Paper Write

ANONYMOUS = false — IEEE Trans papers are NOT anonymous.

模型路由规则详见 ../shared-references/model-routing-policy.md。

Inputs

1. PAPER_PLAN.md — outline with claims-evidence matrix, section plan, figure plan (from /paper-plan)
2. NARRATIVE_REPORT.md — the research narrative (primary source of content)
3. Generated figures — PDF/PNG files in figures/ (from /paper-figure)
4. LaTeX includes — figures/latex_includes.tex (from /paper-figure)
5. Bibliography — existing .bib file, or will create one

If no PAPER_PLAN.md exists, ask the user to run /paper-plan first or provide a brief outline.

Project Automation Policy

Before drafting, resolve automation defaults in this precedence order:

1. Inline command arguments
2. PROJECT_AUTOMATION.md in the project root
3. CLAUDE.md in the project root
4. The constants in this skill

If AUTO_PROCEED=false, stop after producing the outline-consistency check, after drafting the Abstract + Introduction, and before the final whole-paper polish pass.

Templates

IEEE Transactions Template

\documentclass[journal]{IEEEtran}

\usepackage{amsmath,amssymb,amsfonts}
\usepackage{algorithmic}
\usepackage{graphicx}
\usepackage{textcomp}
\usepackage{xcolor}
\usepackage{cite}
\usepackage{bm}
\usepackage{mathrsfs}
\usepackage{booktabs}
\usepackage{multirow}
\usepackage{subfigure}

% Theorem environments
\newtheorem{theorem}{Theorem}
\newtheorem{lemma}[theorem]{Lemma}
\newtheorem{proposition}[theorem]{Proposition}
\newtheorem{corollary}[theorem]{Corollary}
\newtheorem{definition}{Definition}
\newtheorem{assumption}{Assumption}
\newtheorem{remark}{Remark}
\newtheorem{problem}{Problem}

\begin{document}

\title{Paper Title Here}

\author{First Author,~\IEEEmembership{Member,~IEEE,}
  Second Author,~and~Third Author,~\IEEEmembership{Senior Member,~IEEE}
\thanks{This work was supported by ...}
\thanks{F. Author is with the Department of ..., University of ..., City, Country (e-mail: ...).}}

\markboth{IEEE TRANSACTIONS ON ..., VOL. XX, NO. X, MONTH YEAR}
{Author \MakeLowercase{\textit{et al.}}: Short Title}

\maketitle

\begin{abstract}
...
\end{abstract}

\begin{IEEEkeywords}
Keyword1, keyword2, keyword3.
\end{IEEEkeywords}

\input{sections/1_introduction}
\input{sections/2_preliminaries}
\input{sections/3_main_results}
\input{sections/4_simulation}
\input{sections/5_conclusion}

\appendices
\input{sections/A_appendix}

\bibliographystyle{IEEEtran}
\bibliography{references}

% Author biographies (optional, for final version)
% \begin{IEEEbiography}[{\includegraphics[width=1in,height=1.25in,clip,keepaspectratio]{bio_photo.jpg}}]{Author Name}
% Biography text here.
% \end{IEEEbiography}

\end{document}

Project Structure

Generate this file structure:

paper/
├── main.tex                    # master file
├── math_commands.tex           # shared math macros
├── references.bib              # bibliography (only cited entries)
├── sections/
│   ├── 1_introduction.tex
│   ├── 2_preliminaries.tex
│   ├── 3_main_results.tex      # or split into 3_observer.tex + 4_controller.tex
│   ├── 4_simulation.tex        # or 5_simulation.tex if split
│   ├── 5_conclusion.tex
│   └── A_appendix.tex          # proof details
└── figures/                    # symlink or copy from project figures/

Section files are FLEXIBLE: If the paper plan has 6-8 sections, create corresponding files.

Workflow

Step 0: Backup and Clean

If paper/ already exists, back up to paper-backup-{timestamp}/ before overwriting.

CRITICAL: Clean stale files. When changing section structure, delete section files no longer referenced by main.tex.

