Auto Paper Citation

Purpose

Support claims, not reference stuffing. Do not rewrite the manuscript in this phase.

Required Inputs

• claim_register.md
• bibliography files from config.yaml
• research_dossier.md
• optional local reference PDFs or external search results
• candidate references extracted from PDF or Markdown notes

Support Grades

Use the grades from .agents/skills/auto-paper/references/citation-support-bank.md:

• strong
• partial
• background
• limiting
• metadata_only
• unsupported

metadata_only is never claim support.

For blogs, reviews, surveys, and tutorials, every external factual claim, named method, named dataset, named system, clinical-effectiveness claim, economic claim, benchmark claim, and history/trend claim needs either a verified citation row or an explicit unsupported / metadata_only row with a revision action. Do not treat "blog" as permission to omit citations.

If the only available reference source is an AI dialogue, a PDF conversation, or notes that quote literature without primary-source verification, extract the candidate paper metadata into citation_support_bank.md with support_grade: metadata_only or partial as appropriate and set needs_user_confirmation: yes.

Outputs

Write:

• citation_support_bank.md
• claim_citation_map.md
• optional citation_audit_report.md

Run the deterministic support-bank gate when the bank is drafted:

• .agents/skills/auto-paper/scripts/citation_bank_check.py auto_paper_output/<paper_id>/citation_support_bank.md

If a core claim has only background, metadata_only, or unsupported support, route back to citation or argument instead of adding a weak citation during patch.

Gate Ledger

Report a Gate ledger entry with commands run, artifacts written, support-bank gate result, any USER_GATE or NOT_RUN reason, and the next owner before handoff.