Boom5426 commited on
Commit
f293103
·
verified ·
1 Parent(s): 521ad73

Upload VCDesign_Manuscript_Revision_Plan_2026-09-25.md

Browse files
code/VCDesign_Manuscript_Revision_Plan_2026-09-25.md ADDED
@@ -0,0 +1,665 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # VCDesign:任务与论证层次升级——逐段修改说明
2
+
3
+ **基准版本:**《main (7).pdf》,39 页;正文结束于第 9 页。
4
+ **修改类型:**论文写作、段落组织与结论措辞;不新增实验,不更改方法或评价定义。
5
+ **交付对象:**可直接交给 CC / Codex 执行的段落级修改稿。中文说明,英文替换正文。
6
+ **日期:**2026-09-25。
7
+
8
+ ## 0. 使用方式与版本边界
9
+
10
+ 本文件的原文定位来自实际读取的《main (7).pdf》。页码与行号指该 PDF 的纸面页码和审稿行号,**不是 LaTeX 源文件行号**。以给出的首尾英文为搜索锚点定位真实源文件;本文未读取你的仓库,不指定未经确认的文件路径。
11
+
12
+ Section 6 在 PDF 中仍是旧版三段内容,而本对话已认可新的两段结尾。执行时使用 M18 的两段同步版本;M18 明列相对已认可文本的口径修正,不把旧 PDF 当作最新结尾的决定依据。其他章节以本 PDF 为定位基准。
13
+
14
+ 英文替换稿中的作者—年份、Figure/Section/Appendix 编号用于对应当前稿件。实际落入 LaTeX 时复用已存在的引用键、标签和数学宏;不要把编号硬编码成新的引用,也不要凭空生成 BibTeX keys。M19 新增附录时,可以按仓库命名规范创建真实 label,再由正文交叉引用。
15
+
16
+ **执行原则:先替换,再压缩重复,不通过追加“高层宣言”来堆长正文。** 每条修改已经给出完整目标文字;除明确标注的单句操作外,不在新段落后保留被替换的旧段落。
17
+
18
+ ## 1. 本轮要形成的论证层次
19
+
20
+ 全文保留既有科学主线:**定义有限预算的干预选择 → 用独立结果评价 → 以 CED 实现响应未测候选的效应迁移 → 通过控制实验区分信息、测量、模型变化及其决策价值。**
21
+
22
+ 引言说明“面向实验的模型需要什么成功标准”;Framework 给出可操作定义;方法解释两条分支学习什么;Results 给出有适用条件的证据;Discussion 才回到更广泛的研究含义。不是每个段落都再次复述 prediction → identification → design。
23
+
24
+ ### 1.1 上一轮概念建议中,不应直接写入论文的三点
25
+
26
+ **不能写成已发现 biological-information scaling law,或已证明信息而非模型是根本瓶颈。** 当前结果显示 natural atlas 扩大时收益提高,删除 STRING 后降低;但 row matching 同时改变 identity coverage,unanchored memory 的负向影响也不小。Table 26 还报告 predictive basis 为 +0.0009 [+0.0006, +0.0013] BU@20。应报告这些受控观察,而不是推导普遍的 scaling law 或相对因果重要性。
27
+
28
+ **不能把 CED 写成与 query 无关的 candidate prior。** CED 预测效应,并通过与目标转变的 alignment 为排序供分。它的机制高度来自“历史响应如何转成未测候选的目标条件化证据”,而不是给回归映射换一个听起来更大的名字。
29
+
30
+ **不能因强调评价贡献而移除 controlled diagnosis。** 保留三项贡献,仍然覆盖任务、评价与 CED、受控诊断。上一轮提出的“任务—评价—模型”三项新划分会遗漏已经确定的第三项贡献,M06 不采用这种改法。
31
+
32
+ ### 1.2 本轮不动的内容
33
+
34
+ - 不改题目、主实验、数据切分、utility、BU@B、预算集合、模型结构、loss 公式、融合权重或已存结果。
35
+ - Related Work 继续只有 Forward / Inverse 两部分,每部分两段;scGPT、CellFM 保留在 Forward。
36
+ - §3.3 保留 68% 的分支暴露事实;联合 unseen 的结果继续在 §5.2 与对应表注/附录,不反复搬移。
37
+ - §5.2 保留“held-out contexts → K562 identities → controls/PHR”的三段结构;§5.3 保留 response availability 和 routing 的实验责任。
38
+ - §5.5 标题保持 **Better Prediction Does Not Ensure Better Design**。
39
+ - 图 1–4 不重新设计、不替换数据。仅 M16/M17 修改图注首句。
40
+ - 正文维持 9 页;不能靠缩小字体、改模板页边距或压缩图片字字号换空间。
41
+
42
+ ## 2. 执行顺序
43
+
44
+ 先执行 M02–M06(引言与贡献)及 M10–M15(方法与结果解释),形成统一论证;随后执行 M01(摘要)、M07–M09(相关工作与 Framework)、M16–M17(图注);最后完成 M18–M19(两段结尾与限制保留)。D01 与 M10/M12 配套执行。
45
+
46
+ **P0** 是本轮核心修改;**P1** 是为了消除重复并使前后口径一致的局部修改。所有修改均限于写作。
47
+
48
+ ## 3. 逐段替换稿
49
+
50
+ ### M01|摘要:把决策、方法和诊断接成一条线
51
+
52
+ **优先级:**P0
53
+ **定位:**第 1 页,Abstract,纸面行号约 011–030
54
+
55
+ **原文起始锚点:**
56
+
57
+ > Virtual-cell modeling aims to predict cellular responses and prioritize interventions for experimental validation.
58
+
59
+ **原文结束锚点:**
60
+
61
+ > VCDesign advances virtual-cell modeling from prediction and intervention identification toward cellular intervention design evaluated by the experimental decisions it supports.
