Paper

Normative authority: docs/SCIPAPER_STANDARD.md. This skill supplies concrete scientific-writing guidance and canonical L0 examples. It does not define an independent paper verdict. If this file, a style profile, or a workflow conflicts with the unified standard, the unified standard wins. Project-specific anchors remain marked [WGL].

Paper Writing Standards / 论文写作标准

Target: Top-tier astrophysics journals (ApJ, MNRAS, A&A level).

General Principles

• Accuracy over elegance. Never sacrifice precision for readability. Every claim must be verifiable.
• Quantitative over qualitative. Replace vague descriptions with numbers. Not "significantly improved" but "AUC increased from 0.834 to 0.927 (+11.1%)".
• Reproducibility. The Methods section must contain enough detail for an independent researcher to reproduce all results. All hyperparameters, data splits, and evaluation protocols must be specified.

Structure and Narrative

A motivation → method → validation arc is a useful default for many empirical papers, not a universal document template. Theory papers, methods papers, data releases, and multi-contribution papers may require a different shape. In every case, make the contribution graph explicit: explain how the claims relate, place definitions before use, and ensure that each conclusion is supported by the presented argument or evidence.

When the three-part arc fits:

1. Motivation: identify the scientific problem, prior limitations, and the specific gap addressed.
2. Method: present physical motivation, mathematical formulation, and implementation in dependency order.
3. Validation: state what the results establish, compare like with like, and delimit the evidence and limitations.

文章主旨 / Thesis spine — one result, everything subordinate to it

Normative authority: docs/SCIPAPER_STANDARD.md §5.4. That section is the policy; this is how to write to it. It is a writing rule, not a measured axis — no linter reports it and none may be built on it without new evidence (EVALUATION §15.1).

The arc above says what order to write in. It does not say what to leave out, and that is where drafts fail: an arc filled with everything the authors did is still an inventory. Choose the one result first, then let the arc carry it.

Before drafting or revising any section, write the thesis line. One sentence, the paper's single central result. Not the topic, not the method, not a list — the finding. Everything in the paper then exists to 铺垫 / 介绍 / 解释 / 推理 / 论证 that sentence. A second result is either subordinate to it or belongs in a second paper.

The reader must finish able to answer three questions in this order:

1. 我们做了什么 — what was done;
2. 成果是什么 — what the result is;
3. 牛逼在哪里 — what it changes that was not true before.

A draft that answers only (1) is an inventory however well written. A draft that answers (1) and (2) but not (3) has reported a measurement and not made a claim. Question (3) is the one that must be argued, not asserted: name what was previously believed, or previously impossible, and what this result replaces it with.

The inventory test. Build this table before editing, and again after:

A section with no such sentence is inventory: cut it, or fold its load-bearing clause into a section that has one. Two sections whose sentences say the same thing are one section. Any row a reader can falsify in seconds is doing its job; a row you cannot fill is the finding.

Do not confuse this with condensing. /sci-paper:condense removes what is repeated; this removes what is unranked. A section can be perfectly non-redundant, fully sourced, and still be an inventory entry.

Rank, do not delete, the supporting results. A result that does not carry the thesis is not thereby worthless: state it where it does work, in one sentence, subordinate to the claim it supports. A number that is the sole support of a claim never leaves the paper (§6 eligibility).

Never cut to sharpen. Conditions, ranges, uncertainties, scope limits, negations, and conceded limitations are load-bearing by definition and are never inventory, however secondary they sound. A paper made punchier by deleting its hedges has been damaged.

Structural Updates · Forward Narrative / 结构式更新 · 正向叙述

每一次写作、修订、纠错都必须把文章重写到"当前真值的最终态"——禁止把"旧态 → 新态"的迁移痕迹留在正文里。 本节是 paper-review §O ("Process-artifact removal") 的写作端对偶： 审查端默认 删除 过程残影；写作端默认 一开始就不写进去。

正向叙述（forward narrative）—— 唯一合法形式

正文只描述当前科学声明的最终态。读者应看到：

• ✅ "We use method B; the result is X."
• ✅ "Equation (3) gives Y, which we then evaluate at $z = 0.5$."
• ✅ "We adopt $H_0 = 67.4,\mathrm{km,s^{-1},Mpc^{-1}}$."

