a new rule, a new node type, or a new replication check in the evidence base, and the error becomes a guardrail, fewer next round. The figure below draws this “error feedback” as a closed loop.

提问也被充裕之后，人退到最后一个问题：该追哪个真相。它看着还是研究，其实已经不在原来的坐标系里了。前面每一步问的都是“这答案对不对”——有证据可摆、可机检、机器迟早追上；而“该追哪个”问的是“这值不值得知道、对谁值得”，它没有一个对所有人都成立的正确答案。同一台再准的 AI 也答不了后者，不是因为它还不够强，而是因为后者根本不在“对不对”那根轴上。坐标系就换了。

Once questioning too is made abundant, the human retreats to one last question: which truth to chase. It still looks like research, yet it no longer lives in the old coordinate system. Every earlier step asked “is this answer correct”: evidence can be laid out, a machine can check it, and the machine catches up eventually. But “which one to chase” asks “is it worth knowing, and worth to whom”: a question with no answer correct for everyone. However accurate the AI, it cannot settle the latter, not because it is not yet strong enough but because the latter simply is not on the “correct-or-not” axis. The coordinate system has changed.

先看一个具体的：对一个把"延长健康寿命"当头等大事的群体，"衰老的分子机制"太值得知道了；对一个把"生态完整"当头等大事的群体，同一笔预算也许更该投去别处。两边都没"判错"——因为"值得"是相对某个价值框架被定出来的，不是被发现的。这一步的特殊就藏在这里：翻到"值不值得知道"这根新轴上，AI 再准，也给不出一个对所有价值框架都对的答案，因为那样的答案在逻辑上不存在。下面把它为什么是一道人接住的判断、而非又一种能力，一层层说清。

Start concrete: to a group that holds “extending healthy lifespan” as its first good, “the molecular mechanism of aging” is well worth knowing; to a group that holds “ecological integrity” as its first good, the same budget might better go elsewhere. Neither “judged wrong”: “worth” is defined relative to a value frame, not discovered. That is where this step is special: once you pivot onto the “worth-knowing” axis, no matter how accurate AI gets, there is no answer correct for all value frames, because logically no such answer exists. What follows unpacks, layer by layer, why this is not one more capability but a judgment a human holds.

异质的"值得"学不到，因为它的样本只有一个

The heterogeneous “worth” cannot be learned, because its sample size is one

为什么 AI 学得到平均的"值得"，学不到异质的"值得"？根子在学习的前提：任何被学的东西，都要有足够多的、可被归纳的样本。平均的"值得"（一个领域里被反复表达、被大量论文共同认可的价值取向）有海量样本，所以可被外化、可被 RLCF 当 reward 学走。但真正承重的那种"值得"，是只对某个个体、某个群体、在某个特定价值框架下才成立的，它的样本量是一。一个研究者基于自己独特的处境、经历、所属共同体的关切，判断"这个真相对我们值得追"，这个判断没有一个可被归纳的训练集，因为它死死绑定在那个独一无二的视角上。

这正好对应 RES 08 讲的同质化机制的镜像：AI 默认拉向均值（regression to a domain prototype），而异质的价值判断按定义就是偏离均值的那部分。偏离均值的"值得"能不能被学走？本卷押"不能"——赌的是它样本量为一、统计学习要的可归纳性在此落空，不是赌模型还不够强。这是押注不是定理：最强的反方就在下面那张图里——RLCF 已能学走"科学品味的社群均值"（arXiv:2603.14473），能否够到偏离均值的前沿价值，至今没有直接实验。若某天 AI 自选、偏离共识的议程能长期稳定长出被独立复现的新框架，这块守地就得让出来。

Why can AI learn the average “worth” but not the heterogeneous “worth”? The root is learning’s premise: anything learned needs enough inducible samples. The average “worth” — a value orientation repeatedly expressed in a field, jointly endorsed by many papers — has vast samples, so it can be externalized and learned by RLCF as reward. But the truly load-bearing kind of “worth” is the one that holds only for a particular individual, a particular group, under a particular value frame: its sample size is essentially one. When a researcher, drawing on their unique situation, history, and community’s concerns, judges “this truth is worth chasing for us,” that judgment has no inducible training set, because it is constitutively bound to that one-of-a-kind vantage point.

This mirrors the homogenization mechanism of RES 08: AI pulls toward the mean by default (regression to a domain prototype), and a heterogeneous value judgment is by definition the part that departs from the mean. Can the off-mean “worth” be learned? This volume bets it cannot, wagering that its sample size is one and the inducibility statistical learning needs has nothing to grip, not that the model is merely too weak yet. This is a bet, not a theorem: its strongest rival sits in the figure below, where RLCF can already learn “the community mean of scientific taste” (arXiv:2603.14473), while whether it reaches the off-mean frontier value has no direct experiment yet. If one day an AI’s self-chosen, off-consensus agenda durably grows independently replicated new frames, this ground has to be ceded.

换的是轴，不是轴上的位置——内核②那条"只能人判"的支，就落在这里。

The axis changes, not the position on it: the kernel’s “human-only” branch of fork ② lands right here.

"值得"由人定义，不是又一种能力

“Worth” is a constitutive stipulation, not one more capability

把"哪个真相值得知道"读成"人比 AI 更会判断价值的一种能力"，是最常见也最危险的误读：能力可以被超越，若"值得"只是能力，够强的模型迟早做得比人好，整卷的人本立论就只是暂时的。命题主张的不是这个："值得"是由人定义的，不是一种能力。区别在于，能力问"谁判得更准"，这里问的是"由谁来定义这件事算不算数"。"这个真相对我们值得知道"这句话里，没有一个外在的、可被更准的判断逼近的"正确答案"；它的真值由提出它的那个价值框架定出来。

Reading “which truth is worth knowing” as “a capability at which humans judge value better than AI” is the most common and most dangerous misreading: capabilities can be surpassed, so if “worth” were only a capability, a strong-enough model would eventually beat humans at it and the volume’s human argument would be only temporary. That is not what the thesis claims: “worth” is not a capability but a constitutive stipulation. The difference: a capability asks “who judges more accurately,” a constitutive matter asks “who gets to define whether this counts at all”. In the sentence “this truth is worth knowing to us” there is no external “right answer” that a more accurate judgment could approach; its truth value is constituted by the value frame that poses it.

反转可能（向上游汇入的承重）：研究的终极问题或许在于"为什么某些真相更值得知道"，而非"怎么发现真相"。这让研究卷天然向上汇入创新（价值发现）：研究在"知识边界上识别空白"，创新在"判断这空白指向的价值"。两者咬合处，正是"值得"这个词从"怎么算真"交到"什么值得"的那一刻。

The reversal (the load-bearing merge upward): research’s ultimate question may never have been “how to find truth” but “why some truths are more worth knowing”. This makes the research volume merge naturally upward into Innovation (value discovery): research spots gaps at the edge of knowledge, innovation judges the value those gaps point to. Their meshing point is exactly the moment the word “worth” passes from epistemology into axiology.

问题选择的品味，是研究里最稀缺的判断

Taste in problem selection is research’s scarcest judgment

把"哪个真相值得知道"再往实操拉一格，它落地成一个具体动作：问题选择。Anthropic 2026 的"自主性阶梯"[R4]把研究 agent 的能力分级，结论很硬——最右端、最难自动化的一阶，恰是 research agenda selection（研究议程选择）。Claude 可以在"执行良定义的实验"上匹敌甚至超过熟练人类，但"选择该做哪些问题、哪些异常值得追、哪个诱人想法其实是死路"这件事，仍有明显差距。chenhaot 的"The Mirage of the AI Scientist"把这条形式化为：人类不可替代的角色是 Selector（选什么做）+ Evaluator（评质量/可信）：科学根本是一个资源分配问题，不是智能问题，产出更多不等于知识更多。Terence Tao 的那句话是同一件事的另一种说法："当想法生成的成本被压到近零，瓶颈就变成 verify / evaluate"，注意力成为知识经济里最稀缺的资源。

Pull “which truth is worth knowing” one notch toward the operational and it lands as a concrete act: problem selection. Anthropic’s 2026 “ladder of autonomy” [R4]grades a research agent’s capabilities, and the conclusion is hard: the rightmost, hardest-to-automate rung is precisely research-agenda selection. Claude can match or exceed skilled humans at “executing a well-defined experiment,” but “choosing which problems to work on, which anomalies are worth chasing, which seductive idea is actually a dead end” still shows a clear gap. chenhaot’s “The Mirage of the AI Scientist” formalizes the irreplaceable human role as Selector (what to do) + Evaluator (quality / credibility): science is fundamentally a resource-allocation problem, not an intelligence problem; producing more is not knowing more. Terence Tao’s line is the same thing in other words: “when the cost of idea generation is driven to near-zero, the bottleneck becomes verify / evaluate”; attention becomes the scarcest resource in the knowledge economy.

品味不是不可讨论的直觉，它是一道可分级的梯度。从"选一个有数据、有 benchmark、稳出论文的问题"（低品味，AI 已能做），到"选一个别人觉得无聊但你直觉有矿的问题"，到"选一个连提出来都需要换框架的问题"（高品味，AI 训练分布之外）。AI 能学到的是这道梯度的左半段：RLCF（用社群偏好当 reward）[R5]已经证明"科学品味的社群均值"可被外化、可被学。但它学到的恰恰是社群均值，而问题选择品味，在于偏离当前社群均值的那部分前沿价值，这正好是同质化研究系统化的东西。下面这张图把问题选择品味画成一道梯度，并标出 AI 能学到哪一段、学不到哪一段。

Taste is not mysticism; it is a gradable gradient. From “pick a problem with data, a benchmark, and a steady paper yield” (low taste, AI already does it), to “pick a problem others find boring but your intuition says holds ore”. At the top: “pick a problem you cannot even pose without changing the frame” (high taste, outside AI’s training distribution). What AI can learn is the left half of this gradient: RLCF (community preference as reward) [R5]has shown that “the community mean of scientific taste” can be externalized and learned. But what it learns is exactly the community mean, whereas real problem-selection taste lies in the frontier value that departs from the current community mean, precisely what homogenized research systematizes away. The figure below draws problem-selection taste as a gradient and marks which segment AI can learn and which it cannot.

人回归意义，在研究面是把研究者还给值得追问的问题

On the research face, “humans return to meaning” means returning researchers to the questions worth asking

内核第④步"人回归意义"不是一句温情的收尾，它在研究面有一个非常具体、可检验的落点：把研究者从"多产论文"的跑步机上解放出来，还给那些值得追问、却不可度量的问题。这正是人本立论在研究卷的全部分量，更便宜的执行，从来不是目的本身。AI 把执行做便宜，目的不是让研究者在单位时间里产出更多论文（那只是把人更深地绑在 exploitation 的轮子上），而是让研究者腾出认知带宽，去做那个机器做不了、也最值得人做的动作：判断哪个真相值得知道、守住一个不被产量绑架的研究方向。一个把 AI 用对了的研究组织，它的研究者花在"追问、判断、整合"上的时间应该变多，而不是花在"赶产出"上的时间变多。

这条人本主线与组织卷"让人回归组织中心"是同一句话的两个面：组织面是让人回到判断节点，研究面是让人回到值得追问的问题。两者咬合处，是贯穿全系列的同一件事：把执行交出去省下的认知带宽，落回人手里去问、去判断、去负责，而不是反过来把人绑在产出指标上。

Kernel step ④, “humans return to meaning,” is not a sentimental closer; on the research face it has a very concrete, testable landing point: freeing researchers from the “more papers” treadmill and returning them to the questions worth asking yet unmeasurable. This is the full weight of the human argument in the research volume — cheaper execution is never the point in itself. AI makes execution cheap not so that researchers produce more papers per unit time (that only binds people deeper to the wheel of exploitation), but so that researchers free up cognitive bandwidth for the act the machine cannot do and that is most worth a human doing: judging which truth is worth knowing, holding a research direction not hijacked by output. In a research organization that uses AI rightly, researchers’ time on “asking, judging, integrating” should increase, not their time on “chasing output.”

This human through-line is two faces of one sentence with the Org volume’s “put people back at the organization’s center”: the org face returns people to the judgment node, the research face returns people to the questions worth asking. Their meshing point is the one thing that runs through the whole series: the cognitive bandwidth freed by handing execution away lands back in human hands — to ask, to judge, to take responsibility — rather than binding people to output metrics.

正交退守——补盲区：当一阶研究被充裕化，人退守的不止"值得知"，还有两个正交方向。其一，退守到设计科学本身：当跑实验近乎免费，"该跑哪个实验、用什么判据算证据、什么样的研究设计能真正区分假设"这套元层方法判断，反而升值。其二，退守到元科学：研究"研究本身怎么被 AI 改写"——RES 03 的 Nature 文献计量正是这一退守的产物。两条都是把判断节点抬高一层，不是"做得更快"。

Orthogonal retreats — filling the blind spot: when first-order research is made abundant, the human retreats not only to “worth knowing” but in two orthogonal directions. First, to the design of science itself: when running an experiment is near-free, the meta-level methodological judgment — “which experiment to run, what counts as evidence, what research design actually discriminates between hypotheses” — appreciates. Second, to meta-science: studying “how research itself is being rewritten by AI”; RES 03’s Nature bibliometrics is a product of exactly this retreat. Neither is “doing it faster”; both raise the judgment node one level.

生成层的保守偏置（再钉一遍：加速 ≠ 进步）：RES 02 的提问聚集、RES 03 的主题收缩与新颖被压分，合起来指向同一件事：生成层默认朝向"安全、数据丰富、与现行框架一致"的方向加速。它让科学跑得更快，却可能跑得更窄。守住研究方向的价值，正是为了抵抗这条偏置：让"值得"由人（在某个价值框架下）来定，而不是由"离已知最近"来定。这就是为什么 第 8 节 必须把价值责任交给组织的治理结构——价值判断没有归属，就会被生成层的默认偏置悄悄替换掉。

The generation layer’s conservative bias (nailed once more: acceleration ≠ progress): RES 02’s question-clustering and RES 03’s topical contraction and novelty-penalty point to one thing: the generation layer accelerates by default toward the “safe, data-rich, paradigm-consistent” direction. It makes science run faster while possibly running narrower. Owning the value of a research direction is exactly how to resist this bias: let “worth” be set by humans (within some value frame), not by “nearest to the known”. This is why Section 8 must hand value accountability to the organization’s governance structure; a value judgment with no owner gets quietly replaced by the generation layer’s default bias.