62
+
63
+ **操作与目的:**整段替换。把“模型支持的选择”作为评价对象,同时保留 CED、信息可得性与 routing 的实际结果;删除没有被比较实验识别出的“主要因果瓶颈”判断。
64
+
65
+ **替换为:**
66
+
67
+ ```text
68
+ Virtual-cell models can predict responses or identify plausible interventions, but experimental planning requires choosing which candidates to test under a limited budget. We introduce VCDesign, which formulates cellular intervention design as finite-budget ranking over a variable candidate set. Given starting and target states, it evaluates the selected interventions by independently measured held-out outcomes rather than recovery of the intervention that generated the query. This decision-level interface accommodates retrieval, forward simulation, optimization, and direct scoring. For candidates whose responses are unavailable to the effect predictor, VCDesign-CED learns a mapping from biological knowledge to perturbational effects using historical measurements and combines effect alignment with direct candidate scoring. Experiments across four CRISPRi Perturb-seq contexts show improved selection on held-out identities, in held-out contexts, and under controlled response masking. When candidate responses are measured in other contexts, profile retrieval outperforms VCDesign-CED, and routing measured candidates to retrieval improves selection from mixed pools. Controlled comparisons show gains from larger perturbation atlases and losses after removing candidate knowledge, whereas the richer effect maps tested offer no clear advantage over the linear reference. Prediction accuracy does not consistently order configurations by design utility. Together, these results connect virtual-cell modeling to experimental prioritization and motivate evaluating progress through the outcomes of finite-budget selections.
69
+ ```
70
+
71
+ **执行注意:**
72
+
73
+ - 四个 context 是数据覆盖范围,不要改成“四个完全独立的 held-out contexts”;K562 与其他三个 context 的角色仍由 §5.1 说明。
74
+ - 保留 `for candidates whose responses are unavailable to the effect predictor`,不要改成整个系统从未接触这些候选。
75
+ - `accommodates optimization` 指任务接口兼容,不声称本文运行过 optimization solver。
76
+ - Routing 的收益是 `improves selection from mixed pools`,不是候选库本身被“改善”,也不是 routing 优于全测量条件下的 retrieval-only。
77
+ - 摘要更新后同步已有 submission metadata 文件;本文没有读取该文件,不指定其路径或擅自更改投稿网站内容。
78
+
79
+ ### M02|引言第一段:从“已有方法不够”改为“实验用途需要额外的成功标准”
80
+
81
+ **优先级:**P0
82
+ **定位:**第 1 页,Introduction 第一段,约 035–047
83
+
84
+ **原文起始锚点:**
85
+
86
+ > Virtual-cell models have advanced rapidly as tools for modeling cellular perturbations, but most work still centers on a forward problem:
87
+
88
+ **原文结束锚点:**
89
+
90
+ > with the model ultimately evaluated by the experimental decisions it supports.
91
+
92
+ **操作与目的:**整段替换。保留 prediction → identification → design 的主线,但限定在“旨在指导实验的模型”,不要求整个虚拟细胞领域放弃预测任务。
93
+
94
+ **替换为:**
95
+
96
+ ```text
97
+ The value of a virtual-cell model for experimental planning depends not only on what it predicts, but on which interventions it helps prioritize. Most perturbation models address a forward problem: given an intervention, predict the cellular response (Lotfollahi et al., 2019; 2023; Bunne et al., 2023; Roohani et al., 2024; Cui et al., 2024; Yu et al., 2025; Adduri et al., 2026). These predictions support the study of perturbational effects, but do not by themselves specify which experiments to perform. Researchers instead start from a desired cellular transition, a feasible candidate library, and a limited experimental budget. For models intended to guide this choice, success must also be assessed by the value of the interventions selected. This motivates extending cellular intervention modeling from prediction and identification to design, with experimental prioritization as the downstream decision.
98
+ ```
99
+
100
+ **执行注意:**
101
+
102
+ - 作者—年份引用对应当前第一段已有文献。落入 LaTeX 时复用原来的 citation keys,不把这里的文字引用直接作为普通正文粘贴,也不创造新的 BibTeX 条目。
103
+ - 不使用 `descriptive prediction`:干预后的响应预测不应被贬低为单纯描述。
104
+ - 不使用 `the objective of virtual-cell modeling must change`:本文支持补充决策层评价,而不是用 design 全面替代其他科学目标。
105
+
106
+ ### M03|引言第二段:把已有路线定位为同一决策问题的求解器
107
+
108
+ **优先级:**P1
109
+ **定位:**第 1 页末至第 2 页,Introduction 第二段,约 048–053、083–088;Figure 1 插在段落中间,不属于替换范围
110
+
111
+ **原文起始锚点:**
112
+
113
+ > This distinction is particularly clear in recent inverse-modeling work.
114
+
115
+ **原文结束锚点:**
116
+
117
+ > but how the method prioritizes experiments from the candidate set that is actually available.
118
+
119
+ **操作与目的:**替换同一个跨页段落。用 CellNavi 说明 identification 与 design 的区别,不把所有逆向方法一概描述成固定分类或身份恢复。
120
+
121
+ **替换为:**
122
+
123
+ ```text
124
+ Inverse methods bring the desired transition into the modeling problem. CellNavi, for example, uses starting and target states to identify genes that may drive a transition (Wang et al., 2025). Identifying a plausible intervention, however, does not fully specify an experimental plan: several interventions may be useful, the feasible library can vary, and only a subset can be tested. Response retrieval, forward enumeration, and intervention optimization offer additional computational routes toward candidate selection (Gonzalez et al., 2025; Nourreddine et al., 2024; Han et al., 2025; Zeng et al., 2025; Cao et al., 2025; Peng et al., 2026; Liang et al., 2026). We view these as complementary solvers for a shared decision problem. Their common objective is to prioritize useful interventions from the available library, regardless of whether the ordering is produced by classification, retrieval, simulation, or optimization.
125
+ ```
126
+
127
+ **执行注意:**
128
+
129
+ - 保留当前段落的两组引用及其对应位置。本轮不重新裁决已有文献的算法归类。
130
+ - `complementary solvers` 是本文的任务组织视角,不等于已对所有路线完成公平的实证比较。
131
+ - 原段落被 Figure 1 打断,编辑时按首尾文字查找同一个段落,不能把图环境一并删掉。
132
+
133
+ ### M04|引言第三段:明确评价对象是所选干预,而非模型的中间输出
134
+
135
+ **优先级:**P0
136
+ **定位:**第 2 页,Introduction 第三段,约 089–101
137
+
138
+ **原文起始锚点:**
139
+
140
+ > Based on this view, we introduce VCDesign (Figure 1),
141
+
142
+ **原文结束锚点:**
143
+
144
+ > whose own perturbational response has not been used for effect learning.