读者绝对不该看到（无论初稿还是修订稿）：

• ❌ "我们一开始尝试了 A，发现 A 不对，所以改用 B"
• ❌ "We did A, found it gave incorrect results, and therefore switched to B"
• ❌ "Method A would have suggested X, but in fact ..."（除非这是本文有意做的 baseline 对照，见下"唯一例外"）
• ❌ "An earlier version of Eq. (5) had a factor-of-2 error; the corrected form is ..."
• ❌ "Note: the previous draft used $H_0 = 70$; the present analysis uses $H_0 = 67.4$."
• ❌ "After several iterations we converged on the following architecture."

Patch（禁）vs Structural（要）—— 修订时的硬区分

发现错误 / 公式不对 / 数字漂移 / reviewer 指出问题时：

• ❌ Patch 式修改：在错误旁注释"原方法 A 错在 …，因此现采用 B"；保留旧公式 + 加 footnote 解释"现已修正"；保留旧数字 + 加括号给"修正后值"；段尾追加"this supersedes our earlier claim that ..."。
• ✅ Structural 式修改：删掉 A 的所有痕迹，从段首重写到段尾，让正文呈现为"从一开始就一直是 B"。读者无法从论文本身推断出 A 曾经存在过。

唯一例外 —— 真正的 baseline 对照

仅当全部三条满足时，可以在正文中保留"方法对照"：

1. 被对照的方法是 领域内已发表的外部 baseline / prior published method——不是作者自己的早期迭代；
2. 本文对该对照做了正式 head-to-head 实验并给出数值；
3. 呈现方式是 "contrast with baseline X (Smith+2020) on protocol P → numerical comparison"，不是 "we initially tried X" 这种第一人称自传式。

判定时若任一条不满足 → 默认按 "Patch（禁）" 处理，重写到只剩当前方法。

写作时强制自检（每段 / 每次修订执行）

1. 这段是不是只描述了当前最终态？（如果包含过去态，删掉过去态。）
2. 一个从未参与本研究的读者，看了这段会觉得有别的方案 / 旧版本存在过吗？如果会 → 重写。
3. 修订时新增的内容是"重写过的最终态"还是"在旧态上贴的补丁"？是补丁 → 撕掉补丁、重写整段，不要"补丁 + 解释为什么打补丁"。
4. 是否出现过程时间词指向本文自身研究进程？任一命中 → 重写。（指外部科学时间维度的不算，例如 "recent supernova observations" / "previously published catalogs"。）词族与 tools/deai_residue.py 同源，validate_plugin 比对两处；检测端（residue-self-history）只在同句带 we / our / this paper 且无引用时报：
   • 高精度（strong）：previous draft、earlier draft、previous version、earlier version、supersedes、superseded、no longer、we tried、we switched、we abandoned、we replaced、we moved away from
   • 普通（ordinary，常见科学用词，只提示；initially/originally/at first 在已发表论文里多是流程顺序而非改稿史）：initially、originally、at first、now、currently、earlier、previously、corrected、revised、updated
5. 是否有句子用缺席来描述本文自己的对象（菜单上写"番茄炒蛋（无东坡肉）"）？"The head never participates in the decision" / "the reference carries no quoted number" / "no support threshold is applied, because …" 说的都是对象不做什么，读者看到的是一个空位而不是那件东西；改成它做什么："The head runs downstream of the decision" / "a labeled reference only"。先删后写：结合上下文判断，相邻句子已经说了对象做什么的，整句或该分句直接删（"downstream of detection" 在一篇稿子里出现五次，就是只改写不删的后果）；只有页面上还没有的信息才改写成正向陈述。删掉的否定词会被 rewrite_reward 的 §6 否定不变量报为 missing，按同一 disposition 记录。物理事实、范围限制、让步性局限（"the B-mode map carries no lensing signal"、"shear alone cannot break the mass-sheet degeneracy"）是承重的，保留并记录 disposition；检测端 residue-absence 对同句带引用的对照不报。词族与 tools/deai_residue.py 同源，validate_plugin 比对两处；分档依据 442 篇 held-out 已发表全文（190 万词）：strong 档在其中每千词 0.008 次，ordinary 档 0.02–0.05 次且多为流程语句：
   • strong：never、nothing is、nothing was、none sees、none enters、no ... is applied（模板：以 no 开头，接一到三个词的对象，再接 is/are/was/were + applied/used/combined/quoted/imposed）
   • ordinary：carries no、carry no、carrying no、holds no、touches no、sees no、enters no、makes no、is not applied、are not applied、is not used、are not used、is not quoted、is not combined、is not treated、are not treated、is not reproduced、does not participate、do not participate、does not enter、does not claim、does not see、do not see