散木的命运：效率会自动吃掉冗余探索

The fate of the useless tree: efficiency eats redundant exploration by default

"谁有权定方向"不是一个抽象的治理问题，它有一个非常具体的失败模式：散木（创新卷的刻度名：暂时无用、却值得留下的探索）被效率悄悄吃掉。多源一致地指向同一件事：AI 放大的是 exploitation（精炼、效率、执行），而不是 exploration。组织结构天然偏 exploitation，因为它可预测、可度量、反馈快；AI 每一次落地都发出一个干净的"进步"信号（CFO 友好），而 exploration 的故事模糊、需要想象力、回报不可度量。March 1991 的框架[R15]仍是底座：探索（搜索/变异/冒险）与利用（精炼/选择/效率）竞争同一份资源，利用倾向于赢。更隐蔽的机制是："省下来的产能不会自动变成 slack"：技术省下的工时通常被重新分配去做更多同样的事（more volume），而不是不同的事。"什么被度量，什么被管理；什么不可度量，什么先被砍"，slack 因不可度量而最先被砍。这就是为什么"守住研究方向的价值"必须是一个有归属人、被刻意保护的治理动作，而不能指望它自然存活。

“Who owns the direction” is not an abstract governance question; it has a very concrete failure mode: the useless tree (the Innovation volume’s gauge for exploration that is useless now but worth keeping) quietly eaten by efficiency. Multiple sources converge on one thing: what AI amplifies is exploitation (refinement, efficiency, execution), not exploration. Organizational structure tilts toward exploitation by nature, because it is predictable, measurable, fast-feedback; every AI rollout emits a clean “progress” signal (CFO-friendly), while exploration’s story is fuzzy, demands imagination, and pays off unmeasurably. March’s 1991 frame [R15]is still the base: exploration (search / variation / risk) and exploitation (refinement / selection / efficiency) compete for one budget, and exploitation tends to win. The subtler mechanism: “freed capacity does not automatically become slack”; saved hours usually get reallocated to more of the same (more volume), not to something different. “What gets measured gets managed; what cannot be measured gets cut” — slack, being unmeasurable, is cut first. This is why “owning the value of a research direction” must be an owned, deliberately protected governance act and cannot be expected to survive on its own.

省下的产能不会自动变成 slack。技术省下的工时，默认被重新分配去做更多同样的事，而不是变成可自由探索的 slack，因为 slack 不可度量，"多产 X% 论文"可度量，不可度量的先被砍。所以守值（守住"哪个方向值得"）与守散木（为不可度量的探索留预算）是同一个治理动作的两面：都得被显式预留、有人为长期回报负责，不能指望它作为省时的副产品自然涌现。

Freed capacity does not automatically become slack. Hours that technology frees get reallocated by default to more of the same, not turned into slack for free exploration — because slack is unmeasurable while “X% more papers” is measurable, and the unmeasurable gets cut first. So owning value (holding “which direction is worth it”) and protecting the useless tree (a budget for unmeasurable exploration) are two faces of one governance act: both must be explicitly reserved and owned by someone accountable for the long-term return, not expected to emerge as a by-product of time-saving.

但散木可被守护，这是条件论，不是宿命。"用 token 替换人＝exploitation；用 token 增强人＝exploration"——同一项 AI 能力，落进不同的激励结构，结局相反。守护冗余探索的具体动作是治理性的：设独立的探索单元、按"学习/新颖"而非"产量"考核、显式给"看上去无用"的方向留预算（Bell Labs、Xerox PARC、早期剑桥 LMB 的历史证据都指向"小团队 + 体制保护"）。把这条接回内核④：人回归意义，在研究面的具体落点，就是把研究者从"多产论文"的 exploitation 跑步机上解放出来，还给那些不可度量、却可能换地图的问题。更便宜的执行从来不是目的本身；真正要的，是让人去问那个值得问的问题。

价值判断没有归属，就被默认偏置悄悄替换。"谁有权定方向"之所以是治理问题而非技术问题，关键在一个容易被忽视的动力学：价值判断不会停留在真空里，它要么有归属人、被有意识地行使，要么被生成层的默认偏置悄悄填补。没有中间态。

当一个研究组织不明确"谁来判这个方向值不值得追"，这个判断不会消失，它会被默认地交给"哪个方向有数据、有 benchmark、能稳出结果"，也就是 RES 02/03/08 反复指认的那条向已知收敛的保守偏置。于是组织以为自己在"中立地跟随数据"，实际上是在无人负责的情况下，让生成层的结构性偏置替它做了方向选择。治理的全部意义，就是把这个判断从默认偏置手里夺回来，交给一个具名的、要为长期后果负责的人。这也是为什么 第 8 节 必须把研究方向的价值责任明确交给组织的治理结构——不是为了多设一个审批岗，是为了堵住"价值判断被偏置接管"这个无声的漏洞。

A value judgment with no owner gets quietly replaced by the default bias. “Who owns the direction” is a governance question, not a technical one, because of a dynamic easily overlooked: a value judgment does not stay in a vacuum: either it has an owner exercising it consciously, or the generation layer’s default bias quietly fills it. There is no middle state.

When a research organization leaves “who judges whether this direction is worth chasing” unspecified, the judgment does not vanish. It is handed by default to “which direction has data, a benchmark, a steady yield,” exactly the conservative bias toward the known that RES 02/03/08 keep naming. So the organization believes it is “neutrally following the data” while in fact, with no one accountable, the generation layer’s structural bias has made the direction choice for it. The whole point of governance is to take that judgment back from the default bias and give it to a named person accountable for the long-term consequences. This is why Section 8 must explicitly hand research-direction value accountability to the organization’s governance structure — not to add an approval seat, but to plug the silent leak of “value judgment captured by the bias.”

But the useless tree can be protected — this is conditional, not fated. “Replacing humans with tokens = exploitation; augmenting humans with tokens = exploration”: the same AI capability, dropped into different incentive structures, ends opposite ways. The concrete acts of protecting redundant exploration are governance acts: stand up independent exploration units, appraise on “learning / novelty” rather than “output,” explicitly budget for “seemingly useless” directions (the historical evidence of Bell Labs, Xerox PARC, the early Cambridge LMB all points to “small teams + institutional protection”). Wire this back to kernel ④: the human’s return to meaning, on the research face, lands concretely as freeing researchers from the exploitation treadmill of “more papers” and returning them to the unmeasurable questions that might redraw the map. Cheaper execution is never the point in itself; what we want is to let people ask the question worth asking.

预测准 ≠ 理解对：太阳系模型没长出"引力"

Accurate prediction ≠ correct understanding: the solar-system model never grew “gravity”

hypernormal science 最容易被低估，是因为它常常伪装成"成功"。一个能精确预测的模型看起来就是好科学，但"预测准"和"理解对"是两回事，而 AI 的目标函数只奖励前者。最干净的例子：在 1000 万个模拟太阳系上训练的基础模型，能把行星轨道预测到极高精度，却从未在内部表示里长出"引力"这个概念：它学到的是一堆能复现观测的统计规律的拼凑，不是那条让所有轨道统一起来的力学定律。从"预测准"这个指标看，它完美；从"理解对"看，它什么都没理解。

这就是 hypernormal 的伪装：当评价标准只剩"在现有 benchmark 上更高分"，一个把框架内预测做到极致、却完全没触及描述层级的系统，会被一路绿灯放行，甚至被当成换框架的突破来庆祝。守住这道分别——预测力的提升不等于理解的加深——是不被 hypernormal science 骗过去的第一道认知防线。它也解释了为什么本卷坚持"产出量/预测精度本身不是指标"：真正该问的是覆盖广度有没有扩、描述层级有没有被挑战。

Hypernormal science is most easily underestimated because it often disguises itself as “success.” A model that predicts accurately looks like good science — but “predicting accurately” and “understanding correctly” are two things, and AI’s objective rewards only the former. The cleanest example: a foundation model trained on 10 million simulated solar systems predicts planetary orbits to very high precision, yet never grows the concept of “gravity” in its internal representation. What it learned is a patchwork of statistical regularities that reproduce the observations, not the one law of mechanics that unifies all orbits. By the metric “predicts accurately,” it is perfect; by “understands correctly,” it understood nothing.

This is hypernormal’s disguise: when the only evaluation criterion left is “a higher score on the existing benchmark,” a system that perfects in-paradigm prediction while never touching the level of description gets waved through, even celebrated as a paradigm-level breakthrough. Holding this distinction — a gain in predictive power is not a deepening of understanding — is the first cognitive defense against being fooled by hypernormal science. It also explains why the volume insists “output volume / prediction accuracy itself is not the metric”: what should be asked is whether topical breadth widened and whether the level of description was challenged.

结构成因 · 为何会这样：机器学习靠"对预先定义好的变量/标签最小化预测误差"，它擅长预测当前数据，却被锁进所学数据的概念词汇（这正是 RES 00 地图隐喻的要害：填满地图 ≠ 重画地图）。AI 能把地图上的空白填满（DeepMind GNoME 发现 220 万新材料，绝大多数是已知结构类型内的元素替换；ESM3 设计新荧光蛋白＝填空，不是画新地图），但它不会去问"现在这套描述层级是不是错的"。

William Farr 的霍乱地图把数据围绕"空气质量"组织，再聪明的 AI 也推不出"水传播微生物"这个没人记录过的变量：germ theory 要靠换显微镜、换仪器、换变量。

Structural cause · why this happens: machine learning works by “minimizing prediction error against pre-defined variables/labels” — good at predicting current data, but locked into the conceptual vocabulary of what it was trained on (the point of RES 00’s map metaphor: filling the map ≠ redrawing it). AI can fill the blanks (DeepMind’s GNoME found 2.2 million new materials, the vast majority element-substitutions inside known structure types; ESM3 designed novel fluorescent proteins: filling gaps, not drawing a new map), but it will not ask “is this whole level of description wrong.”

Farr’s cholera map organized data around “air quality,” and no AI, however clever, could infer “waterborne microbes,” a variable no one had recorded: germ theory required changing the microscope, the instrument, the variable.

暗面发生在判断交还给人之前。前面七张的逻辑是"执行被充裕 → 判断退守给人"，听起来像一条干净的接力：机器跑完执行，人接过判断。但 hypernormal science 揭示了一个前置故障：暗面发生在判断被交还之前。原因在于，生成层不是中立地铺开所有候选再等人来选，它在生成的那一刻就已经带着偏置：它优先生成"安全、数据丰富、与现行框架一致"的候选，把"换框架、换变量"的候选压到尾部甚至根本不生成。于是当人来接判断时，摆在他面前的候选集本身已经被悄悄收窄了。人以为自己在"从所有可能里选最值得的"，实际上在"从一个已被保守偏置过滤过的子集里选"。

这就是为什么不能把研究环简单理解为"生成中立、判断承重"——生成本身就携带价值倾向，而这个倾向恰恰朝着框架内。守值的动作因此必须前移：不只在判断时守，还要在生成时主动要求候选集包含"换框架"的选项，否则人守的是一个已经被做了手脚的菜单。

The dark side happens before judgment is handed back to humans. The logic of the first seven sections is “execution made abundant → judgment retreats to humans,” which sounds like a clean relay: the machine finishes execution, the human takes over judgment. But hypernormal science reveals a pre-failure: the dark side happens before judgment is handed back. The reason: the generation layer does not neutrally spread all candidates and wait for a human to pick; it already carries bias at the moment of generation, and it preferentially produces “safe, data-rich, paradigm-consistent” candidates and pushes “change-the-frame, change-the-variable” candidates to the tail or never generates them. So when the human arrives to judge, the candidate set in front of them has already been quietly narrowed. The human believes they are “selecting the most worthy from all possibilities” while actually “selecting from a subset already filtered by the conservative bias.”

This is why the research loop cannot be read simply as “neutral generation, load-bearing judgment”: generation itself carries a value tilt, and that tilt runs toward the in-paradigm. The act of owning value must therefore move earlier: not only guarding at judgment time but, at generation time, actively demanding that the candidate set include paradigm-level options; otherwise the human guards a menu that has already been rigged.

同质化的机制：写在权重里，推理时救不回

The mechanism of homogenization: written into the weights, unrecoverable at inference

hypernormal science 不只是"AI 偏保守"的软倾向，它有一条写进权重的硬机制。最强的因果锚是 Doshi & Hauser（Science Advances 2024）[R12]：给写作者 LLM 点子，个体故事更"有创意"，但故事彼此更相似：他们明确称之为"社会困境"（个人更好、集体更窄）。同质化是群体层效应（Anderson 等 2024，36 人实验）[R13]：它不来自个体固着，而来自 LLM 向不同用户建议相似点子。更狠的是跨模型同质（"We're Different, We're the Same" 2025）[R14]：控制结构变量后，LLM 之间的相似度远高于人与人之间：换个模型也救不了。机制层的决定性证据：post-training 的 diversity collapse 写在权重里、推理时无法挽回（arXiv:2604.16027，Olmo 3 三条 lineage）；recursive 训练合成数据会导致 model collapse、分布尾部消失（Shumailov 等，Nature 2024）。这条机制把"人类真实交互数据"变成愈发珍贵的资源，它直接为内核④"人是异质性来源"背书。

Hypernormal science is not just a soft “AI leans conservative” tendency; it has a hard mechanism written into the weights. The strongest causal anchor is Doshi & Hauser (Science Advances 2024) [R12]: give a writer LLM ideas and individual stories get more “creative,” yet the stories grow more similar to each other: they explicitly call it a “social dilemma” (better individually, narrower collectively). Homogenization is a group-level effect (Anderson et al. 2024, a 36-person experiment) [R13]: it comes not from individual fixation but from the LLM suggesting similar ideas to different users. Harsher still is cross-model homogeneity (“We’re Different, We’re the Same,” 2025) [R14]: controlling for structural variables, LLMs resemble each other far more than humans resemble each other: switching models does not save you. The decisive mechanism-level evidence: post-training diversity collapse is written into the weights and unrecoverable at inference (arXiv:2604.16027, three Olmo 3 lineages); recursively training on synthetic data causes model collapse with the distribution tails vanishing (Shumailov et al., Nature 2024). This mechanism makes “real human interaction data” an ever more precious resource: directly endorsing kernel ④’s “humans as the source of heterogeneity.”

真实交互数据，因此成了愈发珍贵的反同质化资源。这条机制有一个常被忽略的推论：既然 diversity collapse 写在权重里、且 recursive 训练合成数据会让分布尾部消失（model collapse），那么未被 AI 中介过的、真实的人类交互数据就成了愈发稀缺、愈发珍贵的资源，它是分布尾部、是异质性的最后蓄水池。这对研究组织有直接的操作含义：当所有人都在用同几个模型生成假设、写综述、做评审时，那些仍由人独立产生、未被模型均值拉平的判断与观察，恰恰是组织最该刻意保存、而非急于"用 AI 提效"掉的东西。它也回连内核④"人是异质性来源"：人之所以不可替代，不是因为人比 AI 聪明，而是因为人群携带着 AI 分布尾部已经丢失的多样性。守住这份多样性，需要刻意施力，在流程里留出"不经 AI 中介"的判断节点，在数据上珍惜真实人类信号，在激励上奖励偏离均值的探索。这三条合起来，就是抵抗 hypernormal 的组织级动作。

Real interaction data thus becomes an ever more precious anti-homogenization resource. This mechanism has an often-overlooked corollary: since diversity collapse is written into the weights, and recursively training on synthetic data makes the distribution tails vanish (model collapse), then real human interaction data not mediated by AI becomes an ever scarcer, ever more precious resource. It is the distribution tail, the last reservoir of heterogeneity. This has a direct operational implication for research organizations: when everyone generates hypotheses, writes reviews, and reviews with the same few models, the judgments and observations still produced independently by humans, not flattened to the model mean, are exactly what the organization should deliberately preserve rather than rush to “make efficient with AI.” It also wires back to kernel ④’s “humans as the source of heterogeneity”: humans are irreplaceable not because they are smarter than AI but because the human population carries the diversity AI’s distribution tail has already lost. Holding this diversity takes deliberate force: keeping “non-AI-mediated” judgment nodes in the workflow, treasuring genuine human signal in the data, rewarding off-mean exploration in incentives. Together, these three are the organization-level acts that resist hypernormal.