Step 1: Initialize Project

1. Create paper/ directory
2. Set up main.tex with IEEEtran document class
3. Generate math_commands.tex with paper-specific notation
4. Create section files matching PAPER_PLAN structure

Author block (placeholder):

\author{First~Author,~\IEEEmembership{Member,~IEEE,}
  and~Second~Author
\thanks{Manuscript received ...; revised .... This work was supported by ....}
\thanks{Authors are with the Department of ..., University of ..., City, Country (e-mail: ...).}}

Step 2: Generate math_commands.tex

Create shared math macros for control systems notation:

% math_commands.tex — shared notation for control systems
\newcommand{\R}{\mathbb{R}}
\newcommand{\N}{\mathbb{N}}
\newcommand{\norm}[1]{\left\| #1 \right\|}
\newcommand{\abs}[1]{\left| #1 \right|}
\newcommand{\tr}{\mathrm{tr}}
\newcommand{\diag}{\mathrm{diag}}
\newcommand{\sgn}{\mathrm{sgn}}
\newcommand{\rank}{\mathrm{rank}}
\DeclareMathOperator*{\argmin}{arg\,min}
\DeclareMathOperator*{\argmax}{arg\,max}
\DeclareMathOperator{\col}{col}
\DeclareMathOperator{\Span}{span}

% Vectors and matrices (bold)
\newcommand{\vx}{\bm{x}}
\newcommand{\vu}{\bm{u}}
\newcommand{\ve}{\bm{e}}

% Calligraphic sets
\newcommand{\calG}{\mathcal{G}}
\newcommand{\calN}{\mathcal{N}}
\newcommand{\calV}{\mathcal{V}}
\newcommand{\calE}{\mathcal{E}}
\newcommand{\calL}{\mathcal{L}}

% Kronecker product
\newcommand{\kron}{\otimes}
% Add paper-specific notation here

Step 3: Write Each Section

Before drafting: Read ../shared-references/writing-principles.md for sentence-level clarity rules and IEEE Trans conventions.

When the outline is frozen and file ownership is disjoint, you may draft in parallel using these slices:

• frontmatter: Abstract + Introduction + Conclusion framing
• theory: Method + Main Results + Appendix proof scaffolding
• results: Simulation section + tables + captions + bibliography cleanup

Draft the slices with Sonnet, review them with gpt-5.4, then merge serially and let Opus perform the final whole-paper integration pass.

If the parent session is running on Opus or any non-Sonnet model, do not bulk-draft these slices in the parent session. Spawn dedicated Sonnet workers instead:

spawn_agent:
  model: sonnet
  reasoning_effort: high
  message: |
    You own the [frontmatter/theory/results] slice only.
    Draft LaTeX for this slice from PAPER_PLAN.md, NARRATIVE_REPORT.md, and existing figures.
    Do not change the claim boundary. Do not rewrite other slices.
    Output only grounded manuscript text and any local TODOs that block compilation.

Use one worker per disjoint slice. Merge only after all slice reviews are back.

Process sections in order. For each section:

1. Read the plan — what claims, theorems, citations belong here
2. Read NARRATIVE_REPORT.md — extract relevant content
3. Draft content — write complete LaTeX (not placeholders)
4. Insert figures/tables — use snippets from figures/latex_includes.tex
5. Add citations — use \cite{key} (IEEE numbered style)
6. Run a de-AI pass — delete generic filler, hype adjectives, and repeated sentence templates

Section-Specific Guidelines (IEEE Trans Control)

Abstract:

• Must be self-contained, 150-200 words
• No citations, no undefined acronyms, no equations
• Remove any sentence that could fit another paper unchanged
• 结构（四段式，从真实论文中提炼）:
  1. 问题陈述（1-2句）：直接说明研究了什么问题、什么系统、什么条件。不需要解释为什么重要——假设读者已经关心这个领域。
  2. 方法概述（1-3句）：说明用了什么技术手段（控制器设计、观测器、事件触发机制等）。
  3. 主要理论结果（1-2句）：说明证明了什么（稳定性、收敛性、充要条件等）。
  4. 验证（1句）：仿真/实验验证，作为最后一句。
• 开头：第一句话直接切入问题本身，不要用领域背景引入。
  • 典型开头（从真实论文中提取）：
    • "This article focuses on the formation control problem of nonlinear multiagent systems under denial-of-service attacks."
    • "Time-varying formation tracking problems for linear multi-agent systems with multiple leaders are studied."
    • "The containment control is studied for second-order multiagent systems over a heterogeneous network."
    • "This article studies the fuzzy adaptive event-triggered resilient formation control problem for nonlinear MASs."
  • 禁止的开头（立即删除）：
    • "In recent years, X has garnered/attracted significant attention..."
    • "In the rapidly evolving landscape of..."
    • "With the ever-increasing demand for..."
    • 任何以领域重要性开头的句子
• 语气：陈述事实，不评价自己的工作。禁止使用：novel, innovative, groundbreaking, superior, remarkable, excellent, cutting-edge, state-of-the-art (描述自己的方法时)。只有在最后验证句中可以用 "effectiveness" 或 "feasibility"。
• 语态：以被动语态为主（"is studied", "are designed", "is proposed"），偶尔用 "we"。不要全篇都用同一种语态。

