dsh-plugin-writing-guard

Deterministic manuscript integrity guard for AI-assisted academic revision. Detects mechanical AI writing patterns (defensive writing, template structures) and protects scientific facts during polishing via Scholarship Lock + Epistemic Lock, preserving numbers, citations, claim strength, negation, scope, and evidence status. All rules run locally with regex/statistics, zero LLM/network, returning STYLE / EVIDENCE / JOURNAL layers.

• Detects AI-style patterns: defensive writing, concession stacking, generic value claims, unnecessary epistemic retreat, template phrases
• Scholarship Lock preserves numbers, statistics, p-values, citations, DOIs, and figure references during edits
• Epistemic Lock tracks causal/evidential claim strength, negation, null findings, scope boundaries, and evidence status
• Sub-sentence claim alignment with raw cosine threshold and subject bonus to avoid false drift on full rewrites
• Automatic audit after file writes with pre-execute snapshots, enabling zero-config two-lock protection
• Document profile awareness: manuscript, rebuttal, cover_letter contexts change rule behavior
• Author style profile and journal profile distill writing distributions from historical papers or target journal
• Zero LLM, zero network, deterministic regex/statistical rules; outputs STYLE / EVIDENCE / JOURNAL classification