但它是默认引力，不是铁律。诚实地把反向证据摆上：同质化依任务/prompt/暴露方式而变，在高暴露的动态实验里集体多样性反而能升。所以正确的命题表述不是"AI 必然让科学同质"，而是"AI 默认把研究拉向均值，须刻意施力才能偏离"（regression to a domain prototype）。开放式/QD 算法（novelty-search、MAP-Elites、POET）证明：只要放弃单一目标函数，机器也能产异质。这把命题从"异质性只能来自人"收紧成更稳的版本：异质性的敌人是单一目标的过度优化，不是机器本身：人定义"什么值得不同"，机器在那个定义下产生多样。这条限定既守住人的角色，又抗住"AI 终将学会创意"这个证伪。

But it is a default gravity, not an iron law. Put the counter-evidence on the table honestly: homogenization varies by task / prompt / exposure mode, and in high-exposure dynamic experiments collective diversity can even rise. So the correct statement is not “AI inevitably homogenizes science” but “AI by default pulls research toward the mean and needs deliberate force to depart” (regression to a domain prototype). Open-ended / QD algorithms (novelty-search, MAP-Elites, POET) prove that machines too can produce heterogeneity: as long as the single objective function is abandoned. This tightens the thesis from “heterogeneity can only come from humans” into a sturdier version: the enemy of heterogeneity is the over-optimization of a single objective, not the machine itself: humans define “what is worth differing on,” the machine generates diversity under that definition. This qualifier both holds the human’s role and withstands the falsifier “AI will eventually learn creativity.”

先说清为什么需要一道天平。心理学那批写进教科书的经典结果，大规模复现时只有约三成能重来一遍[R1]——那还是人一篇篇亲手做出来的。截至本版（2026-07），AI 一天能生成上千条同样”看着可信”的主张：你连人写的都核不过来，凭什么信机器批量产的？所以每条主张进门前都得先分诊，而分诊的第一刀，是别把两件性质不同的事搅成一个总分。

First, why a ledger is needed at all. Of the textbook-classic results in psychology, only about a third replicated when tested at scale[R1] — and those were made by humans, one paper at a time. As of this edition (2026-07), AI can generate a thousand equally “credible-looking” claims a day: if you cannot even keep up with what humans wrote, why trust what a machine mass-produces? So every claim must be triaged on the way in, and the first cut of triage is not to stir two things of different nature into one aggregate score.

证据强度可补，框架距离要判——两条轴管两种动作

Evidence strength can be supplemented, paradigm distance must be judged: two axes for two acts

两条轴之所以不能合并，深层原因是它们对应两种性质完全不同的动作。"证据强度"这条轴是可补的、可机检的：一条主张证据弱，处置很清楚——去补证据（多跑复现、找原始数据、查证据链是否完整），这是个有标准答案、可外包给生成与图谱规则的动作。"框架距离"这条轴则要人来判，且没有可补一说：一条主张离现行框架远，这件事本身不是缺陷、也不是优点，它只是一个需要人来定夺的信号——人要判它是框架外的噪声，还是一次重构。

把两条轴合成一个"可信分"，等于把"可补的执行动作"和"不可外包的判断动作"搅成一锅，结果是两种动作都做不好：该补证据的没去补（因为分数已经替它下了结论），该人判的被自动判了（因为分数把"离框架远"直接折算成低可信）。分开记的全部意义，就是让每条轴触发它对应的那种正确动作——证据弱→补证据（左动作），离框架远→人来判（右动作）。这正好是 RES 10 那条"能写可机检验收标准的归左、写不出的归右"判据在单条主张层面的应用。

The deep reason the two axes cannot be merged is that they correspond to two acts of completely different nature. The “evidence strength” axis is supplementable and machine-checkable: if a claim’s evidence is weak, the disposition is clear: go get more evidence (run replications, find raw data, check the evidence chain’s completeness), an act with a right answer, outsourceable to generation and graph rules. The “paradigm distance” axis must be judged by a human, with no “supplementing”: a claim being far from the established paradigm is neither a defect nor a merit in itself, only a signal requiring constitutive judgment — a human must judge whether it is out-of-paradigm noise or a paradigm-level reframing.

Merging the two into one “credibility score” stirs “a supplementable execution act” and “an un-outsourceable judgment act” into one pot, and then both are done badly: the evidence that should be sought is not sought (the score already concluded for it), and what a human should judge is auto-judged (the score converts paradigm-distance straight into low credibility). The whole point of booking them separately is to let each axis trigger its corresponding correct act: weak evidence → seek evidence (the left act), far from paradigm → a human judges (the right act). This is exactly RES 10’s “machine-checkable criterion goes left, otherwise right” test applied at the single-claim level.

为什么"证据弱×框架远"这一格决定整台仪器的价值

Why the “weak × far” cell decides the whole instrument’s value

四象限里有三格是直觉的：证据强×框架内→入库；证据弱×框架内→当噪声；证据强×框架远→人重点看。真正考验一个研究系统成色的，是第四格：证据弱×框架远。直觉和单一可信分都会把它判死："远离现行框架且证据不足，删除。"但科学史上几乎每一次范式转移，诞生时都正落在这一格：爱因斯坦 1905 年的狭义相对论与洛伦兹的以太收缩，起初都只是拟合同一批数据，爱因斯坦的版本既"离框架远"又在当时缺乏决定性实验证据。达尔文的自然选择，核心机制（泛生论 gemmules）后来被证明是错的，但想法本身因为有用而存活。如果当年有一台只算可信分的系统，它会把这些都判为"证据弱×框架远→删"。

所以这一格的正确处置是第三种动作，不是二元的"信/不信"：挂起，并定向去找那个能区分"它是噪声"还是"它是重构"的关键证据。一台仪器值不值得用，全看它对这一格的处置，把它做成"删"，它就是 hypernormal science 的自动扼杀器；把它做成"挂起+定向取证"，它才是真正能接住范式转移的天平。

Three of the four quadrants are intuitive: strong × in-paradigm → integrate; weak × in-paradigm → noise; strong × far → human focuses. What truly tests a research system’s mettle is the fourth: weak evidence × far from paradigm. Intuition and a single credibility score both sentence it to death: “absurd and unsupported, delete.” Yet nearly every paradigm shift in the history of science sat exactly in this cell at birth: Einstein’s 1905 special relativity and Lorentz’s ether contraction both merely fit the same data at first, and Einstein’s version was both “far from paradigm” and lacking decisive experimental evidence at the time. Darwin’s natural selection: its core mechanism (pangenesis, gemmules) later proved wrong, yet the idea survived because it was useful. Had a credibility-score-only system existed then, it would have judged all of these “weak × far → delete.”

So the correct disposition for this cell is not the binary “believe / disbelieve” but a third act: suspend, and go find the decisive evidence that separates “it is noise” from “it is a reframing.” Whether an instrument is worth using turns entirely on its handling of this cell — make it “delete” and it is hypernormal science’s automatic strangler; make it “suspend + targeted evidence-seeking” and it is a ledger that can actually catch a paradigm shift.

天平不是给主张打分，是给"该投多少人类带宽"排序。天平最容易被误用成"给每条主张算一个可信分、按分排序"。这恰恰是它要避免的。它的产物不是分数，是处置（disposition）：对每条主张，回答"接下来该怎么处理它"，而不是"它有多可信"。四象限给的就是四种处置，而不是四档分数：证据强×框架内→直接入库整合（人不必看）；证据弱×框架内→当噪声存疑（人不必看）；证据强×框架远→人来判它是不是重构（值得看）；证据弱×框架远→挂起、定向去找区分证据（最值得看）。这套设计的目的，是把人的稀缺带宽从"逐篇精读"里解放出来，只投到真正吃紧的两格。当生成每小时上千条，逐篇读是物理上不可能的；天平的价值正在于它替你做了"哪些根本不必人看"的分流。

The ledger does not score claims; it ranks how much human bandwidth to spend. The ledger is most easily misused as “compute a credibility score per claim and sort by score.” That is exactly what it avoids. Its real product is not a score but a disposition: for each claim, answering “what to do with it next,” not “how credible it is.” The four quadrants give four dispositions, not four score tiers: strong × in-paradigm → integrate directly (no human needed); weak × in-paradigm → hold as noise (no human needed); strong × far-from-paradigm → a human judges whether it is a reframing (worth looking); weak × far-from-paradigm → suspend and go find discriminating evidence (most worth looking). The design’s purpose is to free the human’s scarce bandwidth from “reading each one closely” and spend it only on the two genuinely tight cells. When generation runs at thousands an hour, reading each is physically impossible; the ledger’s value is precisely that it does the “which ones need no human at all” triage for you.

AI 当协作者，还是当裁判？一道必须先划的界

AI as collaborator, or as judge? a line you must draw first

天平回答"这条主张可不可信"，但它背后压着一个更根本的问题：这一道判断，到底该不该让 AI 来当裁判？把 AI 当协作者（铺候选、查文献、跑实验、提反例）几乎总是安全的，它在执行端，输出还要过人的判断。把 AI 当裁判（让它定"哪个值得信、哪个值得发、哪个该资助"）则是另一回事，因为裁判位是价值与可信度的归属位，一旦交出去，RES 03 的结构性偏置会从"建议"升级成"判决"。这道界不能拍脑袋划，要看三件事：判据能不能机检（能 → AI 可裁）、判断是否价值负载（是 → 留人）、判错的代价可不可逆（不可逆 → 留人）。一个干净的判据：当这道判断的"对"只能诉诸"对谁、在哪个价值框架下"，AI 只能当协作者，不能当裁判。

The ledger answers “is this claim credible,” but underneath it presses a more fundamental question: should AI be the judge of this decision at all? Using AI as a collaborator (spreading candidates, searching literature, running experiments, raising counter-examples) is almost always safe: it sits on the execution side and its output still passes through human judgment. Using AI as a judge (letting it set “which to believe, which to publish, which to fund”) is another matter, because the judge’s seat is the seat of value and credibility ownership, and once handed over, RES 03’s structural bias is promoted from “suggestion” to “verdict.” This line cannot be drawn by gut; it depends on three things: whether the criterion is machine-checkable (yes → AI may judge), whether the judgment is value-laden (yes → keep human), whether the cost of a wrong call is reversible (irreversible → keep human). A clean test: when the “right” of this judgment can only appeal to “for whom, under which value frame,” AI can only be a collaborator, never a judge.

把界划错的两种症状：一种是把裁判位悄悄让出去：团队用"AI 评分高"代替"我读过、我担保"，于是没有人真正为某条主张的可信度负责，偏置在无人察觉中累积（这正是 RES 07"价值判断没有归属就被默认偏置替换"的运行态）。另一种是把协作位也死死攥住：出于不信任，连"铺候选、查文献"这种纯执行都不敢交给 AI，于是没把执行充裕化，团队还困在工时瓶颈里。两种都丢杠杆：前者交出了不该交的判断，后者守住了不必守的执行。正确的姿势是把这两个位子分开：执行位尽量交，裁判位审慎留；审慎与否，由上面那三条（可机检 / 价值负载 / 代价可逆）逐案判定。

Two symptoms of drawing the line wrong: one is quietly vacating the judge’s seat. The team substitutes “AI scored it high” for “I read it, I vouch”. So no one truly owns any claim’s credibility and bias accrues unnoticed (exactly RES 07’s “a value judgment with no owner gets replaced by the default bias,” at runtime). The other is clutching even the collaborator’s seat: out of distrust, the team will not even hand “spreading candidates, searching literature” (pure execution) to AI, so it never makes execution abundant and stays stuck at the hours bottleneck. Both forfeit leverage: the first hands away judgment that should not be handed away, the second holds execution that need not be held. The right posture is to separate the two seats: hand the execution seat freely, keep the judge’s seat with care, and the care is adjudicated case by case by the three above (machine-checkable / value-laden / cost reversible).

一条判据，同时回答三章的问题

One test answers three chapters’ questions at once

这张矩阵不只是一张分诊表，它给了整卷一条能落地的判据：写得出机器可核验的验收标准，归左——交给生成，或交给知识图谱的规则；写不出来、只能诉诸"这对谁成立、在哪套价值框架下成立"，归右——留给人。这条判据的用处在于，它同时回答了三章各自在问的问题，让你不必分别记三套规则。RES 04 问护栏该把关到哪：能机检的，交给证据库自动把关。RES 09 问天平该挂起什么：写不出验收标准、又离范式远的那些，正是该挂起去找证据的一格。RES 13 问研究环该把人放在哪一步：人只接住右格的节点。三章看起来各管一摊——护栏、天平、工作流——其实共用同一条判据。记住它，对每个动作你只需要问一句："这写得出可机检的验收标准吗？"答案会同时告诉你，这个动作该进护栏、进天平，还是进人的判断里。

This matrix is more than a triage table: it hands the whole volume one workable test: if you can write a machine-checkable acceptance criterion, the action goes left, to generation or to the graph’s own rules; if you cannot, if all you can say is “for whom, under which value frame,” it goes right, to a human. The point of this test is that it answers three chapters’ separate questions at once, so you don’t need three rule sets in your head. RES 04 asks where the guardrail should gate: whatever is machine-checkable, the evidence base gates automatically. RES 09 asks what the ledger should suspend: exactly the claims with no acceptance criterion, far from the paradigm; suspend them and go find evidence. RES 13 asks where the research loop should place the human: only at the right-cell nodes. The three chapters look like they each manage something different — guardrail, ledger, workflow — but they share one test. Remember it, and for every action you need ask only one question: can I write a machine-checkable acceptance criterion for this? The answer tells you at once whether the action belongs in the guardrail, the ledger, or human judgment.

怎么接起来：生成那边——检索、提假设、设计实验、跑分析——产出的一切，都落进 RES 04 那个可追溯证据库，每条主张挂着证据边。人这边不用去看全部，只从库里挑那几类必须自己判的：天平挂起的、离范式远的、价值负载的。这不是"先让它生成、我再审一遍"那种流程，是让证据库把框架内的东西自动挡住，省下的带宽全部留给换框架的判断。

How the two sides connect: the generation side — search, hypothesize, design experiments, run analysis — drops everything it produces into RES 04’s traceable evidence base, each claim carrying its evidence edges. The human side doesn’t review all of it; it only pulls the few kinds of node that must be judged by a person: the ones the ledger suspended, the ones far from the paradigm, the value-laden ones. This is letting the evidence base auto-gate everything in-paradigm, so the freed bandwidth goes entirely to paradigm-level judgment.

