dsh-plugin-writing-guard

When AI helps with academic writing, the real risk is not just grammar. It is template-like sentences, defensive phrasing, silent drift in numbers or causal claims, and rejected alternatives leaking into the final artifact. dsh-plugin-writing-guard brings these checks into the DeepSeek Harness workflow, so issues are caught during drafting and revision rather than after the entire manuscript is finished.

It works around four pillars: STYLE reduces mechanical AI patterns without simply hiding AI use; EVIDENCE compares versions to protect numbers, p-values, citations, DOIs, figure and table references, causal strength, null findings, and scope; JOURNAL builds a section-level profile from representative papers of a target journal to compare writing habits and rhetorical structure; DELIVERY detects rejected alternatives, process residue, or context leakage in the final deliverable.

All rules run locally and deterministically, with no network calls and no extra LLM. It is suitable for researchers, authors, and teams writing or revising Chinese and English academic papers inside DeepSeek Harness who want to reduce AI-style writing, preserve evidence, and clean manuscripts before submission.