145
+
146
+ **操作与目的:**整段替换。这里完整定义一次“候选集—预算—排序—独立结果”,其他章节不再反复重述全部接口。
147
+
148
+ **替换为:**
149
+
150
+ ```text
151
+ VCDesign makes this decision explicit (Figure 1). Given starting and target states, a variable set of feasible interventions, and a budget, a solver ranks candidates for experimental validation. The object of evaluation is the resulting finite-budget selection, rather than a predicted profile or a recovered intervention identity. We exclude the intervention that generated the query and evaluate the remaining candidates using independently measured held-out outcomes. This allows multiple candidates to be useful under the same objective and separates the value of a selection from the mechanism that produced it. The interface also makes information availability explicit: transferring to a new cellular context differs from prioritizing a candidate whose own response has not supervised effect learning.
152
+ ```
153
+
154
+ **执行注意:**
155
+
156
+ - `rather than a predicted profile` 指直接评价 solver 的最终选择,不是否认可以在辅助分析中评价 predicted profile。
157
+ - 这里的 measured outcomes 仍受 §3.2 的方向性 utility 定义约束;不能自行换成已验证治疗效果、功能恢复或前瞻性实验成功。
158
+ - 不新增“首个统一接口”“首次提出实验设计”等优先权主张。
159
+
160
+ ### M05|引言第四段:把 CED 写成明确的效应迁移机制,不为简单模型辩护
161
+
162
+ **优先级:**P0
163
+ **定位:**第 2 页末至第 3 页,Introduction 第四段,约 103–113
164
+
165
+ **原文起始锚点:**
166
+
167
+ > Within this formulation, a key question is not only how a solver is parameterized,
168
+
169
+ **原文结束锚点:**
170
+
171
+ > or by the perturbation measurements available for learning?
172
+
173
+ **操作与目的:**整段替换。解释历史响应如何支持未测候选的排序,并自然引出后续控制实验。
174
+
175
+ **替换为:**
176
+
177
+ ```text
178
+ VCDesign-CED instantiates this framework for candidates whose responses are unavailable to the effect predictor. Candidate Effect Distillation (CED) uses historical perturbation measurements to learn a mapping from deployment-available biological knowledge to perturbational effects. The inferred effect is aligned with the requested transition and combined with a direct candidate-conditioned score. Historical measurements thus supervise a transferable mapping rather than supply the candidate's own profile at deployment. We use a linear effect map as a reference for controlled comparisons of perturbation atlases, candidate knowledge, and richer mappings, asking which changes improve finite-budget selection.
179
+ ```
180
+
181
+ **执行注意:**
182
+
183
+ - 删除 `We deliberately keep ... simple`,但保留使用 linear map 的真实研究角色。
184
+ - 不把 CED 重新命名为 `candidate-level prior`。它的 effect score 依赖目标方向,并非忽略 query 的候选先验。
185
+ - CED 仍是知识到效应的回归映射,再与目标匹配;叙事升级不隐藏这个具���实现,也不宣称发明了全新的迁移学习范式。
186
+
187
+ ### M06|贡献列表:强化评价层次,但保留三项贡献和受控诊断
188
+
189
+ **优先级:**P0
190
+ **定位:**第 3 页,Our contributions are threefold 之后的三个项目,约 116–126
191
+
192
+ **原文起始锚点:**
193
+
194
+ > A decision-centric formulation of cellular intervention design.
195
+
196
+ **原文结束锚点:**
197
+
198
+ > clarifying the current limitations of response-unseen transfer.
199
+
200
+ **操作与目的:**替换三个项目的文字,不改变三项贡献的数量。第二项继续包含评价接口与 CED,第三项不能被改成只有模型。
201
+
202
+ **替换为:**
203
+
204
+ ```text
205
+ - **A decision-centric formulation of cellular intervention design.** We formulate design as finite-budget ranking over a supplied, variable candidate library, separating the experimental objective from the computational route used to prioritize interventions.
206
+ - **Decision-aligned evaluation and a response-unseen solver.** We assess selected interventions through independently measured outcomes rather than query-identity recovery, distinguish context transfer from response-unseen effect learning, and introduce VCDesign-CED to combine direct candidate scoring with effects inferred from biological knowledge.
207
+ - **A controlled analysis of transferable design.** We examine how response availability, atlas composition, candidate knowledge, and effect-model changes affect intervention selection, and test whether prediction accuracy reliably orders these configurations by design utility.
208
+ ```
209
+
210
+ **执行注意:**
211
+
212
+ - 第一项回答“决策是什么”;第二项回答“如何评价、如何在未测响应条件下求解”;第三项回答“这些控制实验揭示了什么”。
213
+ - 上一轮建议中的“任务—评价—模型”若被直接改成三个独立项目,会挤掉已冻结的受控诊断贡献;不要这样执行。
214
+ - `decision-aligned` 是评价与所定义的有限预算决策对齐,不等于已证明该 utility 对所有生物学终点充分。
215
+
216
+ ### M07|Related Work:Forward 分支第二段只讨论预测与决策评价的关系
217
+
218
+ **优先级:**P1
219
+ **定位:**第 3 页,Forward perturbation prediction 的第二段,约 142–145
220
+
221
+ **原文起始锚点:**
222
+
223
+ > A forward model can also support intervention design by simulating candidates
224
+
225
+ **原文结束锚点:**
226
+
227
+ > within a shared ranking and outcome-based evaluation interface.
228
+
229
+ **操作与目的:**整段替换,保持这一分支只有两段。第一段综述及其文献不动。
230
+
231
+ **替换为:**
232
+
233
+ ```text
234
+ A forward model can support intervention design by simulating candidates and ranking their predicted outcomes against a target. Predictive accuracy evaluates the response estimates; decision-level utility evaluates the interventions selected from them. VCDesign makes this second criterion explicit through finite-budget ranking and held-out outcome evaluation, allowing forward models to be assessed by the experimental choices they support.