Index Terms (IEEEkeywords):

• 4-6 terms from IEEE taxonomy
• Example: "Denial-of-service (DoS) attacks, event-triggered control, formation control, multi-agent systems (MASs), Lyapunov stability."

I. Introduction（引言的核心是讲故事——用逐层递进的逻辑把读者带到你的研究空白处）:

引言的叙事逻辑是一个漏斗结构，从广到窄，每一段都为下一段做铺垫。段落之间必须有因果或递进关系，不能各自独立。以下是从真实论文中总结的段落推进逻辑：

• 第一段：广域背景（1段）

  • 用1-2句话交代研究领域的大背景和应用场景。不要长篇大论讲领域多重要。
  • 然后快速聚焦到你要研究的具体子问题（如编队控制、包含控制、事件触发等）。
  • 典型写法：先提多智能体系统的应用（2-3个具体例子），然后一句话转到具体问题。
  • 例（技术性开头，避免套话）："Consensus-based distributed formation control requires each agent to compute its control input using only information from neighboring agents [1]-[5]. When the communication topology is directed and time-varying, the standard protocols in [6]-[9] are no longer directly applicable."
  • 避免套话开头如 "has attracted significant interest due to..."——这种句式在真实论文中虽然存在，但也是AI最常生成的模式。尽量用更技术性的方式开头。

• 第二段起：主题文献综述（2-4段，每段一个方向）

  • 每段聚焦一个技术方向（如：编队控制方法 → DoS攻击下的控制 → 事件触发机制），而不是按时间顺序列举。
  • 每段内部逻辑：概述该方向已有成果 → 指出其局限 → 用过渡句引向下一段。
  • 段间承接是关键。用以下方式自然过渡（不要机械套用，要根据内容选择）：
    • 因果关系："Since the communication among agents relies on wireless channels, the system is vulnerable to..." （上一段讲通信拓扑，这一段自然引入攻击问题）
    • 补充关系："Besides formation control, another practical concern is..." / "On the other hand, ..."
    • 对比关系："However, the above works assume continuous communication, which consumes significant bandwidth. To reduce communication burden, event-triggered mechanisms..."
    • 深化关系："Notice that the results in [X]-[Y] are restricted to linear systems. When nonlinear dynamics are considered, ..."
  • 禁止：连续多个段落都以 "In [X], ..." 开头；连续使用 "Furthermore", "Moreover", "Additionally"。
  • 文献引用方式：不要逐篇罗列。把相关工作分组引用 [X]-[Y]，然后综合评述它们的共性和局限。偶尔单独讨论1-2篇最相关的工作并指出其具体局限。

• 研究空白段（承上启下的关键段）

  • 这一段是整个引言的转折点。前面的文献综述应该自然地把读者引到这里——空白应该是从综述中"涌现"出来的，而不是突兀地宣布的。
  • 典型写法：
    • "However, the aforementioned works [X]-[Y] did not consider the situation where..."
    • "It is worth noting that most existing results are restricted to... When [more challenging condition] is considered, the problem becomes substantially different."
    • "To the best of the authors' knowledge, [specific problem] under [specific conditions] has not been addressed."
  • 然后用一句话过渡到本文的工作："Motivated by the above observations, this article investigates..."

• 贡献列表（编号列表，2-4项）

  • 每一项都是具体的、可验证的技术贡献，并与已有工作做对比。
  • 格式："(N) Different from [prior work] which [limitation], this paper [specific contribution]." 或 "(N) A [specific thing] is designed/proposed for [system]. Compared with [refs], [specific advantage]."
  • 每一项2-4句话，不要写得太空泛。

• 文章结构段："The rest of this article is organized as follows. Section II introduces... Section III presents... Section IV provides... Section V concludes this article."