与 "Formula Derivation Standards" 中"射箭画靶"的关系

下节的"No shooting arrow then drawing target / 禁止射箭画靶"是本规则在公式推导场景的窄特例（"我想得到 X 结果，所以改了推导步骤"）。本节是更广的写作准则——覆盖正文叙述、方法描述、结果呈现、讨论、结论：整篇论文的任一段落都适用。

改写不堆叠 / Condense, Do Not Accumulate

规范条文：docs/SCIPAPER_STANDARD.md §5.3。用户规则 2026-07-16： "改写、删减、精简，而不是堆叠！"

• 每次修改的默认方向是更短。优先级：删除 > 原位精简 > 等长改写 > 增长。
• 增长只有两种合法理由：用户要求的新内容，或来源可验证的科学必需 （缺失的假设/定义/单位/caveat 属 integrity 缺陷）。
• 典型违规是解释性补丁：对被标记的句子追加从句、句子或脚注去"解释"， 而不是重写句子本身。上节 forward narrative 禁止堆叠状态；本节禁止 堆叠字数。
• 每处修改报告字数差；靠加字消除 detector 信号是缺陷，不是修复。
• 机械执行（标准 §5.3 v3.3）：改写候选经 rewrite_reward.py --original 硬门（超长即 -inf）；编辑循环收尾经 length_gate.py delta 门 （无理由净增长 = exit 1，循环不得收尾）。增长的唯一合法路径是 --allow/--allow-growth 记录的作者批准理由。

Formula Derivation Standards / 公式推导规范

• Multi-line derivations: Use align/gather environments for complete mathematical derivations, not single-line equations. Show the logical chain: a = b (1), then a = c (2), therefore b = c (3).
• Definition completeness: Every variable, compound term, or logical construct appearing in a formula MUST be either (a) previously defined in the text, or (b) defined/derived immediately near the formula. Never introduce undefined symbols.
• No inline formulas for complex expressions: Any formula longer than ~30 characters must be a displayed equation, not inline text. Short expressions (e.g., $\kappa \ll 1$) can remain inline.
• Logical flow over format: Don't force a rigid template. Derivations should flow naturally — define when needed, derive when needed, summarize at the end. The priority is that reasoning is clear and logically connected.
• No "shooting arrow then drawing target" / 禁止射箭画靶: Never write "we wanted X result so we changed to Y approach" or reference historical/deprecated formulas. Present: method → result → conclusion. Do not discuss the iterative path that led to the current approach. （公式推导场景的窄特例；广义写作准则见上节 "Structural Updates · Forward Narrative"。）
• No outdated content: Only present current formulas and label definitions. Do not reference deprecated versions in the paper body — at most a brief footnote if essential for context.
• Summary block: After a derivation chain, include a brief summary: "We therefore obtain [final formula], where [key quantities] are [definitions]. This shows [physical conclusion]."

Physics Descriptions [WGL]

• Describe weak lensing physics using standard notation (Bartelmann & Schneider 2001 conventions).
• Clearly distinguish between: convergence (kappa), shear (gamma), reduced shear (g), aperture mass (M_ap), and S/N maps.
• When describing the detection pipeline, maintain the distinction between signal (E-mode) and noise (B-mode).
• All filter functions must be written with explicit mathematical definitions, not just names.

Method Descriptions

• For each ML model, specify: architecture, input representation, loss function, training procedure, and evaluation metric.
• For ensembles, explain the aggregation strategy and why it is appropriate given any class imbalance.
• Clearly state any transfer learning protocol: what was pretrained, on what data, and how fine-tuning differs.
• For grouped/leave-one-out CV: explain why grouping is necessary and how group IDs prevent leakage.

Results Presentation

• All performance metrics must include: (a) the metric name and definition, (b) the evaluation protocol, (c) uncertainty estimates where possible.
• Tables should be self-contained — a reader should understand the table without reading the text.
• Figures should have: descriptive captions, labeled axes with units, legends, and consistent color schemes.
• When comparing methods, use the same evaluation protocol for all. Never compare training metrics of one model to validation metrics of another.