235
+ ```
236
+
237
+ **执行注意:**
238
+
239
+ - 不将所有 forward models 概括成“不能设计”;它们可以成为本任务的 solver。
240
+ - scGPT 和 CellFM 留在现有 Forward 第一段,不移动到 Inverse。
241
+
242
+ ### M08|Related Work:Inverse 分支第二段只讨论可比较的决策与信息条件
243
+
244
+ **优先级:**P1
245
+ **定位:**第 3 页末至第 4 页,Inverse perturbation and design 的第二段,约 159–163
246
+
247
+ **原文起始锚点:**
248
+
249
+ > These inverse approaches differ in their output spaces, access to candidate responses, and evaluation objectives.
250
+
251
+ **原文结束锚点:**
252
+
253
+ > have not supervised the effect predictor in any training context.
254
+
255
+ **操作与目的:**整段替换,保持 Inverse 第一段文献综述原样。
256
+
257
+ **替换为:**
258
+
259
+ ```text
260
+ These approaches differ in their output spaces, access to candidate responses, and evaluation objectives. VCDesign separates the decision being evaluated from the mechanism used to produce it: methods rank a supplied library under a budget, and selected interventions are assessed by held-out outcomes. Its response-unseen setting additionally removes a candidate's responses from effect learning across training contexts, distinguishing transfer across contexts from transfer without the candidate's own response supervision.
261
+ ```
262
+
263
+ **执行注意:**
264
+
265
+ - 不添加“现有逆向工作都只做身份识别”这类笼统批评。
266
+ - 正文 §3.3 的 68% 事实仍必须保留,避免这里的 `effect learning` 被理解为整个系统从未使用相关候选。
267
+
268
+ ### M09|Framework 开头:用一句话分开决策、机制和评价
269
+
270
+ **优先级:**P1
271
+ **定位:**第 4 页,Section 3 标题后的引导句,约 168–169
272
+
273
+ **原文起始锚点:**
274
+
275
+ > VCDesign connects the experiments to prioritize, the solver that ranks them,
276
+
277
+ **原文结束锚点:**
278
+
279
+ > and the outcomes used to judge that choice (Figure 1).
280
+
281
+ **操作与目的:**只替换这一句;§3.1–§3.3 的公式、数据边界和组织不变。
282
+
283
+ **替换为:**
284
+
285
+ ```text
286
+ VCDesign separates the experimental decision, the scoring mechanism, and the measured utility used to evaluate the selected interventions (Figure 1).
287
+ ```
288
+
289
+ **执行注意:**
290
+
291
+ - §3 不再重复一次引言的领域宣言。这里需要的是清楚的形式化。
292
+ - 不移动 §3.3 的 68%,不把 +0.033 移回设定段,不重新定义 response-unseen。
293
+
294
+ ### M10|Section 4 开头:让两种监督的分工进入主文
295
+
296
+ **优先级:**P0
297
+ **定位:**第 5 页,Section 4 标题后的整段,约 218–221
298
+
299
+ **原文起始锚点:**
300
+
301
+ > When candidate responses are unavailable, the solver must infer which interventions could produce the desired transition.
302
+
303
+ **原文结束锚点:**
304
+
305
+ > It combines an effect predictor learned from candidate knowledge with a direct candidate-conditioned scorer (Figure 1b,c).
306
+
307
+ **操作与目的:**整段替换。主文解释“为什么是 direct scorer + effect branch”,但不新增 loss、模块或训练阶段。
308
+
309
+ **替换为:**
310
+
311
+ ```text
312
+ Response-unseen ranking requires transferring information from profiled interventions to candidates whose own responses are absent from the effect-learning atlas. VCDesign-CED combines two forms of supervision: the K562 base scorer learns candidate orderings from cross-view utility labels, whereas CED learns a mapping from candidate knowledge to measured effects. Predicted effects are then aligned with the requested transition and fused with direct compatibility scores, providing complementary, goal-conditioned evidence for ranking (Figure 1b,c).
313
+ ```
314
+
315
+ **执行注意:**
316
+
317
+ - 依据:当前 §4 和 Appendix B,尤其第 16–17 页的 base-scorer training 与 CED 定义。
318
+ - 不声称两个分支端到端联合训练;它们的现有训练及融合规则不变。
319
+ - 不把 CED 包装为不含正向预测的纯逆向算法;framework 的逆向任务与 effect branch 的正向映射是不同层次。
320
+
321
+ ### M11|Section 4.1:线性形式是参考实现,不是信息上界证明
322
+
323
+ **优先级:**P1
324
+ **定位:**第 5 页,§4.1 中 deployment 段落的最后一句,约 238–239
325
+
326
+ **原文起始锚点:**
327
+
328
+ > Its linear form tests how much transferable information candidate features carry
329
+
330
+ **原文结束锚点:**
331
+
332
+ > (richer maps: Section 5.4).
333
+
334
+ **操作与目的:**只替换最后一句;privileged information、zero-shot learning 的原引用以及 Eq. (3)–(4) 全部保留。
335
+
336
+ **替换为:**
337
+
338
+ ```text
339
+ Ridge regression serves as the reference effect-transfer map; richer alternatives are compared under the same evaluation setting in Section 5.4.
