Building a Reviewable Chinese-to-English Translation and Editing Workflow for STEM Papers with sci-translation-skill

Project Overview

sci-translation-skill is a set of instructions for paper translation, terminology control, and translation review that can be loaded into a coding agent. It is designed for Chinese-to-English translation of papers in computer science, medicine, engineering, materials science, and other STEM fields. It is not a new translation model; instead, it combines workflow specifications, role contracts, terminology management, and local deterministic checking scripts.

The project aims not simply to generate English that sounds more “SCI-like,” but to reduce terminology drift, distortion of technical meaning, loss of conditions and qualifications, incorrect numeric associations, and formatting damage while preserving natural expression. The result should be closer to a submission-ready version that authors can review item by item.

Fluency does not equal accuracy. Scientific translation should first protect the scope of claims, strength of evidence, causal relationships, negation, conditions, quantifiers, and technical boundaries in the source text.

Core Workflow

1. Establish a structural inventory: Preserve the source text, record the paper type, sections, target readers, and editing scope, and review the boundaries of each proposition paragraph by paragraph.
2. Freeze project terminology: Create concept IDs, preferred translations, contexts, preferred word forms, prohibited aliases, evidence, and open questions for key concepts.
3. Produce a faithful first translation: Accurately convey the source text first, then perform reader-oriented editing, such as clarifying action agents, breaking up high-load sentences, and improving paragraph cohesion, without adding arguments.
4. Conduct bidirectional review: Compare the source, first translation, and edited draft together to check whether the editing process introduced new technical errors.
5. Revise by issue ID: Record each issue, revision, and rationale, then review the final differences and affected terms.
6. Deliver review materials: Provide the translation, glossary, review log, and author confirmation items, and state the actual scope of checks performed.

When processing a new paper, the README recommends first translating 2 to 3 paragraphs from the Methods or Results section, then expanding to the main text, figures, tables, and appendices, and finalizing the abstract last. This allows the terminology and workflow to be validated before expanding the processing scope.

Key Capabilities and Boundaries

Terminology and Technical Fidelity

The project locks project terminology through concept IDs, frozen versions, and context records. It does not require every paper to use the same fixed glossary; instead, it establishes an independent terminology contract for each paper. Paragraph-level review also focuses on negation, conditions, quantifiers, evidence strength, numbers, units, and metric associations.

Reader-Oriented Editing

Reader-oriented editing organizes existing information according to the section, clarifies subjects and references, reduces sentence load, and uses more natural semantic verbs and word order to reduce empty formulaic language. It should not remove necessary hedging, add meanings absent from the source text, or move arguments across paragraphs.

Multiple Agents and Honest Degradation

When genuine tool support exists, terminology research, translation, and independent review can be assigned to different task contexts. Here, “multiple agents” means multiple independent task contexts; it does not automatically imply the use of different models or human experts. Without real tools, the workflow should switch to serial self-review and disclose this explicitly.

run-manifest records actual role IDs and known models; it should not fabricate nonexistent team logs. Complex formatting and medical manuscripts still require human or rendering validation, and medical content requires final review by a qualified professional.

Domain Priorities

• Computer Science, Security, and AI: Focus on threat models, access control, KEMs and payload decryption, mathematical notation, and technical boundaries.
• Medicine and Biology: Focus on study design, association versus causation, effect sizes, intervals, samples, dosage, endpoints, and time points.
• Engineering and Materials: Focus on experimental conditions, units and dimensions, measurements versus simulations, errors, and applicability.

Installation and Runtime Environment

Place the sci-translation-skill directory containing SKILL.md in a skill directory actually supported by the host, and avoid redundant nesting. You can also provide the repository URL directly to the agent and let the host handle installation.

An example location for local Codex is ~/.codex/skills/sci-translation-skill/SKILL.md. If the current directory is the repository root, use the following commands to install it. If a version already exists at the target location, back it up first.

mkdir -p ~/.codex/skills/sci-translation-skill
rsync -a --exclude '.git/' --exclude '.DS_Store' --exclude '__pycache__/' ./ ~/.codex/skills/sci-translation-skill/

After installation, refresh the skill in the host or start a new session, and confirm that it can read SKILL.md. This file does not replace the host's own configuration documentation.

When migrating to another host, first confirm its skill or rules loading interface. Claude Code must load the skill directory supported by the current version; Antigravity, Cursor, Windsurf, and Aider require confirmation of their rules interfaces; Dify, FastGPT, Coze, and self-built Harness systems require their own implementation of role nodes, file transfer, dependency management, and run records. Uploading a knowledge base alone does not establish a multi-agent pipeline.

The skill does not require the fixed use of Sonnet or Opus, nor does it modify host configuration without permission. The checking scripts require only the Python 3 standard library.

Basic Usage

1. Translate a Sample in pilot Mode

For first-time use, ask the agent to establish a glossary and then translate 2 to 3 specified paragraphs as a pilot. The prompt should specify the field, target journal, and input format, and require the preservation of causality, negation, qualifications, and numerical values.

