Prose Cluster Hybrid Citation Fix

| | halmagyi02 | vHIT / vestibular testing | Introduction → Background | "video head impulse testing (vHIT)" | | curthoys09 | Semicircular canal anatomy | Introduction → Background | "vestibular testing including caloric irrigation" | | parnes99 | Endolymphatic sac treatment | Introduction → Background | "endolymphatic sac and stria vascularis" | | iahn17 | Meniere's classification | Introduction → Background | "chronic vestibular disorder characterized by" | | headimpulse | Head impulse modeling | Introduction → Gap | "The head impulse test computational modeling" | | raissi19 | PINN framework | Methods → PINN Architecture | "\subsection{PINN Architecture}" | | chen18 | Neural ODE | Methods → PINN Architecture | "\subsection{PINN Architecture}" | | sanchez22 | Biochemically informed NeuralODE | Methods → PINN Architecture | "\subsection{PINN Architecture}" | | jagtap22 | Conservative PINNs | Methods → PINN Architecture | "\subsection{PINN Architecture}" |

Python batch for section-grouped anchors:

additions = [
    # Group: vestibular testing → Introduction
    ("video head impulse testing (vHIT); and (c) audiometric assessment",
     "video head impulse testing (vHIT)\\cite{halmagyi02,curthoys09}; and (c) audiometric assessment"),
    # Group: endolymphatic sac → Background
    ("within the endolymphatic sac and stria vascularis, with a baseline production rate",
     "within the endolymphatic sac and stria vascularis\\cite{parnes99}, with a baseline production rate"),
    # Group: Meniere's → Background
    ("a chronic vestibular disorder characterized by episodic vertigo",
     "a chronic vestibular disorder\\cite{iahn17} characterized by episodic vertigo"),
    # Group: head impulse → Gap
    ("The head impulse test computational modeling represents the only existing",
     "The head impulse test computational modeling\\cite{headimpulse} represents the only existing"),
    # Group: PINN methods → PINN Architecture
    ("\\subsection{PINN Architecture}",
     "\\subsection{PINN Architecture}\n\nWe follow the physics-informed neural network framework\\cite{raissi19,chen18, sanchez22, jagtap22} for solving the inverse ODE problem."),
]
for anchor, new_text in additions:
    tex = tex.replace(anchor, new_text)

Verification

python3 -c "
import re
with open('01-manuscript/paper.tex') as f: tex = f.read()
bibitems = set(re.findall(r'\\\\bibitem\{([^}]+)\}', tex))
cite_keys = re.findall(r'\\\\cite\{([^}]+)\}', tex)
cites = set()
for ck in cite_keys:
    for k in ck.split(','): cites.add(k.strip())
orphans = bibitems - cites
print(f'D10a: {len(cites & bibitems)}/{len(bibitems)} = {(len(cites & bibitems)/len(bibitems))*100:.0f}%')
print(f'Orphans: {len(orphans)}' + ('' if not orphans else ': ' + ', '.join(sorted(orphans))))
"

Compile Verification

cd 01-manuscript/
pdflatex -interaction=nonstopmode paper.tex
pdflatex -interaction=nonstopmode paper.tex  # second pass for cross-refs
grep -c "undefined" paper.log  # should be 0 (or just cross-ref warnings)
grep -oP "Output written on paper\.pdf \(\d+ pages" paper.log

When This Pattern Occurs

This pattern is most common in papers where:

1. The clinical/domain calibration paragraph is a single paragraph with dense prose author-year references (pressure ranges, clinical thresholds)
2. The remaining bibitems cover foundational literature (methodology, anatomy, clinical tests, classification systems) that the LLM generated as a reference block but never anchored in text
3. The paper has a standard IMRaD structure but the Introduction/Methods sections contain mostly domain description without formal citations

Related Patterns

• Pure prose variant (0 \cite{}, bibitems present, no prose author-year refs): see references/orphan-bibliography-trap.md — map each bibitem to its contextual location in the paper
• Pure inline variant (0 prose refs, bibitems have no anchor): see ref/orphan-bibliography-inline-manual-fix.md