Discussion and Limitations

• Honestly discuss limitations. Acknowledge sample-size and selection-effect limits.
• Distinguish between: limitations of the method vs. limitations of the data.
• For detection-boundary analyses: discuss where sim-to-real transfer breaks down and why.
• Avoid overclaiming. Detection frameworks are not definitive physical measurement tools.

Anti-AI-isms / 去 AI 表达规范

LLM 生成的学术写作有一组明显的 tell；本节保留既有 L0 词汇与标点 目标，同时把结构、信息分布和 learned field-similarity 信号纳入统一反馈协议。 style-profile/<field>/style_dossier.md、lexicon 和 baseline 是可更新的经验 证据，不是独立政策，也不能把论文判为 AI 或非 AI。后续动作由 docs/SCIPAPER_STANDARD.md 的 consequence class、measurement state、ranking 和 disposition 规则决定。

根本层（fundamental）—— 结构性 AI 味，关键词 lint 抓不到。 本节后面 列出的 Tier A/B 是词汇层（lexical），必要但不充分：一篇文章可以 0 关键词命中、 甚至逐段读着都像人，却仍通篇 AI 味。真正的 tell 活在结构里，分两个尺度：

1. 信息分布层（token / 句长）：过度均匀的信息密度、句长同质、重复的 signposting 和缺少局部节奏变化。由 distribution 与 UID axes 度量。
2. 句式与文档形状层：句子或段落的构造被重复模板化。需要重点检查：
   • 报数式枚举：rests on five elements. First, ... Fifth, ... / there are three reasons。
   • 先设数目 → 列举 → 收尾：inherits three obligations. [A][B][C]. These three requirements ...。
   • 排比 / 首语重复：≥3 句同一开头或同一模态（must ... must ... must）。
   • 对称收尾：A is one limit of it, and B another。
   • 段落或章节同形：多个段落重复相同的主题句、展开和收束几何。

这些模式不是单次出现即错误。它们在适用 baseline 下构成测量证据；阈值、样本量、 置信度和效应量属于 EVALUATION.md 或 profile calibration，不写死在规范里。

这一层由 de-AI 子系统统一度量（docs/architecture/DEAI_SUBSYSTEM.md）：

python tools/ai_ism_lint.py <file> --field <field> \
  --structure --distribution --document-structure --oracle --voice \
  --format json --output <scratch>/writing-feedback.json

• Tier A、em-dash 和超过每节每词 cap 的 Tier B 是 l0_target。
• 句式模板、burstiness、UID、document shape 与 learned field-similarity 是 advisory；必须保留 measured / degraded / unmeasured / not_applicable 区别，不能把缺失测量当作零命中。
• 命中的段可用 /sci-paper:de-ai（Pass 3）从 claim graph 重建，而不是做 同义词替换。任何候选先通过 scientific-fidelity eligibility，再比较风格证据。
• 强 advisory 必须行动或显式 disposition；普通 advisory 可以保留并报告。

因此，Tier A / em-dash 清零与 Tier B cap 是 L0 地板；结构和信息分布信号用于 排序后续动作，不构成必须全部归零的通用 prose gate。

最强 L0 标点目标：em-dash (— / \textemdash / ---)

• 正文目标为 0。插入语改用逗号、括号、分号或独立句；范围使用 --。
• 该规则是项目锁定的 L0 policy。当前 corpus 频率及比较值只在 profile 和 EVALUATION.md 中维护，避免把会漂移的测量写进规范。

Tier A — L0 target

正文命中必须重写。canonical set：

2026-07-16 扩充（underscore*, pivotal, tapestry, testament, realm*） 采自 academic-humanizer 词表（MIT，见文末 Provenance），并经两域 curated corpus 复核为 0 出现后才入 Tier A；landscape 虽在其词表中，但它是本领域 正当术语（detection landscape 等，corpus 高频），不入表。

替换原则：使用直接、具体、可核验的动词或范围，不做机械同义词交换。例如 leverages X to Y → uses X to Y，pave the way for → enable， comprehensively → 明确列出覆盖范围。

