Read txt

Check Writing Quality

Scan the full paper text for writing issues: grammar errors, typos, unclear phrasing, inconsistent terminology, overclaim wording, and logical flow gaps.

Workflow

- [ ] Step 1: Read full paper text
- [ ] Step 2: Scan for writing issues
- [ ] Step 3: Save report

---

Step 1: Read Full Paper Text

Read {paper_dir}/metadata/{arxiv_id}/mathpix/{arxiv_id}.md

Read the entire document — do not split by section.

---

Step 2: Scan for Writing Issues

Scan every section for:

Grammar

• Subject-verb disagreement ("this method can effectively balances")
• Wrong tense ("which we illustrated" when describing a figure → "illustrate")
• Missing articles ("available in community" → "in the community")
• Singular/plural mismatch ("three device" → "three devices")

Typos & Formatting

• Misspelled words ("aling" → "allowing", "chocies" → "choices")
• Missing spaces ("inputsor" → "inputs or")
• Double brackets, broken formatting in tables/figures
• Wrong reference numbers

Unclear Phrasing

• Vague language ("comparable results" without threshold)
• Awkward phrasing ("speeds up the performance")
• Incomplete sentences, missing objects ("surpasses in X" — surpasses what?)

Inconsistent Terminology

• Same concept called different names in different sections
• Inconsistent capitalization of terms

Overclaim Wording

• "significantly improves" without statistical test
• "universally effective" tested on only a few benchmarks
• "inherent property" from a single observation

Logical Flow

• Conclusions not supported by the presented evidence
• Missing transitions between sections
• Claims in abstract/conclusion not backed by experiments

Each issue:

{
  "page": 3,
  "quote": "this method can effectively balances component number",
  "status": "warning",
  "reason": "Subject-verb disagreement: 'can effectively balances' should be 'can effectively balance'.",
  "suggest": "Change 'balances' to 'balance'."
}

status:

• warning — grammar error, typo, unclear phrasing, minor inconsistency, or anything that MIGHT be a PDF parsing artifact
• error — only for issues you are 100% certain exist in the actual paper, not caused by PDF parsing. Examples: duplicated words ("that that"), wrong citation (citing paper A for paper B's result), clearly wrong terminology used consistently. If there is ANY chance the issue comes from PDF extraction, use warning instead

---

Step 3: Save Report

Output: {paper_dir}/reports/check_txt.json

{
  "summary": {
    "total_issues": 25,
    "warning_issues": 22,
    "error_issues": 3
  },
  "results": [
    {
      "page": 3,
      "quote": "this method can effectively balances component number",
      "status": "warning",
      "reason": "Subject-verb disagreement: 'can effectively balances' should be 'can effectively balance'.",
      "suggest": "Change 'balances' to 'balance'."
    }
  ]
}

Tips

• Scan every section including abstract, appendix, table/figure captions, and references
• Non-native English writing often has systematic errors (e.g., missing articles throughout) — report the first 2-3 instances, then note "pattern continues throughout"
• Distinguish style preferences from actual errors — only flag clear mistakes or genuinely confusing phrasing
• Overclaim wording is error only if it misrepresents the actual results; stylistic exaggeration is warning

Ignore PDF Parsing Artifacts

The input text is extracted from PDF via Mathpix/MinerU. Do NOT flag these as writing issues:

• Broken hyphens: "ex-\nploratory" or "trajec-\ntory" — these are line-break artifacts, not typos
• Missing hyphens from line breaks: "turnlevel" instead of "turn-level" — if the compound word clearly needs a hyphen but the text was joined across a line break, skip it
• LaTeX math notation: $14.9 \%$, $\beta$, \mathbb{E} — these are correct LaTeX, not formatting errors
• Image references: ![](images/img_001.jpg) — skip these
• Footnote markers: [^0], ${ }^{1}$ — PDF extraction artifacts
• Garbled capitalization: "AgentFLOW", "AgENTFlow" — PDF extraction often mangles mixed-case words. Skip these entirely unless the same inconsistency appears multiple times in clearly different contexts
• Grammar around LaTeX: "achieving average accuracy by $14.9 \%$" — the awkwardness may come from LaTeX rendering, not the actual paper. If the sentence structure is otherwise sound, skip it

Only flag issues that exist in the actual paper, not artifacts from the PDF-to-markdown conversion. When in doubt, downgrade to warning or skip entirely.