340
+ ```
341
+
342
+ **执行注意:**
343
+
344
+ - 线性模型有效或复杂变体无增益,都不能证明候选知识的信息已被完全提取。
345
+ - raw/unit target 的分工说明保持原样,不能为追求统一叙事而隐藏不同结果使用的目标。
346
+
347
+ ### M12|Section 4.2:补清 base scorer 怎样对准有限预算排序
348
+
349
+ **优先级:**P0
350
+ **定位:**第 5 页,§4.2 第一段、Eq. (5) 之前,约 245–249
351
+
352
+ **原文起始锚点:**
353
+
354
+ > A dual encoder maps (xs, dq) and kc to unit vectors uq and vc, respectively
355
+
356
+ **原文结束锚点:**
357
+
358
+ > and add them with equal weight:
359
+
360
+ **操作与目的:**替换 Eq. (5) 前的文字段落。训练目标只作一句概括,完整公式仍在 Appendix B。
361
+
362
+ **替换为:**
363
+
364
+ ```text
365
+ A dual encoder maps $(x_s,d_q)$ and $k_c$ to unit vectors $u_q$ and $v_c$, respectively (Radford et al., 2021; Karpukhin et al., 2020), giving $s_{\mathrm{base}}(q,c)=\tau\langle u_q,v_c\rangle$. The K562 scorer learns from cross-view utility labels using a replicate-informed pairwise loss and a top-20 set loss (Appendix B). We standardize each branch across the current candidate pool and add them with equal weight:
366
+ ```
367
+
368
+ **执行注意:**
369
+
370
+ - 复用本文已有数学宏和 citation keys;这里只展示可读的数学记号,不要求重命名任何宏。
371
+ - `replicate-informed` 对应已有 replicate-disagreement margin;不要换成未经证明的 `noise-optimal` 或 `unbiased ranking loss`。
372
+ - Eq. (5) 后原有一句解释两分支职责,与 M10 重复,可按下面 D01 删除。
373
+ - 上述标准化描述保留当前 §4.2 原文口径。本轮不修改 normalization 的实现;Appendix D.3 对 routing 的标准化集合另有具体说明,不能顺手把两者写成同一新规则。
374
+
375
+ ### D01|删除重复:Eq. (5) 后不再解释一次分支分工
376
+
377
+ **优先级:**P1
378
+ **定位:**第 5 页,Eq. (5) 后、cross-context deployment 句之前,约 251–252
379
+
380
+ **原文起始锚点:**
381
+
382
+ > The direct branch learns which candidate features mark useful interventions;
383
+
384
+ **原文结束锚点:**
385
+
386
+ > (details in Section B).
387
+
388
+ **操作与目的:**整句删除;其内容已在 M10 和 M12 中更具体地表达。
389
+
390
+ **替换为:**删除,不补写新句。
391
+
392
+ **执行注意:**
393
+
394
+ 保留下一个以 `In a new context, CED pools responses...` 开头的段落,不把“held context 不供训练响应”和“base scorer 固定”一并删掉。
395
+
396
+ ### M13��Section 5.4 开头:从解释小增益改为研究迁移的决定条件
397
+
398
+ **优先级:**P0
399
+ **定位:**第 7 页末,§5.4 最开始的两句,约 377;后面的 Across the tested controls 保留
400
+
401
+ **原文起始锚点:**
402
+
403
+ > Why is the masked gain modest?
404
+
405
+ **原文结束锚点:**
406
+
407
+ > (Figure 4a,b).
408
+
409
+ **操作与目的:**只替换前两句,保留现有标题 What Limits Response-Unseen Design?。标题已经提出科学问题,不必为了显得更大再换成 scaling law。
410
+
411
+ **替换为:**
412
+
413
+ ```text
414
+ We examine how training measurements, candidate knowledge, and effect-model choices support transfer to response-unseen candidates (Figure 4a,b).
415
+ ```
416
+
417
+ **执行注意:**
418
+
419
+ - 后面 natural atlas 与 row-matched atlas 的所有数值及 uncertainty 原样保留。
420
+ - 必须保留 `Row matching also lowers the number of distinct identities...` 的说明。这组控制没有完全分离测量行数与 identity coverage。
421
+ - same-identity、unmasked、raw-target 结果也保持现有条件,不混写成 masked/unit 的结果。
422
+
423
+ ### M14|Section 5.4 第二段:用具体的控制结果支撑研究含义
424
+
425
+ **优先级:**P0
426
+ **定位:**第 8 页,§5.4 第二段,约 409–414
427
+
428
+ **原文起始锚点:**
429
+
430
+ > Model changes do not raise the gain; candidate information does.
431
+
432
+ **原文结束锚点:**
433
+
434
+ > A factorial probe with a non-deployable measured-response reference agrees (Section D.5).
435
+
436
+ **操作与目的:**整段替换。明确区分 predictive basis 的小幅正向变化、richer effect maps 的无明确优势、unanchored memory 的明显退化,以及知识消融。
437
+
438
+ **替换为:**
439
+
440
+ ```text
441
+ The tested effect-model extensions do not establish an advantage over the linear reference. A predictive response basis changes BU@20 by +0.0009 [+0.0006, +0.0013], whereas the context-invariant adapter and anchored nonlinear memory show no clear gains. Unanchored memory lowers BU@20 by 0.0233 [0.0208, 0.0257] (Figure 4b). Removing STRING reduces it by 0.0211 [0.0177, 0.0235], compared with 0.0013 when MAP-KG is removed. These controls motivate improving the evidence available for transfer alongside the models that use it. A separate candidate- and query-representation probe is reported in Section D.5.
442
+ ```
443
+
444
+ **执行注意:**
445
+
446
+ - 新加入的 predictive-basis 区间已列于当前 PDF 第 37 页 Table 26,不是新实验或推导。
447
+ - `effect-model extensions` 特指相应 effect-map 变体;response basis 的变化在下一句单独交代。
448
+ - 不使用 `No model change improved on the ridge`,因为 +0.0009 的区间为正。
449
+ - 不使用 `model changes have little effect`,因为 unanchored memory 的负向变化并不小。
450
+ - 不把移除 STRING 的损失写成任何新生物学特征都必然带来同等提升;删除实验支持的是当前信息的贡献。
451
+ - Representation probe 的弱训练配方不应被提升为全模型的性能上界,故保留引用但不再用 `agrees` 直接为“瓶颈已识别”背书。
452
+
453
+ ### M15|Section 5.5:把相关性现象落到模型选择,而非泛泛否定预测指标
454
+
455
+ **优先级:**P0
456
+ **定位:**第 8 页,§5.5 整段,约 418–426;标题保持 Better Prediction Does Not Ensure Better Design
457
+
458
+ **原文起始锚点:**
459
+
460
+ > Prediction and design move together across large changes to what the effect model receives:
461
+
462
+ **原文结束锚点:**
463
+
464
+ > so selecting among models requires scoring the finite-budget ranking itself.
