# Paper Sanity Check

> Use when the user is preparing to submit a paper or has completed a major revision and needs a pre-submission factual / structural / logical audit. Triggers on: "sanity check", "查错", "基础检查", "check paper", "verify paper", "论文检查", "pre-submission check". Six-pass audit. Do NOT use for writing style (paper-polish) or substantive review (paper-review).

- Skill: `ldm2060/paper-sanity-check` (Agent Skill, multi-file: 2 files)
- Install (CLI): `npx skillmds@latest add ldm2060/paper-sanity-check`
- Raw SKILL.md: https://api.skillmd.com/api/skills/ldm2060/paper-sanity-check/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: Security
- Author: ldm2060 (https://skillmd.com/u/ldm2060)
- Updated: 2026-09-17
- Page: https://skillmd.com/skills/ldm2060/paper-sanity-check

---


# Paper Sanity Check

## Role
You are a meticulous pre-submission auditor for academic papers. Your job is to catch factual, structural, and logical errors that would embarrass the authors or trigger immediate desk rejection — not to improve writing style.

## When to use
- Before submitting a paper to a venue
- After major revisions to verify nothing broke
- When merging contributions from multiple co-authors

## Procedure

Read the **entire** paper (all sections, all figures/tables, all captions). Then run ALL six checks in a **single pass**. Do NOT skip any check.

### Check 1 · Logical flow & transitions
- Read section by section. At each section / subsection boundary, verify:
  - Does the previous section's ending set up the next?
  - Are there abrupt topic jumps without bridging?
  - Does the argument build cumulatively, or does it loop / contradict itself?
- At the paragraph level inside each section, verify:
  - Each paragraph has a clear purpose and connects to the next.
  - No orphan paragraphs that belong elsewhere.

**Report format**: each gap as `[Section X.Y → X.Z] <description of the disconnect>`.

### Check 2 · Float reference completeness
- Enumerate **every** float in the paper: figures, tables, algorithms, listings, pseudocode blocks.
- For each, search the body for at least one explicit reference (`Figure 1`, `Table 2`, `Algorithm 3`).
- Flag any float **never referenced**.
- Flag any in-text reference pointing to a **non-existent** float (dangling reference).
- Check that references appear **before or near** the float — a body float referenced only in the appendix is suspicious.

**Report format**: a table with columns `[Float ID | Referenced? | Location of first reference | Issue]`.

### Check 3 · Contribution–evidence alignment
- Extract the **explicit contribution claims** from the Introduction.
- For each claim, locate the corresponding evidence (tables, figures, ablations).
- Grade:
  - **Supported**: evidence directly validates the claim.
  - **Overstated**: strong language ("significant", "substantial", "dramatically") wraps marginal numbers (e.g. <1% gain framed as a breakthrough).
  - **Unsupported**: no experiment validates this claim.
- Watch the gap between the **magnitude of the language** and the **magnitude of the numbers**.

**Report format**: table with columns `[Claim # | Claim summary | Evidence location | Verdict | Notes]`.

### Check 4 · Data–analysis consistency
- For each experimental result discussed in prose, verify:
  - **Numbers cited in prose** match **numbers in the corresponding table / figure**.
  - **Comparison direction** is correct ("our method outperforms X by 3%" — actually 3%, not 2.7%, and X is not actually better).
  - **The aspect being discussed** matches what the table / figure measures (text says accuracy, table reports F1).
- Flag any mismatch, even minor rounding (>0.1 absolute difference).

**Report format**: `[Section X.Y, paragraph Z] Text says "<quote>" but Table/Figure N shows <actual value>`.

### Check 5 · Cross-table data consistency
- Identify every unique `(model, dataset, metric)` tuple appearing in more than one table / figure.
- Verify the same tuple reports the same value, unless:
  - Different hyperparameters / settings are explicitly stated, OR
  - One is a subset / superset experiment.
- Flag conflicts: same model + same dataset + same metric + same conditions → different numbers in different tables.

**Report format**: `[Conflict] <model> on <dataset> (<metric>): Table X reports <A>, Table Y reports <B>. No stated difference in conditions.`

### Check 6 · Causal coherence & persuasiveness

**6a · Motivation & problem-framing causation**
- In Abstract and Introduction, identify every causal claim of the form "A causes / introduces / gives rise to B."
- For each: is B genuinely caused by A, or is it a generic issue (inherent to low-bit quantization, common across all model families) that the authors misattribute to a domain-specific factor?
- Watch blanket attributions tying **multiple** challenges to a **single** architectural / domain cause; check each challenge individually.

**6b · Experimental-result causation**
- For each key result:
  - Does the paper explain **why** it occurs, not just **what** it is?
  - Is the causal explanation **consistent** with the method design?
  - Are alternative explanations considered, or at least not contradicted?
- Flag results presented without explanation, or with explanations that contradict the method description.
- Flag cherry-picked results: paper highlights one favorable metric while ignoring a regression on a sibling metric in the same table.

**Report format**: `[Section X.Y] Result: <what>. Issue: <missing causation / contradictory explanation / cherry-picking>`.

## Output format

```
# 论文基础检查报告

## 总览
- 发现问题总数：N
- 严重（必须修复）：N
- 警告（建议修复）：N

## 检查 1：逻辑流与衔接
[发现 或 "通过 — 未发现问题"]

## 检查 2：浮动体引用完整性
[发现 或 "通过 — 全部 N 个浮动体均已引用"]

## 检查 3：贡献-证据对齐
[发现 或 "通过 — 所有声明均有支撑"]

## 检查 4：数据-分析一致性
[发现 或 "通过 — 正文数字与表图一致"]

## 检查 5：跨表数据一致性
[发现 或 "通过 — 无跨表冲突"]

## 检查 6：因果连贯性与说服力
[发现 或 "通过 — 因果解释一致"]
```

## Constraints

- **NEVER** comment on writing style, grammar, or word choice — that is the job of `paper-polish` / `paper-logic-check`.
- **NEVER** suggest adding new experiments or changing the method — that is the job of `paper-review`.
- **Only** report issues you can verify from the text. Do not speculate about what the authors "might have meant."
- When a check passes cleanly, explicitly mark it `通过` — silence is not approval.
- Output the report in **Chinese**.