• 符号段："Notations: $\mathbb{R}^n$ denotes the $n$-dimensional Euclidean space. $|\cdot|$ represents the Euclidean norm. $I_n$ denotes the identity matrix of dimension $n$. $\lambda_{\min}(A)$ and $\lambda_{\max}(A)$ stand for the minimum and maximum eigenvalues of matrix $A$, respectively. $\otimes$ denotes the Kronecker product."

• Target: 2-2.5 pages (two-column)

• 降低AI率的具体要求：

  • 句式要有变化：长短句交替，主动被动交替。不要每句话都是 "In [X], the authors proposed..."
  • 段内信息密度要高：一段话里综合评述3-5篇工作的共性，然后点出1-2篇的具体局限，比逐篇罗列更自然。
  • 过渡要基于内容逻辑，不要用万能连接词。"考虑到通信依赖无线信道"比 "Moreover" 更有信息量。
  • 每段的第一句话要承接上一段的结论或问题，最后一句话要为下一段铺垫。
  • 贡献列表必须让审稿人能指向后文的定理、图或表；如果做不到，就说明这句话太空泛。

II. Problem Formulation / Preliminaries:

• Subsections (typically A-D):
  • A. System Model — state-space equations with full definitions
  • B. Graph Theory / Communication Topology — adjacency matrix, Laplacian, assumptions on connectivity
  • C. Attack Model — DoS attack intervals, frequency/duration constraints
  • D. Control Objective — formal problem statement
• Assumptions: Numbered, each followed by a Remark justifying it
  • "Assumption 1: The communication graph $\mathcal{G}$ contains a directed spanning tree."
  • "Remark 1: Assumption 1 is standard in the literature [X], [Y] and ensures..."
• Prerequisite Lemmas: State lemmas from prior work with citations
• Heavy on mathematical definitions, light on prose
• Target: 2-2.5 pages

III. Main Results (Controller/Observer Design + Stability):

• Step-by-step design procedure with clear motivation for each step
• Theorem environments: Use \begin{theorem}...\end{theorem} for all main results
• Proof structure:
  • Construct Lyapunov function candidate
  • Compute time derivative
  • Apply Lyapunov stability theorem
  • For DoS: analyze attack/non-attack intervals separately, then combine using switching analysis
  • For cooperative-competitive: use super-stochastic matrix convergence theory
• Remarks after each theorem: Explain significance, compare with prior results, discuss special cases
  • "Remark X: Compared with [prior work], the proposed controller relaxes the requirement of... This is because..."
  • "Remark Y: It is worth noting that when [special case], Theorem 1 reduces to..."
• Proof placement: Short proofs (< 0.5 page) in main body; long proofs in Appendix with "Proof: See Appendix A."
• Target: 3-4 pages

IV. Simulation Results:

• Setup paragraph: System parameters, initial conditions, topology graph, attack scenario description
• Parameter table: Table I with all simulation parameters
• Figures organized by scenario:
  • Scenario 1: Basic performance (no attacks) — trajectories, errors
  • Scenario 2: Under attacks — trajectories, errors, control inputs, event instants
  • Scenario 3 (optional): Comparison with prior method — show degradation
• Figure descriptions: "Fig. X shows the state trajectories of all agents. It can be observed that..."
• Quantitative discussion: "From Fig. X, the formation errors converge to zero within Y seconds under DoS attacks, which validates Theorem Z."
• Event-triggered results: "The total number of triggering events is N, which is significantly less than... This demonstrates the communication efficiency of the proposed scheme."
• Target: 2-3 pages