465
+
466
+ **操作与目的:**整段替换。保留大范围同向与局部背离两层,先给出固定 atlas 和 features 的干净 rank 对比,再交代 modality ablations。
467
+
468
+ **替换为:**
469
+
470
+ ```text
471
+ Prediction accuracy is informative for design-model selection only insofar as it identifies estimators that select higher-utility interventions. Across masked atlas and input configurations, candidate-specific prediction accuracy and effect-only design utility are positively rank-correlated within each held context (Spearman $\rho=0.31$, $0.69$, and $0.71$; stratified permutation test, $p<10^{-4}$; Figure 4c). This association does not reliably order nearby configurations. With the atlas and candidate features fixed, basis ranks 64 and 128 improve prediction but lower design utility, whereas rank 512 does the reverse. Removing MAP-KG also worsens prediction while improving design; only removing STRING lowers both. Thus, four of five configuration changes move prediction and design in opposite directions in every held context. Selecting an effect predictor for intervention design therefore requires evaluating the finite-budget ranking it induces, not prediction accuracy alone.
472
+ ```
473
+
474
+ **执行注意:**
475
+
476
+ - 使用 `configuration changes` 而不是把删除知识模态也叫作“相同信息下的 estimator changes”。
477
+ - 明确是 `candidate-specific prediction accuracy` 与 `effect-only design utility`,不声称已经系统比较所有 forward metrics、所有 virtual-cell models 或不同训练目标。
478
+ - 一个 stratified permutation test 不写成三个独��的 context-specific p-values。
479
+ - 这一节可以支撑“预测指标不足以单独选出更好的 design estimator”,不能支撑“所有预测 benchmark 都无效”或“优化预测 loss 必然损害设计”。
480
+
481
+ ### M16|Figure 1 图注:突出任务、求解器与评价的分离
482
+
483
+ **优先级:**P1
484
+ **定位:**第 2 页,Figure 1 caption 的加粗首句,约 076
485
+
486
+ **原文起始锚点:**
487
+
488
+ > The VCDesign framework and its VCDesign-CED reference realization.
489
+
490
+ **原文结束锚点:**
491
+
492
+ > The VCDesign framework and its VCDesign-CED reference realization.
493
+
494
+ **操作与目的:**只替换图注首句。a、b、c 的技术说明和图片内容不改。
495
+
496
+ **替换为:**
497
+
498
+ ```text
499
+ VCDesign separates the intervention-design decision from its solvers and evaluates the resulting finite-budget selections.
500
+ ```
501
+
502
+ **执行注意:**
503
+
504
+ - 不改变图像、面板布局、配色和数据。
505
+ - b 中 held-out replicate、query identity excluded 的说明必须保留;它们使“decision-level evaluation”具有具体可检验含义。
506
+ - 不把 a 中的 identification 示意推广成所有既有逆向模型只能固定分类的文献结论。
507
+
508
+ ### M17|Figure 4 图注:避免图题比正文结论更强
509
+
510
+ **优先级:**P0
511
+ **定位:**第 9 页,Figure 4 caption 的加粗首句,约 454
512
+
513
+ **原文起始锚点:**
514
+
515
+ > Measurements and candidate information, not the tested model changes, limit response-unseen design; prediction accuracy does not rank nearby estimators.
516
+
517
+ **原文结束锚点:**
518
+
519
+ > Measurements and candidate information, not the tested model changes, limit response-unseen design; prediction accuracy does not rank nearby estimators.
520
+
521
+ **操作与目的:**只替换图注首句;数值、误差线、面板和剩余图注保持不变。
522
+
523
+ **替换为:**
524
+
525
+ ```text
526
+ Atlas composition and candidate knowledge shape response-unseen ranking; prediction accuracy does not reliably order nearby configurations by design utility.
527
+ ```
528
+
529
+ **执行注意:**
530
+
531
+ - 不再用 `not model changes` 暗示已经排除了模型瓶颈。
532
+ - 本轮不要求改图内文字。后续若同步 panel c 文本,推荐 `Prediction–design agreement depends on the comparison`;只能改已有绘图源中的文字,不改数据或重新计算指标。
533
+
534
+ ### M18|Section 6:沿用已认可的两段结构,只同步证据边界
535
+
536
+ **优先级:**P0/已讨论文本同步
537
+ **定位:**第 8 页末至第 9 页,Section 6 全部正文,约 431、461–485;Figure 4 不属于替换范围
538
+
539
+ **原文起始锚点:**
540
+
541
+ > The results give a rule for choosing a solver by response availability.
542
+
543
+ **原文结束锚点:**
544
+
545
+ > Genome-scale libraries, merging retrieval with VCDesign-CED without dilution, and context-aware maps (Section D.14) remain open.
546
+
547
+ **操作与目的:**替换为两段,不新增第三段。第一段采用对话中已认可的结论逻辑;第二段仅对“相对瓶颈”措辞及必要 scope 作显式同步。执行时同时完成 M19,不能只删除原限制。
548
+
549
+ **替换为:**
550
+
551
+ ```text
552
+ This work introduces a decision-centric framework for cellular intervention design and shows that virtual-cell models should be evaluated not only by their ability to predict responses or identify interventions, but also by the quality of the experimental decisions they support. By formulating intervention design as finite-budget candidate ranking and evaluating selected candidates through independently measured outcomes, VCDesign provides a common interface for comparing retrieval, forward simulation, and direct scoring under the same downstream objective. Our results further show that prediction accuracy and identity recovery, while valuable intermediate capabilities, are incomplete proxies for design quality. For virtual-cell models intended to guide experiments, the central question is therefore not only whether a model can predict cellular transitions, but whether it can prioritize high-value interventions within a specified candidate library and experimental budget.
553
+
554
+ The evidence is currently limited to single-gene CRISPRi in four cell lines, and directional ranking gains have not yet established improvements in realized endpoints. Within this setting, larger perturbation atlases improve selection and removing candidate knowledge reduces it, whereas the tested richer effect maps offer no clear advantage over the linear reference. Gains also need not persist in genome-scale libraries, as the K562 stress test demonstrates (Appendix F). Future work should therefore address generalization beyond close biological neighbors, reliable prioritization at genome scale, and principled integration of measured responses with model-based inference. Independent experimental validation should determine whether better rankings translate into biological benefit. Ultimately, virtual-cell systems should be judged not only by the states they predict, but by the outcomes of the experiments they help prioritize.