Use $sci-translation-skill to translate the paper in pilot mode.
Field: [discipline]; target journal: [name or undecided]; format: [LaTeX/Word/plain text].
First build a glossary from the main text, equation descriptions, figures, tables, and appendices, then translate the 2–3 paragraphs I specify.
Translate faithfully first, then organize information within paragraphs by section and reader; clarify action agents and references without forcing sentence lengths.
Preserve causality, negation, qualifications, and numerical values from the source text. Do not add experiments, citations, or security guarantees.
Use separate agents for cross-checking when independent agent tools are available; otherwise disclose that the review was performed serially.
Output the translation, frozen terminology, located revision notes, and author confirmation items.

The pilot result should include at least the translation, frozen terminology, located revision notes, and items requiring author confirmation. After confirming the terminology and working method, expand the workflow to the full paper.

2. Review an Existing Chinese-English Parallel Draft

If the Chinese source and English translation already exist, use review mode. Review technical meaning, terminology, claim strength, numbers, units, and citations before reviewing linguistic naturalness.

Use $sci-translation-skill in review mode to review the English paragraph by paragraph against the Chinese.
First check technical meaning, terminology, claim strength, numbers/units, and citations, then check linguistic naturalness.
Fix only issues supported by evidence; distinguish preferences such as “uses→adopts” from substantive errors.
Check for reading obstacles before revising sentences; compare the first translation and edited draft to recheck condition scope, causality, metric associations, and citation bindings.
For each item, list the source text, current translation, revision, and rationale, then review it again after editing.

This mode separates substantive errors from wording preferences. For example, changing uses to adopts alone does not necessarily indicate a technical issue, so not every stylistic difference should be logged as an error.

3. Translate the Full Paper While Protecting the Source

Full-paper mode is suitable after the terminology and pilot results have been confirmed. Require the agent to deliver the work in a new directory without overwriting the source. All sections should share the same glossary version and research context, covering the main text, figure captions, tables, equation descriptions, algorithms, and appendices.

Use $sci-translation-skill to translate the full paper according to the confirmed glossary.
Deliver the work in a new directory without overwriting the source. All blocks must share the same glossary version and research context.
Cover the main text, figure captions, tables, equation descriptions, algorithms, and appendices; check the scope of claims in the abstract last.
Deliver the English manuscript, glossary, paragraph-level alignment, review-closure log, and items requiring confirmation.

Using Deterministic Checking Tools

The repository provides local scripts for checking observable differences in translations. First run the skill integrity tests:

bash references/test-skill-integrity.sh

Then run a translation audit on the source and target files:

python3 scripts/audit_translation.py source.tex target.tex \
  --glossary terminology-glossary.json \
  --alignment alignment.json \
  --output qa-report.json

The audit can check numbers, protected fragments, citation counts, basic structure, alignment information, and registered terminology. You can also continue using the existing shell entry point:

bash references/test-translation-structure.sh source.tex target.tex glossary.json

The exit codes are as follows:

• 0: None of the implemented checks found differences.
• 1: Differences requiring further review were found.
• 2: Input or parsing error.

An exit code of 0 does not mean that the translation is ready for submission. Without alignment, you cannot claim complete paragraph-level coverage or terminology compliance checks. Some correct changes in number order or formatting may also trigger differences and require human recording and disposition.

These scripts do not detect “AI-like” writing or certify domain accuracy. They can only help identify deterministic structural differences and cannot replace semantic review, human confirmation, or professional final review.

Advanced Practices

Freeze Glossary Versions and Record Context

Do not apply the sample glossary directly to every paper. The provided sample glossary is for demonstrating fields only and has not been externally verified. Actual projects should establish their own terminology records based on concepts, sections, and context, then review affected terms after revisions.

Separate Fidelity Thresholds from Language Quality

First determine whether the translation preserves the scientific meaning, then evaluate whether the sentences are natural. Zero detected errors does not automatically mean a perfect score, because check coverage may be limited or unidentified issues may remain. Language editing must not use more ornate wording to conceal insufficient evidence.

Protect Formatting and Citation Bindings

Establish a structural inventory before translating, recording formulas, macros, citations, table coordinates, figure captions, and appendices in particular. After editing, check not only whether numbers are present, but also whether they remain bound to the correct metrics, units, and conditions. An unchanged formula count does not prove that the variables remain unchanged.

Close the Revision Loop by Issue Number

Keep a stable issue ID for each issue and record the source text, current translation, revision, rationale, and location. After revisions, compare the source, first translation, and edited draft again, focusing on condition scope, causality, metrics, citations, and claim strength.

Project File Guide

• SKILL.md: Skill entry point and release conditions.
• Terminology Contract: Unified glossary schema and alignment format.
• Role Contracts: Requirements for real tools and file ownership.
• Language Guide: Faithful, concise, and natural expression.
• Reader-Oriented Editing: Section routing, sentence-level load, cohesion, and post-edit fidelity checks.
• Domain Profiles: Key risks in different fields.
• Behavioral Evaluation: Model behavior acceptance plan.
• Translation Example: Constructive difference cases, not proof of published effectiveness.

Conclusion

The value of sci-translation-skill lies in transforming paper translation from one-time polishing into a traceable workflow that includes terminology freezing, faithful first translation, independent review, difference checking, and author confirmation. Using it in the order of pilot, review, full-paper delivery, and final audit can reveal technical and formatting risks earlier.

However, it is not an automatic certification tool for domain accuracy and cannot replace the author's confirmation of scientific conclusions. For complex formatting, medical manuscripts, and any translation prepared for submission, rendering checks, human review, and final review by a qualified professional should still be combined.