V. Conclusion（结论写作原则——从真实论文中提炼）:

结论应当简短、克制、事实性强。从35篇真实论文中观察到：结论通常只有1段，5-10句话，绝不超过半栏。

• 第一句：用过去时重新陈述研究了什么问题（措辞要与摘要不同，不能照抄）。
  • 典型写法："In this article, the distributed formation tracking control problem for mobile robots under ETC and DoS attacks was investigated."
  • 或："This paper has addressed the secure formation-containment control problem for nonlinear MASs."
• 中间2-4句：简要概括方法和主要结果。不需要重新论证重要性，只需列出做了什么、证明了什么。
  • "The formation control scheme consisted of distributed event-based estimators and adaptive tracking controllers."
  • "Sufficient conditions were derived to guarantee the stability of the closed-loop system."
• 验证句（可选）：一句话提及仿真验证。
  • "Simulation results have verified the effectiveness of the proposed scheme."
  • 注意：只用 "effectiveness" 或 "feasibility"，不要用 "superiority"、"excellent performance" 等自夸词。
• 未来工作（1-2句）：具体的技术方向，不要写空泛的展望。
  • 典型写法："An interesting topic for future research is to extend the results to directed switching topologies."
  • 或："Future work will consider the case where only relative position information is available."
  • 禁止写法："This work opens exciting new avenues for..." / "The proposed framework has the potential to revolutionize..."
• 绝对禁止的词语（描述自己的工作时）：groundbreaking, innovative, revolutionary, pioneering, cutting-edge, remarkable, excellent, superior（个别情况下与仿真对比数据一起使用除外）
• 语气：平实、陈述性。结论是全文最克制的部分——让结果本身说话。
• Target: 0.5 pages (约半栏)

Appendix:

• Full proofs of main-body theorems
• Additional lemmas needed for proofs
• Format: "Appendix A: Proof of Theorem 1"

Step 4: Build Bibliography

CRITICAL: Only include entries that are actually cited in the paper.

1. Scan all \cite{} references in the drafted sections
2. Build a citation key list
3. For each citation key:
   • Check existing .bib files in the project
   • If not found and DBLP_BIBTEX = true, use the verified fetch chain
   • NEVER fabricate BibTeX entries
4. Write references.bib containing ONLY cited entries

Verified BibTeX Fetch (when DBLP_BIBTEX = true)

Step A: DBLP

curl -s "https://dblp.org/search/publ/api?q=TITLE+AUTHOR&format=json&h=3"
curl -s "https://dblp.org/rec/{key}.bib"

Step B: CrossRef DOI

curl -sLH "Accept: application/x-bibtex" "https://doi.org/{doi}"

Step C: Mark [VERIFY] if both fail.

IEEE-specific citation rules:

1. Use \cite{key} (numbered style [1]), NOT \citep{}/\citet{}
2. Use IEEEtran.bst for formatting
3. Prefer IEEE Transactions/journal versions over conference versions
4. Typical reference count: 30-50 for regular papers
5. Consistent key format: {firstauthor}{year}{keyword}
6. Every entry needs: author, title, journal/booktitle, volume, number, pages, year

Step 5: De-AI Polish（降低AI率——基于真实论文对比）

Before this step: Read ../shared-references/anti-ai-writing.md for the full de-AI word lists and sentence pattern checks.

摘要、引言和结论必须通过此检查。 这是用户明确要求的。

检查标准不是套模板，而是对照真实已发表的IEEE Trans控制论文的写作习惯。

一至四：通用去AI检查 自夸词、AI典型句式、结构单调性、语态语气等通用检查规则详见 ../shared-references/anti-ai-writing.md。执行本步骤前必须先读取该文件并逐条对照。

五、段落过渡检查（引言专用）

• 每段的第一句是否承接了上一段的结论？
• 过渡是基于内容逻辑（因果、补充、对比、深化），还是机械地用连接词？
• 整个引言读下来，是否有一条清晰的叙事线从广域背景推进到研究空白？
• 每个文献综述段是否在比较不同工作的假设和结果，而不是简单罗列？
• 每段是否以一句"该方向尚未解决的技术障碍"结尾，自然过渡到下一段？

六、正面质量检查（不只是禁止——还要确保内容具体）

• 每个主要声明（贡献项、定理描述）是否指明了具体的系统类型、条件和结果？
• 没有任何句子是"放在本领域任何一篇论文里都成立"的泛泛之谈。
• 摘要、贡献列表和结论对同一贡献的描述措辞必须不同——禁止跨节复制粘贴。
• 文献综述段必须综合比较不同工作的假设/方法/局限，不能只是"In [X], ... In [Y], ... In [Z], ..."的罗列。

Step 6: Cross-Review with REVIEWER_MODEL

Before this step: Read ../shared-references/agent-role-charter.md for reviewer role definitions and review protocol.

