This article is derived from the plugin's already-indexed fields (wiki / faq / compatibility_json / readme), not freshly generated by AI. Source field is noted at the end of each section.

Quick start

dsh-plugin-writing-guard

— source: plugin_wiki.wiki_content

Install & verify

dsh plugin --profile web add dsh-plugin-writing-guard

Run the command above in your DSH Web Profile. Then enable the plugin in the plugin list.

— source: plugins.install

Key points

• 写作前加载 writing_rules
• 写作 / 修改时由 writing_audit 自动检查
• 修改后自动对比前后版本，保护 Scholarship / Epistemic invariants
• 修改过程残留：revised、as requested、本轮、审稿人要求
• 防御性写作：concession stacking、limitation pre-emption、generic value claim、unnecessary epistemic retreat

— source: plugin_wiki.readme_en (fallback readme_raw)

FAQ

What is the essential difference between it and traditional "AI Humanizer"?

Humanizer typically hands the entire text to another model for rewriting; Writing Guard does not rewrite, it continuously scans for mechanical sentence patterns, self-deprecating disclaimers, and modification process residues during the DSH writing/revision process, and marks the hits by severity and confidence for the author to revise themselves.

Does it support Chinese papers? Will the rules accidentally flag technical terms?

Yes, it supports both. Chinese and English rules run in parallel with independently calculated density thresholds (English by word count, Chinese by CJK character count), and use "double gating" (meeting both absolute count and per-thousand-unit density) to avoid false positives on domain terminology like robust regression and coupling mechanisms.

Can it detect when numbers, p-values, or references get messed up during revision?

Yes, it can. When passing original=pre-modification text, Scholarship Lock is enabled (detecting drift in numbers/percentages/confidence intervals/references/figure numbers/DOIs) plus Epistemic Lock (claim strength, negation/null results, scope boundaries, evidence status drift), the automatic path will capture pre-modification content based on preimage.

How do I uninstall it? What data does it leave behind?

Simply run DSH's standard plugin uninstall. The plugin persists an incremental audit state to ~/.dsh/plugins/dsh-plugin-writing-guard/state.json (fingerprint + baseline cache of ≤20 files ≤4MB). Uninstalling deletes this directory, completely removing all data.

Can it directly audit paper files in .docx or .pdf format?

No, the plugin only reads .txt/.md/.tex/.markdown and other text files. When encountering .docx/.doc/.pdf, it will directly throw an error, prompting to first use the anydoc tool to convert the document to Markdown before auditing.

Can it only be used for English journals? Does it work for Chinese SCI writing?

Yes, it works. The plugin distills distributions based on "target journal representative papers" (writing_journal_profile), supporting any language; Chinese and English writing share STYLE/EVIDENCE/JOURNAL three-axis rules, and author style profiles and journal profiles also support mixed Chinese-English content.

Does it require internet connection, call large models, or upload paper content?

No, it does not. All rules are implemented locally with deterministic regex/statistics, zero network, zero LLM, zero subprocesses; it only reads the paper file currently being written, and all state resides locally in ~/.dsh/plugins/dsh-plugin-writing-guard/.

— source: plugin_wiki.faq_json

Compatibility

• DSH: 0.1.0-rc.6+
• Node: >=18

— source: plugin_wiki.compatibility_json

Pitfalls

Review the upstream repo before installing. This guide is auto-derived from indexed fields and may lag the latest release. If anything contradicts the official docs, treat the upstream source as authoritative.

— source: general rule