citation-anchoring_skill

Overview

This skill detects regression in citation anchoring to ensure citation markers remain inside the same H3 subsection after polishing. It compares a saved baseline of per-subsection citation sets to the current draft and produces a concise PASS/FAIL report with examples of any drift. The skill is analysis-only and never modifies content.

How this skill works

The auditor loads the baseline anchors file produced by the initial draft-polisher run and parses the current output/DRAFT.md into its ### subsections. For each subsection it extracts citation keys and compares the current sets to the baseline to find added, removed, or migrated keys. The skill writes output/CITATION_ANCHORING_REPORT.md summarizing status and including a short diff and representative examples when failures occur.

When to use it

• After editing or polishing a draft to confirm citations did not move between ### subsections
• Before finalizing a draft to preserve claim→evidence alignment
• As a regression check when running automated polishing pipelines
• When multiple contributors edit the same document and you need citation stability
• When you maintain long documents with strict subsection evidence mapping

Best practices

• Keep an explicit baseline: run draft-polisher once to generate output/citation_anchors.prepolish.jsonl before future checks
• Run citation anchoring after any rewrite or automated polishing step but before publication
• If you intentionally restructure subsections, delete the baseline and regenerate it to avoid false positives
• Treat the auditor as read-only: address reported issues by editing the draft or regenerating the baseline, not by changing the report
• Include representative citation examples in the report to speed triage

Example use cases

• CI job that runs after automated polishing to block merges if citations drift across subsections
• Manual QA step for research papers to confirm claims retain their original evidence
• Post-edit audit when multiple passes of copyediting risk moving inline citations
• Batch checking of multiple drafts to detect systemic polish-induced citation migrations