同一个词，左边右边说的是两件事

The same word means two different things on each side

这张矩阵最容易被读错成"有些动作归 AI，有些动作归人"。不是这样。真正承重的一点是：同一个动作词，落在左格和右格里，指的根本不是一回事。"提假设"在左格是"在已有框架里找下一个可检验的空白"——最近邻搜索，AI 干得好；在右格是"提一个旧框架里根本立不住的假设"——换描述层级，这在 AI 训练分布之外。"分析"在左格是"对预先定好的变量做拟合"——AI Feynman 重发现已知方程就是这个；在右格是"判断现在用的这套变量本身对不对"——Farr 把空气质量换成水传播微生物，就是这个。所以分诊切的不是"动作类型"，切的是"这一刀下去，落在能机检的一侧，还是只能人定的一侧"。一个动作词常常横跨两格，得先把它劈成两半，才知道哪半交给生成、哪半接住给人。

This matrix is most easily misread as “some actions go to AI, some stay with humans.” That’s not it. What’s actually load-bearing: the same action word means a different thing depending on which cell it lands in. “Hypothesize” in the left cell means finding the next checkable gap inside a frame you already have: nearest-neighbor search, AI is good at this. In the right cell it means posing a hypothesis the old frame couldn’t even hold: changing the level of description, which sits outside AI’s training distribution. “Analyze” in the left cell means fitting predefined variables: AI Feynman re-discovering known equations is exactly this. In the right cell it means judging whether the variables themselves are wrong: Farr swapping air quality for waterborne microbes is exactly this. So the cut triage makes isn’t by action type; it’s by where a given cut lands: the machine-checkable side, or the side only a human can call. A single action word usually straddles both cells; you have to split it before you know which half goes to generation and which half a human has to catch.

这张矩阵真正防的，是两个方向相反、却一样致命的错。第一种：把该换框架的动作硬塞进左格。典型说法是"让 AI 决定研究该往哪个方向走"——把"选方向"这个最该由人定的判断，当成可以生成的执行交出去了。后果是 RES 06 的价值判断、RES 12 的方向选择被默认偏置接管，组织在所有指标都绿的情况下越走越窄。第二种：把该交出去的框架内动作死死攥在右格。典型说法是人还在手动追引文、手动扫参数、手动做文献综述——不是因为不该自动化，是出于不信任或者习惯，守着早就能充裕化的执行不放。后果是执行没被充裕，团队照样卡在工时瓶颈上，②省下的时间也没能投回③④。这两种错刚好对称：前一种交出了不该交的东西，后一种守住了不必守的东西；前一种丢的是方向，后一种丢的是杠杆。矩阵的用处，就是给每个动作一条清楚的归格判据，让这两种错都无处可藏。

What this matrix really guards against is two opposite yet equally fatal mistakes. First: forcing a paradigm-level action into the left cell. The classic version sounds like “let AI decide which direction the research should go”: handing away “choosing direction,” the judgment that should sit most firmly with a human, as if it were generatable execution. The result: RES 06’s value judgment and RES 12’s direction-setting get captured by default bias, and the organization narrows while every dashboard reads green. Second: clutching an in-paradigm action in the right cell that should have been handed over. The classic version is people still tracing citations by hand, sweeping parameters by hand, writing literature summaries by hand — not because it shouldn’t be automated, but out of distrust or habit, holding on to execution that’s long been ready to become abundant. The result: execution never gets made abundant, the team stays stuck at the same hours bottleneck, and the time ② would have freed never gets reinvested into ③④. The two errors are exact mirrors: the first hands away what it shouldn’t, the second holds on to what it needn’t; the first costs you direction, the second costs you leverage. What the matrix is for is giving every action one clear test for where it belongs, so neither mistake has anywhere to hide.

schema 越自洽，盲区越难被看见——这是它最危险的地方

The more self-consistent the schema, the harder its blind spot is to see: its most dangerous trait

护栏变笼子这件事有一个反直觉的地方：schema 做得越完整、越自洽，它的盲区反而越难被人发现。一个漏洞百出的图谱经常报错、经常有东西塞不进去，而"塞不进去"恰好是在提醒你："框架可能错了。"可一个设计精良、覆盖全面、查得又快的图谱，会让所有观测都顺滑地落进已有的节点——它从不报错，因为它把每一次观测都成功解释进了现有的变量空间。问题是，"成功解释进现有空间"跟"解释对了"，是两件事。

Farr 的霍乱图就是这样一个"完美自洽"的笼子：每一例死亡都被归因到空气质量的某个梯度，模型拟合得很好，预测也不算差，整套系统运转得毫无破绽——正因为如此，没人会去怀疑"空气质量"这个描述层级本身就错了。这也是为什么 RES 08 那条覆盖广度的反指标必须当成例行体检来跑：一个系统的每个局部指标都健康，却系统性地看不见某一类东西，唯一能发现盲区的办法，是主动去问"我们的图谱在结构上不可能表示什么"——而不是等它报错，因为它永远不会报错。

There’s something counter-intuitive about how a guardrail turns into a cage: the more complete and self-consistent the schema, the harder its blind spot is to spot. A crude, leaky graph errors constantly, has things that just won’t fit, and “won’t fit” is exactly the signal telling you the frame might be wrong. But a well-designed, fully covering, fast-querying graph lets every observation slide smoothly into an existing node. It never errors, because it has successfully explained every observation into the variable space it already has. The catch: “successfully explained into the existing space” and “explained correctly” are two different things.

Farr’s cholera map is exactly this kind of “perfectly self-consistent” cage: every death got attributed to some gradient of air quality, the model fit well, the predictions weren’t bad, the whole system ran without a single visible flaw, and precisely because of that, no one thought to question whether “air quality” was the wrong level of description to begin with. This is why RES 08’s breadth counter-indicator has to run as a routine checkup: when every local metric on a system reads healthy yet it’s systematically blind to a whole class of things, the only way to find the blind spot is to actively ask what the graph is structurally incapable of representing — not wait for it to error, because it never will.

机理 · 护栏为何会变成笼子：说穿了很直白：知识图谱要把"主张—证据—来源"结构化，前提是这些节点和边的类型已经被定义好了。schema 一旦定下来，生成就只能在这个既有的变量空间里排列组合——这正是它高效的原因，也正是它的盲区所在。Farr 的数据是围着 miasma（瘴气）组织的，schema 里压根没有"病原体"这个节点，于是再聪明的查询也问不出一个 schema 外的变量。AI 在这上面加速，只会把"空气质量"这个错误层级钻得更深、更自洽、更难被推翻——护栏把框架内的可信度，筑成了一堵连换框架都翻不过去的墙。

Mechanism · why a guardrail becomes a cage: the mechanism is plain: a knowledge graph structures claim–evidence–source on the premise that its node and edge types are already defined. Once the schema is set, generation can only recombine within that existing variable space: exactly why it’s efficient, and exactly where its blind spot sits. Farr’s data was organized around miasma; the schema had no “pathogen” node at all, so no query, however clever, could ask for a variable the schema didn’t have. AI accelerating on top of this only drills the wrong “air quality” level deeper, more self-consistent, harder to overturn. The guardrail builds in-paradigm credibility into a wall that even a reframing can’t climb.

• 留"换变量"的口子：schema 不能只读不写，给"提出新节点类型/新关系类型"留一条人发起的通道，并定期复审 schema 本身是否还是对的描述层级。
• Leave a “change the variable” door: the schema must be writable, not read-only; keep a human-initiated channel to propose new node/edge types, and periodically review whether the schema is still the right level of description.
• 把"未被任何节点解释的异常"显形：不要让无法入图的观测被默默丢弃；恰是这些落在 schema 外的残差，往往是换地图的入口（霍乱在 schema 外的死亡聚集）。
• Surface “anomalies no node explains”: do not let observations that fail to enter the graph be silently dropped; these out-of-schema residuals are often the doorway to a new map (the cholera death-clusters that the schema could not place).
• 给护栏配一个"反护栏"复盘：定期问"我们的图谱在系统性地看不见什么"，把 RES 08 的覆盖广度反指标接进来当例行体检。
• Pair the guardrail with an “anti-guardrail” review: regularly ask “what is our graph systematically blind to”, wiring RES 08’s breadth counter-indicator in as a routine checkup.

残差才是入口：把"图谱解释不了的"专门留住。一个健康研究系统跟一个完美笼子之间的全部区别，就在于怎么对待残差——那些任何现有节点都解释不了、进不了图的观测。笼子的本能是把残差当噪声扫掉，因为它们弄乱了图的整洁，拉低了"主张落进可追溯链的比例"这个看着很正的指标。但科学史一次次证明：换地图的入口，几乎总是从残差进的。霍乱在 miasma schema 之外的死亡聚集是残差；迈克尔逊-莫雷"测不到以太风"是残差；黑体辐射在经典理论下的"紫外灾难"也是残差。每一个后来引发范式转移的东西，最初都长得像"现有框架解释不了、看上去是误差"的样子。

所以护栏设计里最反直觉、也最承重的一条，是专门给残差建一个不会被自动清扫的收容区，定期让人去复审：这些残差是测量误差，还是在合起来指向一个 schema 外的新变量？这正是 RES 05 那道"揭示盲区 → 造新 eval"的复现之墙，落到知识图谱上的具体样子，也是 RES 11 给护栏留的"换变量口子"真正运行起来的样子。一个系统每清扫掉一批残差，可能就在清掉它下一次范式转移的种子。

The residual is the doorway: deliberately keep what the graph cannot explain. The entire difference between a healthy research system and a perfect cage comes down to how each treats the residual: observations that no existing node can explain, that fail to enter the graph at all. A cage’s instinct is to sweep residuals away as noise, because they mess up the graph’s tidiness and drag down that virtuous-looking metric, “share of claims landing in the traceable chain.” But the history of science says otherwise, again and again: the doorway to a new map is almost always the residual. The cholera death-clusters outside the miasma schema were a residual. Michelson–Morley’s undetected ether wind was a residual. Black-body radiation’s ultraviolet catastrophe under classical theory was a residual. Everything that later triggered a paradigm shift started out looking like error the current frame couldn’t explain.

So the most counter-intuitive, most load-bearing rule in guardrail design is this: build a holding area for residuals that the auto-sweep doesn’t touch, and have a human review it on a schedule — are these residuals measurement error, or are they collectively pointing at a variable outside the schema? This is exactly RES 05’s wall (reveal a blind spot, build a new eval) landing on the knowledge graph; it’s also RES 11’s change-the-variable door, actually running. Every batch of residuals a system sweeps away might be the seed of its next paradigm shift.

护栏跟笼子，是同一件东西的两种用法

The guardrail and the cage are the same thing, used two ways

RES 04 说知识图谱是好护栏，这一节又说它会变笼子——听着矛盾，其实是同一套机制在两种使用次序下的两副面孔。schema 服务于生成、并且对人保持可写的时候，它是护栏：让海量生成落进可追溯的结构里，冲突能显形，工作能被复现。schema 反过来定义了什么算"研究"、并且只读不可写的时候，它是笼子：生成只能在既有变量空间里打转，schema 外的观测被悄悄丢掉。决定它是护栏还是笼子的，不是图谱做得精不精巧——恰恰相反，图谱越精巧、越自洽，它当笼子时就锁得越死。Farr 的霍乱图正是一个"完美的笼子"：结构完整、数据可追溯、查询又快，唯独没有"病原体"这个节点——于是它把整个研究共同体钉在"空气质量"这个错误的层级上，钉得严丝合缝。

RES 04 says the knowledge graph is a good guardrail; this section says it turns into a cage. That sounds like a contradiction, but it’s one mechanism wearing two faces depending on the order you use it in. When the schema serves generation and stays writable by humans, it’s a guardrail: mass generation lands in a traceable structure, conflicts surface, work can be replicated. When the schema instead defines what counts as “research”, and is read-only, it’s a cage: generation can only circle within the existing variable space, and out-of-schema observations quietly get dropped. What decides guardrail or cage isn’t how sophisticated the graph is; quite the opposite: the more sophisticated and self-consistent it is, the tighter it locks as a cage. Farr’s cholera map is exactly a “perfect cage”: structurally complete, data traceable, queries fast, missing only a “pathogen” node, and so it pinned the whole research community to the wrong “air quality” layer, pinned it flawlessly.

第一次能对"科学制度"本身做对照实验

For the first time, a controlled experiment on scientific institutions themselves

元科学一直是一门没有实验台的学问：你没法把同一个科学共同体复制十份、每份换一套激励，再看哪份更容易出颠覆性的东西——现实里只有一个共同体，跑得又慢，变量又多。AI 科学家可能第一次给元科学一个模式生物：让一群 AI agent 在不同研究条件下并行跑——这一群按产出考核，那一群按新颖考核；这一群层级森严，那一群扁平自治；这一群只准待在框架内，那一群被明着鼓励换框架——然后细看哪种条件下涌现更多概念重组。这是历史上头一回，"什么样的组织结构孕育颠覆"从一个只能靠案例回答的问题（贝尔实验室、施乐 PARC、剑桥 LMB 这些"小团队 + 体制保护"的轶事），变成一个可以做对照实验的问题。

它和 AlphaZero 独立自博弈下走出人类从没下过的原创棋（21.Bg5）是同一个道理：把系统从现有框架的训练数据里放出来，让它在受保护的环境里自由探索。〔这是 Asimov 一篇观点文的推演，证据级 Ⅴ；模式生物这个说法尚未被大规模实证，标为前沿命题〕[a projection from an Asimov Press essay, grade Ⅴ; the model organism itself has not been demonstrated at scale, flagged as a frontier claim]

Meta-science has long been a discipline with no bench: you can’t copy one scientific community ten times, hand each a different incentive structure, and see which breeds more disruption. In reality there is only one community, it runs too slowly, and it has too many variables. The AI scientist may give meta-science its first model organism: run populations of AI agents in parallel under different research conditions: this group scored on output, that one on novelty; this group steeply hierarchical, that one flat and autonomous; this group confined to the paradigm, that one openly encouraged to switch frames; then measure closely which conditions yield more conceptual recombination. For the first time, “what organizational structure breeds disruption” turns from a question you can only answer with anecdotes (Bell Labs, Xerox PARC, the Cambridge LMB — small teams under institutional protection) into one you can run a controlled experiment on.

It is the same logic as AlphaZero playing an original move no human ever had (21.Bg5) under independent self-play: free the system from the existing paradigm’s training data and let it explore in a protected space.