Tier B — per-section/per-word cap

Tier B 可以使用，但同一个 Tier B 词在同一 section 最多出现 1 次。第 2 次及 以后是 l0_target；cap 内的出现不是 finding。当前词表由 linter 与 profile 共同维护，常见项包括 Furthermore, Moreover, Additionally, robust/robustly, comprehensive, utilize/utilized, leverage, Importantly, Interestingly, Notably, intricate, foster/fosters/fostering/fostered（后两组 2026-07-16 加入；curated corpus 各有 1 次出现，非零故不入 Tier A）。经验频率只从当前 profile 读取，不在 本文件复制。优先用直接陈述或可验证数字，但不要为了避词而损害准确性。

模糊量化 / 修饰词

• a wide range of, a variety of, a number of, several, numerous, many：有可核验数量时写数量；没有时检查该模糊程度是否科学必要。
• cutting-edge, state-of-the-art, novel, powerful：需要明确比较对象和证据， 否则删除。它们是 claim-quality advisories，不因单词本身自动成为 L0 target。

自指与套话

• This paper presents... / In this work, we... 类 boilerplate 每段最多 1 处。
• 删除：In summary,, To summarize,, In conclusion,（除 conclusion 节外）。

LLM 高频动词替换（dossier 未实证但语法层面是 tell）

• facilitate → enable
• In order to → To
• aim to → we [verb]（直接动词）
• serves as → is（copula 回避；linter style-substitution advisory）

结构 tell（L2 advisory）

• repeated parallel frames，例如连续三句相同首语或 modal；
• announced enumeration 与 setup/list/wrap-up symmetry；
• 多段重复相同开场、展开和收束几何；
• X — that is, Y 同时触发 em-dash L0 target 与可能的冗余 advisory；
• 分词尾巴（-ing tail）：..., highlighting/underscoring/demonstrating X 把解读挂在句尾冒充分析深度。改写为带主语和证据的独立句，或删除 （linter ing-tail advisory）；
• 阐释式冒号（colon-appositive）：X: the rule that ... 这类 "名词: 展开" 结构是 X — that is, Y 的冒号变体。改写为限定从句、 两个句子或 ", so ..."；caption 标签（Left: ...）与真正的列表规格 说明可保留（linter colon-elaboration advisory；用户规则 2026-07-16）。

结构模式必须结合 section、样本量、calibration 和科学功能判断。技术列表若编码真实 分类，不应为了制造参差而破坏可读性。

punctuation / 排版

• 数字与单位之间用 \,（thin space），不要 LLM 习惯的普通空格。
• 千分位用 \,（thin space）或 ,，不要无分隔。
• 不要在文中写 etc.（学术写作可接受但 LLM 滥用），改为完整列举或具体范围。

review 阶段的强制 grep

# em-dash：必须 0
grep -n -E '—|---|\\textemdash' main.tex

# Tier A（必删；正文中不允许出现，包括变体）
grep -n -E -i '(delve|leveraged|leverages|leveraging|paved?|paves|paving|shed[s]?|shedding|showcase[sd]?|showcasing|seamless(ly)?|holistic(ally)?|comprehensively|crucially|utilizes|utilizing|underscor(e|es|ed|ing)|tapestry|testament|pivotal|realms?|recent advances|despite significant|with the advent|in recent years|it is worth)' main.tex

# Tier B（定位；是否超过每节每词 cap 由 linter 按 section 计算）
grep -n -E -i '^\s*(Furthermore|Moreover|Additionally|Importantly|Interestingly|Notably),' main.tex
grep -n -E -i '\b(robust|robustly|comprehensive|utilize|utilized|leverage|intricate|foster(s|ing|ed)?)\b' main.tex

# 顽固替换组（不分级）
grep -n -E -i '\b(in order to|aim to|facilitate|serves as)\b' main.tex

# 结构 advisory 定位（-ing 尾巴 / 阐释式冒号）
grep -n -E -i ',\s+(highlighting|underscoring|showcasing|emphasi[sz]ing|illustrating|demonstrating|signal[l]?ing|revealing|reflecting)\b' main.tex
grep -n -E '([A-Za-z0-9]|\}): [a-z$\\]' main.tex

Tier A / em-dash 残留 = l0_target。 Tier B 超频 = 同词在同 section 的第 2 次及以后为 l0_target。

Companion evidence from /sci-paper:de-ai calibration: corpus assets supply descriptive frequencies and calibration. Re-run python tools/extract_style.py when the corpus changes. They may suggest future policy changes, but do not silently redefine the current consequence classes or cap.