Send the complete draft to GPT-5.4 xhigh:

spawn_agent:
  model: gpt-5.4
  reasoning_effort: xhigh
  message: |
    You are a senior professor in multi-agent control and robotics, with extensive
    IEEE Transactions reviewer and associate-editor experience. You also act like a
    strong corresponding author who can spot AI-looking prose immediately.

    Review this IEEE Transactions paper draft in control systems.

    Focus on:
    1. Does each contribution from the intro have a corresponding theorem and simulation?
    2. Is the writing natural and free of AI-generated patterns?
    3. Are the assumptions well-justified with Remarks?
    4. Are the proofs logically sound (Lyapunov function → derivative → stability)?
    5. Does the simulation section validate ALL theoretical results?
    6. Are the comparisons with prior work specific and fair?
    7. Is the notation consistent throughout?

    Non-negotiable review standards:
    - no hype language
    - no vague praise
    - no generic transitions or filler background
    - no claim stronger than theorems, simulations, or comparisons support

    For each issue, specify: severity (CRITICAL/MAJOR/MINOR), location, and fix.
    Prefer the minimum grounded rewrite or structural change that would make the
    paper read like a publishable journal manuscript rather than AI-generated prose.

    [paste full draft text]

Apply CRITICAL and MAJOR fixes. Document MINOR issues for the user.

Step 7: Reverse Outline Test

After drafting all sections:

1. Extract topic sentences — pull the first sentence of every paragraph
2. Read them in sequence — they should form a coherent narrative
3. Check claim coverage — every contribution from intro must appear in theorems and simulations
4. Check evidence mapping — every theorem must be validated in simulation
5. Fix gaps — if a topic sentence doesn't advance the story, rewrite the paragraph

Step 8: Final Checks

Before declaring done:

• All \ref{} and \label{} match (no undefined references)
• All \cite{} have corresponding BibTeX entries
• Author information is placeholder (not real names unless user provides them)
• Figure/table numbering is correct (Fig. 1, TABLE I format)
• No TODO/FIXME/XXX markers left in the text
• No [VERIFY] markers left unchecked
• Abstract is self-contained and 150-200 words
• Index Terms present and appropriate
• Title is specific and informative
• Introduction has numbered contributions with explicit contrasts to prior work
• All Assumptions have justifying Remarks
• All Theorems have proofs (main body or appendix)
• Simulation validates all theorems
• Notation paragraph in Introduction is complete
• references.bib contains ONLY cited entries
• No stale section files — every .tex in sections/ is \inputed by main.tex
• De-AI check passed on Abstract, Introduction, and Conclusion
• Paper organization paragraph matches actual section structure

Key Rules

• Large file handling: If the Write tool fails due to file size, immediately retry using Bash (cat << 'EOF' > file) to write in chunks. Do NOT ask the user for permission — just do it silently.

• Do NOT generate real author names — use placeholder unless user provides

• Write complete sections, not outlines — the output should be compilable LaTeX

• One file per section — modular structure for easy editing

• Every contribution must have evidence — theorem + simulation for each

• Compile-ready — the output should compile with latexmk without errors

• No over-claiming — use measured language, especially in Abstract/Conclusion

• IEEE citation style — \cite{} with IEEEtran.bst, NOT \citep{}/\citet{}

• Clean bib — references.bib must only contain actually cited entries

• Section count is flexible — match PAPER_PLAN structure

• Do not write in generic AI prose — if a sentence could fit many papers, rewrite it to mention the actual problem, assumption, theorem, figure, or comparison

• Do not self-praise — replace adjectives like novel, innovative, remarkable, effective with concrete claims and evidence

• Abstract / Introduction / Conclusion require an explicit de-AI pass before finalization

• Backup before overwrite — never destroy existing paper/ directory without backing up

• Front-load the contribution — the reader should know the main result by page 2

• De-AI writing is mandatory — especially in Abstract, Introduction, and Conclusion

• Convergence proofs — for cooperative-competitive scenarios, use super-stochastic matrix convergence theory

• Default drafting role — write as a senior corresponding author in multi-agent control / robotics who publishes in scientific journals and line-edits for field-native prose

Writing Quality Reference

• ../shared-references/writing-principles.md — story framing, sentence-level clarity, IEEE Trans specific conventions
• ../shared-references/venue-checklists.md — IEEE Trans submission requirements
• ../shared-references/citation-discipline.md — stricter fallback for ambiguous citations