把两条退路并排放，不是让你二选一。这一卷特意画出两条退路而不是一条，因为只画"退到价值——哪个真相值得"会漏掉一半。第一条是 RES 06 的价值论退路：当提问被充裕，人退到"定哪个真相重要"。第二条是这一节的方法论退路：当跑实验也近乎免费，人退到"设计科学本身——该跑哪个实验、什么算证据、什么样的研究设计真能把假设区分开"。这两条不冲突，是正交的：一条定方向（往哪走值得），一条定制度（什么样的科学机器能让值得的方向被生出来）。一个只退到价值、没退到方法的研究者，知道该追什么真相，却没有一套能生出颠覆的科学制度；一个只退到方法、没退到价值的研究者，有一台精良的方法机器，却不知道该指向哪里。两条都退了，才补全内核④"人回归意义"在研究这一面的全貌：意义既是方向的意义，也是制度的意义。

Placing the two retreat lines side by side is not offering a choice between them. This volume deliberately draws two lines, not one, because drawing only “retreat to value: which truth is worth it” misses half the picture. The first is RES 06’s axiological retreat: as questioning turns abundant, the human retreats to setting which truth matters. The second is this section’s methodological retreat: as running experiments turns near-free too, the human retreats to designing science itself: which experiment to run, what counts as evidence, what research design actually discriminates between hypotheses. The two don’t conflict; they are orthogonal. One sets direction: where it is worth going. The other sets institution: what scientific machine lets worthy directions get born in the first place. A researcher who retreats only to value and not to method knows which truth to chase but has no institution that breeds disruption; one who retreats only to method and not to value has a fine method-machine but no idea where to point it. Retreating on both is what completes kernel ④’s “humans return to meaning” on the research side: meaning here is both the meaning of direction and the meaning of institution.

我们还不知道"什么规则让框架更优"——所以跑得快不等于走得对

We still don’t know what makes a paradigm better — so faster isn’t automatically righter

元科学之所以变得值钱，根子在一个不太舒服的事实：直到今天，我们都没有一条完备的判据，能说清"一个框架为什么比另一个更好"。历史上提出过的候选都不完备。"简单"是一个——符号回归的 AI Feynman 把 100 条费曼方程全找了出来（旧软件只找到 71 条），最小描述长度原理也把"更简单大概率更对"部分地形式化了；但简单从来不保证真——J.J. Thomson 的葡萄干布丁原子模型又简单又优雅，却整个错了。

"类比"是另一个候选——好的框架常常善于在不相干的领域之间打比方（爱因斯坦借光的图像、达尔文借 Lyell 的地质学和 Malthus 的经济学）；但类比同样会骗人：把表面像、实际不相干的两件事当成一回事。既然连"什么让框架更优"都还没有判据，那"把执行加速一万倍"就不会自动生出更好的框架——它只会在当前框架里跑得更快。这正是"加速不等于进步"最深的根据，也是元科学（研究"什么样的科学制度能生出更优的框架"）从一件奢侈品变成一件必需品的原因。互联网就是前车之鉴：它让知识变得可搜索，却因为职业激励这类结构性低效，没能在规模上带来更快的科学，反而让引用变窄了。

Meta-science appreciates for an uncomfortable reason: to this day we don’t have a complete criterion for why one paradigm is better than another. Every historical candidate is incomplete. Simplicity is one — symbolic regression’s AI Feynman recovered all 100 Feynman equations (an older tool got only 71), and the minimum-description-length principle partly formalizes “simpler is more likely right.” But simplicity guarantees nothing: J.J. Thomson’s plum-pudding atom was simple and elegant and completely wrong.

Analogy is another candidate — good paradigms are often good at drawing analogies across unrelated domains (Einstein borrowing the image of light, Darwin borrowing Lyell’s geology and Malthus’s economics). But analogy just as easily deceives: it can pass off two things that look alike but aren’t related as the same thing. Since we don’t even have a criterion for what makes a paradigm better, “accelerating execution ten-thousand-fold” won’t automatically produce better paradigms; it will just run the current one faster. This is the deepest reason “acceleration isn’t progress” holds, and why meta-science — studying what scientific institution breeds better paradigms — moves from a luxury to a necessity. The internet is the cautionary tale here: it made knowledge searchable, yet because of structural inefficiencies like career incentives, it didn’t bring faster science at scale; it narrowed citation instead.

自主性阶梯：选方向是最后一阶，也是最难的一阶

The autonomy ladder: agenda selection is the last rung, and the hardest

把"元科学在升值"这件事画得更准，要借 Anthropic 2026 年那道自主性阶梯。它把研究 agent 的能力从左到右排成一道梯子：执行良定义的实验（最左，已经能匹敌人类）→ 设计实验 → 综合发现 → 选研究议程（最右，最难）。这道梯子跟 RES 02 的可验证性梯度是同一个形状——越往右，可机检的对错代理越薄，剩下的越是"无最近邻可循"、只能由人来定的判断。下面这张图把它和"生成充裕 vs 判断稀缺"叠在一起看：横轴是自主性从执行走到方向，竖轴是这项能力当前的充裕度，曲线显示：生成那一侧（执行、设计实验）已经逼近饱和，判断那一侧（综合、选方向）还陡峭地稀缺着。这正是命题的形状：能力沿着梯子往上爬，但最右边那一阶——选方向——的稀缺，不会因为模型变强而自动消失。

To draw “meta-science appreciates” more precisely, borrow Anthropic’s 2026 ladder of autonomy. It arranges a research agent’s capability left to right: execute a well-defined experiment (leftmost, already matching humans) → design experiments → synthesize findings → select the research agenda (rightmost, hardest). This ladder has the same shape as RES 02’s verifiability gradient: the further right you go, the thinner the machine-checkable proxy for right gets, and the more what’s left is constitutive judgment with no neighbor to follow. The figure below overlays it with “generation abundant vs judgment scarce”: the x-axis runs from execution to direction, the y-axis is that capability’s current abundance, and the curve shows the generation side (execution, experiment design) nearing saturation while the judgment side (synthesis, direction) stays steeply scarce. That’s the shape of the thesis: capability climbs the ladder, but the rightmost rung, agenda selection, doesn’t lose its scarcity just because models get stronger.

这五个装置为什么会一起坏——它们共用一个刚被打穿的前提

Why all five break together: they share one assumption that just got punctured

单看这五件事，像五个互不相关的毛病；放到一起看，才发现它们共用同一个前提——而这个前提刚刚被 AI 打穿了。这个前提是：产出量是诚实的信号。写出一篇论文、攒够一批引用、跑完一个项目，过去都贵到足以证明背后有真实的智力投入，所以拿"量"当"质"的代理，误差还能接受。整套建制的代理链都搭在这块地基上：发表数代理生产力，引用数代理影响力，h 指数把两者打包代理"学者价值"，影响因子代理"期刊质量"，经费规模代理"研究重要性"。每一环都是用一个便宜、可数的量，去代理一个昂贵、难判的质。执行贵的时候，这条代理链的误差有限，因为刷不动——你没法低成本地伪造一百篇看起来像样的论文。

AI 恰好把这件事变便宜了。代理链最怕的不是有人作弊，是作弊的边际成本趋于零：一旦"看起来像样的产出"能近乎免费地批量生成，所有靠"量"当代理的指标会同时失去鉴别力——这是 Goodhart 定律的极端形态（一个度量一旦成为目标，就不再是好度量），只不过 AI 把"成为目标后失效"这件事，从原本的数年压缩到了数周。所以下面这五件不是五个孤立的故障，是同一块地基塌了之后，盖在上面的五个房间一起裂。

Seen one at a time, these five look like five unrelated ailments. Put them side by side and you find they share one assumption, and that assumption has just been punctured by AI. The assumption: output volume is an honest signal. Writing a paper, accruing a batch of citations, finishing a project used to be expensive enough to prove real intellectual work behind it, so quantity standing in for quality carried tolerable error. The whole establishment’s proxy chain sits on this bedrock: publication count proxies productivity, citation count proxies influence, the h-index bundles both to proxy a scholar’s worth, the impact factor proxies a journal’s quality, grant size proxies a study’s importance. Every link is a cheap, countable quantity standing in for an expensive, hard-to-judge quality. When execution was expensive, this chain’s error was bounded, because you couldn’t game it. You couldn’t cheaply fake a hundred plausible-looking papers.

AI is precisely what made that cheap. What a proxy chain fears most isn’t a cheater; it’s the marginal cost of cheating falling to zero. Once “plausible-looking output” can be mass-produced near-free, every metric that proxies via quantity loses its power to discriminate, all at once. That’s Goodhart’s law in its most extreme form (a measure that becomes a target stops being a good measure), except AI compresses the timeline from years to weeks. So the five below aren’t five isolated failures; they’re five rooms cracking together after the one foundation under them gave way.

① "不发表就出局" + h 指数：拿产量当价值，正中高速生成器的下怀

① “Publish or perish” + the h-index: counting output as worth — exactly what a fast generator wants

装置原意："publish or perish"和 h 指数〔Hirsch 2005，PNAS，R21，证据级 Ⅱ〕都拿"可数的产出"当价值的代理：评委读不完每个人的全部工作，就假定发得多、被引得多≈又高产又有影响力。执行昂贵的年代这个误差还能忍，因为产量本身就是努力的证据。失效机理：h 指数把"高被引论文的数量"压成一个数字，而这个数字有两条都能刷的路——多发（分母）、多被引（分子）。Goodhart 定律〔Strathern 1997 对 Goodhart 的转述，R22，证据级 Ⅳ〕讲得很干脆：这个数字一旦成为目标，学者就会去优化数字本身，而不是它原本想代理的东西。切香肠式发表（把一个研究拆成好几篇最小可发表单元）、引用环（互相引用刷分）、自引，都是对着 h 指数做的理性优化。

AI 怎么把它推到极端：过去刷 h 指数受限于"写论文很贵"；现在 LLM 能近免费地批量产出格式齐全、看起来像样的稿件。当"看起来像研究的东西"能批量生成，任何拿产量当代理的指标都会瞬间失去鉴别力——它本来是用来挡"没干活的人"，现在反倒最方便"用机器刷量的人"。这正是 RES 03 那条"AI 拿与既有分布的距离当唯一代理"在激励层的镜像：指标越是奖励"可数的像样产出"，高速生成器就越是它的最优解。

What the device meant: “publish or perish” and the h-index 〔Hirsch 2005, PNAS, R21, grade Ⅱ〕 both use countable output as a stand-in for worth: a committee can’t read everyone’s complete body of work, so it assumes prolific and highly cited roughly equals productive and influential. When execution was expensive this error was tolerable, because volume itself was evidence of effort. The failure mechanism: the h-index compresses “how many highly cited papers” into one number, and that number has two gameable routes: publish more (the denominator), get cited more (the numerator). Goodhart’s law 〔Strathern 1997’s restatement of Goodhart, R22, grade Ⅳ〕 puts it bluntly: once a number becomes a target, scholars optimize the number, not whatever it was meant to stand in for. Salami-slicing (splitting one study into several least-publishable units), citation rings (mutual citation for score), and self-citation are all rational responses to the h-index.

How AI pushes it to the extreme: gaming the h-index used to be capped by “writing a paper is expensive”; now an LLM mass-produces near-free manuscripts that look properly formatted and plausible. Once “things that look like research” can be batch-generated, any metric proxying via volume loses its discrimination instantly: built to keep out people who did no work, it now most conveniences people gaming volume with a machine. This is the incentive-side mirror of RES 03’s “AI uses distance from the existing distribution as its only proxy”: the more a metric rewards countable, plausible-looking output, the more a fast generator becomes its optimal solution.

② 影响因子：拿期刊均值代理单篇质量，把"追热点"焊进了激励

② The impact factor: proxying a single paper’s quality by a journal mean, welding “hype-chasing” into incentives

装置原意：期刊影响因子（Garfield 1955 提出，后成 JCR 商业指标，R23，证据级 Ⅳ）本来是给图书馆挑订阅用的——它是一份期刊近两年文章的平均被引数，从来没打算拿来评单篇论文或单个学者，Garfield 本人也多次警告过这种误用。失效机理：把"期刊均值"当"单篇质量"，是个统计学上的范畴错误：期刊被引分布极度长尾（少数文章贡献了绝大多数引用），拿均值去代理任意一篇的质量，误差大到没什么意义。但因为影响因子可数、可排序、跨学科能比，它被招聘、评职称、经费评审广泛拿来用，学者的理性反应就是"往高影响因子的期刊投"，而高影响因子期刊系统性地偏好新颖、热点、阳性结果：可靠但不性感的工作——复现、阴性结果、方法学订正——被结构性地挤了出去。这正是 RES 08 hypernormalization 在激励侧的来源：不是有人存心想做窄，是指标在奖励"热"，不是"对"。

AI 怎么放大它：生成一旦近免费，"追当前热点、批量产出符合高影响因子期刊口味的稿件"就成了一套可以自动化的策略。AI 最擅长的恰是拟合已有分布（RES 06），影响因子奖励的恰恰是"贴近当前热点分布"——两者一拍即合，把科学进一步推向"所有人都在追同一批热问题"的同质化深渊。可靠性和新颖性本来是两件事，影响因子把它们混成一个"高被引=好"的单一信号，而 AI 让追这个信号的成本趋于零。

What the device meant: the journal impact factor (proposed by Garfield in 1955, later a JCR commercial metric, R23, grade Ⅳ) was built for librarians choosing subscriptions — it’s the mean citation count of a journal’s articles over the prior two years, never meant to judge a single paper or scholar, a misuse Garfield himself warned against repeatedly. The failure mechanism: treating a journal’s mean as one paper’s quality is a statistical category error: journal citation distributions are extremely long-tailed (a handful of articles account for most citations), so a mean is a meaningless proxy for any given paper. But because the impact factor is countable, sortable, and comparable across fields, it got widely adopted in hiring, tenure, and grant review, so a scholar’s rational response is to submit to high-IF journals, which systematically prefer novelty, hype, and positive results. Reliable-but-unsexy work (replication, null results, methodological corrections) gets structurally squeezed out. This is where RES 08’s hypernormalization comes from on the incentive side: no one is trying to go narrow; the metric is rewarding “hot,” not “right.”

How AI amplifies it: once generation is near-free, “chase the current hot topic, mass-produce manuscripts to high-IF taste” becomes an automatable strategy. What AI is best at is fitting the existing distribution (RES 06), and what the impact factor rewards is exactly hugging the current hot distribution; the two click together, pushing science further into the pit where everyone chases the same hot questions. Reliability and novelty used to be two different things; the impact factor blends them into one “highly cited equals good” signal, and AI drives the cost of chasing that signal toward zero.