Claim–Evidence Discipline / 声明-证据纪律

QD 类规则（claim-evidence defects 在 SCIPAPER_STANDARD §2 QD 下是 integrity_blocker）。本节给出操作化检查；条目改编自 academic-humanizer Layer 4（MIT，见文末 Provenance），并按天体物理语料重校准。

对每个经验性声明检查两件事：(a) 它是否有正文内的数字、图、表或引用支撑； (b) 动词强度是否不超过证据强度。

• 无支撑声明 → 补证据指针或降级。 ❌ Our method is more robust. ✅ Our method's accuracy drops by 2 points under distribution shift, versus 11 points for the baseline (Figure 3).
• 动词强于证据 → 降级。 ❌ This demonstrates that our method is universally superior. ✅ On these three datasets, our method matches or exceeds the strongest baseline (Table 2).
• 模糊量级 → 有归属的数字或区间。 ❌ a large improvement. ✅ a 2--6% improvement in balanced accuracy over the strongest baseline. 区间优于单一均值（除非均值方法已声明）；每个数字注明方法、指标、基线。 做比较时先打最强对手，不打平凡基线。
• significantly 必须有伴随检验或数字；孤立的 "significantly better" 是声明缺陷。注意这是证据条件规则，不是词法禁令：astro curated corpus 中 demonstrate*（0.147/1k）与 significantly（0.274/1k，合并语料实测 2026-07-16）都是正常用词，禁词式移植（ML 会议口味）会误伤本领域写作—— 只有"动词/副词超出证据"才构成 finding。

防过度纠正 / Preserve List（rewrite 护栏）

改编自 academic-humanizer Layer 3（MIT）。De-AI 重写循环的反向风险是 "把校准的 hedging 改强"——这会制造 over-claiming，比留下 AI 词更糟。 与 SCIPAPER_STANDARD §6 rewrite eligibility（stance/modality/qualifier 不可变）同源；此处是写作端明细。

• 证据绑定的 hedging 是正确且必需的。 suggests, is consistent with, we hypothesize that, may indicate, appears to 在声明真有 不确定性时保留。把 "the results suggest X" 改成 "the results prove X" 是制造 over-claim，属 rewrite eligibility 违规。
• 被动语态在施动者无关时合法："Samples were normalized to total protein." 不要为主动而主动。
• 第一人称复数 "we" 是学术标准，不改写回避。
• 分号与偶发三联适度可用；em-dash 是唯一零容忍标点。
• 正式定义、命名方法/指标、术语、公式、符号逐字保留。
• 数字、公式、引用永不发明、丢弃或改动；cite key 全保留。

Citation Standards

• Only cite papers that genuinely support the claim being made.
• For established results, cite the original paper, not a review (unless the review adds value).
• For software: cite the primary paper for each library.
• Never fabricate or hallucinate citations. If unsure, flag with [CITATION NEEDED].

Key References / 关键参考文献 [WGL]

• Weak lensing formalism: Bartelmann & Schneider (2001), Schneider et al. (1998)
• Aperture mass / Schirmer filter: Schirmer et al. (2007), Schneider (1996)
• NFW profile: Navarro, Frenk & White (1996, 1997)
• E(2)-equivariant CNNs: Weiler & Cesa (2019), escnn library
• Swin Transformer: Liu et al. (2021)
• SBI / Neural Posterior Estimation: Cranmer, Brehmer & Louppe (2020)
• Persistent homology / TDA: Edelsbrunner & Harer (2010)
• LoVoCCS survey: Fu et al. (2022)

Provenance / 借鉴出处

The 2026-07-16 additions (Tier A/B word extensions, serves as, the -ing-tail and colon-elaboration structure tells, the Claim–Evidence Discipline section, and the Preserve List) adapt material from academic-humanizer (MIT License, Copyright (c) 2026 AIScientists-Dev, itself building on blader/humanizer, MIT). Every lexical adoption was re-verified against the curated field corpora before tier assignment; venue-specific rules that conflict with astro usage (landscape, blanket demonstrate/ significantly bans) were deliberately not adopted.