555
+ ```
556
+
557
+ **执行注意:**
558
+
559
+ **与用户已认可的对话版���相比,仅有以下口径性调整:**
560
+ 1. 将 `more strongly associated with ... than with ... capacity` 改成实际观察:“更大 atlas 有益、去掉知识有损、tested richer maps 没有明确优势”。没有做过统一量纲的相对重要性检验,换成 associated 并不能自动使比较成立。
561
+ 2. 用当前四个细胞系、directional 与 realized endpoints 的区分,以及 K562 大库失效概括适用边界;细节和数字移到 M19,不再用空泛的 `Several challenges remain` 代替它们。
562
+ 3. 将最后一句落在 `outcomes of the experiments`,而不仅是“能帮科学家选实验”。这是对原结尾的评价对象精确化,并不宣称已完成前瞻性验证。
563
+ 4. 将 `realistic candidate spaces` 写成 `a specified candidate library`,避免把当前 retrospective setting 描述成已经验证完整现实实验约束。
564
+
565
+ 这一版不重新讲 retrieval/CED 的使用规则,也不重新介绍领域历史。`Future work` 的目标是扩展当前证据,不应写成已具备跨模块、跨机制或全基因组可靠设计能力。
566
+
567
+ ### M19|将被两段结尾压缩掉的具体限制集中保留在附录
568
+
569
+ **优先级:**P0,与 M18 配套。
570
+ **定位:**当前附录以 E 结束;在其后新增 **F Scope and Limitations**。这不是新增正文第三段,也不增加正文页数。
571
+
572
+ **操作与目的:**旧 §6 的限制不能因结尾改短而从论文中消失。正文保留影响核心解释的短说明,附录保留数字、适用人群和具体失败条件。若仓库中已经有等效的 scope 小节,应更新并复用,不重复新增 F。
573
+
574
+ **新增/整合文字:**
575
+
576
+ ```text
577
+ Our evidence concerns single-gene CRISPRi in four cell lines with transcriptomic readouts. Evaluation is restricted to queries whose transitions reproduce across replicate views, representing 23.5–36.8% of usable identities; this restriction concerns query eligibility, not a claim that all other candidates are excluded. Directional BU@B is the primary endpoint because the measurements support reproducible directional evaluation. Realized-outcome analyses remain mixed: PHR@20 improves in mean but not median, and the K562 realized-objective gain is +0.020 at B = 10 but slightly negative at larger budgets (Sections D.6 and D.7). Directional ranking gains are therefore not yet established as realized endpoint improvements.
578
+
579
+ The informative response-unseen comparisons rely on internal baselines; the published-method runs did not provide an informative masked-candidate comparator under the implemented protocol (Section D.6). Random gene folds leave most complex partners in the atlas, and STRING accounts for most of the observed knowledge-feature contribution, so transfer may partly exploit close biological neighbors. Module-level masking would test this possibility. In the K562 stress test, whose context trained the base scorer, the candidate library contains 6,255 identities absent from the other screens. The gain declines from +0.047 at 1,000 candidates to +0.006 at 3,000 and becomes negative at full size: −0.0103 [−0.0154, −0.0048] BU with raw targets and −0.0250 with unit targets (Figure 5). Preserving retrieval's advantage when combining it with CED, extending reliable ranking to genome-scale libraries, and learning context-aware effect maps remain directions for future work (Sections D.3 and D.14).
580
+ ```
581
+
582
+ **必须保留的实质内容:**四细胞系、单基因 CRISPRi、转录组终点;23.5–36.8% query eligibility;PHR 均值与中位数不一致;K562 realized objective 的预算依赖;response-unseen 外部比较不足;近邻/复合物伙伴可能帮助迁移;K562 +0.047 → +0.006 → 负值的尺度轨迹、raw 与 unit 区别;retrieval 融合稀释与 context-aware maps 的后续问题。
583
+
584
+ `published-method runs did not provide an informative ... comparator` 描述本文运行结果。不要改成“所有已发表方法都不支持 masked regime”:GEARS 的接口兼容性与本次拟合退化是两回事。
585
+
586
+ ## 4. 各节的最终职责与避免重复
587
+
588
+ | 位置 | 这一处只需要读者记住什么 | 不在这里重复什么 |
589
+ |---|---|---|
590
+ | 摘要 | 决策接口、CED、代表性发现及其适用层次 | 全部控制实验、全部限制或完整泛化分类 |
591
+ | 引言第一段 | 用于实验规划的模型还需要决策层成功标准 | 逐项比较 baseline 分数 |
592
+ | 引言第二段 | 既有算法是不同求解路线,不能混同任务与求解器 | 再完整定义一遍所有指标 |
593
+ | 引言第三段 | 有限预算选择与独立 outcome 评价 | 方法结构和所有 solver 的第二次长枚举 |
594
+ | 引言第四段/§4 | 历史响应如何支持未测候选,两个分支如何互补 | 把线性回归写成新学习理论 |
595
+ | §5.2 | CED 的选择收益、有效范围及对照 | 范式宣言 |
596
+ | §5.3 | response availability 与 routing 的实证规则 | 用经验规则替代全文最高层贡献 |
597