③ 同行评审：一道为"稿件稀缺"设计的串行闸，正被无限投稿淹没

③ Peer review: a serial gate designed for scarce manuscripts, now drowning in unbounded submissions

装置原意：同行评审是科学质量的承重墙，让领域同行在发表前把关，挡掉错误、夸大、不可靠的工作。它整套的吞吐假设是"稿件按人类写作的速度到达"、评审按人类阅读的速度处理，两者大致匹配（中位审稿周期常以月计〔Björk & Solomon，R24，证据级 Ⅱ〕，稿件稀缺时这还能接受）。失效机理：同行评审是一道串行闸：每篇稿子要占用 2–4 位领域专家各自几个小时，而合格评审人的总带宽是固定又稀缺的（还是无偿的）。这道闸的吞吐上限由人类专家的数量决定，不是由投稿量决定；投稿量一旦暴涨，闸不会变快，只会排出更长的队、评得更草率。

AI 怎么把它压垮：生成端能近免费地把投稿量翻十倍、百倍，评审端的人类专家带宽却一点没变——这正是 RES 05"剪刀差"在评审环节的具体爆发。更糟的是，AI 还能批量生成"看起来该认真审"的稿件，逼真到评审人必须真花时间才能辨真假，于是本就稀缺的评审带宽，大量被"识别 AI 垃圾"消耗掉了。指望同行评审能拦住一切，在投稿无限的世界里已经不现实；出路是把"判可信"从"逐篇串行精读"改成 RES 09 那种"按证据强度×框架距离分诊、人只投到吃紧的两格"的并行分流——这需要重构激励，不是让评审人加班。

What the device meant: peer review is science’s quality load-bearing wall: domain peers gatekeep before publication, blocking errors, exaggeration, unreliable work. Its whole throughput assumption is that manuscripts arrive at human writing speed and get processed at human reading speed, the two roughly matched (median review cycles often run in months 〔Björk & Solomon, R24, grade Ⅱ〕, tolerable when manuscripts were scarce). The failure mechanism: peer review is a serial gate: each manuscript ties up 2–4 domain experts for hours each, and the total bandwidth of qualified, unpaid reviewers is fixed and scarce. This gate’s throughput ceiling is set by the number of human experts, not by how many submissions come in; when submissions surge, the gate doesn’t speed up, it just produces longer queues and sloppier reviews.

How AI crushes it: the generation side can near-free multiply submissions tenfold, a hundredfold, while the review side’s human-expert bandwidth hasn’t moved at all — RES 05’s scissors gap erupting right at the review stage. Worse, AI can mass-produce manuscripts that look worth taking seriously, lifelike enough that a reviewer has to spend real time telling real from fake, so scarce review bandwidth gets eaten identifying AI slop. Expecting peer review to catch everything is no longer realistic once submissions are unbounded; the way out is shifting “judging credibility” from serial close-reading of every manuscript to RES 09’s parallel triage: sort by evidence strength times paradigm distance, spend the human only on the two tight cells. That takes reworking incentives, not asking reviewers to work overtime.

④ 经费周期：为"实验很贵"调的保守偏好，正好杀死现在最廉价的探索

④ The grant cycle: a conservatism tuned for expensive experiments now kills the cheapest exploration

装置原意：经费评审的保守偏好不是恶意，是理性的风险管理：一个实验要花数百万、跑数年，评委有责任把钱投给"大概率能成"的项目，要求充分的前期数据、清晰的可行性、跟既有文献连得上。失效机理：但这套机制把"跟现行框架连得上"焊成了硬门槛，系统性地偏向框架内的渐进工作，把换框架的重构——按定义离既有文献远、前期数据必然薄，见 RES 09 那个"证据弱×框架远"格——挡在了门外。这正是 RES 06/07 反复说的：新颖诞生时一定"看上去不靠谱"，而经费机制把"看上去不靠谱"直接判了死刑。March 的探索/利用框架〔R15，证据级 Ⅱ〕落到制度层就是这条：利用（渐进、可预测）在争资源的时候，总是赢过探索（冒险、可能颗粒无收）。

AI 怎么改变了这笔账：充裕化恰恰把探索的成本压下来了——很多过去要花数月、数十万才能试一试的想法，现在近免费就能先跑一轮验证。经费机制那条核心假设——"探索很贵，所以要保守"——正在因此失效：探索变便宜了，理性的探索配比理应往上调（RES 07），保守偏好反倒成了把最便宜的新颖机会挡在门外的结构性浪费。经费周期这个旋钮是为"实验很贵的年代"调的，AI 把实验变便宜了，旋钮却没跟着拧——结果就是制度在最该放手探索的时候，反而收得更紧。贝尔实验室、施乐 PARC、剑桥 LMB〔R19〕之所以高产，正是因为它们用制度性保护把这种保守偏好对冲掉了。

What the device meant: a grant committee’s conservatism isn’t malice, it’s rational risk management: when an experiment costs millions and takes years, reviewers have a duty to fund what’s likely to succeed, demanding solid preliminary data, clear feasibility, continuity with the existing literature. The failure mechanism: but this welds “continuity with the current paradigm” into a hard bar, systematically favoring incremental in-paradigm work while barring paradigm-level reframings — which, by definition, sit far from the existing literature and are necessarily thin on preliminary data (RES 09’s “weak evidence, far paradigm” cell). This is exactly what RES 06/07 keep saying: real novelty looks unreliable at birth, and the grant mechanism sentences “looks unreliable” to death outright. March’s explore/exploit frame 〔R15, grade Ⅱ〕 at the institutional level is just this: exploitation (incremental, predictable) always beats exploration (risky, possibly nothing) when the two compete for resources.

How AI changes the arithmetic: abundance is precisely what drives exploration’s cost down: many ideas that used to take months and a hundred thousand dollars to try can now be validated in a first pass, near-free. So the grant mechanism’s core assumption (exploration is expensive, so be conservative) is failing on its own terms: exploration got cheap, so the rational explore-share should move up (RES 07), and conservatism becomes structural waste that bars the cheapest novelty opportunities. Put differently, the grant cycle is a dial tuned for an era when experiments were expensive; AI made experiments cheap, and the dial never got retuned — so the institution tightens exactly when it should let exploration loose. Bell Labs, Xerox PARC, and the Cambridge LMB 〔R19〕 were prolific precisely because they used institutional protection to hedge against this conservatism.

⑤ PI/课题组金字塔 + "复现吃力不讨好"：把承重验证器留在没人愿意干的位置

⑤ The PI/lab pyramid + “replication is thankless”: leaving the load-bearing verifier where no one will do it

装置原意：PI（首席研究员）/课题组的金字塔结构，是为"训练+分工"设计的：资深 PI 定方向、拉经费、担署名责任，博士生博后做执行。这在执行昂贵的时候是高效的——执行是稀缺资源，集中在受训的年轻人手上，由经验把关方向，是合理的分工。同时这套激励把"原创新发现"放在金字塔顶端，把"复现别人的工作"放在没有奖励的位置：复现拿不到经费、发不了高影响因子、不算原创贡献——吃力不讨好，于是几乎没人做。失效机理：这就造成了一个致命错配：RES 00/13 反复论证，独立复现是把"研究环"和"高速生成器"分开的唯一承重验证器（Open Science Collaboration 2015：97 项显著结果仅 36% 复现，R1；Baker 2016：逾 70% 科学家复现他人失败，R2）。也就是说，整个科学最关键的那个质量动作，恰好被激励结构放在了没人愿意干的位置上。

AI 怎么把它从"能忍"变成"致命"：执行昂贵的时候这个错配还能忍——反正复现也贵。当生成端近免费地把待验证的主张翻上百倍，而验证端（复现）还困在"吃力不讨好、没人做"的激励洼地里，缺口就直接爆了，这正是 RES 05 剪刀差最尖锐的样子。讽刺的是，AI 本可以承担复现里的大部分执行——跑代码、重算、交叉核对数据——但只要激励结构还把复现放在没有奖励的位置，执行变便宜也没用：没人有动机去按那个按钮。这就是 RES 07"省下的产能不会自动变成 slack"落在复现上的样子：技术上能复现，不等于制度上有人真去复现。要修的不是技术，是激励——得把"担保可信"（复现、验证、整合）从金字塔底端那个没有奖励的位置，提到跟"原创发现"同等的奖励位上。这正是 RES 00"科学社区的价值从产生知识转向担保可信"落在制度层的意思。

What the device meant: the PI/lab pyramid is built for training plus division of labor: a senior PI sets direction, raises funding, bears authorship responsibility, while PhD students and postdocs execute. This was efficient when execution was expensive: the scarce resource concentrated in trainees, direction gatekept by experience, a reasonable division of labor. At the same time this incentive puts “an original new discovery” at the pyramid’s apex and “replicating someone else’s work” in a position with no reward at all: replication wins no grants, no high-IF publication, counts as no original contribution, and is thankless enough that almost no one does it. The failure mechanism: this creates a fatal mismatch. RES 00/13 argue repeatedly that independent replication is the one load-bearing verifier separating the research loop from a fast generator (Open Science Collaboration 2015: only 36% of 97 significant results replicated, R1; Baker 2016: over 70% of scientists failed to reproduce others’, R2). That is, science’s single most critical quality act sits exactly where the incentive structure ensures no one wants to do it.

How AI turns it from bearable to fatal: when execution was expensive this mismatch was bearable, since replication was expensive too. Once the generation side near-free multiplies claims-to-verify a hundredfold while verification (replication) stays stuck in a thankless, no-one-does-it incentive sink, the gap simply explodes: RES 05’s scissors gap at its sharpest. The irony is that AI could shoulder most of replication’s actual work (running the code, recomputing, cross-checking data), but as long as the incentive structure leaves replication with no reward, cheaper execution doesn’t help: no one has a reason to press the button. This is RES 07’s “freed capacity doesn’t automatically become slack,” landing on replication specifically: being technically able to replicate isn’t the same as someone institutionally doing it. What needs fixing isn’t the technology, it’s the incentive: vouching for credibility (replication, verification, integration) has to move from the pyramid’s unrewarded base up to a reward position equal to original discovery. That’s exactly what RES 00’s “the scientific community’s value shifts from producing knowledge to vouching for credibility” means at the institutional level.

案例一 · 提问分诊：同一个材料发现项目，两支问题走向相反的两侧

Case 1 · Question-triage: in one materials-discovery project, two questions go to opposite sides

情形：一个固态电解质材料组，手里有一个待解的问题包。AI 已能近免费地跑高通量筛选与性质预测（GNoME 类工作，R16），于是瓶颈不在算，而在"哪个问题值得问"。组里把问题包摊开，逐条过 RES 10 的判据，能写出可机检验收标准的归左（交 AI），只能诉诸"换什么框架"的归右（留人）。结果两支问题走向了相反的两侧。

留在人这一侧的判断：是"哪一支问题值得用稀缺的人类带宽去追"，而非"哪个答案对"。分诊的价值正在于：它没有把第二支当噪声删掉（单一可信分会），而是认出它是"证据弱×框架远"格——挂起、定向取证，而非否决。如果当初把两支都丢给 AI，AI 会在第一支上高效地内卷，而第二支（换框架机会）会因为"离文献分布远"被自动降权。这就是 FIG 14.0 决策树要可视化的那条岔路。

The judgment that stayed human: not “which answer is right” but “which branch is worth chasing with scarce human bandwidth.” The triage’s value is precisely that it did not delete the second branch as noise (a single credibility score would) but recognized it as the “weak × far” cell: suspend and seek targeted evidence, not reject. Had both branches gone to AI, AI would have efficiently churned on the first while the second (the real paradigm-level opportunity) would have been auto-downweighted for being “far from the literature distribution.” This is the fork that FIG 14.0’s decision tree visualizes.

把这棵树做成可拨的：下面这台分诊器让你对一个具体问题逐道答"是/否"，它实时给出路由判词——交 AI、还是留人，以及为什么。试着把案例一的两支分别拨进去，看它们怎么走向相反的终点。

The tree, made adjustable: the decider below lets you answer “yes/no” to each gate for a concrete question, and gives a live routing verdict: to AI, or to a human, and why. Try dialing in Case 1’s two branches and watch them reach opposite ends.

案例二 · 可信度天平：一条"AI 设计的新抗生素"主张，两条轴必须分开记

Case 2 · The believability ledger: an “AI-designed new antibiotic” claim, two axes booked separately

情形：一个团队收到一条 AI 生成的主张——"模型在已知抗生素骨架外，设计出一类全新机理的候选分子，体外实验显示对耐药菌有效"。这条主张同时踩了天平的两条轴，而把它压成单一可信分会犯致命错误。逐轴来记：

两条轴若合并会怎样：假设把"证据弱（扣分）"和"离框架远（又扣分）"合成一个可信分，这条主张会拿到极低分，被自动归入"不可信、删"。于是两件本该分开做的事全做错了：该去补的独立复现没人去补（分数已替它下了结论），该由人判的机理重构被自动判死（分数把"离框架远"折算成低可信）。分开记的全部意义，就在于让证据弱触发"补证据"，让"离框架远"触发"人来判"——两个完全不同性质的动作，各归各位。

What happens if the axes merge: suppose “weak evidence (minus points)” and “far from paradigm (minus more)” are fused into one credibility score; this claim scores extremely low and is auto-binned as “not credible, delete.” Then both things that should have been done separately are done wrong: the independent replication that should be sought is not sought (the score already concluded for it), and the mechanistic reframing a human should judge is auto-killed (the score converted paradigm-distance into low credibility). The whole point of booking separately is to let weak evidence trigger “seek evidence” and paradigm distance trigger “a human judges”: two acts of completely different nature, each in its place.

案例三 · 图谱护栏失误：一张本草知识图谱，把"化学成分"锁成了唯一描述层级

Case 3 · A guardrail failure: a materia-medica knowledge graph locked “chemical constituent” as the only level of description

情形：一个团队为天然药物建了一张大规模知识图谱当护栏（RES 04 的思路，让海量 AI 生成留在可追溯结构里）。图谱的本体（ontology）是这样设计的：每味药材 → 其化学成分 → 成分的分子靶点 → 通路。AI 在这张图谱上做候选发现，效率极高，产出可追溯、可验证。一切看起来都对。直到他们撞上一类反复出现却怎么也解释不了的药效。

这个失误的普遍形态：护栏（知识图谱）守住了"框架内的可追溯"，代价是把描述层级冻结在建图谱时的那一层。它的危险恰恰在于它看起来全对：产出可追溯、可验证、效率高，所有读数全绿，但它在悄悄地把"换描述层级"这种换框架动作排除在可能性之外（这正是 RES 11 的核心警告）。护栏的正确用法不是"建一次、永久信任"，而是定期问一句：这张图谱的本体，有没有把某个描述层级锁成唯一？谁来问这一句、谁有权给本体加一层，这又落回 RES 07 的治理问题：本体的边界就是能被表达的框架的边界。

The general form of this failure: the guardrail (knowledge graph) preserves “in-paradigm traceability” at the cost of freezing the level of description at the layer present when the graph was built. Its danger is precisely that it looks entirely correct — traceable, verifiable, efficient output, every reading green, while quietly excluding “switching the level of description,” a paradigm-level act, from the space of possibilities (exactly RES 11’s core warning). The correct use of a guardrail is not “build once, trust forever” but to periodically ask: has this graph’s ontology locked some level of description as the only one? Who asks this, and who has the authority to add an ontology layer: this falls back to RES 07’s governance question: the ontology’s boundary is the boundary of the paradigm that can be expressed.

