Deweight Anchor Rewrite

Purpose

改写目标是降低 AI 模式残留，不是润色成统一顺滑风格。所有改写必须追溯到：

paragraph_id -> hit_modes -> practice_cards -> rewrite_recipe

Required Inputs

• Original paragraph text.
• Semantic Anchor.
• Human Anchor Binding prescription.
• Rewrite Prescription following ../../_shared/references/rewrite-prescription-rubric.md.
• Integrity Snapshot for immutable token restore.

Hard Gate

If a suspect paragraph has no valid human-anchor binding, return:

blocked: missing_human_anchor_binding

Do not rewrite.

If the binding has no human_anchor_evidence.quote_exact, return:

blocked: missing_human_quote_evidence

Rewrite Rules

• Preserve language of the input.
• Preserve claim polarity and evidence strength.
• Preserve citation keys, author/year mentions, numbers, units, formulas, table/figure labels.
• Change structure before vocabulary.
• Read human_anchor_evidence.quote_exact before rewriting; imitate the quoted author's writing move, not the surface wording.
• Use observed_human_move, transferable_structure, and adaptation_plan as the concrete operation.
• Do not copy quote_exact wording, examples, entities, data, or topic-specific phrasing into the user's rewrite.
• Do not copy OCR/MinerU noise from practice cards.
• Do not add first-person experience, new examples, new data, or fabricated sources.
• Do not use SQL/database retrieval results as authority by themselves; retrieved cards are candidates and still need real practice-card binding.

Output Contract

For each changed paragraph:

paragraph_id:
before_modes:
practice_cards_used:
human_quote_used:
human_move_transferred:
rewrite_recipe_applied:
rewritten_text:
integrity_notes:

If report=rewrite_only, output only the final rewritten text after verification.

The rewrite prescription consumed by this skill should pass:

python3 ../../_shared/scripts/check_rewrite_prescription.py <rewrite-prescription.md>