+ | §5.4 | 测量、知识、模型控制分别支持什么观察 | scaling law、信息上界���“模型不重要” |
598
+ | §5.5 | 预测质量不足以单独选择更好的 design estimator | 对所有 forward metrics 的普遍否定 |
599
+ | §6 | 总结已完成贡献,并提出后续决策系统的科学问题 | 重新写引言或再次列出全部分数 |
600
+
601
+ ## 5. 证据口径检查:允许写什么,不允许写什么
602
+
603
+ | 当前证据 | 本轮采用的表达 | 不采用的更强表达 |
604
+ |---|---|---|
605
+ | Task 与评价是有限预算候选排序、独立 replicate outcome,query identity 排除 | `decision-level evaluation` / `finite-budget selection` | 已证明真实实验效率、治疗效果或任意起终点控制能力 |
606
+ | CED 由历史响应监督 knowledge-to-effect map,并做目标方向匹配 | 目标条件化的效应迁移证据 | query-independent prior;全新的 distillation 理论 |
607
+ | masked 效应分支中 68% 候选出现过 base-scorer 训练 | effect-learning response-unseen;更严格子集另报 | 整个系统在主 masked pool 上完全未见所有候选 |
608
+ | 更多 natural-atlas 测量有益,row matching 混有 identity coverage | 更大 atlas 改善本组选择结果 | 已独立识别纯 measurement-volume 的 scaling law |
609
+ | predictive basis 有小幅正向增益;部分 richer maps 无明确优势;unanchored memory 明显退化 | 分开报告 small gain、no clear advantage、degradation | 所有 model changes 无效,或模型改动影响小 |
610
+ | 去掉 STRING 的损失大于去掉 MAP-KG | STRING 对当前模型的候选知识贡献更大 | 再加入任意知识都会产生相同幅度的收益 |
611
+ | prediction 与 effect-only ranking 在配置间可反向 | 用 prediction alone 选择 design estimator 不充分 | 所有预测指标均无效;预测目标与设计目标普遍冲突 |
612
+ | 全测量时 retrieval-only 更好,混合库 routing 优于 CED-for-all | 在各自信息条件和对照下描述收益 | routing 是普遍最优 solver;CED 优于已测响应检索 |
613
+ | directional gain 正向,realized endpoints 较弱且混合 | 区分方向性选择与实际终点改善 | 排序优势已证明更好的生物学终点 |
614
+
615
+ 这张表只约束本轮新文本,不授权重跑或重新裁决已冻结实验。
616
+
617
+ ## 6. 篇幅处理
618
+
619
+ 净压缩主要来自引言四段、删除重复的分支说明,以及以两段结尾替换原 §6。腾出的空间用于 §4 中解释监督分工、§5.4 中精确区分变体结果,以及保留主文中的必要边界。
620
+
621
+ 下表按当前 PDF 中被替换的正文与本文件英文替换稿做空白分词计数;引用、数学表达和连字符会使其与专门的英文词数工具略有差异。**它是正文文本长度对照,不是已经完成的 LaTeX 排版验证。** 新增 Appendix F 不计入正文长度变化。
622
+
623
+ | 修改组 | 原文字数 | 替换文字数 | 净变化 |
624
+ |---|---:|---:|---:|
625
+ | 摘要 | 210 | 205 | -5 |
626
+ | 引言四段与贡献 | 828 | 584 | -244 |
627
+ | Related Work 与 Framework 引导 | 156 | 146 | -10 |
628
+ | 方法说明与删除重复 | 136 | 150 | +14 |
629
+ | 结果解释 | 225 | 232 | +7 |
630
+ | 图注首句 | 27 | 33 | +6 |
631
+ | 两段结尾 | 378 | 257 | -121 |
632
+ | **合计(仅涉及替换的主文/图注)** | **1960** | **1607** | **-353** |
633
+
634
+
635
+ 编译后若正文仍超过 9 页,先删除全文中重复的 solver 枚举和相邻重复的结果解释,不删除源/目标定义、masking 边界、realized-outcome 限制、外部比较范围或 K562 大库负向结果的主文提示。不要为了多放一句“高度”而压缩图中字或修改会议模板。
636
+
637
+ 同时检查 §6 的起始位置:当前 PDF 在第 8 页末只留出开头两行,再被第 9 页顶部 Figure 4 打断。优先通过正常 float 与段落安排消除这种阅读中断,不使用负间距硬压。
638
+
639
+ ## 7. 执行后的验收与交付
640
+
641
+ ### 7.1 内容验收
642
+
643
+ 1. 所有新增的实证句都可追溯到当前结果;不出现新的数值、模型能力、实验结论或未核实文献。
644
+ 2. 三项贡献完整,CED 没有从正文贡献中消失;controlled diagnosis 也没有被删除。
645
+ 3. §5.2 三段、§3.3 的 68%、§5.5 标题与两段 §6 保持指定结构。
646
+ 4. 摘要、§5.4、Figure 4、§6 对 model/data/information 的结论强度一致。
647
+ 5. 当前所有限制在正文短说明或附录中仍可找到;M18 与 M19 必须成套落地。
648
+ 6. `prediction accuracy` 在 §5.5 指本文实际分析的 candidate-specific metric,`design utility` 指相应 effect-only ranking;不扩展成另一 benchmark 分支的研究结论。
649
+ 7. 优化求解器只出现在接口兼容性层面,不声称实证运行或击败了 optimization methods。
650
+
651
+ ### 7.2 技术验收
652
+
653
+ 在真实仓库中先确认项目已有构建命令,再运行原有编译/检查流程。检查未定义引用、参考文献、溢出行、图注与段落跨页,以及正文是否仍止于第 9 页。不得声称本文件已经完成这些源码级检查。
654
+
655
+ ### 7.3 CC / Codex 的最终交付
656
+
657
+ 提交更新后的 LaTeX 源码、编译 PDF,以及一份简短变更记录:实际修改文件与段落、M01–M19/D01 的完成情况、正文页数与编译结果、未执行项及其原因。需要额外科学判断的部分单独报告,不通过改写悄悄做决定。
658
+
659
+ ## 8. 本文件依据与未核实范围
660
+
661
+ **已读取的依据:**《main (7).pdf》正文第 1–9 页;与本轮表述直接相关的 Appendix B、D.3–D.7、E.2–E.3 和第 37 页 Table 26;本对话关于两段 Discussion 的最新决定。正文页面与 Table 26 同时检查了渲染图,未只依赖自动提取的表格数字。
662
+
663
+ **引用位置速查:**任务与 utility:第 4 页;模型与两种监督:第 5、16–17 页;masked gain 与 PHR:第 6、26 页;routing:第 7、22–23 页;严格候选子集:第 23–24 页;atlas/model/knowledge controls:第 8、35–37 页;prediction–design:第 8、19、27 页;scope 与 K562 大库失败:第 9 页及其所引附录。
664
+
665
+ **没有读取或运行:**论文仓库、LaTeX 源码、`derived_numbers.md`、结果 JSON/CSV、训练或评分脚本、submission metadata,以及任何新实验。本文件提供可执行的写作修改,不声称重新验证了原始数值、完成了文献核验或编译通过。