案例四 · 加速却变窄：一个把 AI 用满的领域，三年内更高产也更同质

Case 4 · Accelerated yet narrowed: a field that maxed out AI grew more prolific and more homogeneous in three years

情形：一个计算驱动的子领域（可类比 Hao 等横跨约 4129.8 万篇论文的文献计量所刻画的形态，R9），从早期就把 AI 写作、文献综述、点子生成用满。三年后回看，所有"读数"都在变好：人均发表数上升、个人被引上升、项目周期缩短。按旧学术机器的每一个指标，这是个高歌猛进的领域。但把镜头拉到领域整体，出现了一组相反的信号。

这就是 hypernormalization 的运行态（RES 08）：领域是变窄了，不是变差了：更高效、更稳定、所有指标更好看，但探索的方差在塌缩，大家都在用同一个模型、追同一批热点、收敛到同一个均值。最危险之处在于它没有任何报警：每个个体读数都是绿的，旧学术机器的每个指标都在说"一切向好"。它正是 RES 13 那条"指标奖励可数产出、AI 是其最优解"的领域级后果。什么判断本该留在人这一侧却没留住："这个领域是不是在变窄"这个问题，没有任何一个个人 KPI 会问它，它只能由人在领域层主动问，并主动用制度去对冲（保护偏离均值的探索、奖励复现与反共识工作、给不可度量的 slack 留生存空间，RES 07）。把这个判断也外包给"指标自动监控"，等于让正在制造同质化的那套机制来诊断同质化。要看见变窄，得先有人愿意去看那些不会让自己 KPI 变好看的信号。

This is hypernormalization at runtime (RES 08): the field did not get worse, it got narrower: more efficient, more stable, every metric prettier, but the variance of exploration is collapsing, everyone on the same model, chasing the same hot topics, converging to the same mean. The most dangerous part is that it sets off no alarm: every individual reading is green, every metric of the old academic machine says “all is improving.” It is the field-level consequence of RES 13’s “metrics reward countable output, and AI is their optimal solution.” What judgment should have stayed human but did not: the question “is this field narrowing” is asked by no individual KPI. It can only be asked actively by a human at the field level, and actively hedged by institutions (protect off-mean exploration, reward replication and anti-consensus work, leave survival space for unmeasurable slack, RES 07). Outsourcing this judgment too, to “automatic metric monitoring,” is letting the very mechanism that is producing the homogenization diagnose the homogenization. To see the narrowing, someone must first be willing to look at the signals that will not make their own KPI prettier.

两个正在发生的外部案例：拿判据回真实世界再对一遍

Two real-world cases in progress: checking the tests against the world again

上面四个是把内核压到可照做的细节上；下面这两个是正在发生的外部案例——一反一正，各自给命题划出一条能被独立核验的边界：执行可以被充裕，但"整合与担保"这一步，外包不出去。

The four cases above press the kernel onto copyable detail; the two below are external cases already in progress (one a counter-example, one a positive one), each drawing the thesis a boundary you can independently check: execution can be made abundant, but “integration and vouching” cannot be outsourced.

把省下的工时投回哪里，是这个环最容易垮的一步。这个环里藏着一个看不见的决定，它决定这个环到底带来进步，还是带来 hypernormal：②生成省下来的工时，到底投回了哪里。默认会发生什么，RES 07 早讲过——省下的产能不会自动变成 slack，它会被重新分配去做更多同样的事。落在研究环里，就是拿②省下的时间去多产论文、多跑实验、多生成假设，于是产出曲线更陡，可③判断、④整合、⑤守值的带宽一点没涨。这条失效路径特别隐蔽，因为它在每个局部指标上都像"进步"：产量涨了、影响力涨了、团队看着更高产了——这正是 Hao 等那 4129.8 万篇论文里，那批"个人影响力上升"的科学家的处境。

正确的做法反直觉：把②省下的工时显式地、刻意地投回③④⑤，让判断/复现/整合占研究者时间的比例上升，而不是让产出量上升。判据很简单：一个团队上了 AI 之后产量暴涨，但判断/整合占时间的比例没变，它就不是在跑这个环，它是在跑一台更快的 hypernormal 机器。

Where to reinvest the saved hours is this loop’s easiest step to get wrong. This loop hides one invisible decision that determines whether it delivers progress or hypernormal: where the hours ② saves actually go. RES 07 already told you what happens by default: freed capacity doesn’t become slack on its own, it gets reallocated to more of the same. Inside the research loop that means spending the time ② saved on more papers, more experiments, more hypotheses, so output climbs steeper while ③ judgment, ④ integration, ⑤ owning value gain no bandwidth at all. This way of failing is unusually hard to see, because on every local metric it looks like progress (output up, impact up, the team looking more productive), exactly the position of the “individual impact up” scientists in Hao et al.’s 41.3-million-paper study.

The right move is counter-intuitive: reinvest the hours ② saves, explicitly and deliberately, into ③④⑤: raise the share of researcher time on judgment, replication, and integration, not the volume of output. One simple test: if a team’s output spikes after adopting AI but its share of time on judgment and integration hasn’t moved, it isn’t running this loop. It’s running a faster hypernormal machine.

框定先行、整合优先，是这个环的两处承重。这个环跟工程那套规格驱动环（框定→计划→执行→验证→整合→学习）是同一个道理，但研究版有两处特意拧紧的地方，照抄的时候不能松。第一处是①框定要排在②生成前面：先立好可追溯的证据库、写下"什么算值得相信"的判据，再开始生成。理由 RES 04 早说过——证据库是研究的规格，不是事后归档；次序一反，你得到的是一座整合不动的垃圾山。很多团队把研究环抄成"先让 agent 狂产、再想办法管"，漏掉的正是这一拧。第二处是④整合要排在检索前面：当 RES 02 的生成把产出推向近乎无限，这个环里最容易堵的不是生成，是消化。如果团队把②省下的工时拿去多产而不是投回④整合，生成和整合之间就会越积越高——产出曲线陡升，理解曲线趴着不动。所以这个环的瓶颈阀门在④，不在②。

“Frame first” and “integration first” are this loop’s two load-bearing joints. This loop is the same idea as engineering’s spec-driven loop (specify, plan, execute, verify, integrate, learn), but the research version tightens two joints on purpose, and you can’t let them loosen when you copy it. The first: ① FRAME comes before ② GENERATE. Stand up a traceable evidence base and write down the criteria for “worth believing” before you open generation. RES 04 already gave the reason: the base is research’s spec, not after-the-fact filing; reverse the order and you get a garbage mountain nothing can integrate. A lot of teams copy the loop as “let the agent run wild, figure out management later”: that’s exactly this joint, missed. The second: ④ INTEGRATION comes before retrieval. Once RES 02’s generation pushes output toward the near-infinite, the easiest place for this loop to jam isn’t generation, it’s digestion. If a team spends the hours ② saves on more output instead of reinvesting in ④ integration, a backlog piles up between generate and integrate — output climbing steeply, understanding lying flat. So this loop’s real bottleneck valve sits at ④, not ②.

第⑥步——回流——是把"环"和"流水线"分开的那一步。这个工件叫"环"，不叫"流水线"，全靠这一步把它闭合。流水线是单向的：原料进，产品出，错误当废品扔掉。环带反馈：每一次被撤回的、被证伪的、被误杀的新颖，都不是废品，是下一轮的护栏材料。具体怎么流回去？一条复现不出来的主张，变成证据库里一条新的冲突检测规则；一个被现有 schema 判成异常、却反复出现的残差，变成一个新节点类型（RES 11 的换变量口子）；一次"把该重构的东西当噪声删掉"的事故，变成天平里"证据弱×框架远"那一格的处置改进（RES 09）。

没有这一步，前五步就退化成一条更快的流水线——生成、过滤、整合、产出，错误流走了就不回来，系统永远在重复同一个盲区。有了这一步，错误才变成系统的学习信号：每撞一次墙，护栏就长一点，下一轮少犯一点。这正是它跟工程"错误回流成新测试"完全同一招的地方，也是整卷反复盯着"撤回/证伪率应该往下走"这条指标的原因：它量的不是"少犯错"，是"这个环到底在不在学"。

Step ⑥ — feed-back — is what separates a loop from a pipeline. This artifact is called a loop, not a pipeline, entirely because this step closes it. A pipeline runs one way: raw material in, product out, errors thrown away as scrap. A loop has feedback: every retracted, refuted, mistakenly-killed piece of novelty isn’t scrap, it’s guardrail material for next time. Concretely: a claim that doesn’t replicate becomes a new conflict-detection rule in the base. A residual the current schema calls anomalous but that keeps recurring becomes a new node type: RES 11’s change-the-variable door. An incident where a reframing got deleted as noise becomes an improved disposition for the ledger’s “weak evidence, far paradigm” cell (RES 09).

Without this step, the first five degrade into a faster pipeline — generate, filter, integrate, output, errors flow away and never come back, and the system keeps repeating the same blind spot forever. With it, errors become the system’s learning signal: every wall it hits grows the guardrail a little, and it makes that mistake a little less next time. This is exactly where it’s the same move as engineering’s “errors feed back as new tests,” and it’s why the volume keeps watching one metric: the retraction/refutation rate should fall. That metric doesn’t measure fewer mistakes. It measures whether this loop is actually learning.

研究下注是一个组合，不是单点决定

A research bet is a portfolio, not a single call

环一跑起来，第一个反咬你的治理问题就来了：该把多少带宽投在利用（精炼已知、稳出结果），多少投在探索（追不确定的东西，可能颗粒无收）？RES 07 已经说了，效率默认会吃掉探索，所以这个配比不能放任默认，得当成一个可拨、可看的组合来管。这正是生成充裕带来的新自由度：执行近免费之后，跑一次冗余探索的边际成本骤降，理论上你能负担得起更高的探索比例；但激励要是还按产量考核，省下的产能又会被默认推回利用那一边。下面这台仪器把这层张力做成两根可拨的滑杆——拨"探索 vs 利用"的配比、拨"冗余探索"的允许度，看预期新颖跟成本怎么联动，看自己是不是正滑进 hypernormal 那种"指标全绿、覆盖在缩"。

The moment the loop starts running, the first governance question that bites back is: how much bandwidth goes to exploitation — refining the known, steady yields — and how much to exploration — chasing the uncertain, possibly nothing? RES 07 already showed efficiency eats exploration by default, so this ratio can’t be left on autopilot; it has to be managed as a portfolio you can adjust and watch. This is the new degree of freedom abundance creates: once execution is near-free, the marginal cost of a redundant exploration drops sharply, so in principle you can afford a higher explore share — but if incentives still score on output, the freed capacity gets pushed back to exploitation by default anyway. The instrument below turns this tension into two sliders: the explore-vs-exploit mix, and how much redundant exploration you allow; and shows how expected novelty and cost move together, and whether you’re sliding into hypernormal’s “every metric green, coverage shrinking.”

这台仪器不给你"最优解"，它逼你直面权衡。把探索配比拨到 0，预期成本最低、读数全绿，但覆盖广度也最低——这就是 hypernormal：高效、稳定、在缩。拨到 100，覆盖最广、预期新颖最高，但成本陡升，而且大量探索注定颗粒无收（这正是"无用之用"的字面意思）。生成充裕改变的是这条权衡曲线的形状：执行近免费让"高冗余探索"不再像过去那样贵得离谱，理性的探索配比应该往上移——但只有激励结构肯让"不可度量的 slack"活下来，这个上移才真的发生。仪器里"冗余探索允许度"这根杆，模拟的正是组织愿不愿意为看着没用的方向留预算。

This instrument gives you no optimum; it forces you to face the trade-off. Slide the explore share to 0 and expected cost is lowest, everything reads green, but coverage is also lowest. That’s hypernormal: efficient, stable, shrinking. Slide it to 100 and coverage is widest, expected novelty highest, but cost climbs steeply, and much of that exploration is destined to yield nothing (the literal meaning of “the use of the useless”). What abundance changes is the shape of this trade-off curve: near-free execution means high-redundancy exploration is no longer the prohibitive cost it used to be, so the rational explore share should move up, but that only actually happens when the incentive structure lets unmeasurable slack survive. The “redundant-exploration allowance” slider in the instrument models exactly whether an organization will budget for directions that look useless.

起步别一次全建：先挑一个"执行已经充裕、判断还没被外化"的环节——比如文献综合或参数扫描——立一个最小的可追溯证据库，把这个环节的框架内动作交给生成。跑顺了再加③天平、④整合。把②省下的工时显式投回③④，而不是拿去多产论文，这一条最容易垮（见 RES 14 的边界）。三步起步：先立可追溯证据库 → 用可信度天平挑出该注入人类判断的地方 → 把省下的工时投回整合与守值。

Don’t build it all at once: first run a narrow ①+② workflow: pick one step where execution is already abundant but judgment isn’t yet externalized (literature synthesis, parameter sweeps are the usual entry points), stand up a minimal traceable evidence base, and hand that step’s in-paradigm actions to generation. Once it runs, add ③ the ledger and ④ integration. Reinvest the hours ② saves explicitly into ③④, not into producing more papers — this is the easiest place to fail (see RES 14’s boundary). Three starts: stand up a traceable evidence base → use the believability ledger to pick where human judgment gets injected → reinvest saved hours into integration and owning value.

边界的判据：执行是不是真被充裕了，判断是不是真能被外化

The boundary test: is execution truly made abundant, can judgment truly be externalized

把"在哪适用"讲精确，得回到命题的两个前提，一个个去检验它们在具体的域里成不成立。前提一：执行能被大规模充裕。这在计算生物（AlphaFold, Nature 2021）、材料筛选（A-Lab, Nature 2023）、文献综合里成立——跑一次实验/筛选/综述的边际成本趋近于零，还能并行。但在受物理或伦理限速的域里不成立：一个三期临床试验、一次要等三年生长周期的田野采样、一个罕见样本的湿实验，执行本身仍是真瓶颈，"判断被充裕、执行变便宜"这条前提整个塌了，这时候这卷的退守命题增益很小，因为稀缺根本没从执行端搬走。

前提二：判断能被外化。它要求"什么算值得相信、值得知道"这套判据能部分写下来、部分被机检。在判据高度共识的域里（某些结构化预测任务，"对"几乎没有争议），外化容易，但这卷的增益也小，因为压根没有价值分叉可言。这卷增益最大的甜区，恰恰是两个前提同时成立、判断又还没被外化的域：执行已经能充裕了，判断却还困在少数专家脑子里，没被结构化出来。把这两条当成门禁一个案例一个案例地过，比背一张"适用领域清单"可靠得多。

To state where it applies precisely, go back to the thesis’s two premises and test each against a given field. Premise one: execution can be made abundant at massive scale. This holds in computational biology (AlphaFold, Nature 2021), materials screening (A-Lab, Nature 2023), literature synthesis: the marginal cost of one experiment, screen, or review trends to zero and can run in parallel. It fails in fields rate-limited by physics or ethics: a phase-three clinical trial, a field sample that needs a three-year growth cycle, a rare-sample wet lab. Execution is still the real bottleneck there, the premise “judgment made abundant, execution cheap” collapses entirely, and the volume’s retreat thesis adds little, because scarcity never moved off the execution end.