契约层 · BOUNDARY

边界：技能功能边界。

契约层 · IO_CONTRACT

输入：请求描述、上下文信息。 输出：执行结果、状态反馈。

验证清单 · VERIFICATION

• 已建立"引用键-主题-章节-文本锚点"映射表，每个 bibitem 都有明确的章节归属与锚点（PROS-001）
• 同一主题/方法簇的多个 bibitem 在单一锚点合并为一条 \cite{a,b,c}，未逐句重复插入（PROS-002）
• 密集散文段落的引用插入使用了精确短语 anchor 做字符串替换，替换后上下文完整未断裂（PROS-003）
• 方法论框架类引用插入在对应 \subsection 标题后并附解释性句子（PROS-004）
• 批量替换后运行 Python 脚本，\cite 键与 \bibitem 键集合交集覆盖率（D10a）已报告且 Orphans 为 0（PROS-005）
• 执行了双次 pdflatex 编译，paper.log 中 grep -c "undefined" 为 0，页码正常输出（PROS-006）

核心原则 · PRINCIPLES

1. 准确为先: 所有输出必须经过事实核查，不编造数据
2. 证据驱动: 每个结论必须可追溯到具体证据或数据源
3. 可复现性: 每一步操作必须可重复，结果可验证

约束规则 · RULES

1. 输入约束: 参数类型、范围、格式必须校验
2. 输出约束: 返回值结构、编码、命名必须一致
3. 异常约束: 错误信息必须包含上下文和恢复建议
4. 安全约束: 不执行未验证的任意代码，不暴露内部状态

Golden 集合 · GOLDEN SET

• Golden Input: 01-manuscript/paper.tex 含 9+ 个未锚定 bibitem（iahn17、raissi19 等）+ 密集散文段落，IMRaD 结构、Introduction/Methods 无正式引用
• Golden Output: 按 PROS-002/004 在 \subsection{PINN Architecture} 后单锚点合并 \cite{raissi19,chen18,sanchez22,jagtap22}，Python 脚本报告 D10a: N/N = 100%、Orphans: 0，双次 pdflatex 后 grep -c "undefined" 为 0
• Golden Error: 脚本输出 Orphans > 0（如列出未锚定 bibitem 键）→ 触发 PROS-005，须回查映射表（PROS-001）补齐锚点后重跑编译验证

Golden 集合是测试的单一真理来源。所有改进必须通过 golden 测试。

违反规则的操作视为不安全，必须拒绝或隔离。

违反任何原则的输出视为失败。原则优先级：准确 > 证据 > 可复现。

每项验证必须可执行、可记录、可复现。验证失败时记录原因和修复。

Genes (策略基因)

紧凑策略表示。条件→策略。需要深度时参考完整文档。

• [PROS-001] 当存在未引用的参考文献 (bibitems) 时 → 建立“引用键-主题-章节-文本锚点”映射表以定位插入位置
• [PROS-002] 当多个参考文献属于同一主题或方法簇时 → 在单一文本锚点处合并插入多个 \cite 键以减少冗余
• [PROS-003] 当需要在密集散文段落中插入引用时 → 使用精确的短语匹配 (anchor) 进行字符串替换以保留上下文完整性
• [PROS-004] 当引用涉及方法论框架时 → 在对应的 \subsection 标题后插入解释性句子并附带引用以增强逻辑连贯性
• [PROS-005] 当执行批量引用修复后 → 运行脚本验证 \cite 键与 \bibitem 键的集合交集以检测孤儿引用
• [PROS-006] 当验证引用完整性时 → 执行双次 pdflatex 编译并检查日志中无 "undefined" 错误以确保交叉引用解析正确

示例 · EXAMPLES

示例 1（密集散文段锚点插入）

• 输入：Meniere's 段为密集散文，iahn17 等 bibitem 未在正文锚定
• 操作/输出：按 PROS-003 用精确短语 "a chronic vestibular disorder characterized by episodic vertigo" 做字符串替换，插入 "disorder\cite{iahn17} characterized by"
• 验证：替换后上下文完整未断裂（验证清单 PROS-003 项）；编译后 grep -c "undefined" 为 0

示例 2（方法簇合并锚点）

• 输入：raissi19/chen18/sanchez22/jagtap22 四个 PINN 方法框架 bibitem 同属 Methods 簇
• 操作/输出：按 PROS-002 + PROS-004，在 \subsection{PINN Architecture} 后附解释句 "We follow the physics-informed neural network framework\cite{raissi19,chen18,sanchez22,jagtap22}..."，单锚点合并
• 验证：未逐句重复插入；四个键全部出现在 \cite 集合中，Orphans 计数为 0

示例 3（D10a 覆盖率验证）

• 输入：批量 tex.replace 修复后的 paper.tex
• 操作/输出：按 PROS-005 + PROS-006 运行 Verification 节 Python 脚本统计 D10a，再执行双次 pdflatex（Compile Verification 节）
• 验证：输出 D10a: N/N = 100%、Orphans: 0；paper.log 无 undefined、页码正常输出