bilingual-review-auditor

name

bilingual-review-auditor

description

Use when preparing, auditing, or revising peer-review materials for paper submissions that need a copy-pasteable English submission review plus Chinese internal translation, detailed Chinese rationale, and evidence grounded in PDFs or source materials. Trigger for bilingual review drafts, venue or system Markdown reviews, Chinese manual-check sections, evidence tables, rating/text consistency checks, PDF-grounded review claims, or post-processing existing review outputs. Owns bilingual alignment, evidence support, submission/internal separation, and review-material audit; does not own author-side paper revision or field-specific review standards.

Bilingual Review Auditor

Purpose

Prepare peer-review materials that are concise enough to submit in English and detailed enough to audit in Chinese. Treat this skill as a review-material auditor and packager: preserve the final submission boundary, add bilingual checking support, and force evidence-backed reasoning for major claims.

Operating Modes

Detect the mode from the user's latest instruction and inputs.

• Use companion mode when the user starts a paper review and asks for bilingual output, Chinese checking notes, evidence, PDF grounding, or manually verifiable logic.
• Use post-audit mode when the user provides existing review drafts, Markdown files, venue/system fields, or prior reviewer output and asks to translate, tighten, audit, justify, or add evidence.
• Use discussion-only mode when the user asks to discuss the skill, workflow, or design. Do not edit files unless the user also asks to persist the result.

The same output contract applies in companion and post-audit mode.

Non-Ownership

Do not redefine the scholarly review standard, reviewer persona design, field-specific acceptance threshold, or author-side revision plan. If the source review judgment is weak, inconsistent, or unsupported, audit and strengthen the material by grounding claims and making the rating logic explicit.

Do not make review tone stricter or softer by default. Adjust severity only when source evidence, user instruction, official rating semantics, or internal inconsistency supports the change.

Mandatory PDF Gate

When evidence comes from a PDF, use a PDF skill with layout-aware extraction and rendered-page inspection.

Do not use pdf2text, pdftotext, or any plain text dump as the review input, locator, transcript, or evidence source. PDF papers are layout-bearing documents: formulas, figures, tables, captions, references, footnotes, appendices, and two-column reading order must be checked from PDF-skill extraction artifacts, rendered pages, and cropped page regions together.

Save the PDF processing artifacts needed for auditability: extraction outputs, rendered page images or crops, and notes that connect review claims to visible source evidence.

For any claim that affects rating, rejection/acceptance rationale, manuscript integrity, experimental adequacy, reproducibility, benchmark validity, result correctness, or trustworthiness, verify against visible PDF content. If rendered-page verification is not available, mark the claim as unverified or phrase it cautiously; do not silently fall back to plain text extraction.

For detailed PDF checks, read references/pdf-evidence-checklist.md.

Workflow

1. Classify artifacts.

   • Identify submission-facing review fields, internal notes, paper PDFs, supplementary materials, official rating labels, and existing Markdown files.
   • Treat source papers and official materials as evidence. Treat existing reviews as drafts to audit.

2. Preserve the submission boundary.

   • Preserve the official or user-provided English submission template and keep it copy-pasteable.
   • Put Chinese translation, Chinese rationale, and evidence tables outside the submission-facing section, normally below a clear separator.
   • Never mix internal Chinese notes into the submission-facing section.

3. Audit claims before polishing text.

   • List major criticisms and positive claims.
   • Separate observable evidence, defensible inference, and reviewer judgment.
   • Remove or soften unsupported claims. Strengthen supported claims with precise evidence locations.

4. Check rating consistency.

   • Compare written conclusion, rating number, official rating label, confidence, and evidence strength.
   • Flag contradictions such as severe unrecoverable defects paired with a positive score, or mild comments paired with a clear rejection score.
   • For detailed checks, read references/rating-consistency-checks.md.

5. Produce bilingual support.

   • Translate the English submission review into Chinese faithfully, preserving severity and uncertainty.
   • Add detailed Chinese analysis explaining the logic behind the score.
   • Add an evidence table linking each important claim to source evidence and rating impact.
   • For the default structure, read references/review-output-template.md.

6. Verify the final material.

   • Confirm the English submission-facing material contains no internal notes.
   • Confirm the Chinese translation matches the English content without changing severity.
   • Confirm each rating-changing claim has source-backed evidence or is marked as unverified.
   • If files were changed, review the changed files before finishing.

Evidence Discipline

Use evidence records that a human can quickly check:

• PDF page, section, figure, table, equation, caption, reference, appendix, or crop region.
• Exact metric names, dataset names, task counts, split descriptions, baseline names, or reported numbers when visible in the source.
• Short paraphrase of the relevant evidence, not a long copied passage.

Use cautious wording for negative evidence:

• Prefer: "The submitted materials do not appear to specify the random seed setting."
• Avoid: "The authors did not run multiple seeds" unless the source explicitly proves it.

Output Style

Keep the English submission review concise and direct. Put detailed reasoning in Chinese internal sections.

Prefer concrete, evidence-linked criticism over broad adjectives. If a paper is weak enough to reject, say so clearly in the rating logic, but make the rejection rationale traceable to verifiable evidence.

Use Chinese for internal analysis unless the user asks otherwise. Use English for the final submission-facing review unless the target venue requires another language.