Premise two: judgment can be externalized. It requires that the criteria for “worth believing, worth knowing” can be partly written down, partly machine-checked. In fields where those criteria are highly consensual (some structured-prediction tasks where “correct” is nearly uncontested), externalizing is easy, but the volume’s gain is small too, because there’s no value fork to speak of. The volume’s sweet spot is exactly the field where both premises hold and judgment isn’t yet externalized: execution can already be made abundant, yet judgment is still stuck, unstructured, in a handful of experts’ heads. Running these two as a gate, case by case, is far more reliable than memorizing a list of applicable fields.

可读性问题：突破也许要以部分不可读为代价。适用边界还有一条更深、更不舒服的前沿命题，得诚实摆出来：如果真要 AI 出突破，损失一部分可读性可能是躲不开的。类比 AlphaZero（Silver 等，Science 2018）——它下出的某些棋"概念上不透明"，强过任何人类却没人能完整解释为什么。当 AI 在科学上做出类似的事，风险是发现被"搁浅"在一堆没人能看懂的产出里：你拿到一个比现有理论预测更准的模型，却没法把它翻译成人能理解、能据以行动、能排优先级的知识。这是研究卷一个真实的张力：它整个立论建在"人接住判断"上，可要是突破本身部分不可读，人接住的就只是一个黑箱的输出，不是它的理由。

The legibility problem: breakthroughs may cost some unreadability. The applicability boundary carries one deeper, less comfortable frontier claim that has to be put on the table honestly: if you really want AI to produce breakthroughs, some loss of legibility may be unavoidable. By analogy to AlphaZero (Silver et al., Science 2018): some of its moves are conceptually opaque, stronger than any human yet no one can fully explain why. When AI does something similar in science, the risk is that discoveries get stranded in a flood of output no one can parse: you’re holding a model that predicts more accurately than current theory, yet you can’t translate it into knowledge a human can understand, act on, or prioritize. This is a real tension for the research volume: its whole argument rests on humans catching the judgment, but if a breakthrough is itself partly unreadable, what a human catches is only a black box’s output, not its reasons.

一句话边界：别把探索清单当已证去裁人。这里最该当成硬门禁的，不是技术域的划分，是一条诚实纪律：这卷有不少命题标着"探索清单·待坐实"——提问被充裕（RES 02）、整合鸿沟急剧扩大（RES 05）、净知识 −40%（RES 02 的 ODE 预测），它们是有侧证支撑的推演，不是已证的事实。把这些还在探索的命题当成"已证现实"去做组织决策，尤其是拿去裁人，是这卷最危险的误用。一个组织如果以"提问已经被 AI 充裕了，所以不需要这么多研究员"为理由裁员，它其实是在把一条 Ⅴ 级的推演当 Ⅰ 级证据用——而 RES 06/07 反复强调的恰恰是：被充裕的是框架内提问，换框架的重构和价值判断不但没被充裕，反而在升值。把探索清单误当硬锚，结果是裁掉了正该守住的那批判断力。

The boundary in one line: do not cut people on an exploratory ledger. What should be treated as a hard gate here isn’t the partition of technical fields; it’s an honesty discipline. This volume carries several claims tagged “exploratory, to be grounded”: questioning made abundant (RES 02), the integration gap exploding (RES 05), net knowledge −40% (RES 02’s ODE prediction). These are side-evidenced projections, not proven facts. Treating them as proven reality for organizational decisions (especially to cut people) is this volume’s most dangerous misuse. An organization that lays off staff on the grounds that “questioning is already made abundant by AI, so we need fewer researchers” is using a grade-Ⅴ projection as grade-Ⅰ evidence, and what RES 06/07 keep stressing is exactly the opposite: what turns abundant is in-paradigm questioning, while paradigm-level reframing and value judgment don’t turn abundant, they appreciate. Mistake an exploratory ledger for a hard anchor, and you cut away the very judgment you meant to protect.

绿地直接重画，存量先切一条工作流试

Greenfield: redraw directly; incumbent: carve out one workflow first

落到"怎么开始"，适用边界自然分成两条路径，对应组织的两种起点。绿地：一个从零起步的研究项目，可以直接按这卷重画——第一步不是招更多研究员，是立一个最小可追溯证据库，把"什么算值得相信、值得知道"的判据显式写下来；然后用价值分诊（RES 10 的矩阵 + RES 09 的天平）决定哪些动作交给生成、哪些注入人类判断。绿地的好处是没有存量流程的惯性，能一次把次序立对——规格先于生成，整合优先于检索。这也是这卷那个"从头设计"押注真正能被实测的地方：不是把选题-调研-实验-结论那条旧流水线跑得更快，是把机构的记账单位从"发表了多少"换成"沉淀了多少可复现的知识"。

存量改造：一个已经在跑的实验室，绝不能推倒重来，只能一点点改：从一条工作流里切出"执行已充裕、判断还没被外化"的那个环节（文献综合、参数扫描是最常见的入口），只在这一段重画，跑顺了再往外扩。存量改造最要命的陷阱，是把②省下的工时拿去多产论文——RES 13 反复警告过这一条。两条路径共用一句边界判词：这卷适用于"执行已充裕、判断还没被外化"的研究域；执行仍是真瓶颈、或者判断已经高度共识的地方，直说这不是它的目标群体，别硬套。

Down to how to start, the applicability boundary naturally splits into two paths, matching an organization’s two starting points. Greenfield: a program starting from zero can be redrawn by this volume directly: step one is standing up a minimal traceable evidence base and writing down explicit criteria for “worth believing, worth knowing”; then using value-triage (RES 10’s matrix plus RES 09’s ledger) to decide which actions go to generation and which inject human judgment. Greenfield’s advantage is no legacy-process inertia, so you can get the order right in one go: spec before generation, integration before retrieval. This is also where this volume’s redesign bet actually gets tested: not running the old pick-question-survey-experiment-conclude pipeline faster, but swapping the institution’s unit of account from how much got published to how much reproducible knowledge got banked.

Transformation: a lab already running must never be torn down and rebuilt, only changed piece by piece: carve out of one workflow the single step where execution is already abundant but judgment isn’t yet externalized (literature synthesis, parameter sweeps are the usual entry points), redraw only that segment, and expand once it runs. Transformation’s deadliest trap is spending the hours ② saves on more papers: RES 13 warns against this repeatedly. The two paths share one boundary verdict: this volume applies where execution is already abundant but judgment isn’t yet externalized; where execution is still the real bottleneck or judgment is already highly consensual, say plainly this isn’t the target group — don’t force it.

对策是建一层翻译，不是减速。老实说，这是 Asimov Press（The Legibility Problem）等的一个推演，缺工程实证，标为前沿命题。可能的出路：建一层"解释/翻译层"，让 AI 的发现对人可读、能排优先级。这不是要求 AI 只产人能立刻理解的东西（那等于把它又锁回框架内），而是在它产出之后，专门投入去把不可读的发现翻译成可读的知识。这本身就是内核④"人回归意义"的一个新落点：当一阶发现可能不可读，人的稀缺贡献之一，就是去搭那座把黑箱输出译成人类理解的桥。它也回连 RES 05 的整合：可读性翻译是最难的一种整合——把一个借不到任何框架的发现，缝进人类已有的理解结构里。

The remedy is building a translation layer, not slowing down. Honestly, this is a projection from Asimov Press (The Legibility Problem) and others, lacking engineering empirics, flagged as a frontier claim. One possible way out: build an explanation/translation layer that makes AI’s discoveries legible and prioritizable for humans. This doesn’t require AI to produce only what humans can immediately understand (that would just lock it back into the paradigm); it means, after it produces, deliberately investing in translating unreadable discoveries into readable knowledge. This is itself a new landing point for kernel ④’s “humans return to meaning”: when first-order discovery may be unreadable, one of the human’s scarce contributions is building the bridge that translates a black box’s output into human understanding. It also wires back to RES 05’s integration: legibility translation is, in the end, the hardest kind of integration: stitching a discovery with no frame to borrow into humanity’s existing structure of understanding.

三股会聚的力，每股都带一条能推翻它的观测

Three converging forces, each with an observation that would overturn it

推演不是在预言"哪条线一定会发生"，是指出哪些力正在叠加、各自在什么观测下会被判错。要是下面三股力同时成立，研究的样貌会从"人提问、机器执行"，滑向"机器也提框架内的问题，人退到选议程、定何为真"。每股都配了一条先行指标和一条证伪条件——后者才是这股力的命门：看见它，就说明这股力被高估了。

Speculation here isn’t prophesying which line must happen; it’s pointing at which forces are stacking up, and under what observation each would be judged wrong. If the three forces below all hold at once, research’s face slides from “humans ask, machines execute” toward “machines also pose in-paradigm questions, humans retreat to selecting the agenda and defining what’s true.” Each carries a leading indicator and a falsification condition: the latter is the force’s real pressure point: see it, and the force was overrated.

2026→2028→2030→2032：研究的样貌一步步在变

2026→2028→2030→2032: research’s face deforms, step by step

一件虚构的 2031 自主实验室季报

A fictional 2031 autonomous-lab quarterly

只有断言的推演读着空。下面这件是设计虚构：一个明确标注为虚构的 2031 未来物件，把"研究退守到议程与裁判"做成可触摸的样子。它是把命题投影到 2031 的一种方式，不是预测。

Speculation made only of assertions reads thin. The piece below is design fiction: an explicitly fictional 2031 future artifact that makes “research retreating to agenda and refereeing” tangible. It is not a prediction; it is a way of projecting the thesis onto 2031.

必须写下的反方：判断也许只是又一种在等充裕的能力

The counter-bet, on record: judgment may just be one more capability waiting for abundance

诚实要求把最强的反方也记下来，不是只留对自己有利的那条线。这卷的主命题是：由人定义的价值判断——"哪个真相值得知道"——不会因为模型变强就被充裕，它的稀缺是结构性的，不是能力门槛。反方最锋利的一刀是：这条"结构性稀缺"，也许只是当前模型还不够强留下的临时假象。

Honesty requires writing down the strongest counter-argument too, not just the line that flatters this volume. Its central claim is: constitutive value judgment (“which truth is worth knowing”) won’t turn abundant just because models get stronger; its scarcity is structural, not a capability ceiling. The counter-bet’s sharpest cut: that “structural scarcity” may itself just be a temporary illusion left by models not yet being strong enough.

把反方写进正文，不是谦虚的修辞，是这卷方法本身的一部分：一个命题要是连自己都不肯接受一条证伪条件，它就不是知识，只是一个态度。这一章随时准备被 2032 那条引用判据改写——这恰恰是它配得上"研究方法论"这个名字的原因。

Putting the counter-bet in the body isn’t rhetorical modesty; it’s part of this volume’s method itself: a claim with no falsification condition its own author would accept isn’t knowledge, just an attitude. This chapter stands ready to be rewritten by that 2032 citation test, which is exactly why it earns the name “research methodology.”

最后这一层不给你一张一次性的清单，给的是一道动态的三分。三分不是把一个任务归一次类就完事，是随着充裕前线往右移，得不断重新判——同一个动作今天还在"在变"，明年可能已经滑进"不变"之外，或者滑进"不变"之内。

This last layer hands you no one-time checklist; it’s a dynamic three-way split. The split isn’t a single classification you make and forget; as the abundance frontier keeps moving right, you have to keep re-deciding. The same action sitting in “shifting” today may slide out of, or into, “invariant” next year.

生产速度跟能消化的速度之间那道鸿沟，才是这卷真正押的注

The gap between production rate and digestion rate is this volume’s real bet

整卷收成一句话：研究卷押的不是"AI 做不了科学"，是"知识生产的速度会远远甩开人类能消化的速度，稀缺因此从生产端永久搬到了消化端——判断、整合、定值"。这个赌注有数量级撑着：科学文献已经差不多每年 250 万篇、每 9 年翻一倍，AI 又在上面叠了一层质变加速。生产曲线往上指数走，人类的认知带宽却近乎恒定，两条曲线之间那道剪刀差，就是这卷所有命题栖身的地方——RES 05 的整合鸿沟、RES 03 的判可信、RES 06 的定值，都是这道剪刀差的不同切面。有一天这个赌注要是被证伪了（人类的消化速度也能跟着 AI 等比例往上提，或者机器能无损接管消化端的判断），整卷就该退役。写得出退役的条件，它才是一个命题，不是一份信仰。

Collapse the whole volume into one line: the research volume isn’t betting that AI can’t do science. It’s betting that the rate of knowledge production will far outstrip the rate humans can digest it, so scarcity permanently migrates from the production end to the digestion end: judging, integrating, valuing. This bet has orders of magnitude behind it: the scientific literature already runs at roughly 2.5 million papers a year, doubling every 9 years, with AI layering a qualitative acceleration on top. The production curve climbs exponentially while human cognitive bandwidth stays nearly constant, and the scissors-gap between the two curves is where every claim in this volume lives: RES 05’s integration gap, RES 03’s credibility judgment, RES 06’s valuing are all facets of the same gap. If this bet is ever falsified (human digestion scaling proportionally with AI, or machines losslessly taking over digestion-end judgment), the whole volume should retire. Being able to write the retirement condition is what makes it a claim, not a faith.

三条原则内在是同一回事："生成多·验证严""整合优先于检索""守住价值责任"看着是三件事，其实是同一句话投在环的三个位置上。生成多——是承认①执行已经充裕了；验证严——是②判断退守之后那个承重的动作；整合优先——是④人接住带宽瓶颈的稀缺贡献；守值——是④把"值得"留给人，不交给生成层的默认偏置。三者共用一条度量纪律：产出量本身永远不是指标。该往上走的是判断/复现占时间的比例、整合产物相对原始产出的比率、研究方向价值决策的归属清不清楚；该往下走的是被撤回/证伪率、向已知解收敛的比率。把这套指标钉在墙上，你就有了一面镜子，随时照得出自己是不是正滑进 hypernormal。

The three principles’ inner unity: “generate much, verify hard,” “integration over retrieval,” “hold value accountability” look like three things, but they’re one line projected onto three positions of the loop. Generate much admits ① execution is already abundant; verify hard is the load-bearing act after ② judgment retreats; integration first is ④’s human catching the bandwidth bottleneck’s scarce contribution; hold value is ④ keeping “worth” with humans, not the generation layer’s default bias. All three share one measurement discipline: output volume itself is never the metric. What should rise: the share of time spent judging and replicating, the ratio of integration artifacts to raw output, how clearly research-direction value decisions are owned. What should fall: the retraction/refutation rate, the rate of convergence toward known solutions. Pin this set of metrics on the wall and you have a mirror that, at any moment, shows whether you’re sliding into hypernormal.

# 在 Claude Code 里调用invoke inside Claude Code
$ /skill ai-native-research
> "把这 40 篇相互冲突的论文整合成一个判断：这个结论可信吗、值得我们押注吗？""Synthesize these 40 conflicting papers into one judgment: is this claim credible, and is it worth our bet?"

  → 一份研究发现档案 = 发现 + 可信度账（主张分 Ⅰ–Ⅴ、主张与证据不混）+ 知识图谱贡献 + 盲点登记a Research Finding Dossier = the finding + a credibility ledger (claims graded Ⅰ–Ⅴ, claims kept unmixed from evidence) + a knowledge-graph contribution + a blind-spot register